:\n\n> Hi\n>\n> 2017-01-19 13:35 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n>\n>> Pavel Stehule wrote:\n>> > Hi\n>> >\n>> > New update - rebase after yesterday changes.\n>> >\n>> > What you want to change?\n>>\n>> I think the problem might come from the still pending patch on that\n>> series, which Andres posted in\n>> https://www.postgresql.org/message-id/20170118221154.aldebi7\n>> yyjvds5qa@alap3.anarazel.de\n>> As far as I understand, minor details of that patch might change before\n>> commit, but it is pretty much in definitive form.\n>>\n>\n> new rebased update after these changes\n>\n\nfix white spaces\n\npavel\n\n\n>\n> Regards\n>\n> Pavel\n>\n>\n>>\n>> --\n>> Álvaro Herrera https://www.2ndQuadrant.com/\n>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\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","msg_date":"Sat, 21 Jan 2017 10:31:48 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Hi\n\n2017-01-21 10:31 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\n\n>\n>\n> 2017-01-21 9:30 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\n>\n>> Hi\n>>\n>> 2017-01-19 13:35 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n>>\n>>> Pavel Stehule wrote:\n>>> > Hi\n>>> >\n>>> > New update - rebase after yesterday changes.\n>>> >\n>>> > What you want to change?\n>>>\n>>> I think the problem might come from the still pending patch on that\n>>> series, which Andres posted in\n>>> https://www.postgresql.org/message-id/20170118221154.aldebi7\n>>> yyjvds5qa@alap3.anarazel.de\n>>> As far as I understand, minor details of that patch might change before\n>>> commit, but it is pretty much in definitive form.\n>>>\n>>\n>> new rebased update after these changes\n>>\n>\n> fix white spaces\n>\n\nfew fixes:\n\n* SELECT (xmltable(..)).* + regress tests\n* compilation and regress tests without --with-libxml\n\nRegards\n\nPavel\n\n\n>\n> pavel\n>\n>\n>>\n>> Regards\n>>\n>> Pavel\n>>\n>>\n>>>\n>>> --\n>>> Álvaro Herrera https://www.2ndQuadrant.com/\n>>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>>>\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","msg_date":"Sun, 22 Jan 2017 10:57:22 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Pavel Stehule wrote:\n\n> * SELECT (xmltable(..)).* + regress tests\n> * compilation and regress tests without --with-libxml\n\nThanks. I just realized that this is doing more work than necessary --\nI think it would be simpler to have tableexpr fill a tuplestore with the\nresults, instead of just expecting function execution to apply\nExecEvalExpr over and over to obtain the results. So evaluating a\ntableexpr returns just the tuplestore, which function evaluation can\nreturn as-is. That code doesn't use the value-per-call interface\nanyway.\n\nI also realized that the expr context callback is not called if there's\nan error, which leaves us without shutting down libxml properly. I\nadded PG_TRY around the fetchrow calls, but I'm not sure that's correct\neither, because there could be an error raised in other parts of the\ncode, after we've already emitted a few rows (for example out of\nmemory). I think the right way is to have PG_TRY around the execution\nof the whole thing rather than just row at a time; and the tuplestore\nmechanism helps us with that.\n\nI think it would be good to have a more complex test case in regress --\nlet's say there is a table with some simple XML values, then we use\nXMLFOREST (or maybe one of the table_to_xml functions) to generate a\nlarge document, and then XMLTABLE uses that document as input document.\n\nPlease fix.\n\n-- \n�lvaro Herrera https://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":"Tue, 24 Jan 2017 17:38:49 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Hi,\n\nOn 2017-01-24 17:38:49 -0300, Alvaro Herrera wrote:\n> +static Datum ExecEvalTableExpr(TableExprState *tstate, ExprContext *econtext,\n> +\t\t\t\t bool *isnull);\n> +static Datum ExecEvalTableExprFast(TableExprState *exprstate, ExprContext *econtext,\n> +\t\t\t\t\t bool *isNull);\n> +static Datum tabexprFetchRow(TableExprState *tstate, ExprContext *econtext,\n> +\t\t\t\tbool *isNull);\n> +static void tabexprInitialize(TableExprState *tstate, ExprContext *econtext,\n> +\t\t\t\t Datum doc);\n> +static void ShutdownTableExpr(Datum arg);\n\nTo me this (and a lot of the other code) hints quite strongly that\nexpression evalution is the wrong approach to implementing this. What\nyou're essentially doing is building a vulcano style scan node. Even if\nwe can this, we shouldn't double down on the bad decision to have these\nmagic expressions that return multiple rows. There's historical reason\nfor tSRFs, but we shouldn't add more weirdness like this.\n\nAndres\n\n\n-- \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 Jan 2017 14:37:44 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Andres Freund wrote:\n> Hi,\n> \n> On 2017-01-24 17:38:49 -0300, Alvaro Herrera wrote:\n> > +static Datum ExecEvalTableExpr(TableExprState *tstate, ExprContext *econtext,\n> > +\t\t\t\t bool *isnull);\n> > +static Datum ExecEvalTableExprFast(TableExprState *exprstate, ExprContext *econtext,\n> > +\t\t\t\t\t bool *isNull);\n> > +static Datum tabexprFetchRow(TableExprState *tstate, ExprContext *econtext,\n> > +\t\t\t\tbool *isNull);\n> > +static void tabexprInitialize(TableExprState *tstate, ExprContext *econtext,\n> > +\t\t\t\t Datum doc);\n> > +static void ShutdownTableExpr(Datum arg);\n> \n> To me this (and a lot of the other code) hints quite strongly that\n> expression evalution is the wrong approach to implementing this. What\n> you're essentially doing is building a vulcano style scan node. Even if\n> we can this, we shouldn't double down on the bad decision to have these\n> magic expressions that return multiple rows. There's historical reason\n> for tSRFs, but we shouldn't add more weirdness like this.\n\nThanks for giving it a look. I have long thought that this patch would\nbe at odds with your overall executor work.\n\nXMLTABLE is specified by the standard to return multiple rows ... but\nthen as far as my reading goes, it is only supposed to be supported in\nthe range table (FROM clause) not in the target list. I wonder if\nthis would end up better if we only tried to support it in RT. I asked\nPavel to implement it like that a few weeks ago, but ...\n\n-- \n�lvaro Herrera https://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":"Tue, 24 Jan 2017 21:32:56 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\n> Andres Freund wrote:\n> > Hi,\n> > \n> > On 2017-01-24 17:38:49 -0300, Alvaro Herrera wrote:\n> > > +static Datum ExecEvalTableExpr(TableExprState *tstate, ExprContext *econtext,\n> > > +\t\t\t\t bool *isnull);\n> > > +static Datum ExecEvalTableExprFast(TableExprState *exprstate, ExprContext *econtext,\n> > > +\t\t\t\t\t bool *isNull);\n> > > +static Datum tabexprFetchRow(TableExprState *tstate, ExprContext *econtext,\n> > > +\t\t\t\tbool *isNull);\n> > > +static void tabexprInitialize(TableExprState *tstate, ExprContext *econtext,\n> > > +\t\t\t\t Datum doc);\n> > > +static void ShutdownTableExpr(Datum arg);\n> > \n> > To me this (and a lot of the other code) hints quite strongly that\n> > expression evalution is the wrong approach to implementing this. What\n> > you're essentially doing is building a vulcano style scan node. Even if\n> > we can this, we shouldn't double down on the bad decision to have these\n> > magic expressions that return multiple rows. There's historical reason\n> > for tSRFs, but we shouldn't add more weirdness like this.\n> \n> Thanks for giving it a look. I have long thought that this patch would\n> be at odds with your overall executor work.\n\nNot fundamentally, but it makes it harder.\n\n\n> XMLTABLE is specified by the standard to return multiple rows ... but\n> then as far as my reading goes, it is only supposed to be supported in\n> the range table (FROM clause) not in the target list. I wonder if\n> this would end up better if we only tried to support it in RT. I asked\n> Pavel to implement it like that a few weeks ago, but ...\n\nRight - it makes sense in the FROM list - but then it should be an\nexecutor node, instead of some expression thingy.\n\nGreetings,\n\nAndres Freund\n\n\n-- \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 Jan 2017 16:35:11 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\n>> XMLTABLE is specified by the standard to return multiple rows ... but\n>> then as far as my reading goes, it is only supposed to be supported in\n>> the range table (FROM clause) not in the target list. I wonder if\n>> this would end up better if we only tried to support it in RT. I asked\n>> Pavel to implement it like that a few weeks ago, but ...\n\n> Right - it makes sense in the FROM list - but then it should be an\n> executor node, instead of some expression thingy.\n\n+1 --- we're out of the business of having simple expressions that\nreturn rowsets.\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, 24 Jan 2017 23:45:03 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Hi\n\n2017-01-25 1:35 GMT+01:00 Andres Freund <andres@anarazel.de>:\n\n> On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\n> > Andres Freund wrote:\n> > > Hi,\n> > >\n> > > On 2017-01-24 17:38:49 -0300, Alvaro Herrera wrote:\n> > > > +static Datum ExecEvalTableExpr(TableExprState *tstate, ExprContext\n> *econtext,\n> > > > + bool *isnull);\n> > > > +static Datum ExecEvalTableExprFast(TableExprState *exprstate,\n> ExprContext *econtext,\n> > > > + bool *isNull);\n> > > > +static Datum tabexprFetchRow(TableExprState *tstate, ExprContext\n> *econtext,\n> > > > + bool *isNull);\n> > > > +static void tabexprInitialize(TableExprState *tstate, ExprContext\n> *econtext,\n> > > > + Datum doc);\n> > > > +static void ShutdownTableExpr(Datum arg);\n> > >\n> > > To me this (and a lot of the other code) hints quite strongly that\n> > > expression evalution is the wrong approach to implementing this. What\n> > > you're essentially doing is building a vulcano style scan node. Even\n> if\n> > > we can this, we shouldn't double down on the bad decision to have these\n> > > magic expressions that return multiple rows. There's historical reason\n> > > for tSRFs, but we shouldn't add more weirdness like this.\n> >\n> > Thanks for giving it a look. I have long thought that this patch would\n> > be at odds with your overall executor work.\n>\n> Not fundamentally, but it makes it harder.\n>\n\nIf you plan to hold support SRFin target list, then nothing is different.\nIn last patch is executed under nodeProjectSet.\n\n\n>\n>\n> > XMLTABLE is specified by the standard to return multiple rows ... but\n> > then as far as my reading goes, it is only supposed to be supported in\n> > the range table (FROM clause) not in the target list. I wonder if\n> > this would end up better if we only tried to support it in RT. I asked\n> > Pavel to implement it like that a few weeks ago, but ...\n>\n> Right - it makes sense in the FROM list - but then it should be an\n> executor node, instead of some expression thingy.\n>\n\nThe XMLTABLE function is from user perspective, from implementation\nperspective a form of SRF function. I use own executor node, because fcinfo\nis complex already and not too enough to hold all information about result\ncolumns.\n\nThe implementation as RT doesn't reduce code - it is just moving to\ndifferent file.\n\nI'll try to explain my motivation. Please, check it and correct me if I am\nwrong. I don't keep on my implementation - just try to implement XMLTABLE\nbe consistent with another behave and be used all time without any\nsurprise.\n\n1. Any function that produces a content can be used in target list. We\nsupport SRF in target list and in FROM part. Why XMLTABLE should be a\nexception?\n\n2. In standard the XMLTABLE is placed only on FROM part - but standard\ndoesn't need to solve my question - there are not SRF functions allowed in\ntargetlist.\n\nIf there be a common decision so this inconsistency (in behave of this kind\nof functions) is expected, required - then I have not a problem to remove\nthis support from XMLTABLE.\n\nIn this moment I don't see a technical reason for this step - with last\nAndres changes the support of XMLTABLE in target list needs less than 40\nlines and there is not any special path for XMLTABLE only. Andres write\nsupport for SRF functions and SRF operator. TableExpr is third category.\n\nRegards\n\nPavel\n\n\n> Greetings,\n>\n> Andres Freund\n>\n\nHi2017-01-25 1:35 GMT+01:00 Andres Freund <andres@anarazel.de>:On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\n> Andres Freund wrote:\n> > Hi,\n> >\n> > On 2017-01-24 17:38:49 -0300, Alvaro Herrera wrote:\n> > > +static Datum ExecEvalTableExpr(TableExprState *tstate, ExprContext *econtext,\n> > > + bool *isnull);\n> > > +static Datum ExecEvalTableExprFast(TableExprState *exprstate, ExprContext *econtext,\n> > > + bool *isNull);\n> > > +static Datum tabexprFetchRow(TableExprState *tstate, ExprContext *econtext,\n> > > + bool *isNull);\n> > > +static void tabexprInitialize(TableExprState *tstate, ExprContext *econtext,\n> > > + Datum doc);\n> > > +static void ShutdownTableExpr(Datum arg);\n> >\n> > To me this (and a lot of the other code) hints quite strongly that\n> > expression evalution is the wrong approach to implementing this. What\n> > you're essentially doing is building a vulcano style scan node. Even if\n> > we can this, we shouldn't double down on the bad decision to have these\n> > magic expressions that return multiple rows. There's historical reason\n> > for tSRFs, but we shouldn't add more weirdness like this.\n>\n> Thanks for giving it a look. I have long thought that this patch would\n> be at odds with your overall executor work.\n\nNot fundamentally, but it makes it harder.If you plan to hold support SRFin target list, then nothing is different. In last patch is executed under nodeProjectSet. \n\n\n> XMLTABLE is specified by the standard to return multiple rows ... but\n> then as far as my reading goes, it is only supposed to be supported in\n> the range table (FROM clause) not in the target list. I wonder if\n> this would end up better if we only tried to support it in RT. I asked\n> Pavel to implement it like that a few weeks ago, but ...\n\nRight - it makes sense in the FROM list - but then it should be an\nexecutor node, instead of some expression thingy.The XMLTABLE function is from user perspective, from implementation perspective a form of SRF function. I use own executor node, because fcinfo is complex already and not too enough to hold all information about result columns. The implementation as RT doesn't reduce code - it is just moving to different file. I'll try to explain my motivation. Please, check it and correct me if I am wrong. I don't keep on my implementation - just try to implement XMLTABLE be consistent with another behave and be used all time without any surprise. 1. Any function that produces a content can be used in target list. We support SRF in target list and in FROM part. Why XMLTABLE should be a exception?2. In standard the XMLTABLE is placed only on FROM part - but standard doesn't need to solve my question - there are not SRF functions allowed in targetlist.If there be a common decision so this inconsistency (in behave of this kind of functions) is expected, required - then I have not a problem to remove this support from XMLTABLE.In this moment I don't see a technical reason for this step - with last Andres changes the support of XMLTABLE in target list needs less than 40 lines and there is not any special path for XMLTABLE only. Andres write support for SRF functions and SRF operator. TableExpr is third category.RegardsPavel\n\nGreetings,\n\nAndres Freund","msg_date":"Wed, 25 Jan 2017 05:45:24 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"2017-01-25 5:45 GMT+01:00 Tom Lane <tgl@sss.pgh.pa.us>:\n\n> Andres Freund <andres@anarazel.de> writes:\n> > On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\n> >> XMLTABLE is specified by the standard to return multiple rows ... but\n> >> then as far as my reading goes, it is only supposed to be supported in\n> >> the range table (FROM clause) not in the target list. I wonder if\n> >> this would end up better if we only tried to support it in RT. I asked\n> >> Pavel to implement it like that a few weeks ago, but ...\n>\n> > Right - it makes sense in the FROM list - but then it should be an\n> > executor node, instead of some expression thingy.\n>\n> +1 --- we're out of the business of having simple expressions that\n> return rowsets.\n>\n\nIf we do decision so this kind of function will have different behave than\nother SRF functions, then I remove support for this.\n\nThere are not technical reasons (maybe I don't see it) - last Andres\nchanges do well support for my code.\n\nRegards\n\nPavel\n\n\n>\n> regards, tom lane\n>\n\n2017-01-25 5:45 GMT+01:00 Tom Lane <tgl@sss.pgh.pa.us>:Andres Freund <andres@anarazel.de> writes:\n> On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\n>> XMLTABLE is specified by the standard to return multiple rows ... but\n>> then as far as my reading goes, it is only supposed to be supported in\n>> the range table (FROM clause) not in the target list. I wonder if\n>> this would end up better if we only tried to support it in RT. I asked\n>> Pavel to implement it like that a few weeks ago, but ...\n\n> Right - it makes sense in the FROM list - but then it should be an\n> executor node, instead of some expression thingy.\n\n+1 --- we're out of the business of having simple expressions that\nreturn rowsets.If we do decision so this kind of function will have different behave than other SRF functions, then I remove support for this.There are not technical reasons (maybe I don't see it) - last Andres changes do well support for my code.RegardsPavel \n\n regards, tom lane","msg_date":"Wed, 25 Jan 2017 05:48:58 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"2017-01-24 21:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> Pavel Stehule wrote:\n>\n> > * SELECT (xmltable(..)).* + regress tests\n> > * compilation and regress tests without --with-libxml\n>\n> Thanks. I just realized that this is doing more work than necessary --\n>\n\n?? I don't understand?\n\n\n> I think it would be simpler to have tableexpr fill a tuplestore with the\n> results, instead of just expecting function execution to apply\n> ExecEvalExpr over and over to obtain the results. So evaluating a\n> tableexpr returns just the tuplestore, which function evaluation can\n> return as-is. That code doesn't use the value-per-call interface\n> anyway.\n>\n\nok\n\n\n> I also realized that the expr context callback is not called if there's\n> an error, which leaves us without shutting down libxml properly. I\n> added PG_TRY around the fetchrow calls, but I'm not sure that's correct\n> either, because there could be an error raised in other parts of the\n> code, after we've already emitted a few rows (for example out of\n> memory). I think the right way is to have PG_TRY around the execution\n> of the whole thing rather than just row at a time; and the tuplestore\n> mechanism helps us with that.\n>\n\n\nok.\n\n\n\n>\n> I think it would be good to have a more complex test case in regress --\n> let's say there is a table with some simple XML values, then we use\n> XMLFOREST (or maybe one of the table_to_xml functions) to generate a\n> large document, and then XMLTABLE uses that document as input document.\n>\n> Please fix.\n>\n> --\n> Álvaro Herrera https://www.2ndQuadrant.com/\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>\n\n2017-01-24 21:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\n\n> * SELECT (xmltable(..)).* + regress tests\n> * compilation and regress tests without --with-libxml\n\nThanks. I just realized that this is doing more work than necessary --?? I don't understand? \nI think it would be simpler to have tableexpr fill a tuplestore with the\nresults, instead of just expecting function execution to apply\nExecEvalExpr over and over to obtain the results. So evaluating a\ntableexpr returns just the tuplestore, which function evaluation can\nreturn as-is. That code doesn't use the value-per-call interface\nanyway.ok\n\nI also realized that the expr context callback is not called if there's\nan error, which leaves us without shutting down libxml properly. I\nadded PG_TRY around the fetchrow calls, but I'm not sure that's correct\neither, because there could be an error raised in other parts of the\ncode, after we've already emitted a few rows (for example out of\nmemory). I think the right way is to have PG_TRY around the execution\nof the whole thing rather than just row at a time; and the tuplestore\nmechanism helps us with that.ok. \n\nI think it would be good to have a more complex test case in regress --\nlet's say there is a table with some simple XML values, then we use\nXMLFOREST (or maybe one of the table_to_xml functions) to generate a\nlarge document, and then XMLTABLE uses that document as input document.\n\nPlease fix.\n\n--\nÁlvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Wed, 25 Jan 2017 09:26:27 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Tom Lane wrote:\n> Andres Freund <andres@anarazel.de> writes:\n> > On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\n> >> XMLTABLE is specified by the standard to return multiple rows ... but\n> >> then as far as my reading goes, it is only supposed to be supported in\n> >> the range table (FROM clause) not in the target list. I wonder if\n> >> this would end up better if we only tried to support it in RT. I asked\n> >> Pavel to implement it like that a few weeks ago, but ...\n> \n> > Right - it makes sense in the FROM list - but then it should be an\n> > executor node, instead of some expression thingy.\n> \n> +1 --- we're out of the business of having simple expressions that\n> return rowsets.\n\nWell, that's it. I'm not committing this patch against two other\ncommitters' opinion, plus I was already on the fence about the\nimplementation anyway. I think you should just go with the flow and\nimplement this by creating nodeTableexprscan.c. It's not even\ndifficult.\n\n-- \n�lvaro Herrera https://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, 25 Jan 2017 11:07:47 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Hi,\n\nOn 2017-01-25 05:45:24 +0100, Pavel Stehule wrote:\n> 2017-01-25 1:35 GMT+01:00 Andres Freund <andres@anarazel.de>:\n> \n> > On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\n> > > Andres Freund wrote:\n> > > > Hi,\n> > > >\n> > > > On 2017-01-24 17:38:49 -0300, Alvaro Herrera wrote:\n> > > > > +static Datum ExecEvalTableExpr(TableExprState *tstate, ExprContext\n> > *econtext,\n> > > > > + bool *isnull);\n> > > > > +static Datum ExecEvalTableExprFast(TableExprState *exprstate,\n> > ExprContext *econtext,\n> > > > > + bool *isNull);\n> > > > > +static Datum tabexprFetchRow(TableExprState *tstate, ExprContext\n> > *econtext,\n> > > > > + bool *isNull);\n> > > > > +static void tabexprInitialize(TableExprState *tstate, ExprContext\n> > *econtext,\n> > > > > + Datum doc);\n> > > > > +static void ShutdownTableExpr(Datum arg);\n> > > >\n> > > > To me this (and a lot of the other code) hints quite strongly that\n> > > > expression evalution is the wrong approach to implementing this. What\n> > > > you're essentially doing is building a vulcano style scan node. Even\n> > if\n> > > > we can this, we shouldn't double down on the bad decision to have these\n> > > > magic expressions that return multiple rows. There's historical reason\n> > > > for tSRFs, but we shouldn't add more weirdness like this.\n> > >\n> > > Thanks for giving it a look. I have long thought that this patch would\n> > > be at odds with your overall executor work.\n> >\n> > Not fundamentally, but it makes it harder.\n> >\n> \n> If you plan to hold support SRFin target list, then nothing is different.\n> In last patch is executed under nodeProjectSet.\n\nIt is, because we suddenly need to call different functions - and I'm\nrevamping most of execQual to have an opcode dispatch based execution\nmodel (which then also can be JITed).\n\n\n> > > XMLTABLE is specified by the standard to return multiple rows ... but\n> > > then as far as my reading goes, it is only supposed to be supported in\n> > > the range table (FROM clause) not in the target list. I wonder if\n> > > this would end up better if we only tried to support it in RT. I asked\n> > > Pavel to implement it like that a few weeks ago, but ...\n> >\n> > Right - it makes sense in the FROM list - but then it should be an\n> > executor node, instead of some expression thingy.\n> >\n> \n> The XMLTABLE function is from user perspective, from implementation\n> perspective a form of SRF function. I use own executor node, because fcinfo\n> is complex already and not too enough to hold all information about result\n> columns.\n\n\n> The implementation as RT doesn't reduce code - it is just moving to\n> different file.\n\nYou're introducing a wholly separate callback system (TableExprRoutine)\nfor the new functionality. And that stuff is excruciatingly close to\nstuff that the normal executor already knows how to do.\n\n\n\n> I'll try to explain my motivation. Please, check it and correct me if I am\n> wrong. I don't keep on my implementation - just try to implement XMLTABLE\n> be consistent with another behave and be used all time without any\n> surprise.\n> \n> 1. Any function that produces a content can be used in target list. We\n> support SRF in target list and in FROM part. Why XMLTABLE should be a\n> exception?\n\ntargetlist SRFs were a big mistake. They cause a fair number of problems\ncode-wise. They permeated for a long while into bits of both planner and\nexecutor, where they really shouldn't belong. Even after the recent\nchanges there's a fair amount of uglyness associated with them. We\ncan't remove tSRFs for backward compatibility reasons, but that's not\ntrue for XMLTABLE\n\n\nGreetings,\n\nAndres Freund\n\n\n-- \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, 25 Jan 2017 12:31:28 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"> > >\n> >\n> > If you plan to hold support SRFin target list, then nothing is different.\n> > In last patch is executed under nodeProjectSet.\n>\n> It is, because we suddenly need to call different functions - and I'm\n> revamping most of execQual to have an opcode dispatch based execution\n> model (which then also can be JITed).\n>\n\n> > > > XMLTABLE is specified by the standard to return multiple rows ... but\n> > > > then as far as my reading goes, it is only supposed to be supported\n> in\n> > > > the range table (FROM clause) not in the target list. I wonder if\n> > > > this would end up better if we only tried to support it in RT. I\n> asked\n> > > > Pavel to implement it like that a few weeks ago, but ...\n> > >\n> > > Right - it makes sense in the FROM list - but then it should be an\n> > > executor node, instead of some expression thingy.\n> > >\n> >\n> > The XMLTABLE function is from user perspective, from implementation\n> > perspective a form of SRF function. I use own executor node, because\n> fcinfo\n> > is complex already and not too enough to hold all information about\n> result\n> > columns.\n>\n>\n> > The implementation as RT doesn't reduce code - it is just moving to\n> > different file.\n>\n> You're introducing a wholly separate callback system (TableExprRoutine)\n> for the new functionality. And that stuff is excruciatingly close to\n> stuff that the normal executor already knows how to do.\n>\n\nThese callbacks are related to isolation TableExpr infrastructure and\nTableExpr implementation - This design is prepared for reusing for\nJSON_TABLE function.\n\nAny placing of TableExpr code should not impact this callback system (Or I\nam absolutely out and executor is able do some work what is hidden to me).\n\n\n>\n>\n>\n> > I'll try to explain my motivation. Please, check it and correct me if I\n> am\n> > wrong. I don't keep on my implementation - just try to implement XMLTABLE\n> > be consistent with another behave and be used all time without any\n> > surprise.\n> >\n> > 1. Any function that produces a content can be used in target list. We\n> > support SRF in target list and in FROM part. Why XMLTABLE should be a\n> > exception?\n>\n> targetlist SRFs were a big mistake. They cause a fair number of problems\n> code-wise. They permeated for a long while into bits of both planner and\n> executor, where they really shouldn't belong. Even after the recent\n> changes there's a fair amount of uglyness associated with them. We\n> can't remove tSRFs for backward compatibility reasons, but that's not\n> true for XMLTABLE\n>\n>\n>\nok\n\nI afraid when I cannot to reuse a SRF infrastructure, I have to reimplement\nit partially :( - mainly for usage in \"ROWS FROM ()\"\n\n\n\nGreetings,\n>\n> Andres Freund\n>\n\n\n> >\n>\n> If you plan to hold support SRFin target list, then nothing is different.\n> In last patch is executed under nodeProjectSet.\n\nIt is, because we suddenly need to call different functions - and I'm\nrevamping most of execQual to have an opcode dispatch based execution\nmodel (which then also can be JITed). \n\n> > > XMLTABLE is specified by the standard to return multiple rows ... but\n> > > then as far as my reading goes, it is only supposed to be supported in\n> > > the range table (FROM clause) not in the target list. I wonder if\n> > > this would end up better if we only tried to support it in RT. I asked\n> > > Pavel to implement it like that a few weeks ago, but ...\n> >\n> > Right - it makes sense in the FROM list - but then it should be an\n> > executor node, instead of some expression thingy.\n> >\n>\n> The XMLTABLE function is from user perspective, from implementation\n> perspective a form of SRF function. I use own executor node, because fcinfo\n> is complex already and not too enough to hold all information about result\n> columns.\n\n\n> The implementation as RT doesn't reduce code - it is just moving to\n> different file.\n\nYou're introducing a wholly separate callback system (TableExprRoutine)\nfor the new functionality. And that stuff is excruciatingly close to\nstuff that the normal executor already knows how to do.These callbacks are related to isolation TableExpr infrastructure and TableExpr implementation - This design is prepared for reusing for JSON_TABLE function.Any placing of TableExpr code should not impact this callback system (Or I am absolutely out and executor is able do some work what is hidden to me). \n\n\n\n> I'll try to explain my motivation. Please, check it and correct me if I am\n> wrong. I don't keep on my implementation - just try to implement XMLTABLE\n> be consistent with another behave and be used all time without any\n> surprise.\n>\n> 1. Any function that produces a content can be used in target list. We\n> support SRF in target list and in FROM part. Why XMLTABLE should be a\n> exception?\n\ntargetlist SRFs were a big mistake. They cause a fair number of problems\ncode-wise. They permeated for a long while into bits of both planner and\nexecutor, where they really shouldn't belong. Even after the recent\nchanges there's a fair amount of uglyness associated with them. We\ncan't remove tSRFs for backward compatibility reasons, but that's not\ntrue for XMLTABLE\n\nokI afraid when I cannot to reuse a SRF infrastructure, I have to reimplement it partially :( - mainly for usage in \"ROWS FROM ()\" \nGreetings,\n\nAndres Freund","msg_date":"Wed, 25 Jan 2017 22:38:25 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Hi,\n\n> > > I'll try to explain my motivation. Please, check it and correct me if I\n> > am\n> > > wrong. I don't keep on my implementation - just try to implement XMLTABLE\n> > > be consistent with another behave and be used all time without any\n> > > surprise.\n> > >\n> > > 1. Any function that produces a content can be used in target list. We\n> > > support SRF in target list and in FROM part. Why XMLTABLE should be a\n> > > exception?\n> >\n> > targetlist SRFs were a big mistake. They cause a fair number of problems\n> > code-wise. They permeated for a long while into bits of both planner and\n> > executor, where they really shouldn't belong. Even after the recent\n> > changes there's a fair amount of uglyness associated with them. We\n> > can't remove tSRFs for backward compatibility reasons, but that's not\n> > true for XMLTABLE\n> >\n> >\n> >\n> ok\n> \n> I afraid when I cannot to reuse a SRF infrastructure, I have to reimplement\n> it partially :( - mainly for usage in \"ROWS FROM ()\"\n\nHuh?\n\nGreetings,\n\nAndres Freund\n\n\n-- \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, 25 Jan 2017 13:40:22 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"2017-01-25 22:40 GMT+01:00 Andres Freund <andres@anarazel.de>:\n\n> Hi,\n>\n> > > > I'll try to explain my motivation. Please, check it and correct me\n> if I\n> > > am\n> > > > wrong. I don't keep on my implementation - just try to implement\n> XMLTABLE\n> > > > be consistent with another behave and be used all time without any\n> > > > surprise.\n> > > >\n> > > > 1. Any function that produces a content can be used in target list.\n> We\n> > > > support SRF in target list and in FROM part. Why XMLTABLE should be a\n> > > > exception?\n> > >\n> > > targetlist SRFs were a big mistake. They cause a fair number of\n> problems\n> > > code-wise. They permeated for a long while into bits of both planner\n> and\n> > > executor, where they really shouldn't belong. Even after the recent\n> > > changes there's a fair amount of uglyness associated with them. We\n> > > can't remove tSRFs for backward compatibility reasons, but that's not\n> > > true for XMLTABLE\n> > >\n> > >\n> > >\n> > ok\n> >\n> > I afraid when I cannot to reuse a SRF infrastructure, I have to\n> reimplement\n> > it partially :( - mainly for usage in \"ROWS FROM ()\"\n>\n\nThe TableExpr implementation is based on SRF now. You and Alvaro propose\nindependent implementation like generic executor node. I am sceptic so\nFunctionScan supports reading from generic executor node.\n\nRegards\n\nPavel\n\n\n> Huh?\n>\n> Greetings,\n>\n> Andres Freund\n>\n\n2017-01-25 22:40 GMT+01:00 Andres Freund <andres@anarazel.de>:Hi,\n\n> > > I'll try to explain my motivation. Please, check it and correct me if I\n> > am\n> > > wrong. I don't keep on my implementation - just try to implement XMLTABLE\n> > > be consistent with another behave and be used all time without any\n> > > surprise.\n> > >\n> > > 1. Any function that produces a content can be used in target list. We\n> > > support SRF in target list and in FROM part. Why XMLTABLE should be a\n> > > exception?\n> >\n> > targetlist SRFs were a big mistake. They cause a fair number of problems\n> > code-wise. They permeated for a long while into bits of both planner and\n> > executor, where they really shouldn't belong. Even after the recent\n> > changes there's a fair amount of uglyness associated with them. We\n> > can't remove tSRFs for backward compatibility reasons, but that's not\n> > true for XMLTABLE\n> >\n> >\n> >\n> ok\n>\n> I afraid when I cannot to reuse a SRF infrastructure, I have to reimplement\n> it partially :( - mainly for usage in \"ROWS FROM ()\"The TableExpr implementation is based on SRF now. You and Alvaro propose independent implementation like generic executor node. I am sceptic so FunctionScan supports reading from generic executor node.RegardsPavel\n\nHuh?\n\nGreetings,\n\nAndres Freund","msg_date":"Wed, 25 Jan 2017 22:51:37 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"On 2017-01-25 22:51:37 +0100, Pavel Stehule wrote:\n> 2017-01-25 22:40 GMT+01:00 Andres Freund <andres@anarazel.de>:\n> > > I afraid when I cannot to reuse a SRF infrastructure, I have to\n> > reimplement\n> > > it partially :( - mainly for usage in \"ROWS FROM ()\"\n> >\n> \n> The TableExpr implementation is based on SRF now. You and Alvaro propose\n> independent implementation like generic executor node. I am sceptic so\n> FunctionScan supports reading from generic executor node.\n\nWhy would it need to?\n\n\n-- \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, 25 Jan 2017 14:33:56 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"2017-01-25 23:33 GMT+01:00 Andres Freund <andres@anarazel.de>:\n\n> On 2017-01-25 22:51:37 +0100, Pavel Stehule wrote:\n> > 2017-01-25 22:40 GMT+01:00 Andres Freund <andres@anarazel.de>:\n> > > > I afraid when I cannot to reuse a SRF infrastructure, I have to\n> > > reimplement\n> > > > it partially :( - mainly for usage in \"ROWS FROM ()\"\n> > >\n> >\n> > The TableExpr implementation is based on SRF now. You and Alvaro propose\n> > independent implementation like generic executor node. I am sceptic so\n> > FunctionScan supports reading from generic executor node.\n>\n> Why would it need to?\n>\n\nSimply - due consistency with any other functions that can returns rows.\n\nMaybe I don't understand to Alvaro proposal well - I have a XMLTABLE\nfunction - TableExpr that looks like SRF function, has similar behave -\nreturns more rows, but should be significantly different implemented, and\nshould to have different limits - should not be used there and there ... It\nis hard to see consistency there for me.\n\nRegards\n\nPavel\n\n2017-01-25 23:33 GMT+01:00 Andres Freund <andres@anarazel.de>:On 2017-01-25 22:51:37 +0100, Pavel Stehule wrote:\n> 2017-01-25 22:40 GMT+01:00 Andres Freund <andres@anarazel.de>:\n> > > I afraid when I cannot to reuse a SRF infrastructure, I have to\n> > reimplement\n> > > it partially :( - mainly for usage in \"ROWS FROM ()\"\n> >\n>\n> The TableExpr implementation is based on SRF now. You and Alvaro propose\n> independent implementation like generic executor node. I am sceptic so\n> FunctionScan supports reading from generic executor node.\n\nWhy would it need to?Simply - due consistency with any other functions that can returns rows.Maybe I don't understand to Alvaro proposal well - I have a XMLTABLE function - TableExpr that looks like SRF function, has similar behave - returns more rows, but should be significantly different implemented, and should to have different limits - should not be used there and there ... It is hard to see consistency there for me.RegardsPavel","msg_date":"Wed, 25 Jan 2017 23:43:22 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"2017-01-25 15:07 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> Tom Lane wrote:\n> > Andres Freund <andres@anarazel.de> writes:\n> > > On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\n> > >> XMLTABLE is specified by the standard to return multiple rows ... but\n> > >> then as far as my reading goes, it is only supposed to be supported in\n> > >> the range table (FROM clause) not in the target list. I wonder if\n> > >> this would end up better if we only tried to support it in RT. I\n> asked\n> > >> Pavel to implement it like that a few weeks ago, but ...\n> >\n> > > Right - it makes sense in the FROM list - but then it should be an\n> > > executor node, instead of some expression thingy.\n> >\n> > +1 --- we're out of the business of having simple expressions that\n> > return rowsets.\n>\n> Well, that's it. I'm not committing this patch against two other\n> committers' opinion, plus I was already on the fence about the\n> implementation anyway. I think you should just go with the flow and\n> implement this by creating nodeTableexprscan.c. It's not even\n> difficult.\n>\n\nI am playing with this and the patch looks about 15kB longer - just due\nimplementation basic scan functionality - and I didn't touch a planner.\n\nI am not happy from this - still I have a feeling so I try to reimplement\nreduced SRF.\n\nRegards\n\nPavel\n\n>\n> --\n> Álvaro Herrera https://www.2ndQuadrant.com/\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>\n\n2017-01-25 15:07 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Tom Lane wrote:\n> Andres Freund <andres@anarazel.de> writes:\n> > On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\n> >> XMLTABLE is specified by the standard to return multiple rows ... but\n> >> then as far as my reading goes, it is only supposed to be supported in\n> >> the range table (FROM clause) not in the target list. I wonder if\n> >> this would end up better if we only tried to support it in RT. I asked\n> >> Pavel to implement it like that a few weeks ago, but ...\n>\n> > Right - it makes sense in the FROM list - but then it should be an\n> > executor node, instead of some expression thingy.\n>\n> +1 --- we're out of the business of having simple expressions that\n> return rowsets.\n\nWell, that's it. I'm not committing this patch against two other\ncommitters' opinion, plus I was already on the fence about the\nimplementation anyway. I think you should just go with the flow and\nimplement this by creating nodeTableexprscan.c. It's not even\ndifficult.I am playing with this and the patch looks about 15kB longer - just due implementation basic scan functionality - and I didn't touch a planner. I am not happy from this - still I have a feeling so I try to reimplement reduced SRF. RegardsPavel\n\n--\nÁlvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Thu, 26 Jan 2017 01:51:00 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Hi\n\n2017-01-25 15:07 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> Tom Lane wrote:\n> > Andres Freund <andres@anarazel.de> writes:\n> > > On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\n> > >> XMLTABLE is specified by the standard to return multiple rows ... but\n> > >> then as far as my reading goes, it is only supposed to be supported in\n> > >> the range table (FROM clause) not in the target list. I wonder if\n> > >> this would end up better if we only tried to support it in RT. I\n> asked\n> > >> Pavel to implement it like that a few weeks ago, but ...\n> >\n> > > Right - it makes sense in the FROM list - but then it should be an\n> > > executor node, instead of some expression thingy.\n> >\n> > +1 --- we're out of the business of having simple expressions that\n> > return rowsets.\n>\n> Well, that's it. I'm not committing this patch against two other\n> committers' opinion, plus I was already on the fence about the\n> implementation anyway. I think you should just go with the flow and\n> implement this by creating nodeTableexprscan.c. It's not even\n> difficult.\n>\n\nI am sending new version - it is based on own executor scan node and\ntuplestore.\n\nSome now obsolete regress tests removed, some new added.\n\nThe executor code (memory context usage) should be cleaned little bit - but\nother code should be good.\n\nRegards\n\nPavel\n\n\n> --\n> Álvaro Herrera https://www.2ndQuadrant.com/\n> PostgreSQL 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":"Mon, 30 Jan 2017 20:22:04 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Pavel Stehule wrote:\n\n> I am sending new version - it is based on own executor scan node and\n> tuplestore.\n> \n> Some now obsolete regress tests removed, some new added.\n> \n> The executor code (memory context usage) should be cleaned little bit - but\n> other code should be good.\n\nI think you forgot nodeTableFuncscan.c.\n\n-- \n�lvaro Herrera https://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":"Mon, 30 Jan 2017 16:38:59 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Hi\n\n2017-01-30 20:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> Pavel Stehule wrote:\n>\n> > I am sending new version - it is based on own executor scan node and\n> > tuplestore.\n> >\n> > Some now obsolete regress tests removed, some new added.\n> >\n> > The executor code (memory context usage) should be cleaned little bit -\n> but\n> > other code should be good.\n>\n> I think you forgot nodeTableFuncscan.c.\n>\n\ntrue, I am sorry\n\nattached\n\nRegards\n\nPavel\n\n\n>\n> --\n> Álvaro Herrera https://www.2ndQuadrant.com/\n> PostgreSQL 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":"Mon, 30 Jan 2017 21:18:53 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"On Tue, Jan 31, 2017 at 5:18 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:\n> true, I am sorry\n\nLast status is a new patch and no reviews. On top of that this thread\nis quite active. So moved to next CF. Pavel, please be careful about\nthe status of the patch on the CF app, it was set to \"waiting on\nauthor\"...\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":"Tue, 31 Jan 2017 15:33:05 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"2017-01-24 21:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> Pavel Stehule wrote:\n>\n> > * SELECT (xmltable(..)).* + regress tests\n> > * compilation and regress tests without --with-libxml\n>\n> Thanks. I just realized that this is doing more work than necessary --\n> I think it would be simpler to have tableexpr fill a tuplestore with the\n> results, instead of just expecting function execution to apply\n> ExecEvalExpr over and over to obtain the results. So evaluating a\n> tableexpr returns just the tuplestore, which function evaluation can\n> return as-is. That code doesn't use the value-per-call interface\n> anyway.\n>\n> I also realized that the expr context callback is not called if there's\n> an error, which leaves us without shutting down libxml properly. I\n> added PG_TRY around the fetchrow calls, but I'm not sure that's correct\n> either, because there could be an error raised in other parts of the\n> code, after we've already emitted a few rows (for example out of\n> memory). I think the right way is to have PG_TRY around the execution\n> of the whole thing rather than just row at a time; and the tuplestore\n> mechanism helps us with that.\n>\n> I think it would be good to have a more complex test case in regress --\n> let's say there is a table with some simple XML values, then we use\n> XMLFOREST (or maybe one of the table_to_xml functions) to generate a\n> large document, and then XMLTABLE uses that document as input document.\n>\n\nI have a 16K lines long real XML 6.MB. Probably we would not to append it\nto regress tests.\n\nIt is really fast - original customer implementation 20min, nested our\nxpath implementation 10 sec, PLPython xml reader 5 sec, xmltable 400ms\n\nI have a plan to create tests based on pg_proc and CTE - if all works, then\nthe query must be empty\n\nwith x as (select proname, proowner, procost, pronargs,\narray_to_string(proargnames,',') as proargnames,\narray_to_string(proargtypes,',') as proargtypes from pg_proc), y as (select\nxmlelement(name proc, xmlforest(proname, proowner, procost, pronargs,\nproargnames, proargtypes)) as proc from x), z as (select xmltable.* from y,\nlateral xmltable('/proc' passing proc columns proname name, proowner oid,\nprocost float, pronargs int, proargnames text, proargtypes text)) select *\nfrom z except select * from x;\n\n\n\n>\n> Please fix.\n>\n> --\n> Álvaro Herrera https://www.2ndQuadrant.com/\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>\n\n2017-01-24 21:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\n\n> * SELECT (xmltable(..)).* + regress tests\n> * compilation and regress tests without --with-libxml\n\nThanks. I just realized that this is doing more work than necessary --\nI think it would be simpler to have tableexpr fill a tuplestore with the\nresults, instead of just expecting function execution to apply\nExecEvalExpr over and over to obtain the results. So evaluating a\ntableexpr returns just the tuplestore, which function evaluation can\nreturn as-is. That code doesn't use the value-per-call interface\nanyway.\n\nI also realized that the expr context callback is not called if there's\nan error, which leaves us without shutting down libxml properly. I\nadded PG_TRY around the fetchrow calls, but I'm not sure that's correct\neither, because there could be an error raised in other parts of the\ncode, after we've already emitted a few rows (for example out of\nmemory). I think the right way is to have PG_TRY around the execution\nof the whole thing rather than just row at a time; and the tuplestore\nmechanism helps us with that.\n\nI think it would be good to have a more complex test case in regress --\nlet's say there is a table with some simple XML values, then we use\nXMLFOREST (or maybe one of the table_to_xml functions) to generate a\nlarge document, and then XMLTABLE uses that document as input document.I have a 16K lines long real XML 6.MB. Probably we would not to append it to regress tests.It is really fast - original customer implementation 20min, nested our xpath implementation 10 sec, PLPython xml reader 5 sec, xmltable 400msI have a plan to create tests based on pg_proc and CTE - if all works, then the query must be emptywith x as (select proname, proowner, procost, pronargs, array_to_string(proargnames,',') as proargnames, array_to_string(proargtypes,',') as proargtypes from pg_proc), y as (select xmlelement(name proc, xmlforest(proname, proowner, procost, pronargs, proargnames, proargtypes)) as proc from x), z as (select xmltable.* from y, lateral xmltable('/proc' passing proc columns proname name, proowner oid, procost float, pronargs int, proargnames text, proargtypes text)) select * from z except select * from x; \n\nPlease fix.\n\n--\nÁlvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Tue, 31 Jan 2017 11:26:32 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Pavel Stehule wrote:\n> 2017-01-24 21:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> > I think it would be good to have a more complex test case in regress --\n> > let's say there is a table with some simple XML values, then we use\n> > XMLFOREST (or maybe one of the table_to_xml functions) to generate a\n> > large document, and then XMLTABLE uses that document as input document.\n> \n> I have a 16K lines long real XML 6.MB. Probably we would not to append it\n> to regress tests.\n> \n> It is really fast - original customer implementation 20min, nested our\n> xpath implementation 10 sec, PLPython xml reader 5 sec, xmltable 400ms\n\nThat's great numbers, kudos for the hard work here. That will make for\na nice headline in pg10 PR materials. But what I was getting at is that\nI would like to exercise a bit more of the expression handling in\nxmltable execution, to make sure it doesn't handle just string literals.\n\n> I have a plan to create tests based on pg_proc and CTE - if all works, then\n> the query must be empty\n> \n> with x as (select proname, proowner, procost, pronargs,\n> array_to_string(proargnames,',') as proargnames,\n> array_to_string(proargtypes,',') as proargtypes from pg_proc), y as (select\n> xmlelement(name proc, xmlforest(proname, proowner, procost, pronargs,\n> proargnames, proargtypes)) as proc from x), z as (select xmltable.* from y,\n> lateral xmltable('/proc' passing proc columns proname name, proowner oid,\n> procost float, pronargs int, proargnames text, proargtypes text)) select *\n> from z except select * from x;\n\nNice one :-)\n\n-- \n�lvaro Herrera https://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":"Tue, 31 Jan 2017 10:57:07 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"2017-01-31 14:57 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> Pavel Stehule wrote:\n> > 2017-01-24 21:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n>\n> > > I think it would be good to have a more complex test case in regress --\n> > > let's say there is a table with some simple XML values, then we use\n> > > XMLFOREST (or maybe one of the table_to_xml functions) to generate a\n> > > large document, and then XMLTABLE uses that document as input document.\n> >\n> > I have a 16K lines long real XML 6.MB. Probably we would not to append it\n> > to regress tests.\n> >\n> > It is really fast - original customer implementation 20min, nested our\n> > xpath implementation 10 sec, PLPython xml reader 5 sec, xmltable 400ms\n>\n> That's great numbers, kudos for the hard work here. That will make for\n> a nice headline in pg10 PR materials. But what I was getting at is that\n> I would like to exercise a bit more of the expression handling in\n> xmltable execution, to make sure it doesn't handle just string literals.\n>\n\nI'll try to write some more dynamic examples.\n\n\n>\n> > I have a plan to create tests based on pg_proc and CTE - if all works,\n> then\n> > the query must be empty\n> >\n> > with x as (select proname, proowner, procost, pronargs,\n> > array_to_string(proargnames,',') as proargnames,\n> > array_to_string(proargtypes,',') as proargtypes from pg_proc), y as\n> (select\n> > xmlelement(name proc, xmlforest(proname, proowner, procost, pronargs,\n> > proargnames, proargtypes)) as proc from x), z as (select xmltable.* from\n> y,\n> > lateral xmltable('/proc' passing proc columns proname name, proowner oid,\n> > procost float, pronargs int, proargnames text, proargtypes text)) select\n> *\n> > from z except select * from x;\n>\n> Nice one :-)\n>\n> --\n> Álvaro Herrera https://www.2ndQuadrant.com/\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>\n\n2017-01-31 14:57 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\n> 2017-01-24 21:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> > I think it would be good to have a more complex test case in regress --\n> > let's say there is a table with some simple XML values, then we use\n> > XMLFOREST (or maybe one of the table_to_xml functions) to generate a\n> > large document, and then XMLTABLE uses that document as input document.\n>\n> I have a 16K lines long real XML 6.MB. Probably we would not to append it\n> to regress tests.\n>\n> It is really fast - original customer implementation 20min, nested our\n> xpath implementation 10 sec, PLPython xml reader 5 sec, xmltable 400ms\n\nThat's great numbers, kudos for the hard work here. That will make for\na nice headline in pg10 PR materials. But what I was getting at is that\nI would like to exercise a bit more of the expression handling in\nxmltable execution, to make sure it doesn't handle just string literals.I'll try to write some more dynamic examples. \n\n> I have a plan to create tests based on pg_proc and CTE - if all works, then\n> the query must be empty\n>\n> with x as (select proname, proowner, procost, pronargs,\n> array_to_string(proargnames,',') as proargnames,\n> array_to_string(proargtypes,',') as proargtypes from pg_proc), y as (select\n> xmlelement(name proc, xmlforest(proname, proowner, procost, pronargs,\n> proargnames, proargtypes)) as proc from x), z as (select xmltable.* from y,\n> lateral xmltable('/proc' passing proc columns proname name, proowner oid,\n> procost float, pronargs int, proargnames text, proargtypes text)) select *\n> from z except select * from x;\n\nNice one :-)\n\n--\nÁlvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Tue, 31 Jan 2017 16:32:57 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Hi\n\n2017-01-31 14:57 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> Pavel Stehule wrote:\n> > 2017-01-24 21:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n>\n> > > I think it would be good to have a more complex test case in regress --\n> > > let's say there is a table with some simple XML values, then we use\n> > > XMLFOREST (or maybe one of the table_to_xml functions) to generate a\n> > > large document, and then XMLTABLE uses that document as input document.\n> >\n> > I have a 16K lines long real XML 6.MB. Probably we would not to append it\n> > to regress tests.\n> >\n> > It is really fast - original customer implementation 20min, nested our\n> > xpath implementation 10 sec, PLPython xml reader 5 sec, xmltable 400ms\n>\n> That's great numbers, kudos for the hard work here. That will make for\n> a nice headline in pg10 PR materials. But what I was getting at is that\n> I would like to exercise a bit more of the expression handling in\n> xmltable execution, to make sure it doesn't handle just string literals.\n>\n\ndone\n\n\n>\n> > I have a plan to create tests based on pg_proc and CTE - if all works,\n> then\n> > the query must be empty\n> >\n> > with x as (select proname, proowner, procost, pronargs,\n> > array_to_string(proargnames,',') as proargnames,\n> > array_to_string(proargtypes,',') as proargtypes from pg_proc), y as\n> (select\n> > xmlelement(name proc, xmlforest(proname, proowner, procost, pronargs,\n> > proargnames, proargtypes)) as proc from x), z as (select xmltable.* from\n> y,\n> > lateral xmltable('/proc' passing proc columns proname name, proowner oid,\n> > procost float, pronargs int, proargnames text, proargtypes text)) select\n> *\n> > from z except select * from x;\n>\n> Nice one :-)\n>\n\nplease see attached patch\n\n* enhanced regress tests\n* clean memory context work\n\nRegards\n\nPavel\n\n\n>\n> --\n> Álvaro Herrera https://www.2ndQuadrant.com/\n> PostgreSQL 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":"Tue, 31 Jan 2017 22:20:49 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Hi\n\n\n> please see attached patch\n>\n> * enhanced regress tests\n> * clean memory context work\n>\n\nnew update\n\nfix a bug in string compare\nfix some typo and obsolete comments\n\nRegards\n\nPavel\n\n\n>\n> Regards\n>\n> Pavel\n>\n>\n>>\n>> --\n>> Álvaro Herrera https://www.2ndQuadrant.com/\n>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\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","msg_date":"Thu, 16 Feb 2017 06:38:04 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Hi\n\n2017-02-16 6:38 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\n\n> Hi\n>\n>\n>> please see attached patch\n>>\n>> * enhanced regress tests\n>> * clean memory context work\n>>\n>\n> new update\n>\n> fix a bug in string compare\n> fix some typo and obsolete comments\n>\n> Regards\n>\n\nsome minor but interesting fix.\n\nI found so some xml values imported via recv function can have\ninconsistency between header encoding and used encoding. Internally the\nheader encoding is removed - see xml_out function.\n\nSo now, when I have to prepare data for libxml2, I don't do direct cast,\nbut I use xml_out_internal instead. Maybe this technique should be used\nelsewhere? Same issue I see on xpath function.\n\nSolved issue is not too often probably - the some different encoding than\nutf8 should be used in XML document and XML document should be loaded with\nrecv function.\n\nRegards\n\nPavel\n\n\n\n>\n> Pavel\n>\n>\n>>\n>> Regards\n>>\n>> Pavel\n>>\n>>\n>>>\n>>> --\n>>> Álvaro Herrera https://www.2ndQuadrant.com/\n>>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>>>\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","msg_date":"Mon, 20 Feb 2017 19:12:55 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"\nI've been giving this a look. I started by tweaking the docs once\nagain, and while verifying that the example works as expected, I\nreplayed what I have in sgml:\n\n... begin SGML paste ...\n <para>\n For example, given the following XML document:\n <screen><![CDATA[\n<ROWS>\n <ROW id=\"1\">\n <COUNTRY_ID>AU</COUNTRY_ID>\n <COUNTRY_NAME>Australia</COUNTRY_NAME>\n </ROW>\n <ROW id=\"5\">\n <COUNTRY_ID>JP</COUNTRY_ID>\n <COUNTRY_NAME>Japan</COUNTRY_NAME>\n <PREMIER_NAME>Sinzo Abe</PREMIER_NAME>\n </ROW>\n <ROW id=\"6\">\n <COUNTRY_ID>SG</COUNTRY_ID>\n <COUNTRY_NAME>Singapore</COUNTRY_NAME>\n <SIZE unit=\"km\">791</SIZE>\n </ROW>\n</ROWS>\n]]></screen>\n\n the following query produces the result shown below:\n\n<screen><![CDATA[\nSELECT xmltable.*\n FROM (SELECT data FROM xmldata) x,\n LATERAL xmltable('//ROWS/ROW'\n PASSING data\n COLUMNS id int PATH '@id',\n ordinality FOR ORDINALITY,\n country_name text PATH 'COUNTRY_NAME',\n country_id text PATH 'COUNTRY_ID',\n size float PATH 'SIZE[@unit = \"km\"]/text()',\n unit text PATH 'SIZE/@unit',\n premier_name text PATH 'PREMIER_NAME' DEFAULT 'not specified');\n... end SGML paste ...\n\n\nBut the query doesn't actually return a table, but instead it fails with\nthis error:\n ERROR: invalid input syntax for type double precision: \"\"\nThis is because of the \"size\" column (if I remove SIZE from the COLUMNS\nclause, the query returns correctly). Apparently, for the rows where\nSIZE is not given, we try to inssert an empty string instead of a NULL\nvalue, which is what I expected.\n\nI'm using your v44 code, but trimmed both the XML document used in SGML\nas well as modified the query slightly to show additional features. But\nthose changes should not cause the above error ...\n\n-- \n�lvaro Herrera https://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, 1 Mar 2017 21:12:45 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Hi\n\n2017-03-02 1:12 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n>\n> I've been giving this a look. I started by tweaking the docs once\n> again, and while verifying that the example works as expected, I\n> replayed what I have in sgml:\n>\n> ... begin SGML paste ...\n> <para>\n> For example, given the following XML document:\n> <screen><![CDATA[\n> <ROWS>\n> <ROW id=\"1\">\n> <COUNTRY_ID>AU</COUNTRY_ID>\n> <COUNTRY_NAME>Australia</COUNTRY_NAME>\n> </ROW>\n> <ROW id=\"5\">\n> <COUNTRY_ID>JP</COUNTRY_ID>\n> <COUNTRY_NAME>Japan</COUNTRY_NAME>\n> <PREMIER_NAME>Sinzo Abe</PREMIER_NAME>\n> </ROW>\n> <ROW id=\"6\">\n> <COUNTRY_ID>SG</COUNTRY_ID>\n> <COUNTRY_NAME>Singapore</COUNTRY_NAME>\n> <SIZE unit=\"km\">791</SIZE>\n> </ROW>\n> </ROWS>\n> ]]></screen>\n>\n> the following query produces the result shown below:\n>\n> <screen><![CDATA[\n> SELECT xmltable.*\n> FROM (SELECT data FROM xmldata) x,\n> LATERAL xmltable('//ROWS/ROW'\n> PASSING data\n> COLUMNS id int PATH '@id',\n> ordinality FOR ORDINALITY,\n> country_name text PATH 'COUNTRY_NAME',\n> country_id text PATH 'COUNTRY_ID',\n> size float PATH 'SIZE[@unit =\n> \"km\"]/text()',\n> unit text PATH 'SIZE/@unit',\n> premier_name text PATH 'PREMIER_NAME'\n> DEFAULT 'not specified');\n> ... end SGML paste ...\n>\n>\n> But the query doesn't actually return a table, but instead it fails with\n> this error:\n> ERROR: invalid input syntax for type double precision: \"\"\n> This is because of the \"size\" column (if I remove SIZE from the COLUMNS\n> clause, the query returns correctly). Apparently, for the rows where\n> SIZE is not given, we try to inssert an empty string instead of a NULL\n> value, which is what I expected.\n>\n> I'm using your v44 code, but trimmed both the XML document used in SGML\n> as well as modified the query slightly to show additional features. But\n> those changes should not cause the above error ...\n>\n\nThe example in doc is obsolete. Following example works without problems.\n\n SELECT xmltable.*\n\n FROM (SELECT data FROM xmldata) x,\n LATERAL xmltable('//ROWS/ROW'\n PASSING data\n COLUMNS id int PATH '@id',\n ordinality FOR ORDINALITY,\n country_name text PATH 'COUNTRY_NAME',\n country_id text PATH 'COUNTRY_ID',\n size float PATH 'SIZE[@unit = \"km\"]',\n unit text PATH 'SIZE/@unit',\n premier_name text PATH 'PREMIER_NAME'\nDEFAULT 'not specified');\n\n\nIt is related to older variants of this patch, where I explicitly mapped\nempty strings to NULL.\n\nNow, I don't do it - I use libxml2 result with following mapping\n\nNo tag ... NULL\nempty tag ... empty string\n\nImportant question is about mapping empty tags to Postgres. I prefer\ncurrent behave, because I have a possibility to differ between these states\non application level. If we returns NULL for empty tag, then there will not\nbe possible detect if XML has tag (although empty) or not. The change is\nsimple - just one row - but I am thinking so current behave is better.\nThere is possible risk of using /text() somewhere - it enforce a empty tag\nwith all negative impacts.\n\nI prefer to fix doc in conformance with regress tests and append note about\nmapping these corner cases from XML to relations.\n\nWhat do you think about it?\n\nRegards\n\nPavel\n\n\n\n\n\n\n\n\n\n> --\n> Álvaro Herrera https://www.2ndQuadrant.com/\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>\n\nHi2017-03-02 1:12 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\nI've been giving this a look. I started by tweaking the docs once\nagain, and while verifying that the example works as expected, I\nreplayed what I have in sgml:\n\n... begin SGML paste ...\n <para>\n For example, given the following XML document:\n <screen><![CDATA[\n<ROWS>\n <ROW id=\"1\">\n <COUNTRY_ID>AU</COUNTRY_ID>\n <COUNTRY_NAME>Australia</COUNTRY_NAME>\n </ROW>\n <ROW id=\"5\">\n <COUNTRY_ID>JP</COUNTRY_ID>\n <COUNTRY_NAME>Japan</COUNTRY_NAME>\n <PREMIER_NAME>Sinzo Abe</PREMIER_NAME>\n </ROW>\n <ROW id=\"6\">\n <COUNTRY_ID>SG</COUNTRY_ID>\n <COUNTRY_NAME>Singapore</COUNTRY_NAME>\n <SIZE unit=\"km\">791</SIZE>\n </ROW>\n</ROWS>\n]]></screen>\n\n the following query produces the result shown below:\n\n<screen><![CDATA[\nSELECT xmltable.*\n FROM (SELECT data FROM xmldata) x,\n LATERAL xmltable('//ROWS/ROW'\n PASSING data\n COLUMNS id int PATH '@id',\n ordinality FOR ORDINALITY,\n country_name text PATH 'COUNTRY_NAME',\n country_id text PATH 'COUNTRY_ID',\n size float PATH 'SIZE[@unit = \"km\"]/text()',\n unit text PATH 'SIZE/@unit',\n premier_name text PATH 'PREMIER_NAME' DEFAULT 'not specified');\n... end SGML paste ...\n\n\nBut the query doesn't actually return a table, but instead it fails with\nthis error:\n ERROR: invalid input syntax for type double precision: \"\"\nThis is because of the \"size\" column (if I remove SIZE from the COLUMNS\nclause, the query returns correctly). Apparently, for the rows where\nSIZE is not given, we try to inssert an empty string instead of a NULL\nvalue, which is what I expected.\n\nI'm using your v44 code, but trimmed both the XML document used in SGML\nas well as modified the query slightly to show additional features. But\nthose changes should not cause the above error ...The example in doc is obsolete. Following example works without problems. SELECT xmltable.* FROM (SELECT data FROM xmldata) x,\n LATERAL xmltable('//ROWS/ROW'\n PASSING data\n COLUMNS id int PATH '@id',\n ordinality FOR ORDINALITY,\n country_name text PATH 'COUNTRY_NAME',\n country_id text PATH 'COUNTRY_ID',\n size float PATH 'SIZE[@unit = \"km\"]',\n unit text PATH 'SIZE/@unit',\n premier_name text PATH 'PREMIER_NAME' DEFAULT 'not specified');It is related to older variants of this patch, where I explicitly mapped empty strings to NULL. Now, I don't do it - I use libxml2 result with following mappingNo tag ... NULLempty tag ... empty stringImportant question is about mapping empty tags to Postgres. I prefer current behave, because I have a possibility to differ between these states on application level. If we returns NULL for empty tag, then there will not be possible detect if XML has tag (although empty) or not. The change is simple - just one row - but I am thinking so current behave is better. There is possible risk of using /text() somewhere - it enforce a empty tag with all negative impacts.I prefer to fix doc in conformance with regress tests and append note about mapping these corner cases from XML to relations.What do you think about it?RegardsPavel\n\n--\nÁlvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Thu, 2 Mar 2017 08:04:14 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"2017-03-02 8:04 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\n\n> Hi\n>\n> 2017-03-02 1:12 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n>\n>>\n>> I've been giving this a look. I started by tweaking the docs once\n>> again, and while verifying that the example works as expected, I\n>> replayed what I have in sgml:\n>>\n>> ... begin SGML paste ...\n>> <para>\n>> For example, given the following XML document:\n>> <screen><![CDATA[\n>> <ROWS>\n>> <ROW id=\"1\">\n>> <COUNTRY_ID>AU</COUNTRY_ID>\n>> <COUNTRY_NAME>Australia</COUNTRY_NAME>\n>> </ROW>\n>> <ROW id=\"5\">\n>> <COUNTRY_ID>JP</COUNTRY_ID>\n>> <COUNTRY_NAME>Japan</COUNTRY_NAME>\n>> <PREMIER_NAME>Sinzo Abe</PREMIER_NAME>\n>> </ROW>\n>> <ROW id=\"6\">\n>> <COUNTRY_ID>SG</COUNTRY_ID>\n>> <COUNTRY_NAME>Singapore</COUNTRY_NAME>\n>> <SIZE unit=\"km\">791</SIZE>\n>> </ROW>\n>> </ROWS>\n>> ]]></screen>\n>>\n>> the following query produces the result shown below:\n>>\n>> <screen><![CDATA[\n>> SELECT xmltable.*\n>> FROM (SELECT data FROM xmldata) x,\n>> LATERAL xmltable('//ROWS/ROW'\n>> PASSING data\n>> COLUMNS id int PATH '@id',\n>> ordinality FOR ORDINALITY,\n>> country_name text PATH 'COUNTRY_NAME',\n>> country_id text PATH 'COUNTRY_ID',\n>> size float PATH 'SIZE[@unit =\n>> \"km\"]/text()',\n>> unit text PATH 'SIZE/@unit',\n>> premier_name text PATH 'PREMIER_NAME'\n>> DEFAULT 'not specified');\n>> ... end SGML paste ...\n>>\n>>\n>> But the query doesn't actually return a table, but instead it fails with\n>> this error:\n>> ERROR: invalid input syntax for type double precision: \"\"\n>> This is because of the \"size\" column (if I remove SIZE from the COLUMNS\n>> clause, the query returns correctly). Apparently, for the rows where\n>> SIZE is not given, we try to inssert an empty string instead of a NULL\n>> value, which is what I expected.\n>>\n>> I'm using your v44 code, but trimmed both the XML document used in SGML\n>> as well as modified the query slightly to show additional features. But\n>> those changes should not cause the above error ...\n>>\n>\n> The example in doc is obsolete. Following example works without problems.\n>\n> SELECT xmltable.*\n>\n> FROM (SELECT data FROM xmldata) x,\n> LATERAL xmltable('//ROWS/ROW'\n> PASSING data\n> COLUMNS id int PATH '@id',\n> ordinality FOR ORDINALITY,\n> country_name text PATH 'COUNTRY_NAME',\n> country_id text PATH 'COUNTRY_ID',\n> size float PATH 'SIZE[@unit = \"km\"]',\n> unit text PATH 'SIZE/@unit',\n> premier_name text PATH 'PREMIER_NAME' DEFAULT 'not specified');\n>\n>\n> It is related to older variants of this patch, where I explicitly mapped\n> empty strings to NULL.\n>\n> Now, I don't do it - I use libxml2 result with following mapping\n>\n> No tag ... NULL\n> empty tag ... empty string\n>\n> Important question is about mapping empty tags to Postgres. I prefer\n> current behave, because I have a possibility to differ between these states\n> on application level. If we returns NULL for empty tag, then there will not\n> be possible detect if XML has tag (although empty) or not. The change is\n> simple - just one row - but I am thinking so current behave is better.\n> There is possible risk of using /text() somewhere - it enforce a empty tag\n> with all negative impacts.\n>\n> I prefer to fix doc in conformance with regress tests and append note\n> about mapping these corner cases from XML to relations.\n>\n> What do you think about it?\n>\n\nIt is documented already\n\n\"If the <literal>PATH</> matches an empty tag the result is an empty string\"\n\nAttached new patch\n\ncleaned documentation\nregress tests is more robust\nappended comment in src related to generating empty string for empty tag\n\nRegards\n\nPavel\n\n\n>\n> Regards\n>\n> Pavel\n>\n>\n>\n>\n>\n>\n>\n>\n>\n>> --\n>> Álvaro Herrera https://www.2ndQuadrant.com/\n>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\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","msg_date":"Thu, 2 Mar 2017 09:13:02 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Pavel Stehule wrote:\n\n> It is documented already\n> \n> \"If the <literal>PATH</> matches an empty tag the result is an empty string\"\n\nHmm, okay. But what we have here is not an empty tag, but a tag that is\ncompletely missing. I don't think those two cases should be treated in\nthe same way ...\n\n> Attached new patch\n> \n> cleaned documentation\n> regress tests is more robust\n> appended comment in src related to generating empty string for empty tag\n\nThanks, I incorporated those changes. Here's v46. I rewrote the\ndocumentation, and fixed a couple of incorrectly copied&pasted comments\nin the new executor code; I think that one looks good. In the future we\ncould rewrite it to avoid the need for a tuplestore, but I think the\ncurrent approach is good enough for a pg10 implementation.\n\nBarring serious problems, I intend to commit this later today.\n\n-- \n�lvaro Herrera https://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, 2 Mar 2017 14:13:53 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Dne 2. 3. 2017 18:14 napsal uživatel \"Alvaro Herrera\" <\nalvherre@2ndquadrant.com>:\n\nPavel Stehule wrote:\n\n> It is documented already\n>\n> \"If the <literal>PATH</> matches an empty tag the result is an empty\nstring\"\n\nHmm, okay. But what we have here is not an empty tag, but a tag that is\ncompletely missing. I don't think those two cases should be treated in\nthe same way ...\n\n\nthis information is not propagated from libxml2.\n\n\n> Attached new patch\n>\n> cleaned documentation\n> regress tests is more robust\n> appended comment in src related to generating empty string for empty tag\n\nThanks, I incorporated those changes. Here's v46. I rewrote the\ndocumentation, and fixed a couple of incorrectly copied&pasted comments\nin the new executor code; I think that one looks good. In the future we\ncould rewrite it to avoid the need for a tuplestore, but I think the\ncurrent approach is good enough for a pg10 implementation.\n\nBarring serious problems, I intend to commit this later today.\n\n\n\nthank you very much\n\nregards\n\nPavel\n\n\n--\nÁlvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\nDne 2. 3. 2017 18:14 napsal uživatel \"Alvaro Herrera\" <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\n\n> It is documented already\n>\n> \"If the <literal>PATH</> matches an empty tag the result is an empty string\"\n\nHmm, okay. But what we have here is not an empty tag, but a tag that is\ncompletely missing. I don't think those two cases should be treated in\nthe same way ...this information is not propagated from libxml2.\n\n> Attached new patch\n>\n> cleaned documentation\n> regress tests is more robust\n> appended comment in src related to generating empty string for empty tag\n\nThanks, I incorporated those changes. Here's v46. I rewrote the\ndocumentation, and fixed a couple of incorrectly copied&pasted comments\nin the new executor code; I think that one looks good. In the future we\ncould rewrite it to avoid the need for a tuplestore, but I think the\ncurrent approach is good enough for a pg10 implementation.\n\nBarring serious problems, I intend to commit this later today.\nthank you very muchregardsPavel\n--\nÁlvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Thu, 2 Mar 2017 18:46:23 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"So in the old (non-executor-node) implementation, you could attach WITH\nORDINALITY to the xmltable expression and it would count the output\nrows, regardless of which XML document it comes from. With the new\nimplementation, the grammar no longer accepts it. To count output rows,\nyou still need to use row_number(). Maybe this is okay. This is the\nexample from the docs, and I add another XML document with two more rows\nfor xmltable. Look at the three numbering columns ...\n\nCREATE TABLE xmldata AS SELECT\nxml $$\n<ROWS> \n <ROW id=\"1\">\n <COUNTRY_ID>AU</COUNTRY_ID>\n <COUNTRY_NAME>Australia</COUNTRY_NAME>\n </ROW> \n <ROW id=\"5\"> \n <COUNTRY_ID>JP</COUNTRY_ID>\n <COUNTRY_NAME>Japan</COUNTRY_NAME>\n <PREMIER_NAME>Shinzo Abe</PREMIER_NAME>\n <SIZE unit=\"sq_mi\">145935</SIZE>\n </ROW>\n <ROW id=\"6\">\n <COUNTRY_ID>SG</COUNTRY_ID>\n <COUNTRY_NAME>Singapore</COUNTRY_NAME>\n <SIZE unit=\"sq_km\">697</SIZE>\n </ROW>\n</ROWS>\n$$ AS data;\n\n insert into xmldata values ($$\n <ROWS><ROW id=\"2\"><COUNTRY_ID>CL</COUNTRY_ID><COUNTRY_NAME>Chile</COUNTRY_NAME></ROW>\n <ROW id=\"3\"><COUNTRY_ID>AR</COUNTRY_ID><COUNTRY_NAME>Argentina</COUNTRY_NAME></ROW></ROWS>$$);\n\nSELECT ROW_NUMBER() OVER (), xmltable.*\n FROM xmldata, \n XMLTABLE('//ROWS/ROW' \n PASSING data\n COLUMNS id int PATH '@id',\n ordinality FOR ORDINALITY,\n \"COUNTRY_NAME\" text,\n country_id text PATH 'COUNTRY_ID',\n size_sq_km float PATH 'SIZE[@unit = \"sq_km\"]',\n size_other text PATH\n 'concat(SIZE[@unit!=\"sq_km\"], \" \", SIZE[@unit!=\"sq_km\"]/@unit)',\n premier_name text PATH 'PREMIER_NAME' DEFAULT 'not specified')\n;\n\n row_number │ id │ ordinality │ COUNTRY_NAME │ country_id │ size_sq_km │ size_other │ premier_name \n────────────┼────┼────────────┼──────────────┼────────────┼────────────┼──────────────┼───────────────\n 1 │ 1 │ 1 │ Australia │ AU │ │ │ not specified\n 2 │ 5 │ 2 │ Japan │ JP │ │ 145935 sq_mi │ Shinzo Abe\n 3 │ 6 │ 3 │ Singapore │ SG │ 697 │ │ not specified\n 4 │ 2 │ 1 │ Chile │ CL │ │ │ not specified\n 5 │ 3 │ 2 │ Argentina │ AR │ │ │ not specified\n\n\n-- \nÁlvaro Herrera https://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, 2 Mar 2017 15:32:05 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"2017-03-02 19:32 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> So in the old (non-executor-node) implementation, you could attach WITH\n> ORDINALITY to the xmltable expression and it would count the output\n> rows, regardless of which XML document it comes from. With the new\n> implementation, the grammar no longer accepts it. To count output rows,\n> you still need to use row_number(). Maybe this is okay. This is the\n> example from the docs, and I add another XML document with two more rows\n> for xmltable. Look at the three numbering columns ...\n>\n\nIt is expected - now tablefunc are not special case of SRF, so it lost all\nSRF functionality. It is not critical lost - it supports internally FOR\nORDINALITY column, and classic ROW_NUMBER can be used. It can be enhanced\nto support WITH ORDINALITY in future, but I have not any use case for it.\n\nRegards\n\nPavel\n\n\n\n>\n> CREATE TABLE xmldata AS SELECT\n> xml $$\n> <ROWS>\n> <ROW id=\"1\">\n> <COUNTRY_ID>AU</COUNTRY_ID>\n> <COUNTRY_NAME>Australia</COUNTRY_NAME>\n> </ROW>\n> <ROW id=\"5\">\n> <COUNTRY_ID>JP</COUNTRY_ID>\n> <COUNTRY_NAME>Japan</COUNTRY_NAME>\n> <PREMIER_NAME>Shinzo Abe</PREMIER_NAME>\n> <SIZE unit=\"sq_mi\">145935</SIZE>\n> </ROW>\n> <ROW id=\"6\">\n> <COUNTRY_ID>SG</COUNTRY_ID>\n> <COUNTRY_NAME>Singapore</COUNTRY_NAME>\n> <SIZE unit=\"sq_km\">697</SIZE>\n> </ROW>\n> </ROWS>\n> $$ AS data;\n>\n> insert into xmldata values ($$\n> <ROWS><ROW id=\"2\"><COUNTRY_ID>CL</COUNTRY_ID><COUNTRY_NAME>\n> Chile</COUNTRY_NAME></ROW>\n> <ROW id=\"3\"><COUNTRY_ID>AR</COUNTRY_ID><COUNTRY_NAME>\n> Argentina</COUNTRY_NAME></ROW></ROWS>$$);\n>\n> SELECT ROW_NUMBER() OVER (), xmltable.*\n> FROM xmldata,\n> XMLTABLE('//ROWS/ROW'\n> PASSING data\n> COLUMNS id int PATH '@id',\n> ordinality FOR ORDINALITY,\n> \"COUNTRY_NAME\" text,\n> country_id text PATH 'COUNTRY_ID',\n> size_sq_km float PATH 'SIZE[@unit = \"sq_km\"]',\n> size_other text PATH\n> 'concat(SIZE[@unit!=\"sq_km\"], \" \",\n> SIZE[@unit!=\"sq_km\"]/@unit)',\n> premier_name text PATH 'PREMIER_NAME' DEFAULT 'not\n> specified')\n> ;\n>\n> row_number │ id │ ordinality │ COUNTRY_NAME │ country_id │ size_sq_km │\n> size_other │ premier_name\n> ────────────┼────┼────────────┼──────────────┼────────────┼─\n> ───────────┼──────────────┼───────────────\n> 1 │ 1 │ 1 │ Australia │ AU │ │\n> │ not specified\n> 2 │ 5 │ 2 │ Japan │ JP │ │\n> 145935 sq_mi │ Shinzo Abe\n> 3 │ 6 │ 3 │ Singapore │ SG │ 697 │\n> │ not specified\n> 4 │ 2 │ 1 │ Chile │ CL │ │\n> │ not specified\n> 5 │ 3 │ 2 │ Argentina │ AR │ │\n> │ not specified\n>\n>\n> --\n> Álvaro Herrera https://www.2ndQuadrant.com/\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>\n\n2017-03-02 19:32 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:So in the old (non-executor-node) implementation, you could attach WITH\r\nORDINALITY to the xmltable expression and it would count the output\r\nrows, regardless of which XML document it comes from. With the new\r\nimplementation, the grammar no longer accepts it. To count output rows,\r\nyou still need to use row_number(). Maybe this is okay. This is the\r\nexample from the docs, and I add another XML document with two more rows\r\nfor xmltable. Look at the three numbering columns ...It is expected - now tablefunc are not special case of SRF, so it lost all SRF functionality. It is not critical lost - it supports internally FOR ORDINALITY column, and classic ROW_NUMBER can be used. It can be enhanced to support WITH ORDINALITY in future, but I have not any use case for it.RegardsPavel \n\r\nCREATE TABLE xmldata AS SELECT\r\nxml $$\n<ROWS>\r\n <ROW id=\"1\">\r\n <COUNTRY_ID>AU</COUNTRY_ID>\r\n <COUNTRY_NAME>Australia</COUNTRY_NAME>\r\n </ROW>\r\n <ROW id=\"5\">\r\n <COUNTRY_ID>JP</COUNTRY_ID>\r\n <COUNTRY_NAME>Japan</COUNTRY_NAME>\n <PREMIER_NAME>Shinzo Abe</PREMIER_NAME>\r\n <SIZE unit=\"sq_mi\">145935</SIZE>\n </ROW>\r\n <ROW id=\"6\">\r\n <COUNTRY_ID>SG</COUNTRY_ID>\r\n <COUNTRY_NAME>Singapore</COUNTRY_NAME>\n <SIZE unit=\"sq_km\">697</SIZE>\r\n </ROW>\r\n</ROWS>\r\n$$ AS data;\n\r\n insert into xmldata values ($$\r\n <ROWS><ROW id=\"2\"><COUNTRY_ID>CL</COUNTRY_ID><COUNTRY_NAME>Chile</COUNTRY_NAME></ROW>\r\n <ROW id=\"3\"><COUNTRY_ID>AR</COUNTRY_ID><COUNTRY_NAME>Argentina</COUNTRY_NAME></ROW></ROWS>$$);\n\r\nSELECT ROW_NUMBER() OVER (), xmltable.*\r\n FROM xmldata,\r\n XMLTABLE('//ROWS/ROW'\n PASSING data\r\n COLUMNS id int PATH '@id',\r\n ordinality FOR ORDINALITY,\n \"COUNTRY_NAME\" text,\n country_id text PATH 'COUNTRY_ID',\n size_sq_km float PATH 'SIZE[@unit = \"sq_km\"]',\r\n size_other text PATH\r\n 'concat(SIZE[@unit!=\"sq_km\"], \" \", SIZE[@unit!=\"sq_km\"]/@unit)',\n premier_name text PATH 'PREMIER_NAME' DEFAULT 'not specified')\r\n;\n\n row_number │ id │ ordinality │ COUNTRY_NAME │ country_id │ size_sq_km │ size_other │ premier_name\r\n────────────┼────┼────────────┼──────────────┼────────────┼────────────┼──────────────┼───────────────\r\n 1 │ 1 │ 1 │ Australia │ AU │ │ │ not specified\r\n 2 │ 5 │ 2 │ Japan │ JP │ │ 145935 sq_mi │ Shinzo Abe\r\n 3 │ 6 │ 3 │ Singapore │ SG │ 697 │ │ not specified\r\n 4 │ 2 │ 1 │ Chile │ CL │ │ │ not specified\r\n 5 │ 3 │ 2 │ Argentina │ AR │ │ │ not specified\n\n\r\n--\r\nÁlvaro Herrera https://www.2ndQuadrant.com/\r\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Thu, 2 Mar 2017 20:17:12 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Pavel Stehule wrote:\n> 2017-03-02 19:32 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n> \n> > So in the old (non-executor-node) implementation, you could attach WITH\n> > ORDINALITY to the xmltable expression and it would count the output\n> > rows, regardless of which XML document it comes from. With the new\n> > implementation, the grammar no longer accepts it. To count output rows,\n> > you still need to use row_number(). Maybe this is okay. This is the\n> > example from the docs, and I add another XML document with two more rows\n> > for xmltable. Look at the three numbering columns ...\n> >\n> \n> It is expected - now tablefunc are not special case of SRF, so it lost all\n> SRF functionality. It is not critical lost - it supports internally FOR\n> ORDINALITY column, and classic ROW_NUMBER can be used. It can be enhanced\n> to support WITH ORDINALITY in future, but I have not any use case for it.\n\nFine.\n\nAfter looking at the new executor code a bit, I noticed that we don't\nneed the resultSlot anymore; we can use the ss_ScanTupleSlot instead.\nBecause resultSlot was being used in the xml.c code (which already\nappeared a bit dubious to me), I changed the interface so that instead\nthe things that it read from it are passed as parameters -- namely, in\nInitBuilder we pass natts, and in GetValue we pass typid and typmod.\n\nSecondly, I noticed we have the FetchRow routine produce a minimal\ntuple, put it in a slot; then its caller takes the slot and put the\ntuple in the tuplestore. This is pointless; we can just have FetchRow\nput the tuple in the tuplestore directly and not bother with any slot\nmanipulations there. This simplifies the code a bit.\n\nHere's v47 with those changes.\n\n-- \n�lvaro Herrera https://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, 2 Mar 2017 18:35:44 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"2017-03-02 22:35 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> Pavel Stehule wrote:\n> > 2017-03-02 19:32 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n> >\n> > > So in the old (non-executor-node) implementation, you could attach WITH\n> > > ORDINALITY to the xmltable expression and it would count the output\n> > > rows, regardless of which XML document it comes from. With the new\n> > > implementation, the grammar no longer accepts it. To count output\n> rows,\n> > > you still need to use row_number(). Maybe this is okay. This is the\n> > > example from the docs, and I add another XML document with two more\n> rows\n> > > for xmltable. Look at the three numbering columns ...\n> > >\n> >\n> > It is expected - now tablefunc are not special case of SRF, so it lost\n> all\n> > SRF functionality. It is not critical lost - it supports internally FOR\n> > ORDINALITY column, and classic ROW_NUMBER can be used. It can be enhanced\n> > to support WITH ORDINALITY in future, but I have not any use case for it.\n>\n> Fine.\n>\n> After looking at the new executor code a bit, I noticed that we don't\n> need the resultSlot anymore; we can use the ss_ScanTupleSlot instead.\n> Because resultSlot was being used in the xml.c code (which already\n> appeared a bit dubious to me), I changed the interface so that instead\n> the things that it read from it are passed as parameters -- namely, in\n> InitBuilder we pass natts, and in GetValue we pass typid and typmod.\n>\n\nI had similar feeling\n\n>\n> Secondly, I noticed we have the FetchRow routine produce a minimal\n> tuple, put it in a slot; then its caller takes the slot and put the\n> tuple in the tuplestore. This is pointless; we can just have FetchRow\n> put the tuple in the tuplestore directly and not bother with any slot\n> manipulations there. This simplifies the code a bit.\n>\n>\nhas sense\n\nattached update with fixed tests\n\nRegards\n\nPavel\n\n\n\n> Here's v47 with those changes.\n>\n> --\n> Álvaro Herrera https://www.2ndQuadrant.com/\n> PostgreSQL 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":"Fri, 3 Mar 2017 10:28:54 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Pavel Stehule wrote:\n\n> attached update with fixed tests\n\nHeh, I noticed that you removed the libxml \"context\" lines that\ndifferentiate xml.out from xml_2.out when doing this. My implementation\nemits those lines, so it was failing for me. I restored them.\n\nI also changed a few things to avoid copying into TableFuncScanState\nthings that come from the TableFunc itself, since the executor state\nnode can grab them from the plan node. Let's do that. So instead of\n\"evalcols\" the code now checks that the column list is empty; and also,\nread the ordinality column number from the plan node.\n\nI have to bounce this back to you one more time, hopefully the last one\nI hope. Two things:\n\n1. Please verify that pg_stat_statements behaves correctly. The patch\ndoesn't have changes to contrib/ so without testing I'm guessing that it\ndoesn't work. I think something very simple should do.\n\n2. As I've complained many times, I find the way we manage an empty\nCOLUMNS clause pretty bad. The standard doesn't require that syntax\n(COLUMNS is required), and I don't like the implementation, so why not\nprovide the feature in a different way? My proposal is to change the\ncolumn options in gram.y to be something like this:\n\nxmltable_column_option_el:\n\t\t\tIDENT b_expr\n\t\t\t\t{ $$ = makeDefElem($1, $2, @1); }\n\t\t\t| DEFAULT b_expr\n\t\t\t\t{ $$ = makeDefElem(\"default\", $2, @1); }\n\t\t\t| FULL VALUE_P\n\t\t\t\t{ $$ = makeDefElem(\"full_value\", NULL, @1); }\n\t\t\t| NOT NULL_P\n\t\t\t\t{ $$ = makeDefElem(\"is_not_null\", (Node *) makeInteger(true), @1); }\n\t\t\t| NULL_P\n\t\t\t\t{ $$ = makeDefElem(\"is_not_null\", (Node *) makeInteger(false), @1); }\n\t\t;\n\nNote the FULL VALUE. Then we can process it like\n\n\t\t\telse if (strcmp(defel->defname, \"full_value\") == 0)\n\t\t\t{\n\t\t\t\tif (fc->colexpr != NULL)\n\t\t\t\t\tereport(ERROR,\n\t\t\t\t\t\t\t(errcode(ERRCODE_SYNTAX_ERROR),\n\t\t\t\t\t\t\t errmsg(\"FULL ROW may not be specified together with PATH\"),\n\t\t\t\t\t\t\t parser_errposition(defel->location)));\n\t\t\t\tfc->full_row = true;\n\t\t\t}\n\nSo if you want the full XML value of the row, you have to specify it,\n\n .. XMLTABLE ( ... COLUMNS ..., whole_row xml FULL VALUE, ... )\n\nThis has the extra feature that you can add, say, an ORDINALITY column\ntogether with the XML value, something that you cannot do with the\ncurrent implementation.\n\nIt doesn't have to be FULL VALUE, but I couldn't think of anything\nbetter. (I didn't want to add any new keywords for this.) If you have\na better idea, let's discuss.\n\nCode-wise, this completely removes the \"else\" block in transformRangeTableFunc\nwhich I marked with an XXX comment. That's a good thing -- let's get\nrid of that. Also, it should remove the need for the separate \"if\n!columns\" case in tfuncLoadRows. All those cases would become part of\nthe normal code path instead of special cases. I think\nXmlTableSetColumnFilter doesn't need any change (we just don't call if\nfor the FULL VALUE row); and XmlTableGetValue needs a special case that\nif the column filter is NULL (i.e. SetColumnFilter wasn't called for\nthat column) then return the whole row.\n\n\nOf course, this opens an implementation issue: how do you annotate\nthings from parse analysis till execution? The current TableFunc\nstructure doesn't help, because there are only lists of column names and\nexpressions; and we can't use the case of a NULL colexpr, because that\ncase is already used by the column filter being the column name (a\nfeature required by the standard). A simple way would be to have a new\n\"colno\" struct member, to store a column number for the column marked\nFULL VALUE (just like ordinalitycol). This means you can't have more\nthan one of those FULL VALUE columns, but that seems okay.\n\n\n(Of course, this means that the two cases that have no COLUMNS in the\n\"xmltable\" production in gram.y should go away).\n\n-- \n�lvaro Herrera https://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":"Fri, 3 Mar 2017 15:15:11 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"2017-03-03 19:15 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> Pavel Stehule wrote:\n>\n> > attached update with fixed tests\n>\n> Heh, I noticed that you removed the libxml \"context\" lines that\n> differentiate xml.out from xml_2.out when doing this. My implementation\n> emits those lines, so it was failing for me. I restored them.\n>\n> I also changed a few things to avoid copying into TableFuncScanState\n> things that come from the TableFunc itself, since the executor state\n> node can grab them from the plan node. Let's do that. So instead of\n> \"evalcols\" the code now checks that the column list is empty; and also,\n> read the ordinality column number from the plan node.\n>\n> I have to bounce this back to you one more time, hopefully the last one\n> I hope. Two things:\n>\n> 1. Please verify that pg_stat_statements behaves correctly. The patch\n> doesn't have changes to contrib/ so without testing I'm guessing that it\n> doesn't work. I think something very simple should do.\n>\n> 2. As I've complained many times, I find the way we manage an empty\n> COLUMNS clause pretty bad. The standard doesn't require that syntax\n> (COLUMNS is required), and I don't like the implementation, so why not\n> provide the feature in a different way? My proposal is to change the\n> column options in gram.y to be something like this:\n\n\nThe clause COLUMNS is optional on Oracle and DB2\n\nSo I prefer a Oracle, DB2 design. If you are strongly against it, then we\ncan remove it to be ANSI/SQL only.\n\nI am don't see an good idea to introduce third syntax.\n\n\n> xmltable_column_option_el:\n> IDENT b_expr\n> { $$ = makeDefElem($1, $2, @1); }\n> | DEFAULT b_expr\n> { $$ = makeDefElem(\"default\", $2, @1); }\n> | FULL VALUE_P\n> { $$ = makeDefElem(\"full_value\", NULL,\n> @1); }\n> | NOT NULL_P\n> { $$ = makeDefElem(\"is_not_null\", (Node *)\n> makeInteger(true), @1); }\n> | NULL_P\n> { $$ = makeDefElem(\"is_not_null\", (Node *)\n> makeInteger(false), @1); }\n> ;\n>\n> Note the FULL VALUE. Then we can process it like\n>\n> else if (strcmp(defel->defname, \"full_value\") == 0)\n> {\n> if (fc->colexpr != NULL)\n> ereport(ERROR,\n>\n> (errcode(ERRCODE_SYNTAX_ERROR),\n> errmsg(\"FULL ROW\n> may not be specified together with PATH\"),\n>\n> parser_errposition(defel->location)));\n> fc->full_row = true;\n> }\n>\n> So if you want the full XML value of the row, you have to specify it,\n>\n> .. XMLTABLE ( ... COLUMNS ..., whole_row xml FULL VALUE, ... )\n>\n> This has the extra feature that you can add, say, an ORDINALITY column\n> together with the XML value, something that you cannot do with the\n> current implementation.\n>\n> It doesn't have to be FULL VALUE, but I couldn't think of anything\n> better. (I didn't want to add any new keywords for this.) If you have\n> a better idea, let's discuss.\n>\n\nI don't see a introduction own syntax as necessary solution here - use\nOracle, DB2 compatible syntax, or ANSI.\n\nIt is partially corner case - the benefit of this case is almost bigger\ncompatibility with mentioned databases.\n\n\n>\n> Code-wise, this completely removes the \"else\" block in\n> transformRangeTableFunc\n> which I marked with an XXX comment. That's a good thing -- let's get\n> rid of that. Also, it should remove the need for the separate \"if\n> !columns\" case in tfuncLoadRows. All those cases would become part of\n> the normal code path instead of special cases. I think\n> XmlTableSetColumnFilter doesn't need any change (we just don't call if\n> for the FULL VALUE row); and XmlTableGetValue needs a special case that\n> if the column filter is NULL (i.e. SetColumnFilter wasn't called for\n> that column) then return the whole row.\n>\n>\n> Of course, this opens an implementation issue: how do you annotate\n> things from parse analysis till execution? The current TableFunc\n> structure doesn't help, because there are only lists of column names and\n> expressions; and we can't use the case of a NULL colexpr, because that\n> case is already used by the column filter being the column name (a\n> feature required by the standard). A simple way would be to have a new\n> \"colno\" struct member, to store a column number for the column marked\n> FULL VALUE (just like ordinalitycol). This means you can't have more\n> than one of those FULL VALUE columns, but that seems okay.\n>\n>\n> (Of course, this means that the two cases that have no COLUMNS in the\n> \"xmltable\" production in gram.y should go away).\n>\n\nYou are commiter, and you should to decide - as first I prefer current\nstate, as second a remove this part - it should be good for you too,\nbecause code that you don't like will be left.\n\nI dislike introduce new syntax - this case is not too important for this.\n\nRegards\n\nPavel\n\n\n> --\n> Álvaro Herrera https://www.2ndQuadrant.com/\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>\n\n2017-03-03 19:15 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\n\n> attached update with fixed tests\n\nHeh, I noticed that you removed the libxml \"context\" lines that\ndifferentiate xml.out from xml_2.out when doing this. My implementation\nemits those lines, so it was failing for me. I restored them.\n\nI also changed a few things to avoid copying into TableFuncScanState\nthings that come from the TableFunc itself, since the executor state\nnode can grab them from the plan node. Let's do that. So instead of\n\"evalcols\" the code now checks that the column list is empty; and also,\nread the ordinality column number from the plan node.\n\nI have to bounce this back to you one more time, hopefully the last one\nI hope. Two things:\n\n1. Please verify that pg_stat_statements behaves correctly. The patch\ndoesn't have changes to contrib/ so without testing I'm guessing that it\ndoesn't work. I think something very simple should do.\n\n2. As I've complained many times, I find the way we manage an empty\nCOLUMNS clause pretty bad. The standard doesn't require that syntax\n(COLUMNS is required), and I don't like the implementation, so why not\nprovide the feature in a different way? My proposal is to change the\ncolumn options in gram.y to be something like this:The clause COLUMNS is optional on Oracle and DB2 So I prefer a Oracle, DB2 design. If you are strongly against it, then we can remove it to be ANSI/SQL only.I am don't see an good idea to introduce third syntax.\n\nxmltable_column_option_el:\n IDENT b_expr\n { $$ = makeDefElem($1, $2, @1); }\n | DEFAULT b_expr\n { $$ = makeDefElem(\"default\", $2, @1); }\n | FULL VALUE_P\n { $$ = makeDefElem(\"full_value\", NULL, @1); }\n | NOT NULL_P\n { $$ = makeDefElem(\"is_not_null\", (Node *) makeInteger(true), @1); }\n | NULL_P\n { $$ = makeDefElem(\"is_not_null\", (Node *) makeInteger(false), @1); }\n ;\n\nNote the FULL VALUE. Then we can process it like\n\n else if (strcmp(defel->defname, \"full_value\") == 0)\n {\n if (fc->colexpr != NULL)\n ereport(ERROR,\n (errcode(ERRCODE_SYNTAX_ERROR),\n errmsg(\"FULL ROW may not be specified together with PATH\"),\n parser_errposition(defel->location)));\n fc->full_row = true;\n }\n\nSo if you want the full XML value of the row, you have to specify it,\n\n .. XMLTABLE ( ... COLUMNS ..., whole_row xml FULL VALUE, ... )\n\nThis has the extra feature that you can add, say, an ORDINALITY column\ntogether with the XML value, something that you cannot do with the\ncurrent implementation.\n\nIt doesn't have to be FULL VALUE, but I couldn't think of anything\nbetter. (I didn't want to add any new keywords for this.) If you have\na better idea, let's discuss.I don't see a introduction own syntax as necessary solution here - use Oracle, DB2 compatible syntax, or ANSI. It is partially corner case - the benefit of this case is almost bigger compatibility with mentioned databases. \n\nCode-wise, this completely removes the \"else\" block in transformRangeTableFunc\nwhich I marked with an XXX comment. That's a good thing -- let's get\nrid of that. Also, it should remove the need for the separate \"if\n!columns\" case in tfuncLoadRows. All those cases would become part of\nthe normal code path instead of special cases. I think\nXmlTableSetColumnFilter doesn't need any change (we just don't call if\nfor the FULL VALUE row); and XmlTableGetValue needs a special case that\nif the column filter is NULL (i.e. SetColumnFilter wasn't called for\nthat column) then return the whole row.\n\n\nOf course, this opens an implementation issue: how do you annotate\nthings from parse analysis till execution? The current TableFunc\nstructure doesn't help, because there are only lists of column names and\nexpressions; and we can't use the case of a NULL colexpr, because that\ncase is already used by the column filter being the column name (a\nfeature required by the standard). A simple way would be to have a new\n\"colno\" struct member, to store a column number for the column marked\nFULL VALUE (just like ordinalitycol). This means you can't have more\nthan one of those FULL VALUE columns, but that seems okay.\n\n\n(Of course, this means that the two cases that have no COLUMNS in the\n\"xmltable\" production in gram.y should go away).You are commiter, and you should to decide - as first I prefer current state, as second a remove this part - it should be good for you too, because code that you don't like will be left.I dislike introduce new syntax - this case is not too important for this.RegardsPavel\n\n--\nÁlvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Fri, 3 Mar 2017 19:42:11 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"2017-03-03 19:42 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\n\n>\n>\n> 2017-03-03 19:15 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n>\n>> Pavel Stehule wrote:\n>>\n>> > attached update with fixed tests\n>>\n>> Heh, I noticed that you removed the libxml \"context\" lines that\n>> differentiate xml.out from xml_2.out when doing this. My implementation\n>> emits those lines, so it was failing for me. I restored them.\n>>\n>> I also changed a few things to avoid copying into TableFuncScanState\n>> things that come from the TableFunc itself, since the executor state\n>> node can grab them from the plan node. Let's do that. So instead of\n>> \"evalcols\" the code now checks that the column list is empty; and also,\n>> read the ordinality column number from the plan node.\n>>\n>> I have to bounce this back to you one more time, hopefully the last one\n>> I hope. Two things:\n>>\n>> 1. Please verify that pg_stat_statements behaves correctly. The patch\n>> doesn't have changes to contrib/ so without testing I'm guessing that it\n>> doesn't work. I think something very simple should do.\n>>\n>> 2. As I've complained many times, I find the way we manage an empty\n>> COLUMNS clause pretty bad. The standard doesn't require that syntax\n>> (COLUMNS is required), and I don't like the implementation, so why not\n>> provide the feature in a different way? My proposal is to change the\n>> column options in gram.y to be something like this:\n>\n>\n> The clause COLUMNS is optional on Oracle and DB2\n>\n> So I prefer a Oracle, DB2 design. If you are strongly against it, then we\n> can remove it to be ANSI/SQL only.\n>\n> I am don't see an good idea to introduce third syntax.\n>\n>\n>> xmltable_column_option_el:\n>> IDENT b_expr\n>> { $$ = makeDefElem($1, $2, @1); }\n>> | DEFAULT b_expr\n>> { $$ = makeDefElem(\"default\", $2, @1); }\n>> | FULL VALUE_P\n>> { $$ = makeDefElem(\"full_value\", NULL,\n>> @1); }\n>> | NOT NULL_P\n>> { $$ = makeDefElem(\"is_not_null\", (Node\n>> *) makeInteger(true), @1); }\n>> | NULL_P\n>> { $$ = makeDefElem(\"is_not_null\", (Node\n>> *) makeInteger(false), @1); }\n>> ;\n>>\n>> Note the FULL VALUE. Then we can process it like\n>>\n>> else if (strcmp(defel->defname, \"full_value\") ==\n>> 0)\n>> {\n>> if (fc->colexpr != NULL)\n>> ereport(ERROR,\n>>\n>> (errcode(ERRCODE_SYNTAX_ERROR),\n>> errmsg(\"FULL ROW\n>> may not be specified together with PATH\"),\n>>\n>> parser_errposition(defel->location)));\n>> fc->full_row = true;\n>> }\n>>\n>> So if you want the full XML value of the row, you have to specify it,\n>>\n>> .. XMLTABLE ( ... COLUMNS ..., whole_row xml FULL VALUE, ... )\n>>\n>> This has the extra feature that you can add, say, an ORDINALITY column\n>> together with the XML value, something that you cannot do with the\n>> current implementation.\n>>\n>> It doesn't have to be FULL VALUE, but I couldn't think of anything\n>> better. (I didn't want to add any new keywords for this.) If you have\n>> a better idea, let's discuss.\n>>\n>\n> I don't see a introduction own syntax as necessary solution here - use\n> Oracle, DB2 compatible syntax, or ANSI.\n>\n> It is partially corner case - the benefit of this case is almost bigger\n> compatibility with mentioned databases.\n>\n>\n>>\n>> Code-wise, this completely removes the \"else\" block in\n>> transformRangeTableFunc\n>> which I marked with an XXX comment. That's a good thing -- let's get\n>> rid of that. Also, it should remove the need for the separate \"if\n>> !columns\" case in tfuncLoadRows. All those cases would become part of\n>> the normal code path instead of special cases. I think\n>> XmlTableSetColumnFilter doesn't need any change (we just don't call if\n>> for the FULL VALUE row); and XmlTableGetValue needs a special case that\n>> if the column filter is NULL (i.e. SetColumnFilter wasn't called for\n>> that column) then return the whole row.\n>>\n>>\n>> Of course, this opens an implementation issue: how do you annotate\n>> things from parse analysis till execution? The current TableFunc\n>> structure doesn't help, because there are only lists of column names and\n>> expressions; and we can't use the case of a NULL colexpr, because that\n>> case is already used by the column filter being the column name (a\n>> feature required by the standard). A simple way would be to have a new\n>> \"colno\" struct member, to store a column number for the column marked\n>> FULL VALUE (just like ordinalitycol). This means you can't have more\n>> than one of those FULL VALUE columns, but that seems okay.\n>>\n>>\n>> (Of course, this means that the two cases that have no COLUMNS in the\n>> \"xmltable\" production in gram.y should go away).\n>>\n>\n> You are commiter, and you should to decide - as first I prefer current\n> state, as second a remove this part - it should be good for you too,\n> because code that you don't like will be left.\n>\n> I dislike introduce new syntax - this case is not too important for this.\n>\n>\nI am able to prepare reduced version if we do a agreement\n\nRegards\n\nPavel\n\n\n> Regards\n>\n> Pavel\n>\n>\n>> --\n>> Álvaro Herrera https://www.2ndQuadrant.com/\n>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>>\n>\n>\n\n2017-03-03 19:42 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:2017-03-03 19:15 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\n\n> attached update with fixed tests\n\nHeh, I noticed that you removed the libxml \"context\" lines that\ndifferentiate xml.out from xml_2.out when doing this. My implementation\nemits those lines, so it was failing for me. I restored them.\n\nI also changed a few things to avoid copying into TableFuncScanState\nthings that come from the TableFunc itself, since the executor state\nnode can grab them from the plan node. Let's do that. So instead of\n\"evalcols\" the code now checks that the column list is empty; and also,\nread the ordinality column number from the plan node.\n\nI have to bounce this back to you one more time, hopefully the last one\nI hope. Two things:\n\n1. Please verify that pg_stat_statements behaves correctly. The patch\ndoesn't have changes to contrib/ so without testing I'm guessing that it\ndoesn't work. I think something very simple should do.\n\n2. As I've complained many times, I find the way we manage an empty\nCOLUMNS clause pretty bad. The standard doesn't require that syntax\n(COLUMNS is required), and I don't like the implementation, so why not\nprovide the feature in a different way? My proposal is to change the\ncolumn options in gram.y to be something like this:The clause COLUMNS is optional on Oracle and DB2 So I prefer a Oracle, DB2 design. If you are strongly against it, then we can remove it to be ANSI/SQL only.I am don't see an good idea to introduce third syntax.\n\nxmltable_column_option_el:\n IDENT b_expr\n { $$ = makeDefElem($1, $2, @1); }\n | DEFAULT b_expr\n { $$ = makeDefElem(\"default\", $2, @1); }\n | FULL VALUE_P\n { $$ = makeDefElem(\"full_value\", NULL, @1); }\n | NOT NULL_P\n { $$ = makeDefElem(\"is_not_null\", (Node *) makeInteger(true), @1); }\n | NULL_P\n { $$ = makeDefElem(\"is_not_null\", (Node *) makeInteger(false), @1); }\n ;\n\nNote the FULL VALUE. Then we can process it like\n\n else if (strcmp(defel->defname, \"full_value\") == 0)\n {\n if (fc->colexpr != NULL)\n ereport(ERROR,\n (errcode(ERRCODE_SYNTAX_ERROR),\n errmsg(\"FULL ROW may not be specified together with PATH\"),\n parser_errposition(defel->location)));\n fc->full_row = true;\n }\n\nSo if you want the full XML value of the row, you have to specify it,\n\n .. XMLTABLE ( ... COLUMNS ..., whole_row xml FULL VALUE, ... )\n\nThis has the extra feature that you can add, say, an ORDINALITY column\ntogether with the XML value, something that you cannot do with the\ncurrent implementation.\n\nIt doesn't have to be FULL VALUE, but I couldn't think of anything\nbetter. (I didn't want to add any new keywords for this.) If you have\na better idea, let's discuss.I don't see a introduction own syntax as necessary solution here - use Oracle, DB2 compatible syntax, or ANSI. It is partially corner case - the benefit of this case is almost bigger compatibility with mentioned databases. \n\nCode-wise, this completely removes the \"else\" block in transformRangeTableFunc\nwhich I marked with an XXX comment. That's a good thing -- let's get\nrid of that. Also, it should remove the need for the separate \"if\n!columns\" case in tfuncLoadRows. All those cases would become part of\nthe normal code path instead of special cases. I think\nXmlTableSetColumnFilter doesn't need any change (we just don't call if\nfor the FULL VALUE row); and XmlTableGetValue needs a special case that\nif the column filter is NULL (i.e. SetColumnFilter wasn't called for\nthat column) then return the whole row.\n\n\nOf course, this opens an implementation issue: how do you annotate\nthings from parse analysis till execution? The current TableFunc\nstructure doesn't help, because there are only lists of column names and\nexpressions; and we can't use the case of a NULL colexpr, because that\ncase is already used by the column filter being the column name (a\nfeature required by the standard). A simple way would be to have a new\n\"colno\" struct member, to store a column number for the column marked\nFULL VALUE (just like ordinalitycol). This means you can't have more\nthan one of those FULL VALUE columns, but that seems okay.\n\n\n(Of course, this means that the two cases that have no COLUMNS in the\n\"xmltable\" production in gram.y should go away).You are commiter, and you should to decide - as first I prefer current state, as second a remove this part - it should be good for you too, because code that you don't like will be left.I dislike introduce new syntax - this case is not too important for this.I am able to prepare reduced version if we do a agreement RegardsPavel RegardsPavel\n\n--\nÁlvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Fri, 3 Mar 2017 20:23:03 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Pavel Stehule wrote:\n> 2017-03-03 19:15 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> > 2. As I've complained many times, I find the way we manage an empty\n> > COLUMNS clause pretty bad. The standard doesn't require that syntax\n> > (COLUMNS is required), and I don't like the implementation, so why not\n> > provide the feature in a different way? My proposal is to change the\n> > column options in gram.y to be something like this:\n> \n> The clause COLUMNS is optional on Oracle and DB2\n> \n> So I prefer a Oracle, DB2 design. If you are strongly against it, then we\n> can remove it to be ANSI/SQL only.\n> \n> I am don't see an good idea to introduce third syntax.\n\nOK. I think trying to be syntax compatible with DB2 or Oracle is a lost\ncause, because the syntax used in the XPath expressions seems different\n-- I think Oracle uses XQuery (which we don't support) and DB2 uses ...\nnot sure what it is, but it doesn't work in our implementation\n(stuff like '$d/employees/emp' in the row expression.)\n\nIn existing applications using those Oracle/DB2, is it common to omit\nthe COLUMNS clause? I searched for \"xmltable oracle\" and had a look at\nthe first few hits outside of the oracle docs:\nhttp://viralpatel.net/blogs/oracle-xmltable-tutorial/\nhttp://www.dba-oracle.com/t_xmltable.htm\nhttp://stackoverflow.com/questions/12690868/how-to-use-xmltable-in-oracle\nhttps://asktom.oracle.com/pls/asktom/f?p=100:11:0::::P11_QUESTION_ID:9533111800346252295\nhttp://stackoverflow.com/questions/1222570/what-is-an-xmltable\nhttps://community.oracle.com/thread/3955198\n\nNot a single one of these omit the COLUMNS clause (though the second one\nmentions that the clause can be omitted).\n\nI also looked at a few samples with DB2 -- same thing; it is possible,\nbut is it common?\n\nAnyway, I noticed that \"xml PATH '.'\" can be used to obtain the full XML\nof the row, which I think is the feature I wanted, so I think we're\ncovered and we can omit the case with no COLUMNS, since we already have\nthe feature in another way. No need to implement anything further, and\nwe can rip out the special case I don't like. Example:\n\nCREATE TABLE EMPLOYEES\n(\n id integer,\n data XML\n);\nINSERT INTO EMPLOYEES\n VALUES (1, '<Employees>\n <Employee emplid=\"1111\" type=\"admin\">\n <firstname>John</firstname>\n <lastname>Watson</lastname>\n <age>30</age>\n <email>johnwatson@sh.com</email>\n </Employee>\n <Employee emplid=\"2222\" type=\"admin\">\n <firstname>Sherlock</firstname>\n <lastname>Homes</lastname>\n <age>32</age>\n <email>sherlock@sh.com</email>\n </Employee>\n <Employee emplid=\"3333\" type=\"user\">\n <firstname>Jim</firstname>\n <lastname>Moriarty</lastname>\n <age>52</age>\n <email>jim@sh.com</email>\n </Employee>\n <Employee emplid=\"4444\" type=\"user\">\n <firstname>Mycroft</firstname>\n <lastname>Holmes</lastname>\n <age>41</age>\n <email>mycroft@sh.com</email>\n </Employee>\n</Employees>');\n\nThis is with COLUMNS omitted:\n\nalvherre=# select xmltable.* from employees, xmltable('/Employees/Employee' passing data);\n xmltable \n──────────────────────────────────────────\n <Employee emplid=\"1111\" type=\"admin\"> ↵\n <firstname>John</firstname> ↵\n <lastname>Watson</lastname> ↵\n <age>30</age> ↵\n <email>johnwatson@sh.com</email>↵\n </Employee>\n <Employee emplid=\"2222\" type=\"admin\"> ↵\n <firstname>Sherlock</firstname> ↵\n <lastname>Homes</lastname> ↵\n <age>32</age> ↵\n <email>sherlock@sh.com</email> ↵\n </Employee>\n <Employee emplid=\"3333\" type=\"user\"> ↵\n <firstname>Jim</firstname> ↵\n <lastname>Moriarty</lastname> ↵\n <age>52</age> ↵\n <email>jim@sh.com</email> ↵\n </Employee>\n <Employee emplid=\"4444\" type=\"user\"> ↵\n <firstname>Mycroft</firstname> ↵\n <lastname>Holmes</lastname> ↵\n <age>41</age> ↵\n <email>mycroft@sh.com</email> ↵\n </Employee>\n\nand this is what you get with \"xml PATH '.'\" (I threw in ORDINALITY just\nfor fun):\n\nalvherre=# select xmltable.* from employees, xmltable('/Employees/Employee' passing data columns row_number for ordinality, emp xml path '.');\n row_number │ emp \n────────────┼──────────────────────────────────────────\n 1 │ <Employee emplid=\"1111\" type=\"admin\"> ↵\n │ <firstname>John</firstname> ↵\n │ <lastname>Watson</lastname> ↵\n │ <age>30</age> ↵\n │ <email>johnwatson@sh.com</email>↵\n │ </Employee>\n 2 │ <Employee emplid=\"2222\" type=\"admin\"> ↵\n │ <firstname>Sherlock</firstname> ↵\n │ <lastname>Homes</lastname> ↵\n │ <age>32</age> ↵\n │ <email>sherlock@sh.com</email> ↵\n │ </Employee>\n 3 │ <Employee emplid=\"3333\" type=\"user\"> ↵\n │ <firstname>Jim</firstname> ↵\n │ <lastname>Moriarty</lastname> ↵\n │ <age>52</age> ↵\n │ <email>jim@sh.com</email> ↵\n │ </Employee>\n 4 │ <Employee emplid=\"4444\" type=\"user\"> ↵\n │ <firstname>Mycroft</firstname> ↵\n │ <lastname>Holmes</lastname> ↵\n │ <age>41</age> ↵\n │ <email>mycroft@sh.com</email> ↵\n │ </Employee>\n\n-- \nÁlvaro Herrera https://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":"Fri, 3 Mar 2017 17:04:28 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"2017-03-03 21:04 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> Pavel Stehule wrote:\n> > 2017-03-03 19:15 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n>\n> > > 2. As I've complained many times, I find the way we manage an empty\n> > > COLUMNS clause pretty bad. The standard doesn't require that syntax\n> > > (COLUMNS is required), and I don't like the implementation, so why not\n> > > provide the feature in a different way? My proposal is to change the\n> > > column options in gram.y to be something like this:\n> >\n> > The clause COLUMNS is optional on Oracle and DB2\n> >\n> > So I prefer a Oracle, DB2 design. If you are strongly against it, then we\n> > can remove it to be ANSI/SQL only.\n> >\n> > I am don't see an good idea to introduce third syntax.\n>\n> OK. I think trying to be syntax compatible with DB2 or Oracle is a lost\n> cause, because the syntax used in the XPath expressions seems different\n> -- I think Oracle uses XQuery (which we don't support) and DB2 uses ...\n> not sure what it is, but it doesn't work in our implementation\n> (stuff like '$d/employees/emp' in the row expression.)\n>\n\n100% compatibility is not possible - but XPath is subset of XQuery and in\nreality - the full XQuery examples of XMLTABLE is not often.\n\nAlmost all examples of usage XMLTABLE, what I found in blogs, uses XPath\nonly\n\n\n>\n> In existing applications using those Oracle/DB2, is it common to omit\n> the COLUMNS clause? I searched for \"xmltable oracle\" and had a look at\n> the first few hits outside of the oracle docs:\n> http://viralpatel.net/blogs/oracle-xmltable-tutorial/\n> http://www.dba-oracle.com/t_xmltable.htm\n> http://stackoverflow.com/questions/12690868/how-to-use-xmltable-in-oracle\n> https://asktom.oracle.com/pls/asktom/f?p=100:11:0::::P11_QUESTION_ID:\n> 9533111800346252295\n> http://stackoverflow.com/questions/1222570/what-is-an-xmltable\n> https://community.oracle.com/thread/3955198\n>\n> Not a single one of these omit the COLUMNS clause (though the second one\n> mentions that the clause can be omitted).\n>\n> I also looked at a few samples with DB2 -- same thing; it is possible,\n> but is it common?\n>\n\nI don't think so it is common - it is corner case - and I can live without\nit well\n\n\n>\n> Anyway, I noticed that \"xml PATH '.'\" can be used to obtain the full XML\n> of the row, which I think is the feature I wanted, so I think we're\n> covered and we can omit the case with no COLUMNS, since we already have\n> the feature in another way. No need to implement anything further, and\n> we can rip out the special case I don't like. Example:\n>\n\nyes,\n\n\n>\n> CREATE TABLE EMPLOYEES\n> (\n> id integer,\n> data XML\n> );\n> INSERT INTO EMPLOYEES\n> VALUES (1, '<Employees>\n> <Employee emplid=\"1111\" type=\"admin\">\n> <firstname>John</firstname>\n> <lastname>Watson</lastname>\n> <age>30</age>\n> <email>johnwatson@sh.com</email>\n> </Employee>\n> <Employee emplid=\"2222\" type=\"admin\">\n> <firstname>Sherlock</firstname>\n> <lastname>Homes</lastname>\n> <age>32</age>\n> <email>sherlock@sh.com</email>\n> </Employee>\n> <Employee emplid=\"3333\" type=\"user\">\n> <firstname>Jim</firstname>\n> <lastname>Moriarty</lastname>\n> <age>52</age>\n> <email>jim@sh.com</email>\n> </Employee>\n> <Employee emplid=\"4444\" type=\"user\">\n> <firstname>Mycroft</firstname>\n> <lastname>Holmes</lastname>\n> <age>41</age>\n> <email>mycroft@sh.com</email>\n> </Employee>\n> </Employees>');\n>\n> This is with COLUMNS omitted:\n>\n> alvherre=# select xmltable.* from employees,\n> xmltable('/Employees/Employee' passing data);\n> xmltable\n> ──────────────────────────────────────────\n> <Employee emplid=\"1111\" type=\"admin\"> ↵\n> <firstname>John</firstname> ↵\n> <lastname>Watson</lastname> ↵\n> <age>30</age> ↵\n> <email>johnwatson@sh.com</email>↵\n> </Employee>\n> <Employee emplid=\"2222\" type=\"admin\"> ↵\n> <firstname>Sherlock</firstname> ↵\n> <lastname>Homes</lastname> ↵\n> <age>32</age> ↵\n> <email>sherlock@sh.com</email> ↵\n> </Employee>\n> <Employee emplid=\"3333\" type=\"user\"> ↵\n> <firstname>Jim</firstname> ↵\n> <lastname>Moriarty</lastname> ↵\n> <age>52</age> ↵\n> <email>jim@sh.com</email> ↵\n> </Employee>\n> <Employee emplid=\"4444\" type=\"user\"> ↵\n> <firstname>Mycroft</firstname> ↵\n> <lastname>Holmes</lastname> ↵\n> <age>41</age> ↵\n> <email>mycroft@sh.com</email> ↵\n> </Employee>\n>\n> and this is what you get with \"xml PATH '.'\" (I threw in ORDINALITY just\n> for fun):\n>\n> alvherre=# select xmltable.* from employees,\n> xmltable('/Employees/Employee' passing data columns row_number for\n> ordinality, emp xml path '.');\n> row_number │ emp\n> ────────────┼──────────────────────────────────────────\n> 1 │ <Employee emplid=\"1111\" type=\"admin\"> ↵\n> │ <firstname>John</firstname> ↵\n> │ <lastname>Watson</lastname> ↵\n> │ <age>30</age> ↵\n> │ <email>johnwatson@sh.com</email>↵\n> │ </Employee>\n> 2 │ <Employee emplid=\"2222\" type=\"admin\"> ↵\n> │ <firstname>Sherlock</firstname> ↵\n> │ <lastname>Homes</lastname> ↵\n> │ <age>32</age> ↵\n> │ <email>sherlock@sh.com</email> ↵\n> │ </Employee>\n> 3 │ <Employee emplid=\"3333\" type=\"user\"> ↵\n> │ <firstname>Jim</firstname> ↵\n> │ <lastname>Moriarty</lastname> ↵\n> │ <age>52</age> ↵\n> │ <email>jim@sh.com</email> ↵\n> │ </Employee>\n> 4 │ <Employee emplid=\"4444\" type=\"user\"> ↵\n> │ <firstname>Mycroft</firstname> ↵\n> │ <lastname>Holmes</lastname> ↵\n> │ <age>41</age> ↵\n> │ <email>mycroft@sh.com</email> ↵\n> │ </Employee>\n>\n> --\n> Álvaro Herrera https://www.2ndQuadrant.com/\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>\n\n2017-03-03 21:04 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\n> 2017-03-03 19:15 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> > 2. As I've complained many times, I find the way we manage an empty\n> > COLUMNS clause pretty bad. The standard doesn't require that syntax\n> > (COLUMNS is required), and I don't like the implementation, so why not\n> > provide the feature in a different way? My proposal is to change the\n> > column options in gram.y to be something like this:\n>\n> The clause COLUMNS is optional on Oracle and DB2\n>\n> So I prefer a Oracle, DB2 design. If you are strongly against it, then we\n> can remove it to be ANSI/SQL only.\n>\n> I am don't see an good idea to introduce third syntax.\n\nOK. I think trying to be syntax compatible with DB2 or Oracle is a lost\ncause, because the syntax used in the XPath expressions seems different\n-- I think Oracle uses XQuery (which we don't support) and DB2 uses ...\nnot sure what it is, but it doesn't work in our implementation\n(stuff like '$d/employees/emp' in the row expression.)100% compatibility is not possible - but XPath is subset of XQuery and in reality - the full XQuery examples of XMLTABLE is not often.Almost all examples of usage XMLTABLE, what I found in blogs, uses XPath only \n\nIn existing applications using those Oracle/DB2, is it common to omit\nthe COLUMNS clause? I searched for \"xmltable oracle\" and had a look at\nthe first few hits outside of the oracle docs:\nhttp://viralpatel.net/blogs/oracle-xmltable-tutorial/\nhttp://www.dba-oracle.com/t_xmltable.htm\nhttp://stackoverflow.com/questions/12690868/how-to-use-xmltable-in-oracle\nhttps://asktom.oracle.com/pls/asktom/f?p=100:11:0::::P11_QUESTION_ID:9533111800346252295\nhttp://stackoverflow.com/questions/1222570/what-is-an-xmltable\nhttps://community.oracle.com/thread/3955198\n\nNot a single one of these omit the COLUMNS clause (though the second one\nmentions that the clause can be omitted).\n\nI also looked at a few samples with DB2 -- same thing; it is possible,\nbut is it common?I don't think so it is common - it is corner case - and I can live without it well \n\nAnyway, I noticed that \"xml PATH '.'\" can be used to obtain the full XML\nof the row, which I think is the feature I wanted, so I think we're\ncovered and we can omit the case with no COLUMNS, since we already have\nthe feature in another way. No need to implement anything further, and\nwe can rip out the special case I don't like. Example:yes, \n\nCREATE TABLE EMPLOYEES\n(\n id integer,\n data XML\n);\nINSERT INTO EMPLOYEES\n VALUES (1, '<Employees>\n <Employee emplid=\"1111\" type=\"admin\">\n <firstname>John</firstname>\n <lastname>Watson</lastname>\n <age>30</age>\n <email>johnwatson@sh.com</email>\n </Employee>\n <Employee emplid=\"2222\" type=\"admin\">\n <firstname>Sherlock</firstname>\n <lastname>Homes</lastname>\n <age>32</age>\n <email>sherlock@sh.com</email>\n </Employee>\n <Employee emplid=\"3333\" type=\"user\">\n <firstname>Jim</firstname>\n <lastname>Moriarty</lastname>\n <age>52</age>\n <email>jim@sh.com</email>\n </Employee>\n <Employee emplid=\"4444\" type=\"user\">\n <firstname>Mycroft</firstname>\n <lastname>Holmes</lastname>\n <age>41</age>\n <email>mycroft@sh.com</email>\n </Employee>\n</Employees>');\n\nThis is with COLUMNS omitted:\n\nalvherre=# select xmltable.* from employees, xmltable('/Employees/Employee' passing data);\n xmltable\n──────────────────────────────────────────\n <Employee emplid=\"1111\" type=\"admin\"> ↵\n <firstname>John</firstname> ↵\n <lastname>Watson</lastname> ↵\n <age>30</age> ↵\n <email>johnwatson@sh.com</email>↵\n </Employee>\n <Employee emplid=\"2222\" type=\"admin\"> ↵\n <firstname>Sherlock</firstname> ↵\n <lastname>Homes</lastname> ↵\n <age>32</age> ↵\n <email>sherlock@sh.com</email> ↵\n </Employee>\n <Employee emplid=\"3333\" type=\"user\"> ↵\n <firstname>Jim</firstname> ↵\n <lastname>Moriarty</lastname> ↵\n <age>52</age> ↵\n <email>jim@sh.com</email> ↵\n </Employee>\n <Employee emplid=\"4444\" type=\"user\"> ↵\n <firstname>Mycroft</firstname> ↵\n <lastname>Holmes</lastname> ↵\n <age>41</age> ↵\n <email>mycroft@sh.com</email> ↵\n </Employee>\n\nand this is what you get with \"xml PATH '.'\" (I threw in ORDINALITY just\nfor fun):\n\nalvherre=# select xmltable.* from employees, xmltable('/Employees/Employee' passing data columns row_number for ordinality, emp xml path '.');\n row_number │ emp\n────────────┼──────────────────────────────────────────\n 1 │ <Employee emplid=\"1111\" type=\"admin\"> ↵\n │ <firstname>John</firstname> ↵\n │ <lastname>Watson</lastname> ↵\n │ <age>30</age> ↵\n │ <email>johnwatson@sh.com</email>↵\n │ </Employee>\n 2 │ <Employee emplid=\"2222\" type=\"admin\"> ↵\n │ <firstname>Sherlock</firstname> ↵\n │ <lastname>Homes</lastname> ↵\n │ <age>32</age> ↵\n │ <email>sherlock@sh.com</email> ↵\n │ </Employee>\n 3 │ <Employee emplid=\"3333\" type=\"user\"> ↵\n │ <firstname>Jim</firstname> ↵\n │ <lastname>Moriarty</lastname> ↵\n │ <age>52</age> ↵\n │ <email>jim@sh.com</email> ↵\n │ </Employee>\n 4 │ <Employee emplid=\"4444\" type=\"user\"> ↵\n │ <firstname>Mycroft</firstname> ↵\n │ <lastname>Holmes</lastname> ↵\n │ <age>41</age> ↵\n │ <email>mycroft@sh.com</email> ↵\n │ </Employee>\n\n--\nÁlvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Fri, 3 Mar 2017 22:41:57 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Hi\n\nI used your idea about special columns when COLUMNS are not explicitly\ndefined.\n\nAll lines that you are dislike removed.\n\nNow, almost all code, related to this behave, is in next few lines.\n\n+ /*\n+ * Use implicit column when it is necessary. The COLUMNS clause is\noptional\n+ * on Oracle and DB2. In this case a result is complete row of XML type.\n+ */\n+ if (rtf->columns == NIL)\n+ {\n+ RangeTableFuncCol *fc = makeNode(RangeTableFuncCol);\n+ A_Const *n = makeNode(A_Const);\n+\n+ fc->colname = \"xmltable\";\n+ fc->typeName = makeTypeNameFromOid(XMLOID, -1);\n+ n->val.type = T_String;\n+ n->val.val.str = \".\";\n+ n->location = -1;\n+\n+ fc->colexpr = (Node *) n;\n+ rtf->columns = list_make1(fc);\n+ }\n\nall regress tests passing.\n\nRegards\n\nPavel\n\n\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, 5 Mar 2017 10:32:18 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Pavel Stehule wrote:\n> Hi\n> \n> I used your idea about special columns when COLUMNS are not explicitly\n> defined.\n> \n> All lines that you are dislike removed.\n\nI just pushed XMLTABLE, after some additional changes. Please test it\nthoroughly and report any problems.\n\nI didn't add the change you proposed here to keep COLUMNS optional;\ninstead, I just made COLUMNS mandatory. I think what you propose here\nis not entirely out of the question, but you left out ruleutils.c\nsupport for it, so I decided to leave it aside for now so that I could\nget this patch out of my plate once and for all. If you really want\nthat feature, you can submit another patch for it and discuss with the\nRMT whether it belongs in PG10 or not.\n\nSome changes I made:\n* I added some pg_stat_statements support. It works fine for simple\ntests, but deeper testing of it would be appreciated.\n\n* I removed the \"buildercxt\" memory context. It seemed mostly\npointless, and I was disturbed by the MemoryContextResetOnly().\nPer-value memory still uses the per-value memory context, but the rest\nof the stuff is in the per-query context, which should be pretty much\nthe same.\n\n* Desultory stylistic changes\n\n-- \n�lvaro Herrera https://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, 8 Mar 2017 13:01:01 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"2017-03-08 17:01 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> Pavel Stehule wrote:\n> > Hi\n> >\n> > I used your idea about special columns when COLUMNS are not explicitly\n> > defined.\n> >\n> > All lines that you are dislike removed.\n>\n> I just pushed XMLTABLE, after some additional changes. Please test it\n> thoroughly and report any problems.\n>\n\nThank you\n\n>\n> I didn't add the change you proposed here to keep COLUMNS optional;\n> instead, I just made COLUMNS mandatory. I think what you propose here\n> is not entirely out of the question, but you left out ruleutils.c\n> support for it, so I decided to leave it aside for now so that I could\n> get this patch out of my plate once and for all. If you really want\n> that feature, you can submit another patch for it and discuss with the\n> RMT whether it belongs in PG10 or not.\n>\n\nIt is interesting feature - because it replaces XPATH function, but not\nimportant enough.\n\nFor daily work the default schema support is much more interesting.\n\n\n>\n> Some changes I made:\n> * I added some pg_stat_statements support. It works fine for simple\n> tests, but deeper testing of it would be appreciated.\n>\n> * I removed the \"buildercxt\" memory context. It seemed mostly\n> pointless, and I was disturbed by the MemoryContextResetOnly().\n> Per-value memory still uses the per-value memory context, but the rest\n> of the stuff is in the per-query context, which should be pretty much\n> the same.\n>\n> * Desultory stylistic changes\n>\n\nok\n\nRegards\n\nPavel\n\n>\n> --\n> Álvaro Herrera https://www.2ndQuadrant.com/\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>\n\n2017-03-08 17:01 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\n> Hi\n>\n> I used your idea about special columns when COLUMNS are not explicitly\n> defined.\n>\n> All lines that you are dislike removed.\n\nI just pushed XMLTABLE, after some additional changes. Please test it\nthoroughly and report any problems.Thank you \n\nI didn't add the change you proposed here to keep COLUMNS optional;\ninstead, I just made COLUMNS mandatory. I think what you propose here\nis not entirely out of the question, but you left out ruleutils.c\nsupport for it, so I decided to leave it aside for now so that I could\nget this patch out of my plate once and for all. If you really want\nthat feature, you can submit another patch for it and discuss with the\nRMT whether it belongs in PG10 or not.It is interesting feature - because it replaces XPATH function, but not important enough.For daily work the default schema support is much more interesting. \n\nSome changes I made:\n* I added some pg_stat_statements support. It works fine for simple\ntests, but deeper testing of it would be appreciated.\n\n* I removed the \"buildercxt\" memory context. It seemed mostly\npointless, and I was disturbed by the MemoryContextResetOnly().\nPer-value memory still uses the per-value memory context, but the rest\nof the stuff is in the per-query context, which should be pretty much\nthe same.\n\n* Desultory stylistic changesokRegardsPavel \n\n--\nÁlvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Wed, 8 Mar 2017 17:10:53 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Pavel Stehule wrote:\n> 2017-03-08 17:01 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> > I didn't add the change you proposed here to keep COLUMNS optional;\n> > instead, I just made COLUMNS mandatory. I think what you propose here\n> > is not entirely out of the question, but you left out ruleutils.c\n> > support for it, so I decided to leave it aside for now so that I could\n> > get this patch out of my plate once and for all. If you really want\n> > that feature, you can submit another patch for it and discuss with the\n> > RMT whether it belongs in PG10 or not.\n> \n> It is interesting feature - because it replaces XPATH function, but not\n> important enough.\n\nOK.\n\n> For daily work the default schema support is much more interesting.\n\nLet's see that one, then. It was part of the original submission so\ndepending on how the patch we looks can still cram it in. But other\npatches have priority for me now.\n\n-- \n�lvaro Herrera https://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, 8 Mar 2017 13:32:19 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"2017-03-08 17:32 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> Pavel Stehule wrote:\n> > 2017-03-08 17:01 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n>\n> > > I didn't add the change you proposed here to keep COLUMNS optional;\n> > > instead, I just made COLUMNS mandatory. I think what you propose here\n> > > is not entirely out of the question, but you left out ruleutils.c\n> > > support for it, so I decided to leave it aside for now so that I could\n> > > get this patch out of my plate once and for all. If you really want\n> > > that feature, you can submit another patch for it and discuss with the\n> > > RMT whether it belongs in PG10 or not.\n> >\n> > It is interesting feature - because it replaces XPATH function, but not\n> > important enough.\n>\n> OK.\n>\n> > For daily work the default schema support is much more interesting.\n>\n> Let's see that one, then. It was part of the original submission so\n> depending on how the patch we looks can still cram it in. But other\n> patches have priority for me now.\n>\n\nIt is theme for 11\n\nThank you very much\n\nPavel\n\n\n>\n> --\n> Álvaro Herrera https://www.2ndQuadrant.com/\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>\n\n2017-03-08 17:32 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\n> 2017-03-08 17:01 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> > I didn't add the change you proposed here to keep COLUMNS optional;\n> > instead, I just made COLUMNS mandatory. I think what you propose here\n> > is not entirely out of the question, but you left out ruleutils.c\n> > support for it, so I decided to leave it aside for now so that I could\n> > get this patch out of my plate once and for all. If you really want\n> > that feature, you can submit another patch for it and discuss with the\n> > RMT whether it belongs in PG10 or not.\n>\n> It is interesting feature - because it replaces XPATH function, but not\n> important enough.\n\nOK.\n\n> For daily work the default schema support is much more interesting.\n\nLet's see that one, then. It was part of the original submission so\ndepending on how the patch we looks can still cram it in. But other\npatches have priority for me now.It is theme for 11Thank you very muchPavel \n\n--\nÁlvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Wed, 8 Mar 2017 17:46:48 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"Re: patch: function xmltable"},{"msg_contents":"Pavel Stehule wrote:\n> 2017-03-08 17:32 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\n\n> > > For daily work the default schema support is much more interesting.\n> >\n> > Let's see that one, then. It was part of the original submission so\n> > depending on how the patch we looks can still cram it in. But other\n> > patches have priority for me now.\n> \n> It is theme for 11\n\nAh, great.\n\n-- \n�lvaro Herrera https://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, 8 Mar 2017 14:01:37 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: patch: function xmltable"}],"string":"[\n {\n \"msg_contents\": \"Hi\\n\\nI am sending implementation of xmltable function. The code should to have\\nnear to final quality and it is available for testing.\\n\\nI invite any help with documentation and testing.\\n\\nRegards\\n\\nPavel\\n\\n\\n-- \\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 Aug 2016 10:58:52 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2016-08-19 10:58 GMT+02:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n\\n> Hi\\n>\\n> I am sending implementation of xmltable function. The code should to have\\n> near to final quality and it is available for testing.\\n>\\n> I invite any help with documentation and testing.\\n>\\n\\nnew update - the work with nodes is much more correct now.\\n\\nRegards\\n\\nPavel\\n\\n\\n> Regards\\n>\\n> Pavel\\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 Aug 2016 21:00:12 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-08-23 21:00 GMT+02:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n\\n> Hi\\n>\\n> 2016-08-19 10:58 GMT+02:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n>\\n>> Hi\\n>>\\n>> I am sending implementation of xmltable function. The code should to have\\n>> near to final quality and it is available for testing.\\n>>\\n>> I invite any help with documentation and testing.\\n>>\\n>\\n> new update - the work with nodes is much more correct now.\\n>\\n\\nnext update\\n\\nfix memory leak\\n\\nPavel\\n\\n\\n>\\n> Regards\\n>\\n> Pavel\\n>\\n>\\n>> Regards\\n>>\\n>> Pavel\\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\": \"Wed, 24 Aug 2016 06:56:16 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\nminor update - using DefElem instead own private parser type\\n\\nRegards\\n\\nPavel\\n\\n\\n-- \\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 Sep 2016 10:06:55 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On 4 September 2016 at 16:06, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n> Hi\\n>\\n> minor update - using DefElem instead own private parser type\\n\\nI'm really glad that you're doing this and I'll take a look at it for this CF.\\n\\nIt's quite a big patch so I expect this will take a few rounds of\\nreview and updating.\\n\\n\\nPatch applies cleanly and builds cleanly on master both with and\\nwithout --with-xml .\\n\\nOverall, I think this needs to be revised with appropriate comments.\\nWhitespace/formatting needs fixing since it's all over the place.\\nDocumentation is insufficient (per notes below).\\n\\nRe identifier naming, some of this code uses XmlTable naming patterns,\\nsome uses TableExpr prefixes. Is that intended to indicate a bounary\\nbetween things re-usable for other structured data ingesting\\nfunctions? Do you expect a \\\"JSONEXPR\\\" or similar in future? That's\\nalluded to by\\n\\n+/*----------\\n+ * TableExpr - used for XMLTABLE function\\n+ *\\n+ * This can be used for json_table, jsonb_table functions in future\\n+ *----------\\n+ */\\n+typedef struct TableExpr\\n+{\\n...\\n\\nIf so, should this really be two patches, one to add the table\\nexpression infrastructure and another to add XMLTABLE that uses it?\\nAlso, why in that case does so much of the TableExpr code call\\ndirectly into XmlTable code? It doesn't look very generic.\\n\\nOverall I find identifier naming to be a bit inconsisent and think\\nit's necessary to make it clear that all the \\\"TableExpr\\\" stuff is for\\nXMLTABLE specifically, if that's the case, or make the delineation\\nclearer if not.\\n\\nI'd also like to see tests that exercise the ruleutils get_rule_expr\\nparts of the code for the various XMLTABLE variants.\\n\\nSimilarly, since this seems to add a new xpath parser, that needs\\ncomprehensive tests. Maybe re-purpose an existing xpath test data set?\\n\\n\\n\\n\\nMore detailed comments:\\n====\\n\\nDocs comments:\\n\\n The <function>xmltable</function> produces [a] table based on\\n[the] passed XML value.\\n\\nThe docs are pretty minimal and don't explain the various clauses of\\nXMLTABLE. What is \\\"BY REF\\\" ? Is PATH an xpath expression? If so, is\\nthere a good cross reference link available? The PASSING clause? etc.\\n\\nHow does XMLTABLE decide what to iterate over, and how to iterate over it?\\n\\nPresumably the FOR ORDINALITY clause makes a column emit a numeric counter.\\n\\nWhat standard, if any, does this conform to? Does it resemble\\nimplementations elsewhere? What limitations or unsupported features\\ndoes it have relative to those standards?\\n\\n\\n\\nexecEvalTableExpr seems to be defined twice, with a difference in\\ncase. This is probably not going to fly:\\n\\n\\n+static Datum\\n+execEvalTableExpr(TableExprState *tstate,\\n+ ExprContext *econtext,\\n+ bool *isNull, ExprDoneCond *isDone)\\n+{\\n\\n+static Datum\\n+ExecEvalTableExpr(TableExprState *tstate,\\n+ ExprContext *econtext,\\n+ bool *isNull, ExprDoneCond *isDone)\\n+{\\n\\n\\nIt looks like you've split the function into a \\\"guts\\\" and \\\"wrapper\\\"\\npart, with the error handling PG_TRY / PG_CATCH block in the wrapper.\\nThat seems reasonable for readability, but the naming isn't.\\n\\nA comment is needed to explain what ExecEvalTableExpr is / does. If\\nit's XMLTABLE specific (which it looks like based on the code), its\\nname should reflect that. This pattern is repeated elsewhere; e.g.\\nTableExprState is really the state for an XMLTABLE expression. But\\nPostgreSQL actually has TABLE statements, and in future we might want\\nto support table-expressions, so I don't think this naming is\\nappropriate. This is made worse by the lack of comments on things like\\nthe definition of TableExprState. Please use something that makes it\\nclear it's for XMLTABLE and add appropriate comments.\\n\\nFormatting of variables, arguments, function signatures etc is\\nrandom/haphazard and doesn't follow project convention. It's neither\\naligned or unaligned in the normal way, I don't understand the random\\nspacing at all. Maybe you should try to run pgindent and then extract\\njust the changes related to your patch? Or run your IDE/editor's\\nindent function on your changes? Right now it's actually kind of hard\\nto read. Do you edit with tabstop set to 1 normally or something like\\nthat?\\n\\nThere's a general lack of comments throughout the added code.\\n\\nIn execEvalTableExpr, why are we looping over namespaces? What's that\\nfor? Comment would be nice.\\n\\nTypo: Path caclulation => Path calculation\\n\\nWhat does XmlTableSetRowPath() do? It seems to copy its argument.\\nNothing further is done with the row_path argument after it's called\\nby execEvalTableExpr, so what context is that memory in and do we have\\nto worry about it if it's large?\\n\\nexecEvalTableExpr says it's doing \\\"path calculation\\\". What it actually\\nappears to do is evaluate the path expressions, if provided, and\\notherwise use the column name as the implied path expression. (The\\ndocs should mention that).\\n\\nIt's wasn't immediately obvious to me what the branch around\\ntstate->for_ordinality_col is for and what the alternate path's\\npurpose is in terms of XMLTABLE's behaviour, until I read the parser\\ndefinition. That's largely because the behaviour of XMLTABLE is\\nunderspecified in the docs, since once you know ORDINALITY columns\\nexist it's pretty obvious what it's doing.\\n\\nSimilarly, for the alternate branch tstate->ncols , the\\nXmlTableGetRowValue call there is meant to do what exactly, and\\nwhy/under what conditions? Is it for situations where the field type\\nis a whole-row value? a composite type? (I'm deliberately not studying\\nthis too deeply, these are points I'd like to see commented so it can\\nbe understood to some reasonable degree at a skim-read).\\n\\n\\n /* result is one more columns every time */\\n\\\"one or more\\\"\\n\\n\\n\\n /* when typmod is not valid, refresh it */\\n if (te->typmod == -1)\\n\\n\\nIs this a cache? How is it valid or not valid and when? The comment\\n(thanks!) on TableExprGetTupleDesc says:\\n\\n/*\\n * When we skip transform stage (in view), then TableExpr's\\n * TupleDesc should not be valid. Refresh is necessary.\\n */\\n\\nbut I'm not really grasping what you're trying to explain here. What\\ntransform stage? What view? This could well be my ignorance of this\\npart of the code; if it should be understandable by a reader who is\\nappropriately familiar with the executor that's fine, but if it's\\nspecific to how XMLTABLE works some more explanation would be good.\\n\\nGood that you've got all the required node copy/in/out funcs in place.\\n\\nPlease don't use the name \\\"used_dns\\\". Anyone reading that will read it\\nas \\\"domain name service\\\" and that's actually confusing with XML\\nbecause of XML schema lookups. Maybe used_defnamespace ? used\\ndef_ns?\\n\\nI haven't looked closely at keyword/parser changes yet, but it doesn't\\nlook like you added any reserved keywords, which is good. It does add\\nunreserved keywords PATH and COLUMNS ; I'm not sure what policy for\\nunreserved keywords is or the significance of that.\\n\\nNew ereport() calls specify ERRCODEs, which is good.\\n\\nPostgreSQL already has XPATH support in the form of xmlexists(...)\\netc. Why is getXPathToken() etc needed? What re-use is possible here?\\nThere's no explanation in the patch header or comments. Should the new\\nxpath parser be re-used by the existing xpath stuff? Why can't we use\\nlibxml's facilities? etc. This at least needs explaining in the\\nsubmission, and some kind of hint as to why we have two different ways\\nto do it is needed in the code. If we do need a new XML parser, should\\nit be bundled in adt/xml.c along with a lot of user-facing\\nfunctionality, or a separate file?\\n\\n\\nHow does XmlTableGetValue(...) and XmlTableGetRowValue(...) relate to\\nthis? It doesn't look like they're intended to be called directly by\\nthe user, and they're not documented (or commented).\\n\\nI don't understand this at all:\\n\\n\\n\\n+/*\\n+ * There are different requests from XMLTABLE, JSON_TABLE functions\\n+ * on passed data than has CREATE TABLE command. It is reason for\\n+ * introduction special structure instead using ColumnDef.\\n+ */\\n+typedef struct TableExprRawCol\\n+{\\n+ NodeTag type;\\n+ char *colname;\\n+ TypeName *typeName;\\n+ bool for_ordinality;\\n+ bool is_not_null;\\n+ Node *path_expr;\\n+ Node *default_expr;\\n+ int location;\\n+} TableExprRawCol;\\n\\n\\n\\nThat's my first-pass commentary. I'll return to this once you've had a\\nchance to take a look at these and tell me all the places I got it\\nwrong ;)\\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\": \"Tue, 6 Sep 2016 12:54:31 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2016-09-06 6:54 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:\\n\\n> On 4 September 2016 at 16:06, Pavel Stehule <pavel.stehule@gmail.com>\\n> wrote:\\n> > Hi\\n> >\\n> > minor update - using DefElem instead own private parser type\\n>\\n> I'm really glad that you're doing this and I'll take a look at it for this\\n> CF.\\n>\\n> It's quite a big patch so I expect this will take a few rounds of\\n> review and updating.\\n>\\n\\nThank you for review\\n\\n\\n>\\n>\\n> Patch applies cleanly and builds cleanly on master both with and\\n> without --with-xml .\\n>\\n> Overall, I think this needs to be revised with appropriate comments.\\n> Whitespace/formatting needs fixing since it's all over the place.\\n> Documentation is insufficient (per notes below).\\n>\\n\\nI am not able to write documentation in English language :( - This function\\nis pretty complex - so I hope so anybody with better language skills can\\nhelp with this. It respects standard and it respects little bit different\\nOracle's behave too (different order of DEFAULT and PATH parts).\\n\\n\\n>\\n> Re identifier naming, some of this code uses XmlTable naming patterns,\\n> some uses TableExpr prefixes. Is that intended to indicate a bounary\\n> between things re-usable for other structured data ingesting\\n> functions? Do you expect a \\\"JSONEXPR\\\" or similar in future? That's\\n> alluded to by\\n>\\n\\nThis structure should be reused by JSON_TABLE function. Now, it is little\\nbit strange, because there is only XMLTABLE implementation - and I have to\\nchoose between a) using two different names now, b) renaming some part in\\nfuture.\\n\\nAnd although XMLTABLE and JSON_TABLE functions are pretty similar - share\\n90% of data (input value, path, columns definitions), these functions has\\ndifferent syntax - so only middle level code should be shared.\\n\\n\\n>\\n> +/*----------\\n> + * TableExpr - used for XMLTABLE function\\n> + *\\n> + * This can be used for json_table, jsonb_table functions in future\\n> + *----------\\n> + */\\n> +typedef struct TableExpr\\n> +{\\n> ...\\n>\\n> If so, should this really be two patches, one to add the table\\n> expression infrastructure and another to add XMLTABLE that uses it?\\n> Also, why in that case does so much of the TableExpr code call\\n> directly into XmlTable code? It doesn't look very generic.\\n>\\n\\nCurrently the common part is not too big - just the Node related part - I\\nam not sure about necessity of two patches. I am agree, there is missing\\nsome TableExpBuilder, where can be better isolated the XML part.\\n\\n\\n>\\n> Overall I find identifier naming to be a bit inconsisent and think\\n> it's necessary to make it clear that all the \\\"TableExpr\\\" stuff is for\\n> XMLTABLE specifically, if that's the case, or make the delineation\\n> clearer if not.\\n>\\n> I'd also like to see tests that exercise the ruleutils get_rule_expr\\n> parts of the code for the various XMLTABLE variants.\\n>\\n> Similarly, since this seems to add a new xpath parser, that needs\\n> comprehensive tests. Maybe re-purpose an existing xpath test data set?\\n>\\n>\\nsure\\n\\n\\n>\\n>\\n>\\n> More detailed comments:\\n> ====\\n>\\n> Docs comments:\\n>\\n> The <function>xmltable</function> produces [a] table based on\\n> [the] passed XML value.\\n>\\n> The docs are pretty minimal and don't explain the various clauses of\\n> XMLTABLE. What is \\\"BY REF\\\" ? Is PATH an xpath expression? If so, is\\n> there a good cross reference link available? The PASSING clause? etc.\\n>\\n> How does XMLTABLE decide what to iterate over, and how to iterate over it?\\n>\\n> Presumably the FOR ORDINALITY clause makes a column emit a numeric counter.\\n>\\n> What standard, if any, does this conform to? Does it resemble\\n> implementations elsewhere? What limitations or unsupported features\\n> does it have relative to those standards?\\n>\\n>\\n>\\n> execEvalTableExpr seems to be defined twice, with a difference in\\n> case. This is probably not going to fly:\\n>\\n>\\n> +static Datum\\n> +execEvalTableExpr(TableExprState *tstate,\\n> + ExprContext *econtext,\\n> + bool *isNull, ExprDoneCond *isDone)\\n> +{\\n>\\n> +static Datum\\n> +ExecEvalTableExpr(TableExprState *tstate,\\n> + ExprContext *econtext,\\n> + bool *isNull, ExprDoneCond *isDone)\\n> +{\\n>\\n>\\n> It looks like you've split the function into a \\\"guts\\\" and \\\"wrapper\\\"\\n> part, with the error handling PG_TRY / PG_CATCH block in the wrapper.\\n> That seems reasonable for readability, but the naming isn't.\\n>\\n\\nI invite any idea how these functions should be named.\\n\\n\\n>\\n> A comment is needed to explain what ExecEvalTableExpr is / does. If\\n> it's XMLTABLE specific (which it looks like based on the code), its\\n> name should reflect that. This pattern is repeated elsewhere; e.g.\\n> TableExprState is really the state for an XMLTABLE expression. But\\n> PostgreSQL actually has TABLE statements, and in future we might want\\n> to support table-expressions, so I don't think this naming is\\n> appropriate. This is made worse by the lack of comments on things like\\n> the definition of TableExprState. Please use something that makes it\\n> clear it's for XMLTABLE and add appropriate comments.\\n>\\n\\nI understand, so using TableExpr can be strange (for XMLTABLE function).\\nBut when we will have JSON_TABLE function, then it will have a sense.\\n\\n\\\"TableExprState\\\" is consistent with \\\"TableExpr\\\".\\n\\nAny idea how it should be changed?\\n\\n\\n\\n>\\n> Formatting of variables, arguments, function signatures etc is\\n> random/haphazard and doesn't follow project convention. It's neither\\n> aligned or unaligned in the normal way, I don't understand the random\\n> spacing at all. Maybe you should try to run pgindent and then extract\\n> just the changes related to your patch? Or run your IDE/editor's\\n> indent function on your changes? Right now it's actually kind of hard\\n> to read. Do you edit with tabstop set to 1 normally or something like\\n> that?\\n>\\n> There's a general lack of comments throughout the added code.\\n>\\n> In execEvalTableExpr, why are we looping over namespaces? What's that\\n> for? Comment would be nice.\\n>\\n> Typo: Path caclulation => Path calculation\\n>\\n> What does XmlTableSetRowPath() do? It seems to copy its argument.\\n> Nothing further is done with the row_path argument after it's called\\n> by execEvalTableExpr, so what context is that memory in and do we have\\n> to worry about it if it's large?\\n>\\n> execEvalTableExpr says it's doing \\\"path calculation\\\". What it actually\\n> appears to do is evaluate the path expressions, if provided, and\\n> otherwise use the column name as the implied path expression. (The\\n> docs should mention that).\\n>\\n> It's wasn't immediately obvious to me what the branch around\\n> tstate->for_ordinality_col is for and what the alternate path's\\n> purpose is in terms of XMLTABLE's behaviour, until I read the parser\\n> definition. That's largely because the behaviour of XMLTABLE is\\n> underspecified in the docs, since once you know ORDINALITY columns\\n> exist it's pretty obvious what it's doing.\\n>\\n> Similarly, for the alternate branch tstate->ncols , the\\n> XmlTableGetRowValue call there is meant to do what exactly, and\\n> why/under what conditions? Is it for situations where the field type\\n> is a whole-row value? a composite type? (I'm deliberately not studying\\n> this too deeply, these are points I'd like to see commented so it can\\n> be understood to some reasonable degree at a skim-read).\\n>\\n>\\n> /* result is one more columns every time */\\n> \\\"one or more\\\"\\n>\\n>\\n>\\n> /* when typmod is not valid, refresh it */\\n> if (te->typmod == -1)\\n>\\n>\\n> Is this a cache? How is it valid or not valid and when? The comment\\n> (thanks!) on TableExprGetTupleDesc says:\\n>\\n> /*\\n> * When we skip transform stage (in view), then TableExpr's\\n> * TupleDesc should not be valid. Refresh is necessary.\\n> */\\n>\\n> but I'm not really grasping what you're trying to explain here. What\\n> transform stage? What view? This could well be my ignorance of this\\n> part of the code; if it should be understandable by a reader who is\\n> appropriately familiar with the executor that's fine, but if it's\\n> specific to how XMLTABLE works some more explanation would be good.\\n>\\n\\nThis is most difficult part of this patch, and I am not sure it it is fully\\ncorrectly implemented. I use TupleDesc cache. The TupleDesc is created in\\nparser/transform stage. When the XMLTABLE is used in some view, then the\\ntransformed parser tree is materialized - and when the view is used in\\nquery, then this tree is loaded and the parser/transform stage is\\n\\\"skipped\\\". I'll check this code against implementation of ROW constructor\\nand I'll try to do more comments there.\\n\\n\\n\\n>\\n> Good that you've got all the required node copy/in/out funcs in place.\\n>\\n> Please don't use the name \\\"used_dns\\\". Anyone reading that will read it\\n> as \\\"domain name service\\\" and that's actually confusing with XML\\n> because of XML schema lookups. Maybe used_defnamespace ? used\\n> def_ns?\\n>\\n\\ngood idea\\n\\n\\n>\\n> I haven't looked closely at keyword/parser changes yet, but it doesn't\\n> look like you added any reserved keywords, which is good. It does add\\n> unreserved keywords PATH and COLUMNS ; I'm not sure what policy for\\n> unreserved keywords is or the significance of that.\\n>\\n> New ereport() calls specify ERRCODEs, which is good.\\n>\\n> PostgreSQL already has XPATH support in the form of xmlexists(...)\\n> etc. Why is getXPathToken() etc needed? What re-use is possible here?\\n> There's no explanation in the patch header or comments. Should the new\\n> xpath parser be re-used by the existing xpath stuff? Why can't we use\\n> libxml's facilities? etc. This at least needs explaining in the\\n> submission, and some kind of hint as to why we have two different ways\\n> to do it is needed in the code. If we do need a new XML parser, should\\n> it be bundled in adt/xml.c along with a lot of user-facing\\n> functionality, or a separate file?\\n>\\n>\\nlibxml2 and our XPATH function doesn't support default namespace (\\nhttp://plasmasturm.org/log/259/ ). This is pretty useful feature - so I\\nimplemented. This is the mayor issue of libxml2 library. Another difference\\nbetween XPATH function and XMLTABLE function is using two phase searching\\nand implicit prefix \\\"./\\\" and suffix (\\\"/text()\\\") in XMLTABLE. XMLTABLE using\\ntwo XPATH expressions - for row data cutting and next for column data\\ncutting (from row data). The our XPATH functions is pretty simple mapped to\\nlibxml2 XPATH API. But it is not possible with XMLTABLE function - due\\ndesign of this function in standard (it is more user friendly and doesn't\\nrequire exactly correct xpath expressions).\\n\\nI didn't find any API in libxml2 for a work with parsed xpath expressions -\\nI need some info about the first and last token of xpath expression - it is\\nbase for decision about using prefix or suffix.\\n\\nThis functionality (xpath expression parser) cannot be used for our XPATH\\nfunction now - maybe default namespace in future.\\n\\n\\n>\\n> How does XmlTableGetValue(...) and XmlTableGetRowValue(...) relate to\\n> this? It doesn't look like they're intended to be called directly by\\n> the user, and they're not documented (or commented).\\n>\\n\\nProbably I used wrong names. XMLTABLE function is running in two different\\nmodes - with explicitly defined columns (XmlTableGetValue is used), and\\nwithout explicitly defined columns - so result is one XML column and only\\none one step searching is used (there are not column related xpath\\nexpressions) ( XmlTableGetRowValue is used). The function XmlTableGetValue\\nis used for getting one column value, the function XmlTableGetRowValue is\\nused for getting one value too, but in special case, when there are not any\\nother value.\\n\\n>\\n> I don't understand this at all:\\n>\\n>\\n>\\n> +/*\\n> + * There are different requests from XMLTABLE, JSON_TABLE functions\\n> + * on passed data than has CREATE TABLE command. It is reason for\\n> + * introduction special structure instead using ColumnDef.\\n> + */\\n> +typedef struct TableExprRawCol\\n> +{\\n> + NodeTag type;\\n> + char *colname;\\n> + TypeName *typeName;\\n> + bool for_ordinality;\\n> + bool is_not_null;\\n> + Node *path_expr;\\n> + Node *default_expr;\\n> + int location;\\n> +} TableExprRawCol;\\n>\\n>\\n>\\nI am sorry. It is my fault. Now we have very similar node ColumnDef. This\\nnode is designed for usage in utility commands - and it is not designed for\\nusage inside a query. I had to decide between enhancing ColumnDef node or\\nintroduction new special node. Because there are more special attributes\\nand it is hard to serialize current ColumnDef, I decided to use new node.\\n\\n\\n>\\n> That's my first-pass commentary. I'll return to this once you've had a\\n> chance to take a look at these and tell me all the places I got it\\n> wrong ;)\\n>\\n>\\nThank for this\\n\\nRegard\\n\\nPavel\\n\\n>\\n>\\n> --\\n> Craig Ringer http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Training & Services\\n>\\n\\nHi2016-09-06 6:54 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:On 4 September 2016 at 16:06, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n> Hi\\n>\\n> minor update - using DefElem instead own private parser type\\n\\nI'm really glad that you're doing this and I'll take a look at it for this CF.\\n\\nIt's quite a big patch so I expect this will take a few rounds of\\nreview and updating.Thank you for review \\n\\n\\nPatch applies cleanly and builds cleanly on master both with and\\nwithout --with-xml .\\n\\nOverall, I think this needs to be revised with appropriate comments.\\nWhitespace/formatting needs fixing since it's all over the place.\\nDocumentation is insufficient (per notes below).I am not able to write documentation in English language :( - This function is pretty complex - so I hope so anybody with better language skills can help with this. It respects standard and it respects little bit different Oracle's behave too (different order of DEFAULT and PATH parts). \\n\\nRe identifier naming, some of this code uses XmlTable naming patterns,\\nsome uses TableExpr prefixes. Is that intended to indicate a bounary\\nbetween things re-usable for other structured data ingesting\\nfunctions? Do you expect a \\\"JSONEXPR\\\" or similar in future? That's\\nalluded to byThis structure should be reused by JSON_TABLE function. Now, it is little bit strange, because there is only XMLTABLE implementation - and I have to choose between a) using two different names now, b) renaming some part in future. And although XMLTABLE and JSON_TABLE functions are pretty similar - share 90% of data (input value, path, columns definitions), these functions has different syntax - so only middle level code should be shared. \\n\\n+/*----------\\n+ * TableExpr - used for XMLTABLE function\\n+ *\\n+ * This can be used for json_table, jsonb_table functions in future\\n+ *----------\\n+ */\\n+typedef struct TableExpr\\n+{\\n...\\n\\nIf so, should this really be two patches, one to add the table\\nexpression infrastructure and another to add XMLTABLE that uses it?\\nAlso, why in that case does so much of the TableExpr code call\\ndirectly into XmlTable code? It doesn't look very generic.Currently the common part is not too big - just the Node related part - I am not sure about necessity of two patches. I am agree, there is missing some TableExpBuilder, where can be better isolated the XML part. \\n\\nOverall I find identifier naming to be a bit inconsisent and think\\nit's necessary to make it clear that all the \\\"TableExpr\\\" stuff is for\\nXMLTABLE specifically, if that's the case, or make the delineation\\nclearer if not.\\n\\nI'd also like to see tests that exercise the ruleutils get_rule_expr\\nparts of the code for the various XMLTABLE variants.\\n\\nSimilarly, since this seems to add a new xpath parser, that needs\\ncomprehensive tests. Maybe re-purpose an existing xpath test data set?\\nsure \\n\\n\\n\\nMore detailed comments:\\n====\\n\\nDocs comments:\\n\\n The <function>xmltable</function> produces [a] table based on\\n[the] passed XML value.\\n\\nThe docs are pretty minimal and don't explain the various clauses of\\nXMLTABLE. What is \\\"BY REF\\\" ? Is PATH an xpath expression? If so, is\\nthere a good cross reference link available? The PASSING clause? etc.\\n\\nHow does XMLTABLE decide what to iterate over, and how to iterate over it?\\n\\nPresumably the FOR ORDINALITY clause makes a column emit a numeric counter.\\n\\nWhat standard, if any, does this conform to? Does it resemble\\nimplementations elsewhere? What limitations or unsupported features\\ndoes it have relative to those standards?\\n\\n\\n\\nexecEvalTableExpr seems to be defined twice, with a difference in\\ncase. This is probably not going to fly:\\n\\n\\n+static Datum\\n+execEvalTableExpr(TableExprState *tstate,\\n+ ExprContext *econtext,\\n+ bool *isNull, ExprDoneCond *isDone)\\n+{\\n\\n+static Datum\\n+ExecEvalTableExpr(TableExprState *tstate,\\n+ ExprContext *econtext,\\n+ bool *isNull, ExprDoneCond *isDone)\\n+{\\n\\n\\nIt looks like you've split the function into a \\\"guts\\\" and \\\"wrapper\\\"\\npart, with the error handling PG_TRY / PG_CATCH block in the wrapper.\\nThat seems reasonable for readability, but the naming isn't.I invite any idea how these functions should be named. \\n\\nA comment is needed to explain what ExecEvalTableExpr is / does. If\\nit's XMLTABLE specific (which it looks like based on the code), its\\nname should reflect that. This pattern is repeated elsewhere; e.g.\\nTableExprState is really the state for an XMLTABLE expression. But\\nPostgreSQL actually has TABLE statements, and in future we might want\\nto support table-expressions, so I don't think this naming is\\nappropriate. This is made worse by the lack of comments on things like\\nthe definition of TableExprState. Please use something that makes it\\nclear it's for XMLTABLE and add appropriate comments.I understand, so using TableExpr can be strange (for XMLTABLE function). But when we will have JSON_TABLE function, then it will have a sense.\\\"TableExprState\\\" is consistent with \\\"TableExpr\\\".Any idea how it should be changed? \\n\\nFormatting of variables, arguments, function signatures etc is\\nrandom/haphazard and doesn't follow project convention. It's neither\\naligned or unaligned in the normal way, I don't understand the random\\nspacing at all. Maybe you should try to run pgindent and then extract\\njust the changes related to your patch? Or run your IDE/editor's\\nindent function on your changes? Right now it's actually kind of hard\\nto read. Do you edit with tabstop set to 1 normally or something like\\nthat?\\n\\nThere's a general lack of comments throughout the added code.\\n\\nIn execEvalTableExpr, why are we looping over namespaces? What's that\\nfor? Comment would be nice.\\n\\nTypo: Path caclulation => Path calculation\\n\\nWhat does XmlTableSetRowPath() do? It seems to copy its argument.\\nNothing further is done with the row_path argument after it's called\\nby execEvalTableExpr, so what context is that memory in and do we have\\nto worry about it if it's large?\\n\\nexecEvalTableExpr says it's doing \\\"path calculation\\\". What it actually\\nappears to do is evaluate the path expressions, if provided, and\\notherwise use the column name as the implied path expression. (The\\ndocs should mention that).\\n\\nIt's wasn't immediately obvious to me what the branch around\\ntstate->for_ordinality_col is for and what the alternate path's\\npurpose is in terms of XMLTABLE's behaviour, until I read the parser\\ndefinition. That's largely because the behaviour of XMLTABLE is\\nunderspecified in the docs, since once you know ORDINALITY columns\\nexist it's pretty obvious what it's doing.\\n\\nSimilarly, for the alternate branch tstate->ncols , the\\nXmlTableGetRowValue call there is meant to do what exactly, and\\nwhy/under what conditions? Is it for situations where the field type\\nis a whole-row value? a composite type? (I'm deliberately not studying\\nthis too deeply, these are points I'd like to see commented so it can\\nbe understood to some reasonable degree at a skim-read).\\n\\n\\n /* result is one more columns every time */\\n\\\"one or more\\\"\\n\\n\\n\\n /* when typmod is not valid, refresh it */\\n if (te->typmod == -1)\\n\\n\\nIs this a cache? How is it valid or not valid and when? The comment\\n(thanks!) on TableExprGetTupleDesc says:\\n\\n/*\\n * When we skip transform stage (in view), then TableExpr's\\n * TupleDesc should not be valid. Refresh is necessary.\\n */\\n\\nbut I'm not really grasping what you're trying to explain here. What\\ntransform stage? What view? This could well be my ignorance of this\\npart of the code; if it should be understandable by a reader who is\\nappropriately familiar with the executor that's fine, but if it's\\nspecific to how XMLTABLE works some more explanation would be good.This is most difficult part of this patch, and I am not sure it it is fully correctly implemented. I use TupleDesc cache. The TupleDesc is created in parser/transform stage. When the XMLTABLE is used in some view, then the transformed parser tree is materialized - and when the view is used in query, then this tree is loaded and the parser/transform stage is \\\"skipped\\\". I'll check this code against implementation of ROW constructor and I'll try to do more comments there. \\n\\nGood that you've got all the required node copy/in/out funcs in place.\\n\\nPlease don't use the name \\\"used_dns\\\". Anyone reading that will read it\\nas \\\"domain name service\\\" and that's actually confusing with XML\\nbecause of XML schema lookups. Maybe used_defnamespace ? used\\ndef_ns?good idea \\n\\nI haven't looked closely at keyword/parser changes yet, but it doesn't\\nlook like you added any reserved keywords, which is good. It does add\\nunreserved keywords PATH and COLUMNS ; I'm not sure what policy for\\nunreserved keywords is or the significance of that.\\n\\nNew ereport() calls specify ERRCODEs, which is good.\\n\\nPostgreSQL already has XPATH support in the form of xmlexists(...)\\netc. Why is getXPathToken() etc needed? What re-use is possible here?\\nThere's no explanation in the patch header or comments. Should the new\\nxpath parser be re-used by the existing xpath stuff? Why can't we use\\nlibxml's facilities? etc. This at least needs explaining in the\\nsubmission, and some kind of hint as to why we have two different ways\\nto do it is needed in the code. If we do need a new XML parser, should\\nit be bundled in adt/xml.c along with a lot of user-facing\\nfunctionality, or a separate file?\\nlibxml2 and our XPATH function doesn't support default namespace ( http://plasmasturm.org/log/259/ ). This is pretty useful feature - so I implemented. This is the mayor issue of libxml2 library. Another difference between XPATH function and XMLTABLE function is using two phase searching and implicit prefix \\\"./\\\" and suffix (\\\"/text()\\\") in XMLTABLE. XMLTABLE using two XPATH expressions - for row data cutting and next for column data cutting (from row data). The our XPATH functions is pretty simple mapped to libxml2 XPATH API. But it is not possible with XMLTABLE function - due design of this function in standard (it is more user friendly and doesn't require exactly correct xpath expressions).I didn't find any API in libxml2 for a work with parsed xpath expressions - I need some info about the first and last token of xpath expression - it is base for decision about using prefix or suffix. This functionality (xpath expression parser) cannot be used for our XPATH function now - maybe default namespace in future. \\n\\nHow does XmlTableGetValue(...) and XmlTableGetRowValue(...) relate to\\nthis? It doesn't look like they're intended to be called directly by\\nthe user, and they're not documented (or commented).Probably I used wrong names. XMLTABLE function is running in two different modes - with explicitly defined columns (XmlTableGetValue is used), and without explicitly defined columns - so result is one XML column and only one one step searching is used (there are not column related xpath expressions) ( XmlTableGetRowValue is used). The function XmlTableGetValue is used for getting one column value, the function XmlTableGetRowValue is used for getting one value too, but in special case, when there are not any other value.\\n\\nI don't understand this at all:\\n\\n\\n\\n+/*\\n+ * There are different requests from XMLTABLE, JSON_TABLE functions\\n+ * on passed data than has CREATE TABLE command. It is reason for\\n+ * introduction special structure instead using ColumnDef.\\n+ */\\n+typedef struct TableExprRawCol\\n+{\\n+ NodeTag type;\\n+ char *colname;\\n+ TypeName *typeName;\\n+ bool for_ordinality;\\n+ bool is_not_null;\\n+ Node *path_expr;\\n+ Node *default_expr;\\n+ int location;\\n+} TableExprRawCol;\\n\\nI am sorry. It is my fault. Now we have very similar node ColumnDef. This node is designed for usage in utility commands - and it is not designed for usage inside a query. I had to decide between enhancing ColumnDef node or introduction new special node. Because there are more special attributes and it is hard to serialize current ColumnDef, I decided to use new node. \\n\\nThat's my first-pass commentary. I'll return to this once you've had a\\nchance to take a look at these and tell me all the places I got it\\nwrong ;)\\nThank for thisRegard Pavel\\n\\n\\n--\\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Tue, 6 Sep 2016 22:13:15 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"> libxml2 and our XPATH function doesn't support default namespace (\\n> http://plasmasturm.org/log/259/ ). This is pretty useful feature - so I\\n> implemented. This is the mayor issue of libxml2 library. Another difference\\n> between XPATH function and XMLTABLE function is using two phase searching\\n> and implicit prefix \\\"./\\\" and suffix (\\\"/text()\\\") in XMLTABLE. XMLTABLE using\\n> two XPATH expressions - for row data cutting and next for column data\\n> cutting (from row data). The our XPATH functions is pretty simple mapped to\\n> libxml2 XPATH API. But it is not possible with XMLTABLE function - due\\n> design of this function in standard (it is more user friendly and doesn't\\n> require exactly correct xpath expressions).\\n>\\n\\nlibxm2 doesn't support xpath 2.0 where default namespace was introduced.\\n\\n\\nlibxml2 and our XPATH function doesn't support default namespace ( http://plasmasturm.org/log/259/ ). This is pretty useful feature - so I implemented. This is the mayor issue of libxml2 library. Another difference between XPATH function and XMLTABLE function is using two phase searching and implicit prefix \\\"./\\\" and suffix (\\\"/text()\\\") in XMLTABLE. XMLTABLE using two XPATH expressions - for row data cutting and next for column data cutting (from row data). The our XPATH functions is pretty simple mapped to libxml2 XPATH API. But it is not possible with XMLTABLE function - due design of this function in standard (it is more user friendly and doesn't require exactly correct xpath expressions).libxm2 doesn't support xpath 2.0 where default namespace was introduced.\",\n \"msg_date\": \"Tue, 6 Sep 2016 22:21:49 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On 7 September 2016 at 04:13, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n\\n>> Overall, I think this needs to be revised with appropriate comments.\\n>> Whitespace/formatting needs fixing since it's all over the place.\\n>> Documentation is insufficient (per notes below).\\n>\\n>\\n> I am not able to write documentation in English language :( - This function\\n> is pretty complex - so I hope so anybody with better language skills can\\n> help with this. It respects standard and it respects little bit different\\n> Oracle's behave too (different order of DEFAULT and PATH parts).\\n\\nOK, no problem. It can't be committed without more comprehensive docs\\nthough, especially for new and nontrivial functionality.\\n\\nIs there some reference material you can point to so someone else can\\nhelp with docs? And can you describe what differences there are\\nbetween your implementation and the reference?\\n\\nAlternately, if you document it in Czech, do you know of anyone who\\ncould assist in translating to English for the main documentation?\\n\\n>> Re identifier naming, some of this code uses XmlTable naming patterns,\\n>> some uses TableExpr prefixes. Is that intended to indicate a bounary\\n>> between things re-usable for other structured data ingesting\\n>> functions? Do you expect a \\\"JSONEXPR\\\" or similar in future? That's\\n>> alluded to by\\n>\\n>\\n> This structure should be reused by JSON_TABLE function. Now, it is little\\n> bit strange, because there is only XMLTABLE implementation - and I have to\\n> choose between a) using two different names now, b) renaming some part in\\n> future.\\n\\nOK. Are you planning on writing this JSON_TABLE or are you leaving\\nroom for future growth? Either way is fine, just curious.\\n\\n> And although XMLTABLE and JSON_TABLE functions are pretty similar - share\\n> 90% of data (input value, path, columns definitions), these functions has\\n> different syntax - so only middle level code should be shared.\\n\\nThat makes sense.\\n\\nI think it would be best if you separated out the TableExpr\\ninfrastructure from the XMLTABLE implementation though, so we can\\nreview the first level infrastrcture separately and make this a\\n2-patch series. Most importantly, doing it that way will help you find\\nplaces where TableExpr code calls directly into XMLTABLE code. If\\nTableExpr is supposed to be reusable for json etc, it probably\\nshouldn't be calling XmlTable stuff directly.\\n\\nThat also means somewhat smaller simpler patches, which probably isn't bad.\\n\\nI don't necessarily think this needs to be fully pluggable with\\ncallbacks etc. It doesn't sound like you expect this to be used by\\nextensions or to have a lot of users, right? So it probably just needs\\nclearer separation of the infrastructure layer from the xmltable\\nlayer. I think splitting the patch will make that easier to see and\\nmake it easier to find problems.\\n\\nMy biggest complaint at the moment is that execEvalTableExpr calls\\ninitXmlTableContext(...) directly, is aware of XML namespaces\\ndirectly, calls XmlTableSetRowPath() directly, calls\\nXmlTableFetchRow() directly, etc. It is in no way generic/reusable for\\nsome later JSONTABLE feature. That needs to be fixed by:\\n\\n* Renaming it so it's clearly only for XMLTABLE; or\\n* Abstracting the init context, set row path, fetch row etc operations\\nso json ones can be plugged in later\\n\\n\\n\\n> Currently the common part is not too big - just the Node related part - I am\\n> not sure about necessity of two patches.\\n\\nThe problem is that the common part is all mixed in with the\\nXMLTABLE-specific part, so it's not at all clear it can be common with\\nsomething else.\\n\\n> I am agree, there is missing some\\n> TableExpBuilder, where can be better isolated the XML part.\\n\\nYeah, that's sort of what I'm getting at.\\n\\n\\n>> execEvalTableExpr seems to be defined twice, with a difference in\\n>> case. This is probably not going to fly:\\n>>\\n>>\\n>> +static Datum\\n>> +execEvalTableExpr(TableExprState *tstate,\\n>> + ExprContext *econtext,\\n>> + bool *isNull, ExprDoneCond *isDone)\\n>> +{\\n>>\\n>> +static Datum\\n>> +ExecEvalTableExpr(TableExprState *tstate,\\n>> + ExprContext *econtext,\\n>> + bool *isNull, ExprDoneCond *isDone)\\n>> +{\\n>>\\n>>\\n>> It looks like you've split the function into a \\\"guts\\\" and \\\"wrapper\\\"\\n>> part, with the error handling PG_TRY / PG_CATCH block in the wrapper.\\n>> That seems reasonable for readability, but the naming isn't.\\n>\\n>\\n> I invite any idea how these functions should be named.\\n\\nDefinitely not how they are ;) . They really can't differ in a single\\ncharacter's case.\\n\\nI'm not sure if PostgreSQL has any formal convention for this. Some\\nplaces use _impl e.g. pg_read_barrier_impl() but that's in the\\ncontext of an interface-vs-implementation separation, which isn't the\\ncase here.\\n\\nSome places use _internal, like AlterObjectRename_internal(...), but\\nthat's where there's an associated public/external part, which isn't\\nthe case here.\\n\\nSome places use _guts e.g. pg_logical_slot_get_changes_guts(...),\\nlargely where there's common use by several callers.\\n\\nThis is a fairly arbitrary function split for readability/length. Is\\nit actually useful to split this function up at all?\\n\\nAnyone else have an opinion?\\n\\n>> A comment is needed to explain what ExecEvalTableExpr is / does. If\\n>> it's XMLTABLE specific (which it looks like based on the code), its\\n>> name should reflect that. This pattern is repeated elsewhere; e.g.\\n>> TableExprState is really the state for an XMLTABLE expression. But\\n>> PostgreSQL actually has TABLE statements, and in future we might want\\n>> to support table-expressions, so I don't think this naming is\\n>> appropriate. This is made worse by the lack of comments on things like\\n>> the definition of TableExprState. Please use something that makes it\\n>> clear it's for XMLTABLE and add appropriate comments.\\n>\\n>\\n> I understand, so using TableExpr can be strange (for XMLTABLE function). But\\n> when we will have JSON_TABLE function, then it will have a sense.\\n\\nIt's pretty hard to review that as shared infrastructure when it's\\nstill tangled up in xmltable specifics, though.\\n\\n> \\\"TableExprState\\\" is consistent with \\\"TableExpr\\\".\\n>\\n> Any idea how it should be changed?\\n\\nI think if you want it to be shareable infrasructure, you need to\\nwrite it so it can be used as shared infrastructure. Not just name it\\nthat way but then make it XMLTABLE specific in actual functionality.\\n\\n\\n>> /* when typmod is not valid, refresh it */\\n>> if (te->typmod == -1)\\n>>\\n>>\\n>> Is this a cache? How is it valid or not valid and when? The comment\\n>> (thanks!) on TableExprGetTupleDesc says:\\n>>\\n>> /*\\n>> * When we skip transform stage (in view), then TableExpr's\\n>> * TupleDesc should not be valid. Refresh is necessary.\\n>> */\\n>>\\n>> but I'm not really grasping what you're trying to explain here. What\\n>> transform stage? What view? This could well be my ignorance of this\\n>> part of the code; if it should be understandable by a reader who is\\n>> appropriately familiar with the executor that's fine, but if it's\\n>> specific to how XMLTABLE works some more explanation would be good.\\n>\\n>\\n> This is most difficult part of this patch, and I am not sure it it is fully\\n> correctly implemented. I use TupleDesc cache. The TupleDesc is created in\\n> parser/transform stage. When the XMLTABLE is used in some view, then the\\n> transformed parser tree is materialized - and when the view is used in\\n> query, then this tree is loaded and the parser/transform stage is \\\"skipped\\\".\\n> I'll check this code against implementation of ROW constructor and I'll try\\n> to do more comments there.\\n\\nThanks. It would be good to highlight when this does and does not\\nhappen and why. Why is it necessary at all?\\n\\nWhat happens if XMLTABLE is used in a query directly, not part of a\\nview? if XMLTABLE is used in a view? If XMLTABLE is used in a prepared\\nstatement / plpgsql statement / etc? What about a CTE term?\\n\\nNot necessarily list all these cases one by one, just explain what\\nhappens, when and why. Especially if it's complex, so other readers\\ncan understand it and don't have to study it in detail to understand\\nwhat is going on. It does not need to be good public-facing\\ndocumentation, and details of wording, grammar etc can be fixed up\\nlater, it's the ideas that matter.\\n\\nIs this similar to other logic elsewhere? If so, reference that other\\nlogic so readers know where to look. That way if they're\\nchanging/bugfixing/etc one place they know there's another place that\\nmight need changing.\\n\\nI don't know this area of the code well enough to give a solid review\\nof the actual functionality, and I don't yet understand what it's\\ntrying to do so it's hard to review it by studying what it actually\\ndoes vs what it claims to do. Maybe Peter E can help, he said he was\\nthinking of looking at this patch too. But more information on what\\nit's trying to do would be a big help.\\n\\n>> PostgreSQL already has XPATH support in the form of xmlexists(...)\\n>> etc. Why is getXPathToken() etc needed? What re-use is possible here?\\n>> There's no explanation in the patch header or comments. Should the new\\n>> xpath parser be re-used by the existing xpath stuff? Why can't we use\\n>> libxml's facilities? etc. This at least needs explaining in the\\n>> submission, and some kind of hint as to why we have two different ways\\n>> to do it is needed in the code. If we do need a new XML parser, should\\n>> it be bundled in adt/xml.c along with a lot of user-facing\\n>> functionality, or a separate file?\\n>\\n> libxml2 and our XPATH function doesn't support default namespace (\\n> http://plasmasturm.org/log/259/ ). This is pretty useful feature - so I\\n> implemented.\\n\\nOK, that makes sense.\\n\\nFor the purpose of getting this patch in, is it a _necessary_ feature?\\nCan XMLTABLE be usefully implemented without it, and if so, can it be\\nadded in a subsequent patch? It would be nice to simplify this by\\nusing existing libxml2 functionality in the first version rather than\\nadding a whole new xpath as well!\\n\\n> This is the mayor issue of libxml2 library. Another difference\\n> between XPATH function and XMLTABLE function is using two phase searching\\n> and implicit prefix \\\"./\\\" and suffix (\\\"/text()\\\") in XMLTABLE. XMLTABLE using\\n> two XPATH expressions - for row data cutting and next for column data\\n> cutting (from row data). The our XPATH functions is pretty simple mapped to\\n> libxml2 XPATH API. But it is not possible with XMLTABLE function - due\\n> design of this function in standard (it is more user friendly and doesn't\\n> require exactly correct xpath expressions).\\n\\nSo you can't use existing libxml2 xpath support to implement XMLTABLE,\\neven without default namespaces?\\n\\n> I didn't find any API in libxml2 for a work with parsed xpath expressions -\\n> I need some info about the first and last token of xpath expression - it is\\n> base for decision about using prefix or suffix.\\n>\\n> This functionality (xpath expression parser) cannot be used for our XPATH\\n> function now - maybe default namespace in future.\\n\\nSo we'll have two different XPATH implementations for different\\nplaces, with different limitations, different possible bugs, etc?\\n\\nWhat would be needed to make the new XPATH work for our built-in xpath\\nfunctions too?\\n\\n>> How does XmlTableGetValue(...) and XmlTableGetRowValue(...) relate to\\n>> this? It doesn't look like they're intended to be called directly by\\n>> the user, and they're not documented (or commented).\\n>\\n>\\n> Probably I used wrong names. XMLTABLE function is running in two different\\n> modes - with explicitly defined columns (XmlTableGetValue is used), and\\n> without explicitly defined columns - so result is one XML column and only\\n> one one step searching is used (there are not column related xpath\\n> expressions) ( XmlTableGetRowValue is used). The function XmlTableGetValue\\n> is used for getting one column value, the function XmlTableGetRowValue is\\n> used for getting one value too, but in special case, when there are not any\\n> other value.\\n\\nSo both are internal implementation of the parser-level XMLTABLE(...)\\nconstruct and are not intended to be called directly by users - right?\\n\\nComments please! A short comment on the function saying this would be\\na big help.\\n\\nRegarding naming, do we already have a convention for functions that\\nare internal implementation of something the user \\\"spells\\\"\\ndifferently? Where it's transformed by the parser? I couldn't find\\none. I don't much care about the names so long as there are comments\\nexplaining what calls the functions and what the user-facing interface\\nthat matches the function is.\\n\\nIs it safe for users to call these directly? What happens if they do\\nso incorrectly?\\n\\nWhy are they not in pg_proc.h? Do they need to be?\\n\\n>> +/*\\n>> + * There are different requests from XMLTABLE, JSON_TABLE functions\\n>> + * on passed data than has CREATE TABLE command. It is reason for\\n>> + * introduction special structure instead using ColumnDef.\\n>> + */\\n>> +typedef struct TableExprRawCol\\n>> +{\\n>> + NodeTag type;\\n>> + char *colname;\\n>> + TypeName *typeName;\\n>> + bool for_ordinality;\\n>> + bool is_not_null;\\n>> + Node *path_expr;\\n>> + Node *default_expr;\\n>> + int location;\\n>> +} TableExprRawCol;\\n>\\n> I am sorry. It is my fault. Now we have very similar node ColumnDef. This\\n> node is designed for usage in utility commands - and it is not designed for\\n> usage inside a query.\\n\\nMakes sense.\\n\\n> I had to decide between enhancing ColumnDef node or\\n> introduction new special node. Because there are more special attributes and\\n> it is hard to serialize current ColumnDef, I decided to use new node.\\n\\nSeems reasonable. The summary is \\\"this is the parse node for a column\\nof an XMLTABLE expression\\\".\\n\\nSuggested comment:\\n\\n/*\\n * This is the parsenode for a column definition in a table-expression\\nlike XMLTABLE.\\n *\\n * We can't re-use ColumnDef here; the utility command column\\ndefinition has all the\\n * wrong attributes for use in table-expressions and just doesn't make\\nsense here.\\n */\\ntypedef struct TableExprColumn\\n{\\n...\\n};\\n\\n?\\n\\nWhy \\\"RawCol\\\" ? What does it become when it's not \\\"raw\\\" anymore? Is\\nthat a reference to ColumnDef's raw_default and cooked_default for\\nuntransformed vs transformed parse-trees?\\n\\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, 7 Sep 2016 11:03:21 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-09-07 5:03 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:\\n\\n> On 7 September 2016 at 04:13, Pavel Stehule <pavel.stehule@gmail.com>\\n> wrote:\\n>\\n> >> Overall, I think this needs to be revised with appropriate comments.\\n> >> Whitespace/formatting needs fixing since it's all over the place.\\n> >> Documentation is insufficient (per notes below).\\n> >\\n> >\\n> > I am not able to write documentation in English language :( - This\\n> function\\n> > is pretty complex - so I hope so anybody with better language skills can\\n> > help with this. It respects standard and it respects little bit different\\n> > Oracle's behave too (different order of DEFAULT and PATH parts).\\n>\\n> OK, no problem. It can't be committed without more comprehensive docs\\n> though, especially for new and nontrivial functionality.\\n>\\n> Is there some reference material you can point to so someone else can\\n> help with docs? And can you describe what differences there are\\n> between your implementation and the reference?\\n>\\n> Alternately, if you document it in Czech, do you know of anyone who\\n> could assist in translating to English for the main documentation?\\n>\\n> >> Re identifier naming, some of this code uses XmlTable naming patterns,\\n> >> some uses TableExpr prefixes. Is that intended to indicate a bounary\\n> >> between things re-usable for other structured data ingesting\\n> >> functions? Do you expect a \\\"JSONEXPR\\\" or similar in future? That's\\n> >> alluded to by\\n> >\\n> >\\n> > This structure should be reused by JSON_TABLE function. Now, it is little\\n> > bit strange, because there is only XMLTABLE implementation - and I have\\n> to\\n> > choose between a) using two different names now, b) renaming some part in\\n> > future.\\n>\\n> OK. Are you planning on writing this JSON_TABLE or are you leaving\\n> room for future growth? Either way is fine, just curious.\\n>\\n> > And although XMLTABLE and JSON_TABLE functions are pretty similar - share\\n> > 90% of data (input value, path, columns definitions), these functions has\\n> > different syntax - so only middle level code should be shared.\\n>\\n> That makes sense.\\n>\\n> I think it would be best if you separated out the TableExpr\\n> infrastructure from the XMLTABLE implementation though, so we can\\n> review the first level infrastrcture separately and make this a\\n> 2-patch series. Most importantly, doing it that way will help you find\\n> places where TableExpr code calls directly into XMLTABLE code. If\\n> TableExpr is supposed to be reusable for json etc, it probably\\n> shouldn't be calling XmlTable stuff directly.\\n>\\n> That also means somewhat smaller simpler patches, which probably isn't bad.\\n>\\n> I don't necessarily think this needs to be fully pluggable with\\n> callbacks etc. It doesn't sound like you expect this to be used by\\n> extensions or to have a lot of users, right? So it probably just needs\\n> clearer separation of the infrastructure layer from the xmltable\\n> layer. I think splitting the patch will make that easier to see and\\n> make it easier to find problems.\\n>\\n> My biggest complaint at the moment is that execEvalTableExpr calls\\n> initXmlTableContext(...) directly, is aware of XML namespaces\\n> directly, calls XmlTableSetRowPath() directly, calls\\n> XmlTableFetchRow() directly, etc. It is in no way generic/reusable for\\n> some later JSONTABLE feature. That needs to be fixed by:\\n>\\n> * Renaming it so it's clearly only for XMLTABLE; or\\n> * Abstracting the init context, set row path, fetch row etc operations\\n> so json ones can be plugged in later\\n>\\n>\\n>\\n> > Currently the common part is not too big - just the Node related part -\\n> I am\\n> > not sure about necessity of two patches.\\n>\\n> The problem is that the common part is all mixed in with the\\n> XMLTABLE-specific part, so it's not at all clear it can be common with\\n> something else.\\n>\\n> > I am agree, there is missing some\\n> > TableExpBuilder, where can be better isolated the XML part.\\n>\\n> Yeah, that's sort of what I'm getting at.\\n>\\n>\\n> >> execEvalTableExpr seems to be defined twice, with a difference in\\n> >> case. This is probably not going to fly:\\n> >>\\n> >>\\n> >> +static Datum\\n> >> +execEvalTableExpr(TableExprState *tstate,\\n> >> + ExprContext *econtext,\\n> >> + bool *isNull, ExprDoneCond *isDone)\\n> >> +{\\n> >>\\n> >> +static Datum\\n> >> +ExecEvalTableExpr(TableExprState *tstate,\\n> >> + ExprContext *econtext,\\n> >> + bool *isNull, ExprDoneCond *isDone)\\n> >> +{\\n> >>\\n> >>\\n> >> It looks like you've split the function into a \\\"guts\\\" and \\\"wrapper\\\"\\n> >> part, with the error handling PG_TRY / PG_CATCH block in the wrapper.\\n> >> That seems reasonable for readability, but the naming isn't.\\n> >\\n> >\\n> > I invite any idea how these functions should be named.\\n>\\n> Definitely not how they are ;) . They really can't differ in a single\\n> character's case.\\n>\\n> I'm not sure if PostgreSQL has any formal convention for this. Some\\n> places use _impl e.g. pg_read_barrier_impl() but that's in the\\n> context of an interface-vs-implementation separation, which isn't the\\n> case here.\\n>\\n> Some places use _internal, like AlterObjectRename_internal(...), but\\n> that's where there's an associated public/external part, which isn't\\n> the case here.\\n>\\n> Some places use _guts e.g. pg_logical_slot_get_changes_guts(...),\\n> largely where there's common use by several callers.\\n>\\n> This is a fairly arbitrary function split for readability/length. Is\\n> it actually useful to split this function up at all?\\n>\\n> Anyone else have an opinion?\\n>\\n> >> A comment is needed to explain what ExecEvalTableExpr is / does. If\\n> >> it's XMLTABLE specific (which it looks like based on the code), its\\n> >> name should reflect that. This pattern is repeated elsewhere; e.g.\\n> >> TableExprState is really the state for an XMLTABLE expression. But\\n> >> PostgreSQL actually has TABLE statements, and in future we might want\\n> >> to support table-expressions, so I don't think this naming is\\n> >> appropriate. This is made worse by the lack of comments on things like\\n> >> the definition of TableExprState. Please use something that makes it\\n> >> clear it's for XMLTABLE and add appropriate comments.\\n> >\\n> >\\n> > I understand, so using TableExpr can be strange (for XMLTABLE function).\\n> But\\n> > when we will have JSON_TABLE function, then it will have a sense.\\n>\\n> It's pretty hard to review that as shared infrastructure when it's\\n> still tangled up in xmltable specifics, though.\\n>\\n> > \\\"TableExprState\\\" is consistent with \\\"TableExpr\\\".\\n> >\\n> > Any idea how it should be changed?\\n>\\n> I think if you want it to be shareable infrasructure, you need to\\n> write it so it can be used as shared infrastructure. Not just name it\\n> that way but then make it XMLTABLE specific in actual functionality.\\n>\\n>\\n> >> /* when typmod is not valid, refresh it */\\n> >> if (te->typmod == -1)\\n> >>\\n> >>\\n> >> Is this a cache? How is it valid or not valid and when? The comment\\n> >> (thanks!) on TableExprGetTupleDesc says:\\n> >>\\n> >> /*\\n> >> * When we skip transform stage (in view), then TableExpr's\\n> >> * TupleDesc should not be valid. Refresh is necessary.\\n> >> */\\n> >>\\n> >> but I'm not really grasping what you're trying to explain here. What\\n> >> transform stage? What view? This could well be my ignorance of this\\n> >> part of the code; if it should be understandable by a reader who is\\n> >> appropriately familiar with the executor that's fine, but if it's\\n> >> specific to how XMLTABLE works some more explanation would be good.\\n> >\\n> >\\n> > This is most difficult part of this patch, and I am not sure it it is\\n> fully\\n> > correctly implemented. I use TupleDesc cache. The TupleDesc is created in\\n> > parser/transform stage. When the XMLTABLE is used in some view, then the\\n> > transformed parser tree is materialized - and when the view is used in\\n> > query, then this tree is loaded and the parser/transform stage is\\n> \\\"skipped\\\".\\n> > I'll check this code against implementation of ROW constructor and I'll\\n> try\\n> > to do more comments there.\\n>\\n> Thanks. It would be good to highlight when this does and does not\\n> happen and why. Why is it necessary at all?\\n>\\n> What happens if XMLTABLE is used in a query directly, not part of a\\n> view? if XMLTABLE is used in a view? If XMLTABLE is used in a prepared\\n> statement / plpgsql statement / etc? What about a CTE term?\\n>\\n> Not necessarily list all these cases one by one, just explain what\\n> happens, when and why. Especially if it's complex, so other readers\\n> can understand it and don't have to study it in detail to understand\\n> what is going on. It does not need to be good public-facing\\n> documentation, and details of wording, grammar etc can be fixed up\\n> later, it's the ideas that matter.\\n>\\n> Is this similar to other logic elsewhere? If so, reference that other\\n> logic so readers know where to look. That way if they're\\n> changing/bugfixing/etc one place they know there's another place that\\n> might need changing.\\n>\\n> I don't know this area of the code well enough to give a solid review\\n> of the actual functionality, and I don't yet understand what it's\\n> trying to do so it's hard to review it by studying what it actually\\n> does vs what it claims to do. Maybe Peter E can help, he said he was\\n> thinking of looking at this patch too. But more information on what\\n> it's trying to do would be a big help.\\n>\\n> >> PostgreSQL already has XPATH support in the form of xmlexists(...)\\n> >> etc. Why is getXPathToken() etc needed? What re-use is possible here?\\n> >> There's no explanation in the patch header or comments. Should the new\\n> >> xpath parser be re-used by the existing xpath stuff? Why can't we use\\n> >> libxml's facilities? etc. This at least needs explaining in the\\n> >> submission, and some kind of hint as to why we have two different ways\\n> >> to do it is needed in the code. If we do need a new XML parser, should\\n> >> it be bundled in adt/xml.c along with a lot of user-facing\\n> >> functionality, or a separate file?\\n> >\\n> > libxml2 and our XPATH function doesn't support default namespace (\\n> > http://plasmasturm.org/log/259/ ). This is pretty useful feature - so I\\n> > implemented.\\n>\\n> OK, that makes sense.\\n>\\n> For the purpose of getting this patch in, is it a _necessary_ feature?\\n> Can XMLTABLE be usefully implemented without it, and if so, can it be\\n> added in a subsequent patch? It would be nice to simplify this by\\n> using existing libxml2 functionality in the first version rather than\\n> adding a whole new xpath as well!\\n>\\n\\nThis is not a xpath implementation - it is preprocessing of xpath\\nexpression. without it, a users have to set explicitly PATH clause with\\nexplicit prefix \\\"./\\\" and explicit suffix \\\"/text()\\\". The usability will be\\nsignificantly lower, and what is worst - the examples from internet should\\nnot work. Although is is lot of lines, this code is necessary.\\n\\n\\n>\\n> > This is the mayor issue of libxml2 library. Another difference\\n> > between XPATH function and XMLTABLE function is using two phase searching\\n> > and implicit prefix \\\"./\\\" and suffix (\\\"/text()\\\") in XMLTABLE. XMLTABLE\\n> using\\n> > two XPATH expressions - for row data cutting and next for column data\\n> > cutting (from row data). The our XPATH functions is pretty simple mapped\\n> to\\n> > libxml2 XPATH API. But it is not possible with XMLTABLE function - due\\n> > design of this function in standard (it is more user friendly and doesn't\\n> > require exactly correct xpath expressions).\\n>\\n> So you can't use existing libxml2 xpath support to implement XMLTABLE,\\n> even without default namespaces?\\n>\\n> > I didn't find any API in libxml2 for a work with parsed xpath\\n> expressions -\\n> > I need some info about the first and last token of xpath expression - it\\n> is\\n> > base for decision about using prefix or suffix.\\n> >\\n> > This functionality (xpath expression parser) cannot be used for our XPATH\\n> > function now - maybe default namespace in future.\\n>\\n> So we'll have two different XPATH implementations for different\\n> places, with different limitations, different possible bugs, etc?\\n>\\n\\nIt is just preprocessing. The evaluation of xpath expression is part of\\nlibxml2 and it is shared.\\n\\nOur XPATH function is not short, but the reason is reading namespaces data\\nfrom 2D array. The evaluation of xpath expression is on few lines.\\n\\n\\n>\\n> What would be needed to make the new XPATH work for our built-in xpath\\n> functions too?\\n>\\n\\n> >> How does XmlTableGetValue(...) and XmlTableGetRowValue(...) relate to\\n> >> this? It doesn't look like they're intended to be called directly by\\n> >> the user, and they're not documented (or commented).\\n> >\\n> >\\n> > Probably I used wrong names. XMLTABLE function is running in two\\n> different\\n> > modes - with explicitly defined columns (XmlTableGetValue is used), and\\n> > without explicitly defined columns - so result is one XML column and only\\n> > one one step searching is used (there are not column related xpath\\n> > expressions) ( XmlTableGetRowValue is used). The function\\n> XmlTableGetValue\\n> > is used for getting one column value, the function XmlTableGetRowValue is\\n> > used for getting one value too, but in special case, when there are not\\n> any\\n> > other value.\\n>\\n> So both are internal implementation of the parser-level XMLTABLE(...)\\n> construct and are not intended to be called directly by users - right?\\n>\\n\\nNo - it is called from executor - and it should not be called differently.\\nI have to do better separation from executor, and these functions will be\\nprivate.\\n\\n\\n>\\n> Comments please! A short comment on the function saying this would be\\n> a big help.\\n>\\n> Regarding naming, do we already have a convention for functions that\\n> are internal implementation of something the user \\\"spells\\\"\\n> differently? Where it's transformed by the parser? I couldn't find\\n> one. I don't much care about the names so long as there are comments\\n> explaining what calls the functions and what the user-facing interface\\n> that matches the function is.\\n>\\n> Is it safe for users to call these directly? What happens if they do\\n> so incorrectly?\\n>\\n> Why are they not in pg_proc.h? Do they need to be?\\n>\\n> >> +/*\\n> >> + * There are different requests from XMLTABLE, JSON_TABLE functions\\n> >> + * on passed data than has CREATE TABLE command. It is reason for\\n> >> + * introduction special structure instead using ColumnDef.\\n> >> + */\\n> >> +typedef struct TableExprRawCol\\n> >> +{\\n> >> + NodeTag type;\\n> >> + char *colname;\\n> >> + TypeName *typeName;\\n> >> + bool for_ordinality;\\n> >> + bool is_not_null;\\n> >> + Node *path_expr;\\n> >> + Node *default_expr;\\n> >> + int location;\\n> >> +} TableExprRawCol;\\n> >\\n> > I am sorry. It is my fault. Now we have very similar node ColumnDef. This\\n> > node is designed for usage in utility commands - and it is not designed\\n> for\\n> > usage inside a query.\\n>\\n> Makes sense.\\n>\\n> > I had to decide between enhancing ColumnDef node or\\n> > introduction new special node. Because there are more special attributes\\n> and\\n> > it is hard to serialize current ColumnDef, I decided to use new node.\\n>\\n> Seems reasonable. The summary is \\\"this is the parse node for a column\\n> of an XMLTABLE expression\\\".\\n>\\n> Suggested comment:\\n>\\n> /*\\n> * This is the parsenode for a column definition in a table-expression\\n> like XMLTABLE.\\n> *\\n> * We can't re-use ColumnDef here; the utility command column\\n> definition has all the\\n> * wrong attributes for use in table-expressions and just doesn't make\\n> sense here.\\n> */\\n> typedef struct TableExprColumn\\n> {\\n> ...\\n> };\\n>\\n> ?\\n>\\n> Why \\\"RawCol\\\" ? What does it become when it's not \\\"raw\\\" anymore? Is\\n> that a reference to ColumnDef's raw_default and cooked_default for\\n> untransformed vs transformed parse-trees?\\n>\\n\\nTableExprColumn is better\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n>\\n>\\n>\\n> --\\n> Craig Ringer http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Training & Services\\n>\\n\\n2016-09-07 5:03 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:On 7 September 2016 at 04:13, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n\\n>> Overall, I think this needs to be revised with appropriate comments.\\n>> Whitespace/formatting needs fixing since it's all over the place.\\n>> Documentation is insufficient (per notes below).\\n>\\n>\\n> I am not able to write documentation in English language :( - This function\\n> is pretty complex - so I hope so anybody with better language skills can\\n> help with this. It respects standard and it respects little bit different\\n> Oracle's behave too (different order of DEFAULT and PATH parts).\\n\\nOK, no problem. It can't be committed without more comprehensive docs\\nthough, especially for new and nontrivial functionality.\\n\\nIs there some reference material you can point to so someone else can\\nhelp with docs? And can you describe what differences there are\\nbetween your implementation and the reference?\\n\\nAlternately, if you document it in Czech, do you know of anyone who\\ncould assist in translating to English for the main documentation?\\n\\n>> Re identifier naming, some of this code uses XmlTable naming patterns,\\n>> some uses TableExpr prefixes. Is that intended to indicate a bounary\\n>> between things re-usable for other structured data ingesting\\n>> functions? Do you expect a \\\"JSONEXPR\\\" or similar in future? That's\\n>> alluded to by\\n>\\n>\\n> This structure should be reused by JSON_TABLE function. Now, it is little\\n> bit strange, because there is only XMLTABLE implementation - and I have to\\n> choose between a) using two different names now, b) renaming some part in\\n> future.\\n\\nOK. Are you planning on writing this JSON_TABLE or are you leaving\\nroom for future growth? Either way is fine, just curious.\\n\\n> And although XMLTABLE and JSON_TABLE functions are pretty similar - share\\n> 90% of data (input value, path, columns definitions), these functions has\\n> different syntax - so only middle level code should be shared.\\n\\nThat makes sense.\\n\\nI think it would be best if you separated out the TableExpr\\ninfrastructure from the XMLTABLE implementation though, so we can\\nreview the first level infrastrcture separately and make this a\\n2-patch series. Most importantly, doing it that way will help you find\\nplaces where TableExpr code calls directly into XMLTABLE code. If\\nTableExpr is supposed to be reusable for json etc, it probably\\nshouldn't be calling XmlTable stuff directly.\\n\\nThat also means somewhat smaller simpler patches, which probably isn't bad.\\n\\nI don't necessarily think this needs to be fully pluggable with\\ncallbacks etc. It doesn't sound like you expect this to be used by\\nextensions or to have a lot of users, right? So it probably just needs\\nclearer separation of the infrastructure layer from the xmltable\\nlayer. I think splitting the patch will make that easier to see and\\nmake it easier to find problems.\\n\\nMy biggest complaint at the moment is that execEvalTableExpr calls\\ninitXmlTableContext(...) directly, is aware of XML namespaces\\ndirectly, calls XmlTableSetRowPath() directly, calls\\nXmlTableFetchRow() directly, etc. It is in no way generic/reusable for\\nsome later JSONTABLE feature. That needs to be fixed by:\\n\\n* Renaming it so it's clearly only for XMLTABLE; or\\n* Abstracting the init context, set row path, fetch row etc operations\\nso json ones can be plugged in later\\n\\n\\n\\n> Currently the common part is not too big - just the Node related part - I am\\n> not sure about necessity of two patches.\\n\\nThe problem is that the common part is all mixed in with the\\nXMLTABLE-specific part, so it's not at all clear it can be common with\\nsomething else.\\n\\n> I am agree, there is missing some\\n> TableExpBuilder, where can be better isolated the XML part.\\n\\nYeah, that's sort of what I'm getting at.\\n\\n\\n>> execEvalTableExpr seems to be defined twice, with a difference in\\n>> case. This is probably not going to fly:\\n>>\\n>>\\n>> +static Datum\\n>> +execEvalTableExpr(TableExprState *tstate,\\n>> + ExprContext *econtext,\\n>> + bool *isNull, ExprDoneCond *isDone)\\n>> +{\\n>>\\n>> +static Datum\\n>> +ExecEvalTableExpr(TableExprState *tstate,\\n>> + ExprContext *econtext,\\n>> + bool *isNull, ExprDoneCond *isDone)\\n>> +{\\n>>\\n>>\\n>> It looks like you've split the function into a \\\"guts\\\" and \\\"wrapper\\\"\\n>> part, with the error handling PG_TRY / PG_CATCH block in the wrapper.\\n>> That seems reasonable for readability, but the naming isn't.\\n>\\n>\\n> I invite any idea how these functions should be named.\\n\\nDefinitely not how they are ;) . They really can't differ in a single\\ncharacter's case.\\n\\nI'm not sure if PostgreSQL has any formal convention for this. Some\\nplaces use _impl e.g. pg_read_barrier_impl() but that's in the\\ncontext of an interface-vs-implementation separation, which isn't the\\ncase here.\\n\\nSome places use _internal, like AlterObjectRename_internal(...), but\\nthat's where there's an associated public/external part, which isn't\\nthe case here.\\n\\nSome places use _guts e.g. pg_logical_slot_get_changes_guts(...),\\nlargely where there's common use by several callers.\\n\\nThis is a fairly arbitrary function split for readability/length. Is\\nit actually useful to split this function up at all?\\n\\nAnyone else have an opinion?\\n\\n>> A comment is needed to explain what ExecEvalTableExpr is / does. If\\n>> it's XMLTABLE specific (which it looks like based on the code), its\\n>> name should reflect that. This pattern is repeated elsewhere; e.g.\\n>> TableExprState is really the state for an XMLTABLE expression. But\\n>> PostgreSQL actually has TABLE statements, and in future we might want\\n>> to support table-expressions, so I don't think this naming is\\n>> appropriate. This is made worse by the lack of comments on things like\\n>> the definition of TableExprState. Please use something that makes it\\n>> clear it's for XMLTABLE and add appropriate comments.\\n>\\n>\\n> I understand, so using TableExpr can be strange (for XMLTABLE function). But\\n> when we will have JSON_TABLE function, then it will have a sense.\\n\\nIt's pretty hard to review that as shared infrastructure when it's\\nstill tangled up in xmltable specifics, though.\\n\\n> \\\"TableExprState\\\" is consistent with \\\"TableExpr\\\".\\n>\\n> Any idea how it should be changed?\\n\\nI think if you want it to be shareable infrasructure, you need to\\nwrite it so it can be used as shared infrastructure. Not just name it\\nthat way but then make it XMLTABLE specific in actual functionality.\\n\\n\\n>> /* when typmod is not valid, refresh it */\\n>> if (te->typmod == -1)\\n>>\\n>>\\n>> Is this a cache? How is it valid or not valid and when? The comment\\n>> (thanks!) on TableExprGetTupleDesc says:\\n>>\\n>> /*\\n>> * When we skip transform stage (in view), then TableExpr's\\n>> * TupleDesc should not be valid. Refresh is necessary.\\n>> */\\n>>\\n>> but I'm not really grasping what you're trying to explain here. What\\n>> transform stage? What view? This could well be my ignorance of this\\n>> part of the code; if it should be understandable by a reader who is\\n>> appropriately familiar with the executor that's fine, but if it's\\n>> specific to how XMLTABLE works some more explanation would be good.\\n>\\n>\\n> This is most difficult part of this patch, and I am not sure it it is fully\\n> correctly implemented. I use TupleDesc cache. The TupleDesc is created in\\n> parser/transform stage. When the XMLTABLE is used in some view, then the\\n> transformed parser tree is materialized - and when the view is used in\\n> query, then this tree is loaded and the parser/transform stage is \\\"skipped\\\".\\n> I'll check this code against implementation of ROW constructor and I'll try\\n> to do more comments there.\\n\\nThanks. It would be good to highlight when this does and does not\\nhappen and why. Why is it necessary at all?\\n\\nWhat happens if XMLTABLE is used in a query directly, not part of a\\nview? if XMLTABLE is used in a view? If XMLTABLE is used in a prepared\\nstatement / plpgsql statement / etc? What about a CTE term?\\n\\nNot necessarily list all these cases one by one, just explain what\\nhappens, when and why. Especially if it's complex, so other readers\\ncan understand it and don't have to study it in detail to understand\\nwhat is going on. It does not need to be good public-facing\\ndocumentation, and details of wording, grammar etc can be fixed up\\nlater, it's the ideas that matter.\\n\\nIs this similar to other logic elsewhere? If so, reference that other\\nlogic so readers know where to look. That way if they're\\nchanging/bugfixing/etc one place they know there's another place that\\nmight need changing.\\n\\nI don't know this area of the code well enough to give a solid review\\nof the actual functionality, and I don't yet understand what it's\\ntrying to do so it's hard to review it by studying what it actually\\ndoes vs what it claims to do. Maybe Peter E can help, he said he was\\nthinking of looking at this patch too. But more information on what\\nit's trying to do would be a big help.\\n\\n>> PostgreSQL already has XPATH support in the form of xmlexists(...)\\n>> etc. Why is getXPathToken() etc needed? What re-use is possible here?\\n>> There's no explanation in the patch header or comments. Should the new\\n>> xpath parser be re-used by the existing xpath stuff? Why can't we use\\n>> libxml's facilities? etc. This at least needs explaining in the\\n>> submission, and some kind of hint as to why we have two different ways\\n>> to do it is needed in the code. If we do need a new XML parser, should\\n>> it be bundled in adt/xml.c along with a lot of user-facing\\n>> functionality, or a separate file?\\n>\\n> libxml2 and our XPATH function doesn't support default namespace (\\n> http://plasmasturm.org/log/259/ ). This is pretty useful feature - so I\\n> implemented.\\n\\nOK, that makes sense.\\n\\nFor the purpose of getting this patch in, is it a _necessary_ feature?\\nCan XMLTABLE be usefully implemented without it, and if so, can it be\\nadded in a subsequent patch? It would be nice to simplify this by\\nusing existing libxml2 functionality in the first version rather than\\nadding a whole new xpath as well!This is not a xpath implementation - it is preprocessing of xpath expression. without it, a users have to set explicitly PATH clause with explicit prefix \\\"./\\\" and explicit suffix \\\"/text()\\\". The usability will be significantly lower, and what is worst - the examples from internet should not work. Although is is lot of lines, this code is necessary. \\n\\n> This is the mayor issue of libxml2 library. Another difference\\n> between XPATH function and XMLTABLE function is using two phase searching\\n> and implicit prefix \\\"./\\\" and suffix (\\\"/text()\\\") in XMLTABLE. XMLTABLE using\\n> two XPATH expressions - for row data cutting and next for column data\\n> cutting (from row data). The our XPATH functions is pretty simple mapped to\\n> libxml2 XPATH API. But it is not possible with XMLTABLE function - due\\n> design of this function in standard (it is more user friendly and doesn't\\n> require exactly correct xpath expressions).\\n\\nSo you can't use existing libxml2 xpath support to implement XMLTABLE,\\neven without default namespaces?\\n\\n> I didn't find any API in libxml2 for a work with parsed xpath expressions -\\n> I need some info about the first and last token of xpath expression - it is\\n> base for decision about using prefix or suffix.\\n>\\n> This functionality (xpath expression parser) cannot be used for our XPATH\\n> function now - maybe default namespace in future.\\n\\nSo we'll have two different XPATH implementations for different\\nplaces, with different limitations, different possible bugs, etc?It is just preprocessing. The evaluation of xpath expression is part of libxml2 and it is shared.Our XPATH function is not short, but the reason is reading namespaces data from 2D array. The evaluation of xpath expression is on few lines. \\n\\nWhat would be needed to make the new XPATH work for our built-in xpath\\nfunctions too? \\n\\n>> How does XmlTableGetValue(...) and XmlTableGetRowValue(...) relate to\\n>> this? It doesn't look like they're intended to be called directly by\\n>> the user, and they're not documented (or commented).\\n>\\n>\\n> Probably I used wrong names. XMLTABLE function is running in two different\\n> modes - with explicitly defined columns (XmlTableGetValue is used), and\\n> without explicitly defined columns - so result is one XML column and only\\n> one one step searching is used (there are not column related xpath\\n> expressions) ( XmlTableGetRowValue is used). The function XmlTableGetValue\\n> is used for getting one column value, the function XmlTableGetRowValue is\\n> used for getting one value too, but in special case, when there are not any\\n> other value.\\n\\nSo both are internal implementation of the parser-level XMLTABLE(...)\\nconstruct and are not intended to be called directly by users - right?No - it is called from executor - and it should not be called differently. I have to do better separation from executor, and these functions will be private. \\n\\nComments please! A short comment on the function saying this would be\\na big help.\\n\\nRegarding naming, do we already have a convention for functions that\\nare internal implementation of something the user \\\"spells\\\"\\ndifferently? Where it's transformed by the parser? I couldn't find\\none. I don't much care about the names so long as there are comments\\nexplaining what calls the functions and what the user-facing interface\\nthat matches the function is.\\n\\nIs it safe for users to call these directly? What happens if they do\\nso incorrectly?\\n\\nWhy are they not in pg_proc.h? Do they need to be?\\n\\n>> +/*\\n>> + * There are different requests from XMLTABLE, JSON_TABLE functions\\n>> + * on passed data than has CREATE TABLE command. It is reason for\\n>> + * introduction special structure instead using ColumnDef.\\n>> + */\\n>> +typedef struct TableExprRawCol\\n>> +{\\n>> + NodeTag type;\\n>> + char *colname;\\n>> + TypeName *typeName;\\n>> + bool for_ordinality;\\n>> + bool is_not_null;\\n>> + Node *path_expr;\\n>> + Node *default_expr;\\n>> + int location;\\n>> +} TableExprRawCol;\\n>\\n> I am sorry. It is my fault. Now we have very similar node ColumnDef. This\\n> node is designed for usage in utility commands - and it is not designed for\\n> usage inside a query.\\n\\nMakes sense.\\n\\n> I had to decide between enhancing ColumnDef node or\\n> introduction new special node. Because there are more special attributes and\\n> it is hard to serialize current ColumnDef, I decided to use new node.\\n\\nSeems reasonable. The summary is \\\"this is the parse node for a column\\nof an XMLTABLE expression\\\".\\n\\nSuggested comment:\\n\\n/*\\n * This is the parsenode for a column definition in a table-expression\\nlike XMLTABLE.\\n *\\n * We can't re-use ColumnDef here; the utility command column\\ndefinition has all the\\n * wrong attributes for use in table-expressions and just doesn't make\\nsense here.\\n */\\ntypedef struct TableExprColumn\\n{\\n...\\n};\\n\\n?\\n\\nWhy \\\"RawCol\\\" ? What does it become when it's not \\\"raw\\\" anymore? Is\\nthat a reference to ColumnDef's raw_default and cooked_default for\\nuntransformed vs transformed parse-trees?TableExprColumn is betterRegardsPavel \\n\\n\\n\\n\\n--\\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Wed, 7 Sep 2016 07:49:35 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \">\\n>\\n> Suggested comment:\\n>\\n> /*\\n> * This is the parsenode for a column definition in a table-expression\\n> like XMLTABLE.\\n> *\\n> * We can't re-use ColumnDef here; the utility command column\\n> definition has all the\\n> * wrong attributes for use in table-expressions and just doesn't make\\n> sense here.\\n> */\\n> typedef struct TableExprColumn\\n> {\\n> ...\\n> };\\n>\\n> ?\\n>\\n> Why \\\"RawCol\\\" ? What does it become when it's not \\\"raw\\\" anymore? Is\\n> that a reference to ColumnDef's raw_default and cooked_default for\\n> untransformed vs transformed parse-trees?\\n>\\n\\nMy previous reply was wrong - it is used by parser only and holds TypeName\\nfield. The analogy with ColumnDef raw_default is perfect.\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n>\\n>\\n>\\n> --\\n> Craig Ringer http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Training & Services\\n>\\n\\n\\n\\nSuggested comment:\\n\\n/*\\n * This is the parsenode for a column definition in a table-expression\\nlike XMLTABLE.\\n *\\n * We can't re-use ColumnDef here; the utility command column\\ndefinition has all the\\n * wrong attributes for use in table-expressions and just doesn't make\\nsense here.\\n */\\ntypedef struct TableExprColumn\\n{\\n...\\n};\\n\\n?\\n\\nWhy \\\"RawCol\\\" ? What does it become when it's not \\\"raw\\\" anymore? Is\\nthat a reference to ColumnDef's raw_default and cooked_default for\\nuntransformed vs transformed parse-trees?My previous reply was wrong - it is used by parser only and holds TypeName field. The analogy with ColumnDef raw_default is perfect.RegardsPavel \\n\\n\\n\\n\\n--\\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Wed, 7 Sep 2016 08:44:14 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On 7 September 2016 at 14:44, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n>>\\n>> Suggested comment:\\n>>\\n>> /*\\n>> * This is the parsenode for a column definition in a table-expression\\n>> like XMLTABLE.\\n>> *\\n>> * We can't re-use ColumnDef here; the utility command column\\n>> definition has all the\\n>> * wrong attributes for use in table-expressions and just doesn't make\\n>> sense here.\\n>> */\\n>> typedef struct TableExprColumn\\n>> {\\n>> ...\\n>> };\\n>>\\n>> ?\\n>>\\n>> Why \\\"RawCol\\\" ? What does it become when it's not \\\"raw\\\" anymore? Is\\n>> that a reference to ColumnDef's raw_default and cooked_default for\\n>> untransformed vs transformed parse-trees?\\n>\\n>\\n> My previous reply was wrong - it is used by parser only and holds TypeName\\n> field. The analogy with ColumnDef raw_default is perfect.\\n\\nCool, lets just comment that then.\\n\\nI'll wait on an updated patch per discussion to date. Hopefully\\nsomebody else with more of a clue than me can offer better review of\\nthe executor/view/caching part you specifically called out as complex.\\nOtherwise maybe it'll be clearer in a revised version.\\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, 7 Sep 2016 15:23:59 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nI am sending new version of this patch\\n\\n1. now generic TableExpr is better separated from a real content generation\\n2. I removed cached typmod - using row type cache everywhere - it is\\nconsistent with other few places in Pg where dynamic types are used - the\\nresult tupdesc is generated few times more - but it is not on critical path.\\n3. More comments, few more lines in doc.\\n4. Reformated by pgindent\\n\\nRegards\\n\\nPavel\\n\\n\\n-- \\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, 9 Sep 2016 10:35:22 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-09-09 10:35 GMT+02:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n\\n> Hi,\\n>\\n> I am sending new version of this patch\\n>\\n> 1. now generic TableExpr is better separated from a real content generation\\n> 2. I removed cached typmod - using row type cache everywhere - it is\\n> consistent with other few places in Pg where dynamic types are used - the\\n> result tupdesc is generated few times more - but it is not on critical path.\\n> 3. More comments, few more lines in doc.\\n> 4. Reformated by pgindent\\n>\\n\\nnew update\\n\\nmore regress tests\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> Regards\\n>\\n> Pavel\\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\": \"Fri, 9 Sep 2016 15:44:07 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On 9 September 2016 at 21:44, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n>\\n>\\n> 2016-09-09 10:35 GMT+02:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n>>\\n>> Hi,\\n>>\\n>> I am sending new version of this patch\\n>>\\n>> 1. now generic TableExpr is better separated from a real content\\n>> generation\\n>> 2. I removed cached typmod - using row type cache everywhere - it is\\n>> consistent with other few places in Pg where dynamic types are used - the\\n>> result tupdesc is generated few times more - but it is not on critical path.\\n>> 3. More comments, few more lines in doc.\\n>> 4. Reformated by pgindent\\n\\nThanks.\\n\\nI applied this on top of the same base as your prior patch so I could\\ncompare changes.\\n\\nThe new docs look good. Thanks for that, I know it's a pain. It'll\\nneed to cover ORDINAL too, but that's not hard. I'll try to find some\\ntime to help with the docs per the references you sent offlist.\\n\\nOut of interest, should the syntax allow room for future expansion to\\npermit reading from file rather than just string literal / column\\nreference? It'd be ideal to avoid reading big documents wholly into\\nmemory when using INSERT INTO ... SELECT XMLTABLE (...) . I don't\\nsuggest adding that to this patch, just making sure adding it later\\nwould not cause problems.\\n\\nI see you added a builder context abstraction as discussed, so there's\\nno longer any direct reference to XMLTABLE specifics from TableExpr\\ncode. Good, thanks for that. It'll make things much less messy when\\nadding other table expression types as you expressed the desire to do,\\nand means the TableExpr code now makes more sense as generic\\ninfrastructure.\\n\\nExecEvalTableExprProtected and ExecEvalTableExpr are OK with me, or\\nbetter than execEvalTableExpr and ExecEvalTableExpr were anyway.\\nEventual committer will probably have opinions here.\\n\\nMild nitpick: since you can have multiple namespaces, shouldn't\\nbuilder->SetNS be builder->AddNS ?\\n\\nAdded comments are helpful, thanks.\\n\\nOn first read-through this is a big improvement and addresses all the\\nconcerns I raised. Documentation is much much better, thanks, I know\\nthat's a pain.\\n\\nI'll take a closer read-through shortly.\\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\": \"Mon, 12 Sep 2016 09:58:46 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"> I'll take a closer read-through shortly.\\n\\n\\nMissing file. You omitted executor/tableexpr.h from the patch, so I\\ncan't compile.\\n\\nI've expanded and copy-edited the docs. Some of it is guesswork based\\non the references you sent and a glance at the code. Please check my\\nchanges carefully. I found a few surprises, like the fact that DEFAULT\\nisn't a normal literal, it's an xpath expression evaluated at the same\\ntime as the rowexpression.\\n\\nUpdated patch attached as XMLTABLE-v3 includes the docs changes. Note\\nthat it's missing tableexpr.h. For convenient review or to apply to\\nyour working tree I also attach a diff of just my docs changes as\\nproposed-docs-changes.diff.\\n\\n\\nDocs:\\n\\n- Can you send the sample data used to generate the example output?\\nI'd like to include at least a cut down part of it in the docs to make\\nit clear how the input correlates with output, and preferably put the\\nwhole thing in an appendix.\\n\\n- How does it decide what subset of the document to iterate over?\\nThat's presumably rowexpr, which is xpath in postgresql? (I added this\\nto docs).\\n\\n- xmlnamespaces clause in docs needs an example for a non-default namespace.\\n\\n- What effect does xmlnamespaces clause have? Does supplying it allow\\nyou to reference qualified names in xpath? What happens if you don't\\nspecify it for a document that has namespaces or don't define all the\\nnamespaces? What if you reference an undefined namespace in xpath?\\nWhat about if an undefined namespace isn't referenced by xpath, but is\\ninside a node selected by an xpath expression?\\n\\n- What are the rules for converting the matched XML node into a value?\\nIf the matched node is not a simple text node or lacks a text node as\\nits single child, what happens?\\n\\n- What happens if the matched has multiple text node children? This\\ncan happen if, for example, you have something like\\n\\n <matchedNode>\\n some text <!-- comment splits up text node --> other text\\n </matchedNode>\\n\\n- Is there a way to get an attribute as a value? If so, an example\\nshould show this because it's going to be a common need. Presumably\\nyou want node/@attrname ?\\n\\n- What happens if COLUMNS is not specified at all? It looks like it\\nreturns a single column result set with the matched entries as 'xml'\\ntype, so added to docs, please verify.\\n\\n\\n- PASSING clause isn't really defined. You can specify one PASSING\\nentry as a literal/colref/expression, and it's the argument xml\\ndocument, right? The external docs you referred to say that PASSING\\nmay have a BY VALUE keyword, alias its argument with AS, and may have\\nexpressions, e.g.\\n\\n PASSING BY VALUE '<x/>' AS a, '<y/>' AS b\\n\\n Neither aliases nor multiple entries are supported by the code or\\ngrammar. Should this be documented as a restriction? Do you know if\\nthat's an extension by the other implementation or if it's SQL/XML\\nstandard? (I've drafted a docs update to cover this in the updated\\npatch).\\n\\n\\n- What does BY REF mean? Should this just be mentioned with a \\\"see\\nxmlexists(...)\\\" since it seems to be compatibility noise? Is there a\\ncorresponding BY VALUE or similar?\\n\\n\\n- The parser definitions re-use xmlexists_argument . I don't mind\\nthat, but it's worth noting here in case others do.\\n\\n\\n- Why do the expression arguments take c_expr (anything allowed in\\na_expr or b_expr), not b_expr (restricted expression) ?\\n\\n\\n- Column definitions are underdocumented. The grammar says they can be\\nNOT NULL, for example, but I don't see that in any of the references\\nyou mailed me nor in the docs. What behaviour is expected for a NOT\\nNULL column? I've documented my best guess (not checked closely\\nagainst the code, please verify).\\n\\n-\\n\\n\\n\\n\\nTest suggestions:\\n\\n- Coverage of multiple text() node children of an element, where split\\nup by comment or similar\\n\\n- Coverage of xpath that matches a node with child element nodes\\n\\n\\n\\nMore to come. Please review my docs changes in the mean time. I'm\\nspending a lot more time on this than I expected so I might have to\\nget onto other things for a while too.\\n\\n\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\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, 12 Sep 2016 12:28:44 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On 12 September 2016 at 12:28, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>> I'll take a closer read-through shortly.\\n\\n>DEFAULT\\n> isn't a normal literal, it's an xpath expression evaluated at the same\\n> time as the rowexpression.\\n\\nSorry for the spam, but turns out that's not the case as implemented\\nhere. The docs you referenced say it should be an xpath expression,\\nbut the implementation here is of a literal value, and examples\\nelsewhere on the Internet show a literal value. Unclear if the\\nreferenced docs are wrong or what and I don't have anything to test\\nwith.\\n\\nFeel free to fix/trim the DEFAULT related changes in above docs patch as needed.\\n\\nAlso, tests/docs should probably cover what happens when PATH matches\\nmore than one element, i.e. produces a list of more than one match.\\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\": \"Mon, 12 Sep 2016 12:36:28 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-09-12 6:36 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:\\n\\n> On 12 September 2016 at 12:28, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> >> I'll take a closer read-through shortly.\\n>\\n> >DEFAULT\\n> > isn't a normal literal, it's an xpath expression evaluated at the same\\n> > time as the rowexpression.\\n>\\n> Sorry for the spam, but turns out that's not the case as implemented\\n> here. The docs you referenced say it should be an xpath expression,\\n> but the implementation here is of a literal value, and examples\\n> elsewhere on the Internet show a literal value. Unclear if the\\n> referenced docs are wrong or what and I don't have anything to test\\n> with.\\n>\\n\\nIt is not spam. The previous comment was not correct. DEFAULT is a\\nexpression - result of this expression is used, when data is missing.\\n\\nIn standard, and some others implementation, this is literal only. It is\\nsimilar to DEFAULT clause in CREATE STATEMENT. Postgres allows expression\\nthere. Usually Postgres allows expressions everywhere when it has sense,\\nand when it is allowed by bizon parser.\\n\\n\\n\\n>\\n> Feel free to fix/trim the DEFAULT related changes in above docs patch as\\n> needed.\\n>\\n> Also, tests/docs should probably cover what happens when PATH matches\\n> more than one element, i.e. produces a list of more than one match.\\n>\\n> --\\n> Craig Ringer http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Training & Services\\n>\\n\\n2016-09-12 6:36 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:On 12 September 2016 at 12:28, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>> I'll take a closer read-through shortly.\\n\\n>DEFAULT\\n> isn't a normal literal, it's an xpath expression evaluated at the same\\n> time as the rowexpression.\\n\\nSorry for the spam, but turns out that's not the case as implemented\\nhere. The docs you referenced say it should be an xpath expression,\\nbut the implementation here is of a literal value, and examples\\nelsewhere on the Internet show a literal value. Unclear if the\\nreferenced docs are wrong or what and I don't have anything to test\\nwith.It is not spam. The previous comment was not correct. DEFAULT is a expression - result of this expression is used, when data is missing.In standard, and some others implementation, this is literal only. It is similar to DEFAULT clause in CREATE STATEMENT. Postgres allows expression there. Usually Postgres allows expressions everywhere when it has sense, and when it is allowed by bizon parser. \\n\\nFeel free to fix/trim the DEFAULT related changes in above docs patch as needed.\\n\\nAlso, tests/docs should probably cover what happens when PATH matches\\nmore than one element, i.e. produces a list of more than one match.\\n\\n--\\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Mon, 12 Sep 2016 06:42:08 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-09-12 3:58 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:\\n\\n> On 9 September 2016 at 21:44, Pavel Stehule <pavel.stehule@gmail.com>\\n> wrote:\\n> >\\n> >\\n> > 2016-09-09 10:35 GMT+02:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n> >>\\n> >> Hi,\\n> >>\\n> >> I am sending new version of this patch\\n> >>\\n> >> 1. now generic TableExpr is better separated from a real content\\n> >> generation\\n> >> 2. I removed cached typmod - using row type cache everywhere - it is\\n> >> consistent with other few places in Pg where dynamic types are used -\\n> the\\n> >> result tupdesc is generated few times more - but it is not on critical\\n> path.\\n> >> 3. More comments, few more lines in doc.\\n> >> 4. Reformated by pgindent\\n>\\n> Thanks.\\n>\\n> I applied this on top of the same base as your prior patch so I could\\n> compare changes.\\n>\\n> The new docs look good. Thanks for that, I know it's a pain. It'll\\n> need to cover ORDINAL too, but that's not hard. I'll try to find some\\n> time to help with the docs per the references you sent offlist.\\n>\\n> Out of interest, should the syntax allow room for future expansion to\\n> permit reading from file rather than just string literal / column\\n> reference? It'd be ideal to avoid reading big documents wholly into\\n> memory when using INSERT INTO ... SELECT XMLTABLE (...) . I don't\\n> suggest adding that to this patch, just making sure adding it later\\n> would not cause problems.\\n>\\n\\nthis is little bit different question - it is server side function, so\\nfirst question is - how to push usually client side content to server? Next\\nquestion is how to get this content to a executor. Now only COPY statement\\nis able to do.\\n\\nI am thinking so this should not be problem, but it requires maybe some\\nspecial keywords - fileref, local fileref, and some changes in protocol.\\nBecause this function has own implementation in parser/transform stage,\\nthen nothing will be lost in process, and we can implement lazy parameter\\nevaluation. Another question if libxml2 has enough possibility to work with\\nstream.\\n\\nOne idea - we can introduce \\\"external (server side|client side) blobs\\\" with\\nspecial types and special streaming IO. With these types, there no changes\\nare necessary on syntax level. With this, the syntax sugar flag \\\"BY REF\\\"\\ncan be useful.\\n\\n\\n>\\n> I see you added a builder context abstraction as discussed, so there's\\n> no longer any direct reference to XMLTABLE specifics from TableExpr\\n> code. Good, thanks for that. It'll make things much less messy when\\n> adding other table expression types as you expressed the desire to do,\\n> and means the TableExpr code now makes more sense as generic\\n> infrastructure.\\n>\\n> ExecEvalTableExprProtected and ExecEvalTableExpr are OK with me, or\\n> better than execEvalTableExpr and ExecEvalTableExpr were anyway.\\n> Eventual committer will probably have opinions here.\\n>\\n> Mild nitpick: since you can have multiple namespaces, shouldn't\\n> builder->SetNS be builder->AddNS ?\\n>\\n\\ngood idea.\\n\\n\\n>\\n> Added comments are helpful, thanks.\\n>\\n> On first read-through this is a big improvement and addresses all the\\n> concerns I raised. Documentation is much much better, thanks, I know\\n> that's a pain.\\n>\\n> I'll take a closer read-through shortly.\\n>\\n\\n\\nupdated patch attached - with your documentation.\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> --\\n> Craig Ringer http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, 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\": \"Mon, 12 Sep 2016 07:07:18 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-09-12 6:28 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:\\n\\n> > I'll take a closer read-through shortly.\\n>\\n>\\n> Missing file. You omitted executor/tableexpr.h from the patch, so I\\n> can't compile.\\n>\\n> I've expanded and copy-edited the docs. Some of it is guesswork based\\n> on the references you sent and a glance at the code. Please check my\\n> changes carefully. I found a few surprises, like the fact that DEFAULT\\n> isn't a normal literal, it's an xpath expression evaluated at the same\\n> time as the rowexpression.\\n>\\n> Updated patch attached as XMLTABLE-v3 includes the docs changes. Note\\n> that it's missing tableexpr.h. For convenient review or to apply to\\n> your working tree I also attach a diff of just my docs changes as\\n> proposed-docs-changes.diff.\\n>\\n>\\n> Docs:\\n>\\n> - Can you send the sample data used to generate the example output?\\n> I'd like to include at least a cut down part of it in the docs to make\\n> it clear how the input correlates with output, and preferably put the\\n> whole thing in an appendix.\\n>\\n\\nit is in regress tests.\\n\\n\\n>\\n> - How does it decide what subset of the document to iterate over?\\n> That's presumably rowexpr, which is xpath in postgresql? (I added this\\n> to docs).\\n>\\n> - xmlnamespaces clause in docs needs an example for a non-default\\n> namespace.\\n>\\n> - What effect does xmlnamespaces clause have? Does supplying it allow\\n> you to reference qualified names in xpath? What happens if you don't\\n> specify it for a document that has namespaces or don't define all the\\n> namespaces? What if you reference an undefined namespace in xpath?\\n> What about if an undefined namespace isn't referenced by xpath, but is\\n> inside a node selected by an xpath expression?\\n>\\n\\nAll this is under libxml2 control - when you use undefined namespace, then\\nlibxml2 raises a error. The namespaces in document and in XPath queries are\\nabsolutely independent - the relation is a URI. When you use bad URI\\n(referenced by name), then the result will be empty set. When you use\\nundefined name, then you will get a error.\\n\\n\\n>\\n> - What are the rules for converting the matched XML node into a value?\\n> If the matched node is not a simple text node or lacks a text node as\\n> its single child, what happens?\\n>\\n\\nThis process is described and controlled by \\\"XML SQL mapping\\\". The Postgres\\nhas minimalistic implementation without possibility of external control and\\nwithout schema support. The my implementation is simple. When user doesn't\\nspecify result target like explicit using of text() function, then the\\ntext() function is used implicitly when target type is not XML. Then I dump\\nresult to string and I enforce related input functions for target types.\\n\\n\\n\\n>\\n> - What happens if the matched has multiple text node children? This\\n> can happen if, for example, you have something like\\n>\\n> <matchedNode>\\n> some text <!-- comment splits up text node --> other text\\n> </matchedNode>\\n>\\n\\ndepends on target type - it is allowed in XML, and it is disallowed for\\nother types. I though about support of a arrays - but the patch will be\\nmuch more complex - there can be recursion - so I disallowed it. When the\\nuser have to solve this issue, then he can use nested XMLTABLE functions\\nand nested function is working with XML type.\\n\\nJust for record - This issue is solved in JSON_TABLE functions - it allows\\nnested PATHS. But XMLTABLE doesn't allow it.\\n\\n\\n>\\n> - Is there a way to get an attribute as a value? If so, an example\\n> should show this because it's going to be a common need. Presumably\\n> you want node/@attrname ?\\n>\\n\\nyou can use reference to current node \\\".\\\" - so \\\"./@attname\\\" should to work\\n- a example is in regress tests\\n\\n\\n\\n>\\n> - What happens if COLUMNS is not specified at all? It looks like it\\n> returns a single column result set with the matched entries as 'xml'\\n> type, so added to docs, please verify.\\n>\\n\\nsure, that is it\\n\\n\\n>\\n>\\n> - PASSING clause isn't really defined. You can specify one PASSING\\n> entry as a literal/colref/expression, and it's the argument xml\\n> document, right? The external docs you referred to say that PASSING\\n> may have a BY VALUE keyword, alias its argument with AS, and may have\\n> expressions, e.g.\\n>\\n> PASSING BY VALUE '<x/>' AS a, '<y/>' AS b\\n>\\n> Neither aliases nor multiple entries are supported by the code or\\n> grammar. Should this be documented as a restriction? Do you know if\\n> that's an extension by the other implementation or if it's SQL/XML\\n> standard? (I've drafted a docs update to cover this in the updated\\n> patch).\\n>\\n\\nThe ANSI allows to pass more documents - and then do complex queries with\\nXQuery. Passing more than one document has not sense in libxml2 based\\nimplementation, so I didn't supported it. The referenced names can be\\nimplemented later - but it needs to changes in XPATH function too.\\n\\n\\n>\\n>\\n> - What does BY REF mean? Should this just be mentioned with a \\\"see\\n> xmlexists(...)\\\" since it seems to be compatibility noise? Is there a\\n> corresponding BY VALUE or similar?\\n>\\n\\nWhen the XML document is stored as serialized DOM, then by ref means link\\non this DOM. It has not sense in Postgres - because we store XML documents\\nby value only.\\n\\n\\n>\\n>\\n> - The parser definitions re-use xmlexists_argument . I don't mind\\n> that, but it's worth noting here in case others do.\\n>\\n\\nIt is one clause - see SQL/XML doc PASSING <XML table argument passing\\nmechanism>\\n\\n\\n>\\n> - Why do the expression arguments take c_expr (anything allowed in\\n> a_expr or b_expr), not b_expr (restricted expression) ?\\n>\\n\\nI don't know - I expect the problems with parser - because PASSING is\\nrestricted keyword in ANSI/SQL and unreserved keyword in Postgres.\\n\\n\\n>\\n>\\n> - Column definitions are underdocumented. The grammar says they can be\\n> NOT NULL, for example, but I don't see that in any of the references\\n> you mailed me nor in the docs. What behaviour is expected for a NOT\\n> NULL column? I've documented my best guess (not checked closely\\n> against the code, please verify).\\n>\\n>\\nyes - some other databases allows it - I am thinking so it is useful.\\n\\n\\n> -\\n>\\n>\\n>\\n>\\n> Test suggestions:\\n>\\n> - Coverage of multiple text() node children of an element, where split\\n> up by comment or similar\\n>\\n> - Coverage of xpath that matches a node with child element nodes\\n>\\n\\nI'll do it.\\n\\n\\n>\\n>\\n>\\n> More to come. Please review my docs changes in the mean time. I'm\\n> spending a lot more time on this than I expected so I might have to\\n> get onto other things for a while too.\\n>\\n>\\n>\\n> --\\n> Craig Ringer http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Training & Services\\n>\\n\\n2016-09-12 6:28 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:> I'll take a closer read-through shortly.\\n\\n\\nMissing file. You omitted executor/tableexpr.h from the patch, so I\\ncan't compile.\\n\\nI've expanded and copy-edited the docs. Some of it is guesswork based\\non the references you sent and a glance at the code. Please check my\\nchanges carefully. I found a few surprises, like the fact that DEFAULT\\nisn't a normal literal, it's an xpath expression evaluated at the same\\ntime as the rowexpression.\\n\\nUpdated patch attached as XMLTABLE-v3 includes the docs changes. Note\\nthat it's missing tableexpr.h. For convenient review or to apply to\\nyour working tree I also attach a diff of just my docs changes as\\nproposed-docs-changes.diff.\\n\\n\\nDocs:\\n\\n- Can you send the sample data used to generate the example output?\\nI'd like to include at least a cut down part of it in the docs to make\\nit clear how the input correlates with output, and preferably put the\\nwhole thing in an appendix.it is in regress tests. \\n\\n- How does it decide what subset of the document to iterate over?\\nThat's presumably rowexpr, which is xpath in postgresql? (I added this\\nto docs).\\n\\n- xmlnamespaces clause in docs needs an example for a non-default namespace.\\n\\n- What effect does xmlnamespaces clause have? Does supplying it allow\\nyou to reference qualified names in xpath? What happens if you don't\\nspecify it for a document that has namespaces or don't define all the\\nnamespaces? What if you reference an undefined namespace in xpath?\\nWhat about if an undefined namespace isn't referenced by xpath, but is\\ninside a node selected by an xpath expression?All this is under libxml2 control - when you use undefined namespace, then libxml2 raises a error. The namespaces in document and in XPath queries are absolutely independent - the relation is a URI. When you use bad URI (referenced by name), then the result will be empty set. When you use undefined name, then you will get a error. \\n\\n- What are the rules for converting the matched XML node into a value?\\nIf the matched node is not a simple text node or lacks a text node as\\nits single child, what happens?This process is described and controlled by \\\"XML SQL mapping\\\". The Postgres has minimalistic implementation without possibility of external control and without schema support. The my implementation is simple. When user doesn't specify result target like explicit using of text() function, then the text() function is used implicitly when target type is not XML. Then I dump result to string and I enforce related input functions for target types. \\n\\n- What happens if the matched has multiple text node children? This\\ncan happen if, for example, you have something like\\n\\n <matchedNode>\\n some text <!-- comment splits up text node --> other text\\n </matchedNode>depends on target type - it is allowed in XML, and it is disallowed for other types. I though about support of a arrays - but the patch will be much more complex - there can be recursion - so I disallowed it. When the user have to solve this issue, then he can use nested XMLTABLE functions and nested function is working with XML type. Just for record - This issue is solved in JSON_TABLE functions - it allows nested PATHS. But XMLTABLE doesn't allow it. \\n\\n- Is there a way to get an attribute as a value? If so, an example\\nshould show this because it's going to be a common need. Presumably\\nyou want node/@attrname ?you can use reference to current node \\\".\\\" - so \\\"./@attname\\\" should to work - a example is in regress tests \\n\\n- What happens if COLUMNS is not specified at all? It looks like it\\nreturns a single column result set with the matched entries as 'xml'\\ntype, so added to docs, please verify.sure, that is it \\n\\n\\n- PASSING clause isn't really defined. You can specify one PASSING\\nentry as a literal/colref/expression, and it's the argument xml\\ndocument, right? The external docs you referred to say that PASSING\\nmay have a BY VALUE keyword, alias its argument with AS, and may have\\nexpressions, e.g.\\n\\n PASSING BY VALUE '<x/>' AS a, '<y/>' AS b\\n\\n Neither aliases nor multiple entries are supported by the code or\\ngrammar. Should this be documented as a restriction? Do you know if\\nthat's an extension by the other implementation or if it's SQL/XML\\nstandard? (I've drafted a docs update to cover this in the updated\\npatch).The ANSI allows to pass more documents - and then do complex queries with XQuery. Passing more than one document has not sense in libxml2 based implementation, so I didn't supported it. The referenced names can be implemented later - but it needs to changes in XPATH function too. \\n\\n\\n- What does BY REF mean? Should this just be mentioned with a \\\"see\\nxmlexists(...)\\\" since it seems to be compatibility noise? Is there a\\ncorresponding BY VALUE or similar?When the XML document is stored as serialized DOM, then by ref means link on this DOM. It has not sense in Postgres - because we store XML documents by value only. \\n\\n\\n- The parser definitions re-use xmlexists_argument . I don't mind\\nthat, but it's worth noting here in case others do.It is one clause - see SQL/XML doc PASSING <XML table argument passing mechanism> \\n\\n\\n- Why do the expression arguments take c_expr (anything allowed in\\na_expr or b_expr), not b_expr (restricted expression) ?I don't know - I expect the problems with parser - because PASSING is restricted keyword in ANSI/SQL and unreserved keyword in Postgres. \\n\\n\\n- Column definitions are underdocumented. The grammar says they can be\\nNOT NULL, for example, but I don't see that in any of the references\\nyou mailed me nor in the docs. What behaviour is expected for a NOT\\nNULL column? I've documented my best guess (not checked closely\\nagainst the code, please verify).\\nyes - some other databases allows it - I am thinking so it is useful. \\n-\\n\\n\\n\\n\\nTest suggestions:\\n\\n- Coverage of multiple text() node children of an element, where split\\nup by comment or similar\\n\\n- Coverage of xpath that matches a node with child element nodesI'll do it. \\n\\n\\n\\nMore to come. Please review my docs changes in the mean time. I'm\\nspending a lot more time on this than I expected so I might have to\\nget onto other things for a while too.\\n\\n\\n\\n--\\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Mon, 12 Sep 2016 07:48:35 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-09-12 6:36 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:\\n\\n> On 12 September 2016 at 12:28, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> >> I'll take a closer read-through shortly.\\n>\\n> >DEFAULT\\n> > isn't a normal literal, it's an xpath expression evaluated at the same\\n> > time as the rowexpression.\\n>\\n> Sorry for the spam, but turns out that's not the case as implemented\\n> here. The docs you referenced say it should be an xpath expression,\\n> but the implementation here is of a literal value, and examples\\n> elsewhere on the Internet show a literal value. Unclear if the\\n> referenced docs are wrong or what and I don't have anything to test\\n> with.\\n>\\n> Feel free to fix/trim the DEFAULT related changes in above docs patch as\\n> needed.\\n>\\n> Also, tests/docs should probably cover what happens when PATH matches\\n> more than one element, i.e. produces a list of more than one match.\\n>\\n\\nIt is there for case, when this is allowed. When you change the target\\ntype to any non XML type, then a error is raised.\\n\\nI didn't write a negative test cases until the text of messages will be\\nfinal (or checked by native speaker).\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> --\\n> Craig Ringer http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Training & Services\\n>\\n\\n2016-09-12 6:36 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:On 12 September 2016 at 12:28, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>> I'll take a closer read-through shortly.\\n\\n>DEFAULT\\n> isn't a normal literal, it's an xpath expression evaluated at the same\\n> time as the rowexpression.\\n\\nSorry for the spam, but turns out that's not the case as implemented\\nhere. The docs you referenced say it should be an xpath expression,\\nbut the implementation here is of a literal value, and examples\\nelsewhere on the Internet show a literal value. Unclear if the\\nreferenced docs are wrong or what and I don't have anything to test\\nwith.\\n\\nFeel free to fix/trim the DEFAULT related changes in above docs patch as needed.\\n\\nAlso, tests/docs should probably cover what happens when PATH matches\\nmore than one element, i.e. produces a list of more than one match.It is there for case, when this is allowed. When you change the target type to any non XML type, then a error is raised.I didn't write a negative test cases until the text of messages will be final (or checked by native speaker).RegardsPavel \\n\\n--\\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Mon, 12 Sep 2016 07:52:55 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On 12 September 2016 at 13:07, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n\\n>> Out of interest, should the syntax allow room for future expansion to\\n>> permit reading from file rather than just string literal / column\\n>> reference? It'd be ideal to avoid reading big documents wholly into\\n>> memory when using INSERT INTO ... SELECT XMLTABLE (...) . I don't\\n>> suggest adding that to this patch, just making sure adding it later\\n>> would not cause problems.\\n>\\n>\\n> this is little bit different question - it is server side function, so first\\n> question is - how to push usually client side content to server? Next\\n> question is how to get this content to a executor. Now only COPY statement\\n> is able to do.\\n\\nProbably start with support for server-side files. When people are\\ndealing with really big files they'll be more willing to copy files to\\nthe server or bind them into the server file system over the network.\\n\\nThe v3 protocol doesn't really allow any way for client-to-server\\nstreaming during a query, I think that's hopeless until we have a\\nprotocol bump.\\n\\n> updated patch attached - with your documentation.\\n\\nWill take a look and a better read of the code. Likely tomorrow, I've\\ngot work to do as well.\\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\": \"Mon, 12 Sep 2016 14:02:33 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\nThere is some opened questions - the standard (and some other databases)\\nrequires entering XPath expression as string literal.\\n\\nI am thinking so it is too strong not necessary limit - (it enforces\\ndynamic query in more cases), so I allowed the expressions there.\\n\\nAnother questions is when these expressions should be evaluated. There are\\ntwo possibilities - once per query, once per input row. I selected \\\"once\\nper input row mode\\\" - it is simpler to implement it, and it is consistent\\nwith other \\\"similar\\\" generators - see the behave and related discussion to\\n\\\"array_to_string\\\" and evaluation of separator argument. The switch to \\\"once\\nper query\\\" should not be hard - but it can be strange for users, because\\nsome his volatile expression should be stable.\\n\\nRegards\\n\\nPavel\\n\\nHiThere is some opened questions - the standard (and some other databases) requires entering XPath expression as string literal. I am thinking so it is too strong not necessary limit - (it enforces dynamic query in more cases), so I allowed the expressions there. Another questions is when these expressions should be evaluated. There are two possibilities - once per query, once per input row. I selected \\\"once per input row mode\\\" - it is simpler to implement it, and it is consistent with other \\\"similar\\\" generators - see the behave and related discussion to \\\"array_to_string\\\" and evaluation of separator argument. The switch to \\\"once per query\\\" should not be hard - but it can be strange for users, because some his volatile expression should be stable. RegardsPavel\",\n \"msg_date\": \"Mon, 12 Sep 2016 08:02:50 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On 12 September 2016 at 14:02, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n> Hi\\n>\\n> There is some opened questions - the standard (and some other databases)\\n> requires entering XPath expression as string literal.\\n>\\n> I am thinking so it is too strong not necessary limit - (it enforces dynamic\\n> query in more cases), so I allowed the expressions there.\\n\\nI agree. There's no reason not to permit expressions there, and there\\nare many other places where we have similar extensions.\\n\\n> Another questions is when these expressions should be evaluated. There are\\n> two possibilities - once per query, once per input row. I selected \\\"once per\\n> input row mode\\\" - it is simpler to implement it, and it is consistent with\\n> other \\\"similar\\\" generators - see the behave and related discussion to\\n> \\\"array_to_string\\\" and evaluation of separator argument. The switch to \\\"once\\n> per query\\\" should not be hard - but it can be strange for users, because\\n> some his volatile expression should be stable.\\n\\nI would've expected once per query. There's no way the expressions can\\nreference the row data, so there's no reason to evaluate them each\\ntime.\\n\\nThe only use case I see for evaluating them each time is - maybe -\\nDEFAULT. Where maybe there's a use for nextval() or other volatile\\nfunctions. But honestly, I think that's better done explicitly in a\\npost-pass, i.e.\\n\\nselect uuid_generate_v4(), x.*\\nfrom (\\n xmltable(.....) x\\n);\\n\\nin cases where that's what the user actually wants.\\n\\nThere's no other case I can think of where expressions as arguments to\\nset-returning functions are evaluated once per output row.\\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\": \"Mon, 12 Sep 2016 14:14:49 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On 12 September 2016 at 13:07, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n\\n>> Out of interest, should the syntax allow room for future expansion to\\n>> permit reading from file rather than just string literal / column\\n>> reference? It'd be ideal to avoid reading big documents wholly into\\n>> memory when using INSERT INTO ... SELECT XMLTABLE (...) . I don't\\n>> suggest adding that to this patch, just making sure adding it later\\n>> would not cause problems.\\n>\\n>\\n> this is little bit different question - it is server side function, so first\\n> question is - how to push usually client side content to server? Next\\n> question is how to get this content to a executor. Now only COPY statement\\n> is able to do.\\n\\nProbably start with support for server-side files. When people are\\ndealing with really big files they'll be more willing to copy files to\\nthe server or bind them into the server file system over the network.\\n\\nThe v3 protocol doesn't really allow any way for client-to-server\\nstreaming during a query, I think that's hopeless until we have a\\nprotocol bump.\\n\\n> updated patch attached - with your documentation.\\n\\n\\n>> - Can you send the sample data used to generate the example output?\\n>> I'd like to include at least a cut down part of it in the docs to make\\n>> it clear how the input correlates with output, and preferably put the\\n>> whole thing in an appendix.\\n>\\n>\\n> it is in regress tests.\\n\\nMakes sense.\\n\\nIt's not that verbose (for XML) and I wonder if it's just worth\\nincluding it in-line in the docs along with the XMLTABLE example. It'd\\nbe much easier to understand how XMLTABLE works and what it does then.\\n\\n>> - What effect does xmlnamespaces clause have? Does supplying it allow\\n>> you to reference qualified names in xpath? What happens if you don't\\n>> specify it for a document that has namespaces or don't define all the\\n>> namespaces? What if you reference an undefined namespace in xpath?\\n>> What about if an undefined namespace isn't referenced by xpath, but is\\n>> inside a node selected by an xpath expression?\\n>\\n>\\n> All this is under libxml2 control - when you use undefined namespace, then\\n> libxml2 raises a error. The namespaces in document and in XPath queries are\\n> absolutely independent - the relation is a URI. When you use bad URI\\n> (referenced by name), then the result will be empty set. When you use\\n> undefined name, then you will get a error.\\n\\nOK, makes sense.\\n\\n>> - What are the rules for converting the matched XML node into a value?\\n>> If the matched node is not a simple text node or lacks a text node as\\n>> its single child, what happens?\\n>\\n> This process is described and controlled by \\\"XML SQL mapping\\\". The Postgres\\n> has minimalistic implementation without possibility of external control and\\n> without schema support. The my implementation is simple. When user doesn't\\n> specify result target like explicit using of text() function, then the\\n> text() function is used implicitly when target type is not XML. Then I dump\\n> result to string and I enforce related input functions for target types.\\n\\nOK, so a subset of the full spec functionality is provided because of\\nlimitations in Pg and libxml2. Makes sense.\\n\\nMy only big concern here is that use of text() is a common mistake in\\nXSLT, and I think the same thing will happen here. Users expect\\ncomments to be ignored, but in fact a comment inserts a comment node\\ninto the XML DOM, so a comment between two pieces of text produces a\\ntree of\\n\\nelement\\n text()\\n comment()\\n text()\\n\\n\\nIf you match element/text(), you get a 2-node result and will presumably ERROR.\\n\\nThere is no good way to tell this from\\n\\nelement\\n text()\\nelement\\n text()\\n\\nwhen you use an xpath expression like element/text() . So you can't\\nsafely solve it just by concatenating all resulting text() nodes\\nwithout surprising behaviour.\\n\\n>> - What happens if the matched has multiple text node children? This\\n>> can happen if, for example, you have something like\\n>>\\n>> <matchedNode>\\n>> some text <!-- comment splits up text node --> other text\\n>> </matchedNode>\\n>\\n>\\n> depends on target type - it is allowed in XML, and it is disallowed for\\n> other types. I though about support of a arrays - but the patch will be much\\n> more complex - there can be recursion - so I disallowed it. When the user\\n> have to solve this issue, then he can use nested XMLTABLE functions and\\n> nested function is working with XML type.\\n\\nI don't really understand how that'd work.\\n\\nDo you know how other implementations handle this?\\n\\nI think users are going to be VERY surprised when comments in text\\nbreak their XML.\\n\\n>> - Is there a way to get an attribute as a value? If so, an example\\n>> should show this because it's going to be a common need. Presumably\\n>> you want node/@attrname ?\\n>\\n>\\n> you can use reference to current node \\\".\\\" - so \\\"./@attname\\\" should to work -\\n> a example is in regress tests\\n\\ncool, just needs mention in docs then.\\n\\n>> - PASSING clause isn't really defined. You can specify one PASSING\\n>> entry as a literal/colref/expression, and it's the argument xml\\n>> document, right? The external docs you referred to say that PASSING\\n>> may have a BY VALUE keyword, alias its argument with AS, and may have\\n>> expressions, e.g.\\n>>\\n>> PASSING BY VALUE '<x/>' AS a, '<y/>' AS b\\n>>\\n>> Neither aliases nor multiple entries are supported by the code or\\n>> grammar. Should this be documented as a restriction?\\n>\\n>\\n> The ANSI allows to pass more documents - and then do complex queries with\\n> XQuery. Passing more than one document has not sense in libxml2 based\\n> implementation, so I didn't supported it. The referenced names can be\\n> implemented later - but it needs to changes in XPATH function too.\\n\\nOK, so my docs addition that just says they're not supported should be fine.\\n\\n>> - What does BY REF mean? Should this just be mentioned with a \\\"see\\n>> xmlexists(...)\\\" since it seems to be compatibility noise? Is there a\\n>> corresponding BY VALUE or similar?\\n>\\n> When the XML document is stored as serialized DOM, then by ref means link on\\n> this DOM. It has not sense in Postgres - because we store XML documents by\\n> value only.\\n\\nRight. And since there's already precent for xmlexists there's no\\npoint worrying about whether we lose opportunities to implement it\\nlater.\\n\\n>> - Why do the expression arguments take c_expr (anything allowed in\\n>> a_expr or b_expr), not b_expr (restricted expression) ?\\n>\\n>\\n> I don't know - I expect the problems with parser - because PASSING is\\n> restricted keyword in ANSI/SQL and unreserved keyword in Postgres.\\n\\nI mean for the rowpath argument, not the parts within\\nxmlexists_argument. If the rowpath doesn't need to be c_expr\\npresumably it should be a b_expr or even, if it doesn't cause parsing\\nambiguities, an a_expr ? There doesn't seem to be the same issue here\\nas we have with BETWEEN etc.\\n\\n>> - Column definitions are underdocumented. The grammar says they can be\\n>> NOT NULL, for example, but I don't see that in any of the references\\n>> you mailed me nor in the docs. What behaviour is expected for a NOT\\n>> NULL column? I've documented my best guess (not checked closely\\n>> against the code, please verify).\\n>>\\n>\\n> yes - some other databases allows it - I am thinking so it is useful.\\n\\nSure. Sounds like my docs additions are probably right then, except\\nfor incorrect description of DEFAULT.\\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\": \"Mon, 12 Sep 2016 14:46:26 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-09-12 8:46 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:\\n\\n> On 12 September 2016 at 13:07, Pavel Stehule <pavel.stehule@gmail.com>\\n> wrote:\\n>\\n> >> Out of interest, should the syntax allow room for future expansion to\\n> >> permit reading from file rather than just string literal / column\\n> >> reference? It'd be ideal to avoid reading big documents wholly into\\n> >> memory when using INSERT INTO ... SELECT XMLTABLE (...) . I don't\\n> >> suggest adding that to this patch, just making sure adding it later\\n> >> would not cause problems.\\n> >\\n> >\\n> > this is little bit different question - it is server side function, so\\n> first\\n> > question is - how to push usually client side content to server? Next\\n> > question is how to get this content to a executor. Now only COPY\\n> statement\\n> > is able to do.\\n>\\n> Probably start with support for server-side files. When people are\\n> dealing with really big files they'll be more willing to copy files to\\n> the server or bind them into the server file system over the network.\\n>\\n> The v3 protocol doesn't really allow any way for client-to-server\\n> streaming during a query, I think that's hopeless until we have a\\n> protocol bump.\\n>\\n> > updated patch attached - with your documentation.\\n>\\n>\\n> >> - Can you send the sample data used to generate the example output?\\n> >> I'd like to include at least a cut down part of it in the docs to make\\n> >> it clear how the input correlates with output, and preferably put the\\n> >> whole thing in an appendix.\\n> >\\n> >\\n> > it is in regress tests.\\n>\\n> Makes sense.\\n>\\n> It's not that verbose (for XML) and I wonder if it's just worth\\n> including it in-line in the docs along with the XMLTABLE example. It'd\\n> be much easier to understand how XMLTABLE works and what it does then.\\n>\\n> >> - What effect does xmlnamespaces clause have? Does supplying it allow\\n> >> you to reference qualified names in xpath? What happens if you don't\\n> >> specify it for a document that has namespaces or don't define all the\\n> >> namespaces? What if you reference an undefined namespace in xpath?\\n> >> What about if an undefined namespace isn't referenced by xpath, but is\\n> >> inside a node selected by an xpath expression?\\n> >\\n> >\\n> > All this is under libxml2 control - when you use undefined namespace,\\n> then\\n> > libxml2 raises a error. The namespaces in document and in XPath queries\\n> are\\n> > absolutely independent - the relation is a URI. When you use bad URI\\n> > (referenced by name), then the result will be empty set. When you use\\n> > undefined name, then you will get a error.\\n>\\n> OK, makes sense.\\n>\\n> >> - What are the rules for converting the matched XML node into a value?\\n> >> If the matched node is not a simple text node or lacks a text node as\\n> >> its single child, what happens?\\n> >\\n> > This process is described and controlled by \\\"XML SQL mapping\\\". The\\n> Postgres\\n> > has minimalistic implementation without possibility of external control\\n> and\\n> > without schema support. The my implementation is simple. When user\\n> doesn't\\n> > specify result target like explicit using of text() function, then the\\n> > text() function is used implicitly when target type is not XML. Then I\\n> dump\\n> > result to string and I enforce related input functions for target types.\\n>\\n> OK, so a subset of the full spec functionality is provided because of\\n> limitations in Pg and libxml2. Makes sense.\\n>\\n> My only big concern here is that use of text() is a common mistake in\\n> XSLT, and I think the same thing will happen here. Users expect\\n> comments to be ignored, but in fact a comment inserts a comment node\\n> into the XML DOM, so a comment between two pieces of text produces a\\n> tree of\\n>\\n> element\\n> text()\\n> comment()\\n> text()\\n>\\n>\\n> If you match element/text(), you get a 2-node result and will presumably\\n> ERROR.\\n>\\n> There is no good way to tell this from\\n>\\n> element\\n> text()\\n> element\\n> text()\\n>\\n> when you use an xpath expression like element/text() . So you can't\\n> safely solve it just by concatenating all resulting text() nodes\\n> without surprising behaviour.\\n>\\n> >> - What happens if the matched has multiple text node children? This\\n> >> can happen if, for example, you have something like\\n> >>\\n> >> <matchedNode>\\n> >> some text <!-- comment splits up text node --> other text\\n> >> </matchedNode>\\n> >\\n>\\n\\nI fixed this case - new regress tests added\\n\\n\\n> >\\n> > depends on target type - it is allowed in XML, and it is disallowed for\\n> > other types. I though about support of a arrays - but the patch will be\\n> much\\n> > more complex - there can be recursion - so I disallowed it. When the user\\n> > have to solve this issue, then he can use nested XMLTABLE functions and\\n> > nested function is working with XML type.\\n>\\n> I don't really understand how that'd work.\\n>\\n> Do you know how other implementations handle this?\\n>\\n> I think users are going to be VERY surprised when comments in text\\n> break their XML.\\n>\\n> >> - Is there a way to get an attribute as a value? If so, an example\\n> >> should show this because it's going to be a common need. Presumably\\n> >> you want node/@attrname ?\\n> >\\n> >\\n> > you can use reference to current node \\\".\\\" - so \\\"./@attname\\\" should to\\n> work -\\n> > a example is in regress tests\\n>\\n> cool, just needs mention in docs then.\\n>\\n> >> - PASSING clause isn't really defined. You can specify one PASSING\\n> >> entry as a literal/colref/expression, and it's the argument xml\\n> >> document, right? The external docs you referred to say that PASSING\\n> >> may have a BY VALUE keyword, alias its argument with AS, and may have\\n> >> expressions, e.g.\\n> >>\\n> >> PASSING BY VALUE '<x/>' AS a, '<y/>' AS b\\n> >>\\n> >> Neither aliases nor multiple entries are supported by the code or\\n> >> grammar. Should this be documented as a restriction?\\n> >\\n> >\\n> > The ANSI allows to pass more documents - and then do complex queries with\\n> > XQuery. Passing more than one document has not sense in libxml2 based\\n> > implementation, so I didn't supported it. The referenced names can be\\n> > implemented later - but it needs to changes in XPATH function too.\\n>\\n> OK, so my docs addition that just says they're not supported should be\\n> fine.\\n>\\n> >> - What does BY REF mean? Should this just be mentioned with a \\\"see\\n> >> xmlexists(...)\\\" since it seems to be compatibility noise? Is there a\\n> >> corresponding BY VALUE or similar?\\n> >\\n> > When the XML document is stored as serialized DOM, then by ref means\\n> link on\\n> > this DOM. It has not sense in Postgres - because we store XML documents\\n> by\\n> > value only.\\n>\\n> Right. And since there's already precent for xmlexists there's no\\n> point worrying about whether we lose opportunities to implement it\\n> later.\\n>\\n> >> - Why do the expression arguments take c_expr (anything allowed in\\n> >> a_expr or b_expr), not b_expr (restricted expression) ?\\n> >\\n> >\\n> > I don't know - I expect the problems with parser - because PASSING is\\n> > restricted keyword in ANSI/SQL and unreserved keyword in Postgres.\\n>\\n> I mean for the rowpath argument, not the parts within\\n> xmlexists_argument. If the rowpath doesn't need to be c_expr\\n> presumably it should be a b_expr or even, if it doesn't cause parsing\\n> ambiguities, an a_expr ? There doesn't seem to be the same issue here\\n> as we have with BETWEEN etc.\\n>\\n\\nb_expr enforces shift/reduce conflict :(\\n\\n\\n>\\n> >> - Column definitions are underdocumented. The grammar says they can be\\n> >> NOT NULL, for example, but I don't see that in any of the references\\n> >> you mailed me nor in the docs. What behaviour is expected for a NOT\\n> >> NULL column? I've documented my best guess (not checked closely\\n> >> against the code, please verify).\\n> >>\\n> >\\n> > yes - some other databases allows it - I am thinking so it is useful.\\n>\\n> Sure. Sounds like my docs additions are probably right then, except\\n> for incorrect description of DEFAULT.\\n>\\n\\nI found other opened question - how we can translate empty tag to SQL\\nvalue? The Oracle should not to solve this question, but PostgreSQL does.\\nSome databases returns empty string.\\n\\nI prefer return a empty string - not null in this case. The reason is\\nsimple - Empty string is some information - and NULL is less information.\\nWhen it is necessary I can transform empty string to NULL - different\\ndirection is not unique.\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> --\\n> Craig Ringer http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, 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, 15 Sep 2016 13:31:43 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On 15 September 2016 at 19:31, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n\\n> b_expr enforces shift/reduce conflict :(\\n\\nNo problem then. I just thought it'd be worth allowing more if it\\nworked to do so.\\n\\n> I found other opened question - how we can translate empty tag to SQL value?\\n> The Oracle should not to solve this question, but PostgreSQL does. Some\\n> databases returns empty string.\\n\\nOracle doesn't solve the problem? it ERRORs?\\n\\n> I prefer return a empty string - not null in this case.\\n\\nI agree, and that's consistent with how most XML is interpreted. XSLT\\nfor example considers <x></x> and <x/> to be pretty much the same\\nthing.\\n\\n> The reason is simple\\n> - Empty string is some information - and NULL is less information. When it\\n> is necessary I can transform empty string to NULL - different direction is\\n> not unique.\\n\\nYep, I definitely agree. The only issue is if people want a DEFAULT to\\nbe applied for empty tags. But that's something they can do in a\\npost-process pass easily enough, since XMLTABLE is callable as a\\nsubquery / WITH expression / etc.\\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\": \"Fri, 16 Sep 2016 07:44:38 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On 9/12/16 1:14 AM, Craig Ringer wrote:\\n> I would've expected once per query. There's no way the expressions can\\n> reference the row data, so there's no reason to evaluate them each\\n> time.\\n> \\n> The only use case I see for evaluating them each time is - maybe -\\n> DEFAULT. Where maybe there's a use for nextval() or other volatile\\n> functions. But honestly, I think that's better done explicitly in a\\n> post-pass, i.e.\\n> \\n> select uuid_generate_v4(), x.*\\n> from (\\n> xmltable(.....) x\\n> );\\n> \\n> in cases where that's what the user actually wants.\\n> \\n> There's no other case I can think of where expressions as arguments to\\n> set-returning functions are evaluated once per output row.\\n\\nThe SQL standard appears to show what the behavior ought to be:\\n\\n<XML table> is equivalent to\\n\\nLATERAL ( XNDC\\nSELECT SLI1 AS CN1, SLI2 AS CN2, ..., SLINC AS CNNC FROM XMLITERATE (\\nXMLQUERY ( XTRP XQAL\\n RETURNING SEQUENCE BY REF EMPTY ON EMPTY ) )\\n AS I ( V, N )\\n) AS CORR DCLP\\n\\nand SLIj is\\n\\nCASE WHEN XEj\\nTHEN XMLCAST( XQCj AS DTj CPMj )\\nELSE DEFj END\\n\\nwhere DEFj is the default expression.\\n\\nSo simplified it is\\n\\nLATERAL ( SELECT CASE WHEN ... ELSE DEFj END, ... FROM something )\\n\\nwhich indicates that the default expression is evaluated for every row.\\n\\nIf we're not sure about all this, it might be worth restricting the\\ndefault expressions to stable or immutable expressions for the time being.\\n\\n-- \\nPeter Eisentraut 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, 15 Sep 2016 20:22:49 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-09-16 1:44 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:\\n\\n> On 15 September 2016 at 19:31, Pavel Stehule <pavel.stehule@gmail.com>\\n> wrote:\\n>\\n> > b_expr enforces shift/reduce conflict :(\\n>\\n> No problem then. I just thought it'd be worth allowing more if it\\n> worked to do so.\\n>\\n> > I found other opened question - how we can translate empty tag to SQL\\n> value?\\n> > The Oracle should not to solve this question, but PostgreSQL does. Some\\n> > databases returns empty string.\\n>\\n> Oracle doesn't solve the problem? it ERRORs?\\n>\\n\\nOracle returns NULL. But there are not any difference between NULL and\\nempty string\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> > I prefer return a empty string - not null in this case.\\n>\\n> I agree, and that's consistent with how most XML is interpreted. XSLT\\n> for example considers <x></x> and <x/> to be pretty much the same\\n> thing.\\n>\\n> > The reason is simple\\n> > - Empty string is some information - and NULL is less information. When\\n> it\\n> > is necessary I can transform empty string to NULL - different direction\\n> is\\n> > not unique.\\n>\\n> Yep, I definitely agree. The only issue is if people want a DEFAULT to\\n> be applied for empty tags. But that's something they can do in a\\n> post-process pass easily enough, since XMLTABLE is callable as a\\n> subquery / WITH expression / etc.\\n>\\n>\\n> --\\n> Craig Ringer http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Training & Services\\n>\\n\\n2016-09-16 1:44 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:On 15 September 2016 at 19:31, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n\\n> b_expr enforces shift/reduce conflict :(\\n\\nNo problem then. I just thought it'd be worth allowing more if it\\nworked to do so.\\n\\n> I found other opened question - how we can translate empty tag to SQL value?\\n> The Oracle should not to solve this question, but PostgreSQL does. Some\\n> databases returns empty string.\\n\\nOracle doesn't solve the problem? it ERRORs?Oracle returns NULL. But there are not any difference between NULL and empty stringRegardsPavel \\n\\n> I prefer return a empty string - not null in this case.\\n\\nI agree, and that's consistent with how most XML is interpreted. XSLT\\nfor example considers <x></x> and <x/> to be pretty much the same\\nthing.\\n\\n> The reason is simple\\n> - Empty string is some information - and NULL is less information. When it\\n> is necessary I can transform empty string to NULL - different direction is\\n> not unique.\\n\\nYep, I definitely agree. The only issue is if people want a DEFAULT to\\nbe applied for empty tags. But that's something they can do in a\\npost-process pass easily enough, since XMLTABLE is callable as a\\nsubquery / WITH expression / etc.\\n\\n\\n--\\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Fri, 16 Sep 2016 09:20:23 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\nnew update:\\n\\n* doc is moved to better place - xml processing functions\\n* few more regress tests\\n* call forgotten check_srf_call_placement\\n\\nRegards\\n\\nPavel\\n\\n\\n-- \\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, 18 Sep 2016 11:53:47 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-09-18 11:53 GMT+02:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n\\n> Hi\\n>\\n> new update:\\n>\\n> * doc is moved to better place - xml processing functions\\n> * few more regress tests\\n> * call forgotten check_srf_call_placement\\n>\\n\\nanother small update - fix XMLPath parser - support multibytes characters\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> Regards\\n>\\n> Pavel\\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, 21 Sep 2016 20:31:12 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On 22 September 2016 at 02:31, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n\\n> another small update - fix XMLPath parser - support multibytes characters\\n\\nI'm returning for another round of review.\\n\\nThe code doesn't handle the 5 XML built-in entities correctly in\\ntext-typed output. It processes ' and " but not &, < or\\n> . See added test. I have not fixed this, but I think it's clearly\\nbroken:\\n\\n\\n+ -- XML builtin entities\\n+ SELECT * FROM xmltable('/x/a' PASSING\\n'<x><a><ent>'</ent></a><a><ent>"</ent></a><a><ent>&</ent></a><a><ent><</ent></a><a><ent>></ent></a></x>'\\nCOLUMNS ent text);\\n+ ent\\n+ -------\\n+ '\\n+ \\\"\\n+ &\\n+ <\\n+ >\\n+ (5 rows)\\n\\nso I've adjusted the docs to claim that they're expanded. The code\\nneeds fixing to avoid entity-escaping when the output column type is\\nnot 'xml'.\\n\\n\\n' and " entities in xml-typed output are expanded, not\\npreserved. I don't know if this is intended but I suspect it is:\\n\\n+ SELECT * FROM xmltable('/x/a' PASSING\\n'<x><a><ent>'</ent></a><a><ent>"</ent></a><a><ent>&</ent></a><a><ent><</ent></a><a><ent>></ent></a></x>'\\nCOLUMNS ent xml);\\n+ ent\\n+ ------------------\\n+ <ent>'</ent>\\n+ <ent>\\\"</ent>\\n+ <ent>&</ent>\\n+ <ent><</ent>\\n+ <ent>></ent>\\n+ (5 rows)\\n\\n\\nFor the docs changes relevant to the above search for \\\"The five\\npredefined XML entities\\\". Adjust that bit of docs if I guessed wrong\\nabout the intended behaviour.\\n\\nThe tests don't cover CDATA or PCDATA . I didn't try to add that, but\\nthey should.\\n\\n\\nDid some docs copy-editing and integrated some examples. Explained how\\nnested elements work, that multiple top level elements is an error,\\netc. Explained the time-of-evaluation stuff. Pointed out that you can\\nrefer to prior output columns in PATH and DEFAULT, since that's weird\\nand unusual compared to normal SQL. Documented handling of multiple\\nnode matches, including the surprising results of somepath/text() on\\n<somepath>x<!--blah-->y</somepath>. Documented handling of nested\\nelements. Documented that xmltable works only on XML documents, not\\nfragments/forests.\\n\\nRegarding evaluation time, it struck me that evaluating path\\nexpressions once per row means the xpath must be parsed and processed\\nonce per row. Isn't it desirable to store and re-use the preparsed\\nxpath? I don't think this is a major problem, since we can later\\ndetect stable/immutable expressions including constants, evaluate only\\nonce in that case, and cache. It's just worth thinking about.\\n\\nThe docs and tests don't seem to cover XML entities. What's the\\nbehaviour there? Core XML only defines one entity, but if a schema\\ndefines more how are they processed? The tests need to cover the\\npredefined entities " & ' < and > at least.\\n\\nI have no idea whether the current code can fetch a DTD and use any\\n<!ENTITY > declarations to expand entities, but I'm guessing not? If\\nnot, external DTDs, and internal DTDs with external entities should be\\ndocumented as unsupported.\\n\\nIt doesn't seem to cope with internal DTDs at all (libxml2 limitation?):\\n\\nSELECT * FROM xmltable('/' PASSING $XML$<?xml version=\\\"1.0\\\" standalone=\\\"yes\\\" ?>\\n<!DOCTYPE foo [\\n <!ELEMENT foo (#PCDATA)>\\n <!ENTITY pg \\\"PostgreSQL\\\">\\n]>\\n<foo>Hello &pg;.</foo>\\n$XML$ COLUMNS foo text);\\n\\n+ ERROR: invalid XML content\\n+ LINE 1: SELECT * FROM xmltable('/' PASSING $XML$<?xml version=\\\"1.0\\\" ...\\n+ ^\\n+ DETAIL: line 2: StartTag: invalid element name\\n+ <!DOCTYPE foo [\\n+ ^\\n+ line 3: StartTag: invalid element name\\n+ <!ELEMENT foo (#PCDATA)>\\n+ ^\\n+ line 4: StartTag: invalid element name\\n+ <!ENTITY pg \\\"PostgreSQL\\\">\\n+ ^\\n+ line 6: Entity 'pg' not defined\\n+ <foo>Hello &pg;.</foo>\\n+ ^\\n\\n\\nlibxml seems to support documents with internal DTDs:\\n\\n$ xmllint --valid /tmp/x\\n<?xml version=\\\"1.0\\\" standalone=\\\"yes\\\"?>\\n<!DOCTYPE foo [\\n<!ELEMENT foo (#PCDATA)>\\n<!ENTITY pg \\\"PostgreSQL\\\">\\n]>\\n<foo>Hello &pg;.</foo>\\n\\n\\nso presumably the issue lies in the xpath stuff? Note that it's not\\neven ignoring the DTD and choking on the undefined entity, it's\\nchoking on the DTD its self.\\n\\n\\nOK, code comments:\\n\\n\\nIn +ExecEvalTableExpr, shouldn't you be using PG_ENSURE_ERROR_CLEANUP\\ninstead of a PG_TRY() / PG_CATCH() block?\\n\\n\\nI think the new way you handle the type stuff is much, much better,\\nand with comments to explain too. Thanks very much.\\n\\n\\nThere's an oversight in tableexpr vs xmltable separation here:\\n\\n+ case T_TableExpr:\\n+ *name = \\\"xmltable\\\";\\n+ return 2;\\n\\npresumably you need to look at the node and decide what kind of table\\nexpression it is or just use a generic \\\"tableexpr\\\".\\n\\nSame problem here:\\n\\n+ case T_TableExpr:\\n+ {\\n+ TableExpr *te = (TableExpr *) node;\\n+\\n+ /* c_expr shoud be closed in brackets */\\n+ appendStringInfoString(buf, \\\"XMLTABLE(\\\");\\n\\n\\n\\nI don't have the libxml knowledge or remaining brain to usefully\\nevaluate the xpath and xml specifics in xpath.c today. It does strike\\nme that the new xpath parser should probably live in its own file,\\nthough.\\n\\nI think this is all a big improvement. Barring the notes above and my\\nlack of review of the guts of the xml.c parts of it, I'm pretty happy\\nwith what I see now.\\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\": \"Fri, 23 Sep 2016 16:05:18 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"> Did some docs copy-editing and integrated some examples.\\n\\nWhoops, forgot to attach.\\n\\nRather than sending a whole new copy of the patch, here's a diff\\nagainst your patched tree of my changes so you can see what I've done\\nand apply the parts you want.\\n\\nNote that I didn't updated the expected files.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, 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\": \"Fri, 23 Sep 2016 16:07:24 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-09-23 10:05 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:\\n\\n> On 22 September 2016 at 02:31, Pavel Stehule <pavel.stehule@gmail.com>\\n> wrote:\\n>\\n> > another small update - fix XMLPath parser - support multibytes characters\\n>\\n> I'm returning for another round of review.\\n>\\n> The code doesn't handle the 5 XML built-in entities correctly in\\n> text-typed output. It processes ' and " but not &, < or\\n> > . See added test. I have not fixed this, but I think it's clearly\\n> broken:\\n>\\n>\\n> + -- XML builtin entities\\n> + SELECT * FROM xmltable('/x/a' PASSING\\n> '<x><a><ent>'</ent></a><a><ent>"</ent></a><a><\\n> ent>&</ent></a><a><ent><</ent></a><a><ent>></ent></a></x>'\\n> COLUMNS ent text);\\n> + ent\\n> + -------\\n> + '\\n> + \\\"\\n> + &\\n> + <\\n> + >\\n> + (5 rows)\\n>\\n> so I've adjusted the docs to claim that they're expanded. The code\\n> needs fixing to avoid entity-escaping when the output column type is\\n> not 'xml'.\\n>\\n>\\n> ' and " entities in xml-typed output are expanded, not\\n> preserved. I don't know if this is intended but I suspect it is:\\n>\\n> + SELECT * FROM xmltable('/x/a' PASSING\\n> '<x><a><ent>'</ent></a><a><ent>"</ent></a><a><\\n> ent>&</ent></a><a><ent><</ent></a><a><ent>></ent></a></x>'\\n> COLUMNS ent xml);\\n> + ent\\n> + ------------------\\n> + <ent>'</ent>\\n> + <ent>\\\"</ent>\\n> + <ent>&</ent>\\n> + <ent><</ent>\\n> + <ent>></ent>\\n> + (5 rows)\\n>\\n>\\n> For the docs changes relevant to the above search for \\\"The five\\n> predefined XML entities\\\". Adjust that bit of docs if I guessed wrong\\n> about the intended behaviour.\\n>\\n> The tests don't cover CDATA or PCDATA . I didn't try to add that, but\\n> they should.\\n>\\n>\\n> Did some docs copy-editing and integrated some examples. Explained how\\n> nested elements work, that multiple top level elements is an error,\\n> etc. Explained the time-of-evaluation stuff. Pointed out that you can\\n> refer to prior output columns in PATH and DEFAULT, since that's weird\\n> and unusual compared to normal SQL. Documented handling of multiple\\n> node matches, including the surprising results of somepath/text() on\\n> <somepath>x<!--blah-->y</somepath>. Documented handling of nested\\n> elements. Documented that xmltable works only on XML documents, not\\n> fragments/forests.\\n>\\n> Regarding evaluation time, it struck me that evaluating path\\n> expressions once per row means the xpath must be parsed and processed\\n> once per row. Isn't it desirable to store and re-use the preparsed\\n> xpath? I don't think this is a major problem, since we can later\\n> detect stable/immutable expressions including constants, evaluate only\\n> once in that case, and cache. It's just worth thinking about.\\n>\\n> The docs and tests don't seem to cover XML entities. What's the\\n> behaviour there? Core XML only defines one entity, but if a schema\\n> defines more how are they processed? The tests need to cover the\\n> predefined entities " & ' < and > at least.\\n>\\n> I have no idea whether the current code can fetch a DTD and use any\\n> <!ENTITY > declarations to expand entities, but I'm guessing not? If\\n> not, external DTDs, and internal DTDs with external entities should be\\n> documented as unsupported.\\n>\\n> It doesn't seem to cope with internal DTDs at all (libxml2 limitation?):\\n>\\n> SELECT * FROM xmltable('/' PASSING $XML$<?xml version=\\\"1.0\\\"\\n> standalone=\\\"yes\\\" ?>\\n> <!DOCTYPE foo [\\n> <!ELEMENT foo (#PCDATA)>\\n> <!ENTITY pg \\\"PostgreSQL\\\">\\n> ]>\\n> <foo>Hello &pg;.</foo>\\n> $XML$ COLUMNS foo text);\\n>\\n> + ERROR: invalid XML content\\n> + LINE 1: SELECT * FROM xmltable('/' PASSING $XML$<?xml version=\\\"1.0\\\" ...\\n> + ^\\n> + DETAIL: line 2: StartTag: invalid element name\\n> + <!DOCTYPE foo [\\n> + ^\\n> + line 3: StartTag: invalid element name\\n> + <!ELEMENT foo (#PCDATA)>\\n> + ^\\n> + line 4: StartTag: invalid element name\\n> + <!ENTITY pg \\\"PostgreSQL\\\">\\n> + ^\\n> + line 6: Entity 'pg' not defined\\n> + <foo>Hello &pg;.</foo>\\n> + ^\\n>\\n>\\n> libxml seems to support documents with internal DTDs:\\n>\\n> $ xmllint --valid /tmp/x\\n> <?xml version=\\\"1.0\\\" standalone=\\\"yes\\\"?>\\n> <!DOCTYPE foo [\\n> <!ELEMENT foo (#PCDATA)>\\n> <!ENTITY pg \\\"PostgreSQL\\\">\\n> ]>\\n> <foo>Hello &pg;.</foo>\\n>\\n>\\n> so presumably the issue lies in the xpath stuff? Note that it's not\\n> even ignoring the DTD and choking on the undefined entity, it's\\n> choking on the DTD its self.\\n>\\n>\\n> OK, code comments:\\n>\\n>\\n> In +ExecEvalTableExpr, shouldn't you be using PG_ENSURE_ERROR_CLEANUP\\n> instead of a PG_TRY() / PG_CATCH() block?\\n>\\n>\\n> I think the new way you handle the type stuff is much, much better,\\n> and with comments to explain too. Thanks very much.\\n>\\n>\\n> There's an oversight in tableexpr vs xmltable separation here:\\n>\\n> + case T_TableExpr:\\n> + *name = \\\"xmltable\\\";\\n> + return 2;\\n>\\n> presumably you need to look at the node and decide what kind of table\\n> expression it is or just use a generic \\\"tableexpr\\\".\\n>\\n> Same problem here:\\n>\\n> + case T_TableExpr:\\n> + {\\n> + TableExpr *te = (TableExpr *) node;\\n> +\\n> + /* c_expr shoud be closed in brackets */\\n> + appendStringInfoString(buf, \\\"XMLTABLE(\\\");\\n>\\n>\\nThis is correct, but not well commented - looks on XMLEXPR node - TableExpr\\nis a holder, but it is invisible for user. User running a XMLTABLE function\\nand should to see XMLTABLE. It will be more clean when we will support\\nJSON_TABLE function.\\n\\n\\n>\\n>\\n> I don't have the libxml knowledge or remaining brain to usefully\\n> evaluate the xpath and xml specifics in xpath.c today. It does strike\\n> me that the new xpath parser should probably live in its own file,\\n> though.\\n>\\n\\nI'll try move it to separate file\\n\\n\\n>\\n> I think this is all a big improvement. Barring the notes above and my\\n> lack of review of the guts of the xml.c parts of it, I'm pretty happy\\n> with what I see now.\\n>\\n\\nThank you. I hope so all major issues are solved. Probably some XML\\nspecific related issues are there - but I am happy, so you have well XML\\nknowledge and you will test a corner cases.\\n\\nRegards\\n\\nPavel\\n\\n\\n\\n>\\n>\\n> --\\n> Craig Ringer http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Training & Services\\n>\\n\\n2016-09-23 10:05 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:On 22 September 2016 at 02:31, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n\\n> another small update - fix XMLPath parser - support multibytes characters\\n\\nI'm returning for another round of review.\\n\\nThe code doesn't handle the 5 XML built-in entities correctly in\\ntext-typed output. It processes ' and " but not &, < or\\n> . See added test. I have not fixed this, but I think it's clearly\\nbroken:\\n\\n\\n+ -- XML builtin entities\\n+ SELECT * FROM xmltable('/x/a' PASSING\\n'<x><a><ent>'</ent></a><a><ent>"</ent></a><a><ent>&</ent></a><a><ent><</ent></a><a><ent>></ent></a></x>'\\nCOLUMNS ent text);\\n+ ent\\n+ -------\\n+ '\\n+ \\\"\\n+ &\\n+ <\\n+ >\\n+ (5 rows)\\n\\nso I've adjusted the docs to claim that they're expanded. The code\\nneeds fixing to avoid entity-escaping when the output column type is\\nnot 'xml'.\\n\\n\\n' and " entities in xml-typed output are expanded, not\\npreserved. I don't know if this is intended but I suspect it is:\\n\\n+ SELECT * FROM xmltable('/x/a' PASSING\\n'<x><a><ent>'</ent></a><a><ent>"</ent></a><a><ent>&</ent></a><a><ent><</ent></a><a><ent>></ent></a></x>'\\nCOLUMNS ent xml);\\n+ ent\\n+ ------------------\\n+ <ent>'</ent>\\n+ <ent>\\\"</ent>\\n+ <ent>&</ent>\\n+ <ent><</ent>\\n+ <ent>></ent>\\n+ (5 rows)\\n\\n\\nFor the docs changes relevant to the above search for \\\"The five\\npredefined XML entities\\\". Adjust that bit of docs if I guessed wrong\\nabout the intended behaviour.\\n\\nThe tests don't cover CDATA or PCDATA . I didn't try to add that, but\\nthey should.\\n\\n\\nDid some docs copy-editing and integrated some examples. Explained how\\nnested elements work, that multiple top level elements is an error,\\netc. Explained the time-of-evaluation stuff. Pointed out that you can\\nrefer to prior output columns in PATH and DEFAULT, since that's weird\\nand unusual compared to normal SQL. Documented handling of multiple\\nnode matches, including the surprising results of somepath/text() on\\n<somepath>x<!--blah-->y</somepath>. Documented handling of nested\\nelements. Documented that xmltable works only on XML documents, not\\nfragments/forests.\\n\\nRegarding evaluation time, it struck me that evaluating path\\nexpressions once per row means the xpath must be parsed and processed\\nonce per row. Isn't it desirable to store and re-use the preparsed\\nxpath? I don't think this is a major problem, since we can later\\ndetect stable/immutable expressions including constants, evaluate only\\nonce in that case, and cache. It's just worth thinking about.\\n\\nThe docs and tests don't seem to cover XML entities. What's the\\nbehaviour there? Core XML only defines one entity, but if a schema\\ndefines more how are they processed? The tests need to cover the\\npredefined entities " & ' < and > at least.\\n\\nI have no idea whether the current code can fetch a DTD and use any\\n<!ENTITY > declarations to expand entities, but I'm guessing not? If\\nnot, external DTDs, and internal DTDs with external entities should be\\ndocumented as unsupported.\\n\\nIt doesn't seem to cope with internal DTDs at all (libxml2 limitation?):\\n\\nSELECT * FROM xmltable('/' PASSING $XML$<?xml version=\\\"1.0\\\" standalone=\\\"yes\\\" ?>\\n<!DOCTYPE foo [\\n <!ELEMENT foo (#PCDATA)>\\n <!ENTITY pg \\\"PostgreSQL\\\">\\n]>\\n<foo>Hello &pg;.</foo>\\n$XML$ COLUMNS foo text);\\n\\n+ ERROR: invalid XML content\\n+ LINE 1: SELECT * FROM xmltable('/' PASSING $XML$<?xml version=\\\"1.0\\\" ...\\n+ ^\\n+ DETAIL: line 2: StartTag: invalid element name\\n+ <!DOCTYPE foo [\\n+ ^\\n+ line 3: StartTag: invalid element name\\n+ <!ELEMENT foo (#PCDATA)>\\n+ ^\\n+ line 4: StartTag: invalid element name\\n+ <!ENTITY pg \\\"PostgreSQL\\\">\\n+ ^\\n+ line 6: Entity 'pg' not defined\\n+ <foo>Hello &pg;.</foo>\\n+ ^\\n\\n\\nlibxml seems to support documents with internal DTDs:\\n\\n$ xmllint --valid /tmp/x\\n<?xml version=\\\"1.0\\\" standalone=\\\"yes\\\"?>\\n<!DOCTYPE foo [\\n<!ELEMENT foo (#PCDATA)>\\n<!ENTITY pg \\\"PostgreSQL\\\">\\n]>\\n<foo>Hello &pg;.</foo>\\n\\n\\nso presumably the issue lies in the xpath stuff? Note that it's not\\neven ignoring the DTD and choking on the undefined entity, it's\\nchoking on the DTD its self.\\n\\n\\nOK, code comments:\\n\\n\\nIn +ExecEvalTableExpr, shouldn't you be using PG_ENSURE_ERROR_CLEANUP\\ninstead of a PG_TRY() / PG_CATCH() block?\\n\\n\\nI think the new way you handle the type stuff is much, much better,\\nand with comments to explain too. Thanks very much.\\n\\n\\nThere's an oversight in tableexpr vs xmltable separation here:\\n\\n+ case T_TableExpr:\\n+ *name = \\\"xmltable\\\";\\n+ return 2;\\n\\npresumably you need to look at the node and decide what kind of table\\nexpression it is or just use a generic \\\"tableexpr\\\".\\n\\nSame problem here:\\n\\n+ case T_TableExpr:\\n+ {\\n+ TableExpr *te = (TableExpr *) node;\\n+\\n+ /* c_expr shoud be closed in brackets */\\n+ appendStringInfoString(buf, \\\"XMLTABLE(\\\");\\nThis is correct, but not well commented - looks on XMLEXPR node - TableExpr is a holder, but it is invisible for user. User running a XMLTABLE function and should to see XMLTABLE. It will be more clean when we will support JSON_TABLE function. \\n\\n\\nI don't have the libxml knowledge or remaining brain to usefully\\nevaluate the xpath and xml specifics in xpath.c today. It does strike\\nme that the new xpath parser should probably live in its own file,\\nthough.I'll try move it to separate file \\n\\nI think this is all a big improvement. Barring the notes above and my\\nlack of review of the guts of the xml.c parts of it, I'm pretty happy\\nwith what I see now.Thank you. I hope so all major issues are solved. Probably some XML specific related issues are there - but I am happy, so you have well XML knowledge and you will test a corner cases.RegardsPavel \\n\\n\\n--\\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Fri, 23 Sep 2016 10:29:18 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2016-09-23 10:07 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:\\n\\n> > Did some docs copy-editing and integrated some examples.\\n>\\n> Whoops, forgot to attach.\\n>\\n> Rather than sending a whole new copy of the patch, here's a diff\\n> against your patched tree of my changes so you can see what I've done\\n> and apply the parts you want.\\n>\\n> Note that I didn't updated the expected files.\\n>\\n\\nI applied your patch - there is small misunderstanding. The PATH is\\nevaluated once for input row already. It is not clean in code, because it\\nis executor node started and running for all rows. I changed it in your\\npart of doc.\\n\\n to a simple value before calling the function. <literal>PATH</>\\n+ expressions are normally evaluated <emphasis>exactly once per result\\nrow ## per input row\\n+ </emphasis>,\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> --\\n> Craig Ringer http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Training & Services\\n>\\n\\nHi2016-09-23 10:07 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:> Did some docs copy-editing and integrated some examples.\\n\\nWhoops, forgot to attach.\\n\\nRather than sending a whole new copy of the patch, here's a diff\\nagainst your patched tree of my changes so you can see what I've done\\nand apply the parts you want.\\n\\nNote that I didn't updated the expected files.I applied your patch - there is small misunderstanding. The PATH is evaluated once for input row already. It is not clean in code, because it is executor node started and running for all rows. I changed it in your part of doc. to a simple value before calling the function. <literal>PATH</>+ expressions are normally evaluated <emphasis>exactly once per result row ## per input row+ </emphasis>,RegardsPavel \\n\\n--\\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Fri, 23 Sep 2016 22:22:02 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2016-09-23 10:05 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:\\n\\n> On 22 September 2016 at 02:31, Pavel Stehule <pavel.stehule@gmail.com>\\n> wrote:\\n>\\n> > another small update - fix XMLPath parser - support multibytes characters\\n>\\n> I'm returning for another round of review.\\n>\\n> The code doesn't handle the 5 XML built-in entities correctly in\\n> text-typed output. It processes ' and " but not &, < or\\n> > . See added test. I have not fixed this, but I think it's clearly\\n> broken:\\n>\\n>\\n> + -- XML builtin entities\\n> + SELECT * FROM xmltable('/x/a' PASSING\\n> '<x><a><ent>'</ent></a><a><ent>"</ent></a><a><\\n> ent>&</ent></a><a><ent><</ent></a><a><ent>></ent></a></x>'\\n> COLUMNS ent text);\\n> + ent\\n> + -------\\n> + '\\n> + \\\"\\n> + &\\n> + <\\n> + >\\n> + (5 rows)\\n>\\n> so I've adjusted the docs to claim that they're expanded. The code\\n> needs fixing to avoid entity-escaping when the output column type is\\n> not 'xml'.\\n>\\n>\\nfixed\\n\\n\\n>\\n> ' and " entities in xml-typed output are expanded, not\\n> preserved. I don't know if this is intended but I suspect it is:\\n>\\n> + SELECT * FROM xmltable('/x/a' PASSING\\n> '<x><a><ent>'</ent></a><a><ent>"</ent></a><a><\\n> ent>&</ent></a><a><ent><</ent></a><a><ent>></ent></a></x>'\\n> COLUMNS ent xml);\\n> + ent\\n> + ------------------\\n> + <ent>'</ent>\\n> + <ent>\\\"</ent>\\n> + <ent>&</ent>\\n> + <ent><</ent>\\n> + <ent>></ent>\\n> + (5 rows)\\n>\\n>\\n> For the docs changes relevant to the above search for \\\"The five\\n> predefined XML entities\\\". Adjust that bit of docs if I guessed wrong\\n> about the intended behaviour.\\n>\\n> The tests don't cover CDATA or PCDATA . I didn't try to add that, but\\n> they should.\\n>\\n>\\nappended\\n\\n\\n> Did some docs copy-editing and integrated some examples. Explained how\\n> nested elements work, that multiple top level elements is an error,\\n> etc. Explained the time-of-evaluation stuff. Pointed out that you can\\n> refer to prior output columns in PATH and DEFAULT, since that's weird\\n> and unusual compared to normal SQL. Documented handling of multiple\\n> node matches, including the surprising results of somepath/text() on\\n> <somepath>x<!--blah-->y</somepath>. Documented handling of nested\\n> elements. Documented that xmltable works only on XML documents, not\\n> fragments/forests.\\n>\\n\\nI don't understand to this sentence: \\\"It is possible for a PATH expression\\nto reference output columns that appear before it in the column-list, so\\npaths may be dynamically constructed based on other parts of the XML\\ndocument:\\\"\\n\\n\\n> Regarding evaluation time, it struck me that evaluating path\\n> expressions once per row means the xpath must be parsed and processed\\n> once per row. Isn't it desirable to store and re-use the preparsed\\n> xpath? I don't think this is a major problem, since we can later\\n> detect stable/immutable expressions including constants, evaluate only\\n> once in that case, and cache. It's just worth thinking about.\\n>\\n\\nProbably it is possible - it is exactly how you wrote - it needs to check\\nthe change. We can try do some possible performance optimizations later -\\nwithout compatibility issues. Now, I prefer the most simple code.\\n\\na note: PATH expression is evaluated for any **input** row. In same moment\\nis processed row path expression and man XML document DOM parsing. So\\noverhead of PATH expression and PATH parsing should not be dominant.\\n\\n\\n>\\n> The docs and tests don't seem to cover XML entities. What's the\\n> behaviour there? Core XML only defines one entity, but if a schema\\n> defines more how are they processed? The tests need to cover the\\n> predefined entities " & ' < and > at least.\\n>\\n\\nI don't understand, what you are propose here. ?? Please, can you send some\\nexamples.\\n\\n>\\n> I have no idea whether the current code can fetch a DTD and use any\\n> <!ENTITY > declarations to expand entities, but I'm guessing not? If\\n> not, external DTDs, and internal DTDs with external entities should be\\n> documented as unsupported.\\n>\\n> It doesn't seem to cope with internal DTDs at all (libxml2 limitation?):\\n>\\n> SELECT * FROM xmltable('/' PASSING $XML$<?xml version=\\\"1.0\\\"\\n> standalone=\\\"yes\\\" ?>\\n> <!DOCTYPE foo [\\n> <!ELEMENT foo (#PCDATA)>\\n> <!ENTITY pg \\\"PostgreSQL\\\">\\n> ]>\\n> <foo>Hello &pg;.</foo>\\n> $XML$ COLUMNS foo text);\\n>\\n> + ERROR: invalid XML content\\n> + LINE 1: SELECT * FROM xmltable('/' PASSING $XML$<?xml version=\\\"1.0\\\" ...\\n> + ^\\n> + DETAIL: line 2: StartTag: invalid element name\\n> + <!DOCTYPE foo [\\n> + ^\\n> + line 3: StartTag: invalid element name\\n> + <!ELEMENT foo (#PCDATA)>\\n> + ^\\n> + line 4: StartTag: invalid element name\\n> + <!ENTITY pg \\\"PostgreSQL\\\">\\n> + ^\\n> + line 6: Entity 'pg' not defined\\n> + <foo>Hello &pg;.</foo>\\n> + ^\\n>\\n>\\nIt is rejected before XMLTABLE function call\\n\\npostgres=# select $XML$<?xml version=\\\"1.0\\\" standalone=\\\"yes\\\" ?>\\npostgres$# <!DOCTYPE foo [\\npostgres$# <!ELEMENT foo (#PCDATA)>\\npostgres$# <!ENTITY pg \\\"PostgreSQL\\\">\\npostgres$# ]>\\npostgres$# <foo>Hello &pg;.</foo>\\npostgres$# $XML$::xml;\\nERROR: invalid XML content\\nLINE 1: select $XML$<?xml version=\\\"1.0\\\" standalone=\\\"yes\\\" ?>\\n ^\\nDETAIL: line 2: StartTag: invalid element name\\n<!DOCTYPE foo [\\n ^\\nline 3: StartTag: invalid element name\\n <!ELEMENT foo (#PCDATA)>\\n ^\\nline 4: StartTag: invalid element name\\n <!ENTITY pg \\\"PostgreSQL\\\">\\n ^\\nline 6: Entity 'pg' not defined\\n<foo>Hello &pg;.</foo>\\n ^\\nIt is disabled by default in libxml2. I found a function\\nxmlSubstituteEntitiesDefault http://www.xmlsoft.org/entities.html\\nhttp://www.xmlsoft.org/html/libxml-parser.html#xmlSubstituteEntitiesDefault\\n\\nThe default behave should be common for all PostgreSQL's libxml2 based\\nfunction - and then it is different topic - maybe part for PostgreSQL ToDo?\\nBut I don't remember any user requests related to this issue.\\n\\n\\n\\n>\\n> libxml seems to support documents with internal DTDs:\\n>\\n> $ xmllint --valid /tmp/x\\n> <?xml version=\\\"1.0\\\" standalone=\\\"yes\\\"?>\\n> <!DOCTYPE foo [\\n> <!ELEMENT foo (#PCDATA)>\\n> <!ENTITY pg \\\"PostgreSQL\\\">\\n> ]>\\n> <foo>Hello &pg;.</foo>\\n>\\n\\n\\nI removed this tests - it is not related to XMLTABLE function, but to\\ngeneric XML processing/validation.\\n\\n\\n>\\n>\\n> so presumably the issue lies in the xpath stuff? Note that it's not\\n> even ignoring the DTD and choking on the undefined entity, it's\\n> choking on the DTD its self.\\n>\\n>\\n> OK, code comments:\\n>\\n>\\n> In +ExecEvalTableExpr, shouldn't you be using PG_ENSURE_ERROR_CLEANUP\\n> instead of a PG_TRY() / PG_CATCH() block?\\n>\\n\\nIf I understand to doc, the PG_ENSURE_ERROR_CLEANUP should be used, when\\nyou want to catch FATAL errors (and when you want to clean shared memory).\\nXMLTABLE doesn't use shared memory, and doesn't need to catch fatal errors.\\n\\n\\n>\\n> I think the new way you handle the type stuff is much, much better,\\n> and with comments to explain too. Thanks very much.\\n>\\n>\\n> There's an oversight in tableexpr vs xmltable separation here:\\n>\\n> + case T_TableExpr:\\n> + *name = \\\"xmltable\\\";\\n> + return 2;\\n>\\n> presumably you need to look at the node and decide what kind of table\\n> expression it is or just use a generic \\\"tableexpr\\\".\\n>\\n> Same problem here:\\n>\\n> + case T_TableExpr:\\n> + {\\n> + TableExpr *te = (TableExpr *) node;\\n> +\\n> + /* c_expr shoud be closed in brackets */\\n> + appendStringInfoString(buf, \\\"XMLTABLE(\\\");\\n>\\n>\\ncommented\\n\\n\\n>\\n>\\n> I don't have the libxml knowledge or remaining brain to usefully\\n> evaluate the xpath and xml specifics in xpath.c today. It does strike\\n> me that the new xpath parser should probably live in its own file,\\n> though.\\n>\\n\\nmoved\\n\\n\\n>\\n> I think this is all a big improvement. Barring the notes above and my\\n> lack of review of the guts of the xml.c parts of it, I'm pretty happy\\n> with what I see now.\\n>\\n\\nnew version is attached\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n>\\n> --\\n> Craig Ringer http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Training & Services\\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, 24 Sep 2016 08:01:33 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On 24 September 2016 at 14:01, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n\\n>> Did some docs copy-editing and integrated some examples. Explained how\\n>> nested elements work, that multiple top level elements is an error,\\n>> etc. Explained the time-of-evaluation stuff. Pointed out that you can\\n>> refer to prior output columns in PATH and DEFAULT, since that's weird\\n>> and unusual compared to normal SQL. Documented handling of multiple\\n>> node matches, including the surprising results of somepath/text() on\\n>> <somepath>x<!--blah-->y</somepath>. Documented handling of nested\\n>> elements. Documented that xmltable works only on XML documents, not\\n>> fragments/forests.\\n>\\n>\\n> I don't understand to this sentence: \\\"It is possible for a PATH expression\\n> to reference output columns that appear before it in the column-list, so\\n> paths may be dynamically constructed based on other parts of the XML\\n> document:\\\"\\n\\n\\n\\n>> The docs and tests don't seem to cover XML entities. What's the\\n>> behaviour there? Core XML only defines one entity, but if a schema\\n>> defines more how are they processed? The tests need to cover the\\n>> predefined entities " & ' < and > at least.\\n>\\n>\\n> I don't understand, what you are propose here. ?? Please, can you send some\\n> examples.\\n\\nPer below - handling of DTD <!ENTITY> declarations, and the builtin\\nentity tests I already added tests for.\\n\\n\\n>> It doesn't seem to cope with internal DTDs at all (libxml2 limitation?):\\n>>\\n>> SELECT * FROM xmltable('/' PASSING $XML$<?xml version=\\\"1.0\\\"\\n>> standalone=\\\"yes\\\" ?>\\n>> <!DOCTYPE foo [\\n>> <!ELEMENT foo (#PCDATA)>\\n>> <!ENTITY pg \\\"PostgreSQL\\\">\\n>> ]>\\n>> <foo>Hello &pg;.</foo>\\n>> $XML$ COLUMNS foo text);\\n>>\\n>> + ERROR: invalid XML content\\n>> + LINE 1: SELECT * FROM xmltable('/' PASSING $XML$<?xml version=\\\"1.0\\\" ...\\n>> + ^\\n>> + DETAIL: line 2: StartTag: invalid element name\\n>> + <!DOCTYPE foo [\\n>> + ^\\n>> + line 3: StartTag: invalid element name\\n>> + <!ELEMENT foo (#PCDATA)>\\n>> + ^\\n>> + line 4: StartTag: invalid element name\\n>> + <!ENTITY pg \\\"PostgreSQL\\\">\\n>> + ^\\n>> + line 6: Entity 'pg' not defined\\n>> + <foo>Hello &pg;.</foo>\\n>> + ^\\n>>\\n>\\n> It is rejected before XMLTABLE function call\\n>\\n> postgres=# select $XML$<?xml version=\\\"1.0\\\" standalone=\\\"yes\\\" ?>\\n> postgres$# <!DOCTYPE foo [\\n> postgres$# <!ELEMENT foo (#PCDATA)>\\n> postgres$# <!ENTITY pg \\\"PostgreSQL\\\">\\n> postgres$# ]>\\n> postgres$# <foo>Hello &pg;.</foo>\\n> postgres$# $XML$::xml;\\n> ERROR: invalid XML content\\n> LINE 1: select $XML$<?xml version=\\\"1.0\\\" standalone=\\\"yes\\\" ?>\\n> ^\\n> DETAIL: line 2: StartTag: invalid element name\\n> <!DOCTYPE foo [\\n[snip]\\n\\n> It is disabled by default in libxml2. I found a function\\n> xmlSubstituteEntitiesDefault http://www.xmlsoft.org/entities.html\\n> http://www.xmlsoft.org/html/libxml-parser.html#xmlSubstituteEntitiesDefault\\n>\\n> The default behave should be common for all PostgreSQL's libxml2 based\\n> function - and then it is different topic - maybe part for PostgreSQL ToDo?\\n> But I don't remember any user requests related to this issue.\\n\\n\\nOK, so it's not xmltable specific. Fine by me.\\n\\nSomebody who cares can deal with it. There's clearly nobody breaking\\ndown the walls wanting the feature.\\n\\n> I removed this tests - it is not related to XMLTABLE function, but to\\n> generic XML processing/validation.\\n\\n\\nGood plan.\\n\\n>> In +ExecEvalTableExpr, shouldn't you be using PG_ENSURE_ERROR_CLEANUP\\n>> instead of a PG_TRY() / PG_CATCH() block?\\n>\\n>\\n> If I understand to doc, the PG_ENSURE_ERROR_CLEANUP should be used, when you\\n> want to catch FATAL errors (and when you want to clean shared memory).\\n> XMLTABLE doesn't use shared memory, and doesn't need to catch fatal errors.\\n\\nOk, makes sense.\\n\\n\\n>> I don't have the libxml knowledge or remaining brain to usefully\\n>> evaluate the xpath and xml specifics in xpath.c today. It does strike\\n>> me that the new xpath parser should probably live in its own file,\\n>> though.\\n>\\n> moved\\n\\nThanks.\\n\\n\\n> new version is attached\\n\\n\\nGreat.\\n\\nI'm marking this ready for committer at this point.\\n\\nI think the XML parser likely needs a more close reading, so I'll ping\\nPeter E to see if he'll have a chance to check that bit out. But by\\nand large I think the issues have been ironed out - in terms of\\nfunctionality, structure and clarity I think it's looking solid.\\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\": \"Tue, 27 Sep 2016 09:34:47 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-09-27 3:34 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:\\n\\n> On 24 September 2016 at 14:01, Pavel Stehule <pavel.stehule@gmail.com>\\n> wrote:\\n>\\n> >> Did some docs copy-editing and integrated some examples. Explained how\\n> >> nested elements work, that multiple top level elements is an error,\\n> >> etc. Explained the time-of-evaluation stuff. Pointed out that you can\\n> >> refer to prior output columns in PATH and DEFAULT, since that's weird\\n> >> and unusual compared to normal SQL. Documented handling of multiple\\n> >> node matches, including the surprising results of somepath/text() on\\n> >> <somepath>x<!--blah-->y</somepath>. Documented handling of nested\\n> >> elements. Documented that xmltable works only on XML documents, not\\n> >> fragments/forests.\\n> >\\n> >\\n> > I don't understand to this sentence: \\\"It is possible for a PATH\\n> expression\\n> > to reference output columns that appear before it in the column-list, so\\n> > paths may be dynamically constructed based on other parts of the XML\\n> > document:\\\"\\n>\\n>\\n>\\n> >> The docs and tests don't seem to cover XML entities. What's the\\n> >> behaviour there? Core XML only defines one entity, but if a schema\\n> >> defines more how are they processed? The tests need to cover the\\n> >> predefined entities " & ' < and > at least.\\n> >\\n> >\\n> > I don't understand, what you are propose here. ?? Please, can you send\\n> some\\n> > examples.\\n>\\n> Per below - handling of DTD <!ENTITY> declarations, and the builtin\\n> entity tests I already added tests for.\\n>\\n>\\n> >> It doesn't seem to cope with internal DTDs at all (libxml2 limitation?):\\n> >>\\n> >> SELECT * FROM xmltable('/' PASSING $XML$<?xml version=\\\"1.0\\\"\\n> >> standalone=\\\"yes\\\" ?>\\n> >> <!DOCTYPE foo [\\n> >> <!ELEMENT foo (#PCDATA)>\\n> >> <!ENTITY pg \\\"PostgreSQL\\\">\\n> >> ]>\\n> >> <foo>Hello &pg;.</foo>\\n> >> $XML$ COLUMNS foo text);\\n> >>\\n> >> + ERROR: invalid XML content\\n> >> + LINE 1: SELECT * FROM xmltable('/' PASSING $XML$<?xml version=\\\"1.0\\\"\\n> ...\\n> >> + ^\\n> >> + DETAIL: line 2: StartTag: invalid element name\\n> >> + <!DOCTYPE foo [\\n> >> + ^\\n> >> + line 3: StartTag: invalid element name\\n> >> + <!ELEMENT foo (#PCDATA)>\\n> >> + ^\\n> >> + line 4: StartTag: invalid element name\\n> >> + <!ENTITY pg \\\"PostgreSQL\\\">\\n> >> + ^\\n> >> + line 6: Entity 'pg' not defined\\n> >> + <foo>Hello &pg;.</foo>\\n> >> + ^\\n> >>\\n> >\\n> > It is rejected before XMLTABLE function call\\n> >\\n> > postgres=# select $XML$<?xml version=\\\"1.0\\\" standalone=\\\"yes\\\" ?>\\n> > postgres$# <!DOCTYPE foo [\\n> > postgres$# <!ELEMENT foo (#PCDATA)>\\n> > postgres$# <!ENTITY pg \\\"PostgreSQL\\\">\\n> > postgres$# ]>\\n> > postgres$# <foo>Hello &pg;.</foo>\\n> > postgres$# $XML$::xml;\\n> > ERROR: invalid XML content\\n> > LINE 1: select $XML$<?xml version=\\\"1.0\\\" standalone=\\\"yes\\\" ?>\\n> > ^\\n> > DETAIL: line 2: StartTag: invalid element name\\n> > <!DOCTYPE foo [\\n> [snip]\\n>\\n> > It is disabled by default in libxml2. I found a function\\n> > xmlSubstituteEntitiesDefault http://www.xmlsoft.org/entities.html\\n> > http://www.xmlsoft.org/html/libxml-parser.html#\\n> xmlSubstituteEntitiesDefault\\n> >\\n> > The default behave should be common for all PostgreSQL's libxml2 based\\n> > function - and then it is different topic - maybe part for PostgreSQL\\n> ToDo?\\n> > But I don't remember any user requests related to this issue.\\n>\\n>\\n> OK, so it's not xmltable specific. Fine by me.\\n>\\n> Somebody who cares can deal with it. There's clearly nobody breaking\\n> down the walls wanting the feature.\\n>\\n> > I removed this tests - it is not related to XMLTABLE function, but to\\n> > generic XML processing/validation.\\n>\\n>\\n> Good plan.\\n>\\n> >> In +ExecEvalTableExpr, shouldn't you be using PG_ENSURE_ERROR_CLEANUP\\n> >> instead of a PG_TRY() / PG_CATCH() block?\\n> >\\n> >\\n> > If I understand to doc, the PG_ENSURE_ERROR_CLEANUP should be used, when\\n> you\\n> > want to catch FATAL errors (and when you want to clean shared memory).\\n> > XMLTABLE doesn't use shared memory, and doesn't need to catch fatal\\n> errors.\\n>\\n> Ok, makes sense.\\n>\\n>\\n> >> I don't have the libxml knowledge or remaining brain to usefully\\n> >> evaluate the xpath and xml specifics in xpath.c today. It does strike\\n> >> me that the new xpath parser should probably live in its own file,\\n> >> though.\\n> >\\n> > moved\\n>\\n> Thanks.\\n>\\n>\\n> > new version is attached\\n>\\n>\\n> Great.\\n>\\n> I'm marking this ready for committer at this point.\\n>\\n\\nThank you very much\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> I think the XML parser likely needs a more close reading, so I'll ping\\n> Peter E to see if he'll have a chance to check that bit out. But by\\n> and large I think the issues have been ironed out - in terms of\\n> functionality, structure and clarity I think it's looking solid.\\n>\\n> --\\n> Craig Ringer http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Training & Services\\n>\\n\\n2016-09-27 3:34 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:On 24 September 2016 at 14:01, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n\\n>> Did some docs copy-editing and integrated some examples. Explained how\\n>> nested elements work, that multiple top level elements is an error,\\n>> etc. Explained the time-of-evaluation stuff. Pointed out that you can\\n>> refer to prior output columns in PATH and DEFAULT, since that's weird\\n>> and unusual compared to normal SQL. Documented handling of multiple\\n>> node matches, including the surprising results of somepath/text() on\\n>> <somepath>x<!--blah-->y</somepath>. Documented handling of nested\\n>> elements. Documented that xmltable works only on XML documents, not\\n>> fragments/forests.\\n>\\n>\\n> I don't understand to this sentence: \\\"It is possible for a PATH expression\\n> to reference output columns that appear before it in the column-list, so\\n> paths may be dynamically constructed based on other parts of the XML\\n> document:\\\"\\n\\n\\n\\n>> The docs and tests don't seem to cover XML entities. What's the\\n>> behaviour there? Core XML only defines one entity, but if a schema\\n>> defines more how are they processed? The tests need to cover the\\n>> predefined entities " & ' < and > at least.\\n>\\n>\\n> I don't understand, what you are propose here. ?? Please, can you send some\\n> examples.\\n\\nPer below - handling of DTD <!ENTITY> declarations, and the builtin\\nentity tests I already added tests for.\\n\\n\\n>> It doesn't seem to cope with internal DTDs at all (libxml2 limitation?):\\n>>\\n>> SELECT * FROM xmltable('/' PASSING $XML$<?xml version=\\\"1.0\\\"\\n>> standalone=\\\"yes\\\" ?>\\n>> <!DOCTYPE foo [\\n>> <!ELEMENT foo (#PCDATA)>\\n>> <!ENTITY pg \\\"PostgreSQL\\\">\\n>> ]>\\n>> <foo>Hello &pg;.</foo>\\n>> $XML$ COLUMNS foo text);\\n>>\\n>> + ERROR: invalid XML content\\n>> + LINE 1: SELECT * FROM xmltable('/' PASSING $XML$<?xml version=\\\"1.0\\\" ...\\n>> + ^\\n>> + DETAIL: line 2: StartTag: invalid element name\\n>> + <!DOCTYPE foo [\\n>> + ^\\n>> + line 3: StartTag: invalid element name\\n>> + <!ELEMENT foo (#PCDATA)>\\n>> + ^\\n>> + line 4: StartTag: invalid element name\\n>> + <!ENTITY pg \\\"PostgreSQL\\\">\\n>> + ^\\n>> + line 6: Entity 'pg' not defined\\n>> + <foo>Hello &pg;.</foo>\\n>> + ^\\n>>\\n>\\n> It is rejected before XMLTABLE function call\\n>\\n> postgres=# select $XML$<?xml version=\\\"1.0\\\" standalone=\\\"yes\\\" ?>\\n> postgres$# <!DOCTYPE foo [\\n> postgres$# <!ELEMENT foo (#PCDATA)>\\n> postgres$# <!ENTITY pg \\\"PostgreSQL\\\">\\n> postgres$# ]>\\n> postgres$# <foo>Hello &pg;.</foo>\\n> postgres$# $XML$::xml;\\n> ERROR: invalid XML content\\n> LINE 1: select $XML$<?xml version=\\\"1.0\\\" standalone=\\\"yes\\\" ?>\\n> ^\\n> DETAIL: line 2: StartTag: invalid element name\\n> <!DOCTYPE foo [\\n[snip]\\n\\n> It is disabled by default in libxml2. I found a function\\n> xmlSubstituteEntitiesDefault http://www.xmlsoft.org/entities.html\\n> http://www.xmlsoft.org/html/libxml-parser.html#xmlSubstituteEntitiesDefault\\n>\\n> The default behave should be common for all PostgreSQL's libxml2 based\\n> function - and then it is different topic - maybe part for PostgreSQL ToDo?\\n> But I don't remember any user requests related to this issue.\\n\\n\\nOK, so it's not xmltable specific. Fine by me.\\n\\nSomebody who cares can deal with it. There's clearly nobody breaking\\ndown the walls wanting the feature.\\n\\n> I removed this tests - it is not related to XMLTABLE function, but to\\n> generic XML processing/validation.\\n\\n\\nGood plan.\\n\\n>> In +ExecEvalTableExpr, shouldn't you be using PG_ENSURE_ERROR_CLEANUP\\n>> instead of a PG_TRY() / PG_CATCH() block?\\n>\\n>\\n> If I understand to doc, the PG_ENSURE_ERROR_CLEANUP should be used, when you\\n> want to catch FATAL errors (and when you want to clean shared memory).\\n> XMLTABLE doesn't use shared memory, and doesn't need to catch fatal errors.\\n\\nOk, makes sense.\\n\\n\\n>> I don't have the libxml knowledge or remaining brain to usefully\\n>> evaluate the xpath and xml specifics in xpath.c today. It does strike\\n>> me that the new xpath parser should probably live in its own file,\\n>> though.\\n>\\n> moved\\n\\nThanks.\\n\\n\\n> new version is attached\\n\\n\\nGreat.\\n\\nI'm marking this ready for committer at this point.Thank you very muchRegardsPavel \\n\\nI think the XML parser likely needs a more close reading, so I'll ping\\nPeter E to see if he'll have a chance to check that bit out. But by\\nand large I think the issues have been ironed out - in terms of\\nfunctionality, structure and clarity I think it's looking solid.\\n\\n--\\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Tue, 27 Sep 2016 05:40:30 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On 24 September 2016 at 14:01, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n\\n>> Did some docs copy-editing and integrated some examples. Explained how\\n>> nested elements work, that multiple top level elements is an error,\\n>> etc. Explained the time-of-evaluation stuff. Pointed out that you can\\n>> refer to prior output columns in PATH and DEFAULT, since that's weird\\n>> and unusual compared to normal SQL. Documented handling of multiple\\n>> node matches, including the surprising results of somepath/text() on\\n>> <somepath>x<!--blah-->y</somepath>. Documented handling of nested\\n>> elements. Documented that xmltable works only on XML documents, not\\n>> fragments/forests.\\n>\\n>\\n> I don't understand to this sentence: \\\"It is possible for a PATH expression\\n> to reference output columns that appear before it in the column-list, so\\n> paths may be dynamically constructed based on other parts of the XML\\n> document:\\\"\\n\\nThis was based on a misunderstanding of something you said earlier. I\\nthought the idea was to allow this to work:\\n\\nSELECT * FROM xmltable('/x' PASSING\\n'<x><elemName>a</elemName><a>value</a></x>' COLUMNS elemName text,\\nextractedValue text PATH elemName);\\n\\n... but it doesn't:\\n\\n\\nSELECT * FROM xmltable('/x' PASSING\\n'<x><elemName>a</elemName><a>value</a></x>' COLUMNS elemName text,\\nextractedValue text PATH elemName);\\nERROR: column \\\"elemname\\\" does not exist\\nLINE 1: ...' COLUMNS elemName text, extractedValue text PATH elemName);\\n\\n... so please delete that text. I thought I'd tested it but the state\\nof my tests dir says I just got distracted by another task at the\\nwrong time.\\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\": \"Tue, 27 Sep 2016 11:53:06 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-09-27 5:53 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:\\n\\n> On 24 September 2016 at 14:01, Pavel Stehule <pavel.stehule@gmail.com>\\n> wrote:\\n>\\n> >> Did some docs copy-editing and integrated some examples. Explained how\\n> >> nested elements work, that multiple top level elements is an error,\\n> >> etc. Explained the time-of-evaluation stuff. Pointed out that you can\\n> >> refer to prior output columns in PATH and DEFAULT, since that's weird\\n> >> and unusual compared to normal SQL. Documented handling of multiple\\n> >> node matches, including the surprising results of somepath/text() on\\n> >> <somepath>x<!--blah-->y</somepath>. Documented handling of nested\\n> >> elements. Documented that xmltable works only on XML documents, not\\n> >> fragments/forests.\\n> >\\n> >\\n> > I don't understand to this sentence: \\\"It is possible for a PATH\\n> expression\\n> > to reference output columns that appear before it in the column-list, so\\n> > paths may be dynamically constructed based on other parts of the XML\\n> > document:\\\"\\n>\\n> This was based on a misunderstanding of something you said earlier. I\\n> thought the idea was to allow this to work:\\n>\\n> SELECT * FROM xmltable('/x' PASSING\\n> '<x><elemName>a</elemName><a>value</a></x>' COLUMNS elemName text,\\n> extractedValue text PATH elemName);\\n>\\n> ... but it doesn't:\\n>\\n>\\n> SELECT * FROM xmltable('/x' PASSING\\n> '<x><elemName>a</elemName><a>value</a></x>' COLUMNS elemName text,\\n> extractedValue text PATH elemName);\\n> ERROR: column \\\"elemname\\\" does not exist\\n> LINE 1: ...' COLUMNS elemName text, extractedValue text PATH elemName);\\n>\\n> ... so please delete that text. I thought I'd tested it but the state\\n> of my tests dir says I just got distracted by another task at the\\n> wrong time.\\n>\\n\\ndeleted\\n\\nRegards\\n\\nPavel\\n\\n>\\n> --\\n> Craig Ringer http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, 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\": \"Tue, 27 Sep 2016 07:29:03 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On Tue, Sep 27, 2016 at 2:29 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n> 2016-09-27 5:53 GMT+02:00 Craig Ringer <craig@2ndquadrant.com>:\\n>>\\n>> [...]\\n>> ... so please delete that text. I thought I'd tested it but the state\\n>> of my tests dir says I just got distracted by another task at the\\n>> wrong time.\\n\\nMoved patch to next CF with same status: ready for committer.\\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, 3 Oct 2016 11:18:44 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\nnew update - only doc\\n\\n+ <para>\\n+ Only XPath query language is supported. PostgreSQL doesn't support\\nXQuery\\n+ language. Then the syntax of <function>xmltable</function> doesn't\\n+ allow to use XQuery related functionality - the name of xml expression\\n+ (clause <literal>AS</literal>) is not allowed, and only one xml\\nexpression\\n+ should be passed to <function>xmltable</function> function as\\nparameter.\\n+ </para>\\n\\nRegards\\n\\nPavel\\n\\n\\n-- \\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, 14 Oct 2016 06:23:45 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"I've been going over this patch. I think it'd be better to restructure\\nthe <sect2> before adding the docs for this new function; I already\\nsplit it out, so don't do anything about this.\\n\\nNext, looking at struct TableExprBuilder I noticed that the comments are\\nalready obsolete, as they talk about function params that do not exist\\n(missing_columns) and they fail to mention the ones that do exist.\\nAlso, function member SetContent is not documented at all. Overall,\\nthese comments do not convey a lot -- apparently, whoever reads them is\\nalready supposed to know how it works: \\\"xyz sets a row generating\\nfilter\\\" doesn't tell me anything. Since this is API documentation, it\\nneeds to be much clearer.\\n\\nExecEvalTableExpr and ExecEvalTableExprProtected have no comments\\nwhatsoever. Needs fixed.\\n\\nI wonder if it'd be a good idea to install TableExpr first without the\\nimplementing XMLTABLE, so that it's clearer what is API and what is\\nimplementation.\\n\\nThe number of new keywords in this patch is depressing. I suppose\\nthere's no way around that -- as I understand, this is caused by the SQL\\nstandard's definition of the syntax for this feature.\\n\\nHave to go now for a bit -- will continue looking afterwards. Please\\nsubmit delta patches on top of your latest v12 to fix the comments I\\nmentioned.\\n\\n-- \\n�lvaro Herrera https://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, 17 Nov 2016 15:22:26 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2016-11-17 19:22 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> I've been going over this patch. I think it'd be better to restructure\\n> the <sect2> before adding the docs for this new function; I already\\n> split it out, so don't do anything about this.\\n>\\n> Next, looking at struct TableExprBuilder I noticed that the comments are\\n> already obsolete, as they talk about function params that do not exist\\n> (missing_columns) and they fail to mention the ones that do exist.\\n> Also, function member SetContent is not documented at all. Overall,\\n> these comments do not convey a lot -- apparently, whoever reads them is\\n> already supposed to know how it works: \\\"xyz sets a row generating\\n> filter\\\" doesn't tell me anything. Since this is API documentation, it\\n> needs to be much clearer.\\n>\\n> ExecEvalTableExpr and ExecEvalTableExprProtected have no comments\\n> whatsoever. Needs fixed.\\n>\\n\\nI am sending the patch with more comments - but it needs a care someone\\nwith good English skills.\\n\\n\\n>\\n> I wonder if it'd be a good idea to install TableExpr first without the\\n> implementing XMLTABLE, so that it's clearer what is API and what is\\n> implementation.\\n>\\n\\nI am not sure about this step - the API is clean from name. In this moment,\\nfor this API is not any other tests than XMLTABLE implementation.\\n\\n\\n>\\n> The number of new keywords in this patch is depressing. I suppose\\n> there's no way around that -- as I understand, this is caused by the SQL\\n> standard's definition of the syntax for this feature.\\n>\\n> Have to go now for a bit -- will continue looking afterwards. Please\\n> submit delta patches on top of your latest v12 to fix the comments I\\n> mentioned.\\n>\\n>\\nRegards\\n\\nPavel\\n\\n\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL 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\": \"Fri, 18 Nov 2016 21:49:36 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n\\n> 2016-11-17 19:22 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n> \\n> > Next, looking at struct TableExprBuilder I noticed that the comments are\\n> > already obsolete, as they talk about function params that do not exist\\n> > (missing_columns) and they fail to mention the ones that do exist.\\n> > Also, function member SetContent is not documented at all. Overall,\\n> > these comments do not convey a lot -- apparently, whoever reads them is\\n> > already supposed to know how it works: \\\"xyz sets a row generating\\n> > filter\\\" doesn't tell me anything. Since this is API documentation, it\\n> > needs to be much clearer.\\n> >\\n> > ExecEvalTableExpr and ExecEvalTableExprProtected have no comments\\n> > whatsoever. Needs fixed.\\n> \\n> I am sending the patch with more comments - but it needs a care someone\\n> with good English skills.\\n\\nThanks, I can help with that.\\n\\n-- \\n�lvaro Herrera https://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\": \"Fri, 18 Nov 2016 17:53:56 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"The SQL standard seems to require a comma after the XMLNAMESPACES\\nclause:\\n\\n<XML table> ::=\\n XMLTABLE <left paren>\\n [ <XML namespace declaration> <comma> ]\\n <XML table row pattern>\\n [ <XML table argument list> ]\\n COLUMNS <XML table column definitions> <right paren>\\n\\nI don't understand the reason for that, but I have added it:\\n\\n\\t\\t\\t| XMLTABLE '(' XMLNAMESPACES '(' XmlNamespaceList ')' ',' c_expr xmlexists_argument ')'\\n\\t\\t\\t\\t{\\n\\t\\t\\t\\t\\tTableExpr *n = makeNode(TableExpr);\\n\\t\\t\\t\\t\\tn->row_path = $8;\\n\\t\\t\\t\\t\\tn->expr = $9;\\n\\t\\t\\t\\t\\tn->cols = NIL;\\n\\t\\t\\t\\t\\tn->namespaces = $5;\\n\\t\\t\\t\\t\\tn->location = @1;\\n\\t\\t\\t\\t\\t$$ = (Node *)n;\\n\\t\\t\\t\\t}\\n\\t\\t\\t| XMLTABLE '(' XMLNAMESPACES '(' XmlNamespaceList ')' ',' c_expr xmlexists_argument COLUMNS TableExprColList ')'\\n\\t\\t\\t\\t{\\n\\t\\t\\t\\t\\tTableExpr *n = makeNode(TableExpr);\\n\\t\\t\\t\\t\\tn->row_path = $8;\\n\\t\\t\\t\\t\\tn->expr = $9;\\n\\t\\t\\t\\t\\tn->cols = $11;\\n\\t\\t\\t\\t\\tn->namespaces = $5;\\n\\t\\t\\t\\t\\tn->location = @1;\\n\\t\\t\\t\\t\\t$$ = (Node *)n;\\n\\t\\t\\t\\t}\\n\\t\\t;\\n\\nAnother thing I did was remove the TableExprColOptionsOpt production; in\\nits place I added a third rule in TableExprCol for \\\"ColId Typename\\nIsNotNull\\\" (i.e. no options). This seems to reduce the size of the\\ngenerated gram.c a good dozen kB.\\n\\n\\nI didn't like much the use of c_expr in all these productions. As I\\nunderstand it, c_expr is mostly an implementation artifact and we should\\nbe using a_expr or b_expr almost everywhere. I see that XMLEXISTS\\nalready expanded the very limited use of c_expr there was; I would\\nprefer to fix that one too rather than replicate it here. TBH I'm not\\nsure I like that XMLTABLE is re-using xmlexists_argument. \\n\\nActually, is the existing XMLEXISTS production correct? What I see in\\nthe standard is\\n\\n<XML table row pattern> ::= <character string literal>\\n\\n<XML table argument list> ::=\\n PASSING <XML table argument passing mechanism> <XML query argument>\\n [ { <comma> <XML query argument> }... ]\\n\\n<XML table argument passing mechanism> ::= <XML passing mechanism>\\n\\n<XML table column definitions> ::= <XML table column definition> [ { <comma> <XML table column definition> }... ]\\n\\n<XML table column definition> ::=\\n <XML table ordinality column definition>\\n | <XML table regular column definition>\\n\\n<XML table ordinality column definition> ::= <column name> FOR ORDINALITY\\n\\n<XML table regular column definition> ::=\\n <column name> <data type> [ <XML passing mechanism> ]\\n [ <default clause> ]\\n [ PATH <XML table column pattern> ]\\n\\n<XML table column pattern> ::= <character string literal>\\n\\nso I think this resolves \\\"PASSING BY {REF,VALUE} <XML query argument>\\\", but what\\nwe have in gram.y is:\\n\\n/* We allow several variants for SQL and other compatibility. */\\nxmlexists_argument:\\n\\t\\t\\tPASSING c_expr\\n\\t\\t\\t\\t{\\n\\t\\t\\t\\t\\t$$ = $2;\\n\\t\\t\\t\\t}\\n\\t\\t\\t| PASSING c_expr BY REF\\n\\t\\t\\t\\t{\\n\\t\\t\\t\\t\\t$$ = $2;\\n\\t\\t\\t\\t}\\n\\t\\t\\t| PASSING BY REF c_expr\\n\\t\\t\\t\\t{\\n\\t\\t\\t\\t\\t$$ = $4;\\n\\t\\t\\t\\t}\\n\\t\\t\\t| PASSING BY REF c_expr BY REF\\n\\t\\t\\t\\t{\\n\\t\\t\\t\\t\\t$$ = $4;\\n\\t\\t\\t\\t}\\n\\t\\t;\\n\\nI'm not sure why we allow \\\"PASSING c_expr\\\" at all. Maybe if BY VALUE/BY\\nREF is not specified, we should just not have PASSING at all?\\n\\nIf we extended this for XMLEXISTS for compatibility with some other\\nproduct, perhaps we should look into what that product supports for\\nXMLTABLE; maybe XMLTABLE does not need all the same options as\\nXMLEXISTS.\\n\\nThe fourth option seems very dubious to me. This was added by commit\\n641459f26, submitted here:\\nhttps://www.postgresql.org/message-id/4C0F6DBF.9000001@mlfowler.com\\n\\n... Hm, actually some perusal of the XMLEXISTS predicate in the standard\\nshows that it's quite a different thing from XMLTABLE. Maybe we\\nshouldn't reuse xmlexists_argument here at all.\\n\\n-- \\n�lvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 18 Nov 2016 20:42:15 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-11-19 0:42 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> The SQL standard seems to require a comma after the XMLNAMESPACES\\n> clause:\\n>\\n> <XML table> ::=\\n> XMLTABLE <left paren>\\n> [ <XML namespace declaration> <comma> ]\\n> <XML table row pattern>\\n> [ <XML table argument list> ]\\n> COLUMNS <XML table column definitions> <right paren>\\n>\\n> I don't understand the reason for that, but I have added it:\\n>\\n> | XMLTABLE '(' XMLNAMESPACES '(' XmlNamespaceList\\n> ')' ',' c_expr xmlexists_argument ')'\\n> {\\n> TableExpr *n = makeNode(TableExpr);\\n> n->row_path = $8;\\n> n->expr = $9;\\n> n->cols = NIL;\\n> n->namespaces = $5;\\n> n->location = @1;\\n> $$ = (Node *)n;\\n> }\\n> | XMLTABLE '(' XMLNAMESPACES '(' XmlNamespaceList\\n> ')' ',' c_expr xmlexists_argument COLUMNS TableExprColList ')'\\n> {\\n> TableExpr *n = makeNode(TableExpr);\\n> n->row_path = $8;\\n> n->expr = $9;\\n> n->cols = $11;\\n> n->namespaces = $5;\\n> n->location = @1;\\n> $$ = (Node *)n;\\n> }\\n> ;\\n>\\n>\\nyes, looks my oversight - it is better\\n\\n\\n\\n> Another thing I did was remove the TableExprColOptionsOpt production; in\\n> its place I added a third rule in TableExprCol for \\\"ColId Typename\\n> IsNotNull\\\" (i.e. no options). This seems to reduce the size of the\\n> generated gram.c a good dozen kB.\\n>\\n\\nIf I remember well - this was required by better compatibility with Oracle\\n\\nANSI SQL: colname type DEFAULT PATH\\nOracle: colname PATH DEFAULT\\n\\nMy implementation allows both combinations - there are two reasons: 1. one\\nless issue when people does port from Oracle, 2. almost all examples of\\nXMLTABLE on a net are from Oracle - it can be unfriendly, when these\\nexamples would not work on PG - there was discussion about this issue in\\nthis mailing list\\n\\n\\n>\\n>\\n> I didn't like much the use of c_expr in all these productions. As I\\n> understand it, c_expr is mostly an implementation artifact and we should\\n> be using a_expr or b_expr almost everywhere. I see that XMLEXISTS\\n> already expanded the very limited use of c_expr there was; I would\\n> prefer to fix that one too rather than replicate it here. TBH I'm not\\n> sure I like that XMLTABLE is re-using xmlexists_argument.\\n>\\n\\nThere are two situations: c_expr as document content, and c_expr after\\nDEFAULT and PATH keywords. First probably can be fixed, second not, because\\n\\\"PATH\\\" is unreserved keyword only.\\n\\n\\n>\\n> Actually, is the existing XMLEXISTS production correct? What I see in\\n> the standard is\\n>\\n> <XML table row pattern> ::= <character string literal>\\n>\\n> <XML table argument list> ::=\\n> PASSING <XML table argument passing mechanism> <XML query argument>\\n> [ { <comma> <XML query argument> }... ]\\n>\\n> <XML table argument passing mechanism> ::= <XML passing mechanism>\\n>\\n> <XML table column definitions> ::= <XML table column definition> [ {\\n> <comma> <XML table column definition> }... ]\\n>\\n> <XML table column definition> ::=\\n> <XML table ordinality column definition>\\n> | <XML table regular column definition>\\n>\\n> <XML table ordinality column definition> ::= <column name> FOR ORDINALITY\\n>\\n> <XML table regular column definition> ::=\\n> <column name> <data type> [ <XML passing mechanism> ]\\n> [ <default clause> ]\\n> [ PATH <XML table column pattern> ]\\n>\\n> <XML table column pattern> ::= <character string literal>\\n>\\n> so I think this resolves \\\"PASSING BY {REF,VALUE} <XML query argument>\\\",\\n> but what\\n> we have in gram.y is:\\n>\\n> /* We allow several variants for SQL and other compatibility. */\\n> xmlexists_argument:\\n> PASSING c_expr\\n> {\\n> $$ = $2;\\n> }\\n> | PASSING c_expr BY REF\\n> {\\n> $$ = $2;\\n> }\\n> | PASSING BY REF c_expr\\n> {\\n> $$ = $4;\\n> }\\n> | PASSING BY REF c_expr BY REF\\n> {\\n> $$ = $4;\\n> }\\n> ;\\n>\\n> I'm not sure why we allow \\\"PASSING c_expr\\\" at all. Maybe if BY VALUE/BY\\n> REF is not specified, we should just not have PASSING at all?\\n>\\n>\\nIf we extended this for XMLEXISTS for compatibility with some other\\n> product, perhaps we should look into what that product supports for\\n> XMLTABLE; maybe XMLTABLE does not need all the same options as\\n> XMLEXISTS.\\n>\\n>\\nThe reason is a compatibility with other products - DB2. XMLTABLE uses same\\noptions like XMLEXISTS. These options has zero value for Postgres, but its\\nare important - compatibility, and workable examples.\\n\\n\\n> The fourth option seems very dubious to me. This was added by commit\\n> 641459f26, submitted here:\\n> https://www.postgresql.org/message-id/4C0F6DBF.9000001@mlfowler.com\\n>\\n> ... Hm, actually some perusal of the XMLEXISTS predicate in the standard\\n> shows that it's quite a different thing from XMLTABLE. Maybe we\\n> shouldn't reuse xmlexists_argument here at all.\\n>\\n\\nnot sure If I understand\\n\\nRegards\\n\\nPavel\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2016-11-19 0:42 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:The SQL standard seems to require a comma after the XMLNAMESPACES\\nclause:\\n\\n<XML table> ::=\\n XMLTABLE <left paren>\\n [ <XML namespace declaration> <comma> ]\\n <XML table row pattern>\\n [ <XML table argument list> ]\\n COLUMNS <XML table column definitions> <right paren>\\n\\nI don't understand the reason for that, but I have added it:\\n\\n | XMLTABLE '(' XMLNAMESPACES '(' XmlNamespaceList ')' ',' c_expr xmlexists_argument ')'\\n {\\n TableExpr *n = makeNode(TableExpr);\\n n->row_path = $8;\\n n->expr = $9;\\n n->cols = NIL;\\n n->namespaces = $5;\\n n->location = @1;\\n $$ = (Node *)n;\\n }\\n | XMLTABLE '(' XMLNAMESPACES '(' XmlNamespaceList ')' ',' c_expr xmlexists_argument COLUMNS TableExprColList ')'\\n {\\n TableExpr *n = makeNode(TableExpr);\\n n->row_path = $8;\\n n->expr = $9;\\n n->cols = $11;\\n n->namespaces = $5;\\n n->location = @1;\\n $$ = (Node *)n;\\n }\\n ;\\nyes, looks my oversight - it is better \\nAnother thing I did was remove the TableExprColOptionsOpt production; in\\nits place I added a third rule in TableExprCol for \\\"ColId Typename\\nIsNotNull\\\" (i.e. no options). This seems to reduce the size of the\\ngenerated gram.c a good dozen kB.If I remember well - this was required by better compatibility with OracleANSI SQL: colname type DEFAULT PATHOracle: colname PATH DEFAULTMy implementation allows both combinations - there are two reasons: 1. one less issue when people does port from Oracle, 2. almost all examples of XMLTABLE on a net are from Oracle - it can be unfriendly, when these examples would not work on PG - there was discussion about this issue in this mailing list \\n\\n\\nI didn't like much the use of c_expr in all these productions. As I\\nunderstand it, c_expr is mostly an implementation artifact and we should\\nbe using a_expr or b_expr almost everywhere. I see that XMLEXISTS\\nalready expanded the very limited use of c_expr there was; I would\\nprefer to fix that one too rather than replicate it here. TBH I'm not\\nsure I like that XMLTABLE is re-using xmlexists_argument.There are two situations: c_expr as document content, and c_expr after DEFAULT and PATH keywords. First probably can be fixed, second not, because \\\"PATH\\\" is unreserved keyword only. \\n\\nActually, is the existing XMLEXISTS production correct? What I see in\\nthe standard is\\n\\n<XML table row pattern> ::= <character string literal>\\n\\n<XML table argument list> ::=\\n PASSING <XML table argument passing mechanism> <XML query argument>\\n [ { <comma> <XML query argument> }... ]\\n\\n<XML table argument passing mechanism> ::= <XML passing mechanism>\\n\\n<XML table column definitions> ::= <XML table column definition> [ { <comma> <XML table column definition> }... ]\\n\\n<XML table column definition> ::=\\n <XML table ordinality column definition>\\n | <XML table regular column definition>\\n\\n<XML table ordinality column definition> ::= <column name> FOR ORDINALITY\\n\\n<XML table regular column definition> ::=\\n <column name> <data type> [ <XML passing mechanism> ]\\n [ <default clause> ]\\n [ PATH <XML table column pattern> ]\\n\\n<XML table column pattern> ::= <character string literal>\\n\\nso I think this resolves \\\"PASSING BY {REF,VALUE} <XML query argument>\\\", but what\\nwe have in gram.y is:\\n\\n/* We allow several variants for SQL and other compatibility. */\\nxmlexists_argument:\\n PASSING c_expr\\n {\\n $$ = $2;\\n }\\n | PASSING c_expr BY REF\\n {\\n $$ = $2;\\n }\\n | PASSING BY REF c_expr\\n {\\n $$ = $4;\\n }\\n | PASSING BY REF c_expr BY REF\\n {\\n $$ = $4;\\n }\\n ;\\n\\nI'm not sure why we allow \\\"PASSING c_expr\\\" at all. Maybe if BY VALUE/BY\\nREF is not specified, we should just not have PASSING at all? \\n\\nIf we extended this for XMLEXISTS for compatibility with some other\\nproduct, perhaps we should look into what that product supports for\\nXMLTABLE; maybe XMLTABLE does not need all the same options as\\nXMLEXISTS.\\nThe reason is a compatibility with other products - DB2. XMLTABLE uses same options like XMLEXISTS. These options has zero value for Postgres, but its are important - compatibility, and workable examples. \\nThe fourth option seems very dubious to me. This was added by commit\\n641459f26, submitted here:\\nhttps://www.postgresql.org/message-id/4C0F6DBF.9000001@mlfowler.com\\n\\n... Hm, actually some perusal of the XMLEXISTS predicate in the standard\\nshows that it's quite a different thing from XMLTABLE. Maybe we\\nshouldn't reuse xmlexists_argument here at all.not sure If I understand RegardsPavel \\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Sat, 19 Nov 2016 05:19:52 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-11-19 5:19 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n\\n>\\n>\\n> 2016-11-19 0:42 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>\\n>> The SQL standard seems to require a comma after the XMLNAMESPACES\\n>> clause:\\n>>\\n>> <XML table> ::=\\n>> XMLTABLE <left paren>\\n>> [ <XML namespace declaration> <comma> ]\\n>> <XML table row pattern>\\n>> [ <XML table argument list> ]\\n>> COLUMNS <XML table column definitions> <right paren>\\n>>\\n>> I don't understand the reason for that, but I have added it:\\n>>\\n>> | XMLTABLE '(' XMLNAMESPACES '(' XmlNamespaceList\\n>> ')' ',' c_expr xmlexists_argument ')'\\n>> {\\n>> TableExpr *n =\\n>> makeNode(TableExpr);\\n>> n->row_path = $8;\\n>> n->expr = $9;\\n>> n->cols = NIL;\\n>> n->namespaces = $5;\\n>> n->location = @1;\\n>> $$ = (Node *)n;\\n>> }\\n>> | XMLTABLE '(' XMLNAMESPACES '(' XmlNamespaceList\\n>> ')' ',' c_expr xmlexists_argument COLUMNS TableExprColList ')'\\n>> {\\n>> TableExpr *n =\\n>> makeNode(TableExpr);\\n>> n->row_path = $8;\\n>> n->expr = $9;\\n>> n->cols = $11;\\n>> n->namespaces = $5;\\n>> n->location = @1;\\n>> $$ = (Node *)n;\\n>> }\\n>> ;\\n>>\\n>>\\n> yes, looks my oversight - it is better\\n>\\n>\\n>\\n>> Another thing I did was remove the TableExprColOptionsOpt production; in\\n>> its place I added a third rule in TableExprCol for \\\"ColId Typename\\n>> IsNotNull\\\" (i.e. no options). This seems to reduce the size of the\\n>> generated gram.c a good dozen kB.\\n>>\\n>\\n> If I remember well - this was required by better compatibility with Oracle\\n>\\n> ANSI SQL: colname type DEFAULT PATH\\n> Oracle: colname PATH DEFAULT\\n>\\n> My implementation allows both combinations - there are two reasons: 1. one\\n> less issue when people does port from Oracle, 2. almost all examples of\\n> XMLTABLE on a net are from Oracle - it can be unfriendly, when these\\n> examples would not work on PG - there was discussion about this issue in\\n> this mailing list\\n>\\n>\\n>>\\n>>\\n>> I didn't like much the use of c_expr in all these productions. As I\\n>> understand it, c_expr is mostly an implementation artifact and we should\\n>> be using a_expr or b_expr almost everywhere. I see that XMLEXISTS\\n>> already expanded the very limited use of c_expr there was; I would\\n>> prefer to fix that one too rather than replicate it here. TBH I'm not\\n>> sure I like that XMLTABLE is re-using xmlexists_argument.\\n>>\\n>\\n> There are two situations: c_expr as document content, and c_expr after\\n> DEFAULT and PATH keywords. First probably can be fixed, second not, because\\n> \\\"PATH\\\" is unreserved keyword only.\\n>\\n\\nIt is not possible PASSING is unreserved keyword too.\\n\\nRegards\\n\\nPavel\\n\\n>\\n>\\n>>\\n>> Actually, is the existing XMLEXISTS production correct? What I see in\\n>> the standard is\\n>>\\n>> <XML table row pattern> ::= <character string literal>\\n>>\\n>> <XML table argument list> ::=\\n>> PASSING <XML table argument passing mechanism> <XML query argument>\\n>> [ { <comma> <XML query argument> }... ]\\n>>\\n>> <XML table argument passing mechanism> ::= <XML passing mechanism>\\n>>\\n>> <XML table column definitions> ::= <XML table column definition> [ {\\n>> <comma> <XML table column definition> }... ]\\n>>\\n>> <XML table column definition> ::=\\n>> <XML table ordinality column definition>\\n>> | <XML table regular column definition>\\n>>\\n>> <XML table ordinality column definition> ::= <column name> FOR ORDINALITY\\n>>\\n>> <XML table regular column definition> ::=\\n>> <column name> <data type> [ <XML passing mechanism> ]\\n>> [ <default clause> ]\\n>> [ PATH <XML table column pattern> ]\\n>>\\n>> <XML table column pattern> ::= <character string literal>\\n>>\\n>> so I think this resolves \\\"PASSING BY {REF,VALUE} <XML query argument>\\\",\\n>> but what\\n>> we have in gram.y is:\\n>>\\n>> /* We allow several variants for SQL and other compatibility. */\\n>> xmlexists_argument:\\n>> PASSING c_expr\\n>> {\\n>> $$ = $2;\\n>> }\\n>> | PASSING c_expr BY REF\\n>> {\\n>> $$ = $2;\\n>> }\\n>> | PASSING BY REF c_expr\\n>> {\\n>> $$ = $4;\\n>> }\\n>> | PASSING BY REF c_expr BY REF\\n>> {\\n>> $$ = $4;\\n>> }\\n>> ;\\n>>\\n>> I'm not sure why we allow \\\"PASSING c_expr\\\" at all. Maybe if BY VALUE/BY\\n>> REF is not specified, we should just not have PASSING at all?\\n>>\\n>>\\n> If we extended this for XMLEXISTS for compatibility with some other\\n>> product, perhaps we should look into what that product supports for\\n>> XMLTABLE; maybe XMLTABLE does not need all the same options as\\n>> XMLEXISTS.\\n>>\\n>>\\n> The reason is a compatibility with other products - DB2. XMLTABLE uses\\n> same options like XMLEXISTS. These options has zero value for Postgres, but\\n> its are important - compatibility, and workable examples.\\n>\\n>\\n>> The fourth option seems very dubious to me. This was added by commit\\n>> 641459f26, submitted here:\\n>> https://www.postgresql.org/message-id/4C0F6DBF.9000001@mlfowler.com\\n>>\\n>> ... Hm, actually some perusal of the XMLEXISTS predicate in the standard\\n>> shows that it's quite a different thing from XMLTABLE. Maybe we\\n>> shouldn't reuse xmlexists_argument here at all.\\n>>\\n>\\n> not sure If I understand\\n>\\n> Regards\\n>\\n> Pavel\\n>\\n>>\\n>> --\\n>> Álvaro Herrera https://www.2ndQuadrant.com/\\n>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>>\\n>\\n>\\n\\n2016-11-19 5:19 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:2016-11-19 0:42 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:The SQL standard seems to require a comma after the XMLNAMESPACES\\nclause:\\n\\n<XML table> ::=\\n XMLTABLE <left paren>\\n [ <XML namespace declaration> <comma> ]\\n <XML table row pattern>\\n [ <XML table argument list> ]\\n COLUMNS <XML table column definitions> <right paren>\\n\\nI don't understand the reason for that, but I have added it:\\n\\n | XMLTABLE '(' XMLNAMESPACES '(' XmlNamespaceList ')' ',' c_expr xmlexists_argument ')'\\n {\\n TableExpr *n = makeNode(TableExpr);\\n n->row_path = $8;\\n n->expr = $9;\\n n->cols = NIL;\\n n->namespaces = $5;\\n n->location = @1;\\n $$ = (Node *)n;\\n }\\n | XMLTABLE '(' XMLNAMESPACES '(' XmlNamespaceList ')' ',' c_expr xmlexists_argument COLUMNS TableExprColList ')'\\n {\\n TableExpr *n = makeNode(TableExpr);\\n n->row_path = $8;\\n n->expr = $9;\\n n->cols = $11;\\n n->namespaces = $5;\\n n->location = @1;\\n $$ = (Node *)n;\\n }\\n ;\\nyes, looks my oversight - it is better \\nAnother thing I did was remove the TableExprColOptionsOpt production; in\\nits place I added a third rule in TableExprCol for \\\"ColId Typename\\nIsNotNull\\\" (i.e. no options). This seems to reduce the size of the\\ngenerated gram.c a good dozen kB.If I remember well - this was required by better compatibility with OracleANSI SQL: colname type DEFAULT PATHOracle: colname PATH DEFAULTMy implementation allows both combinations - there are two reasons: 1. one less issue when people does port from Oracle, 2. almost all examples of XMLTABLE on a net are from Oracle - it can be unfriendly, when these examples would not work on PG - there was discussion about this issue in this mailing list \\n\\n\\nI didn't like much the use of c_expr in all these productions. As I\\nunderstand it, c_expr is mostly an implementation artifact and we should\\nbe using a_expr or b_expr almost everywhere. I see that XMLEXISTS\\nalready expanded the very limited use of c_expr there was; I would\\nprefer to fix that one too rather than replicate it here. TBH I'm not\\nsure I like that XMLTABLE is re-using xmlexists_argument.There are two situations: c_expr as document content, and c_expr after DEFAULT and PATH keywords. First probably can be fixed, second not, because \\\"PATH\\\" is unreserved keyword only.It is not possible PASSING is unreserved keyword too.RegardsPavel \\n\\nActually, is the existing XMLEXISTS production correct? What I see in\\nthe standard is\\n\\n<XML table row pattern> ::= <character string literal>\\n\\n<XML table argument list> ::=\\n PASSING <XML table argument passing mechanism> <XML query argument>\\n [ { <comma> <XML query argument> }... ]\\n\\n<XML table argument passing mechanism> ::= <XML passing mechanism>\\n\\n<XML table column definitions> ::= <XML table column definition> [ { <comma> <XML table column definition> }... ]\\n\\n<XML table column definition> ::=\\n <XML table ordinality column definition>\\n | <XML table regular column definition>\\n\\n<XML table ordinality column definition> ::= <column name> FOR ORDINALITY\\n\\n<XML table regular column definition> ::=\\n <column name> <data type> [ <XML passing mechanism> ]\\n [ <default clause> ]\\n [ PATH <XML table column pattern> ]\\n\\n<XML table column pattern> ::= <character string literal>\\n\\nso I think this resolves \\\"PASSING BY {REF,VALUE} <XML query argument>\\\", but what\\nwe have in gram.y is:\\n\\n/* We allow several variants for SQL and other compatibility. */\\nxmlexists_argument:\\n PASSING c_expr\\n {\\n $$ = $2;\\n }\\n | PASSING c_expr BY REF\\n {\\n $$ = $2;\\n }\\n | PASSING BY REF c_expr\\n {\\n $$ = $4;\\n }\\n | PASSING BY REF c_expr BY REF\\n {\\n $$ = $4;\\n }\\n ;\\n\\nI'm not sure why we allow \\\"PASSING c_expr\\\" at all. Maybe if BY VALUE/BY\\nREF is not specified, we should just not have PASSING at all? \\n\\nIf we extended this for XMLEXISTS for compatibility with some other\\nproduct, perhaps we should look into what that product supports for\\nXMLTABLE; maybe XMLTABLE does not need all the same options as\\nXMLEXISTS.\\nThe reason is a compatibility with other products - DB2. XMLTABLE uses same options like XMLEXISTS. These options has zero value for Postgres, but its are important - compatibility, and workable examples. \\nThe fourth option seems very dubious to me. This was added by commit\\n641459f26, submitted here:\\nhttps://www.postgresql.org/message-id/4C0F6DBF.9000001@mlfowler.com\\n\\n... Hm, actually some perusal of the XMLEXISTS predicate in the standard\\nshows that it's quite a different thing from XMLTABLE. Maybe we\\nshouldn't reuse xmlexists_argument here at all.not sure If I understand RegardsPavel \\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Sat, 19 Nov 2016 05:25:46 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Something I just noticed is that transformTableExpr takes a TableExpr\\nnode and returns another TableExpr node. That's unlike what we do in\\nother places, where the node returned is of a different type than the\\ninput node. I'm not real clear what happens if you try to re-transform\\na node that was already transformed, but it seems worth thinking about.\\n\\n-- \\n�lvaro Herrera https://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\": \"Mon, 21 Nov 2016 15:12:21 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n> Something I just noticed is that transformTableExpr takes a TableExpr\\n> node and returns another TableExpr node. That's unlike what we do in\\n> other places, where the node returned is of a different type than the\\n> input node. I'm not real clear what happens if you try to re-transform\\n> a node that was already transformed, but it seems worth thinking about.\\n\\nWe're not 100% consistent on that --- there are cases such as RowExpr\\nand CaseExpr where the same struct type is used for pre-parse-analysis\\nand post-parse-analysis nodes. I think it's okay as long as the\\ninformation content isn't markedly different, ie the transformation\\njust consists of transforming all the sub-nodes.\\n\\nBeing able to behave sanely on a re-transformation used to be an\\nissue, but we no longer expect transformExpr to support that.\\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, 21 Nov 2016 15:16:15 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-11-21 21:16 GMT+01:00 Tom Lane <tgl@sss.pgh.pa.us>:\\n\\n> Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n> > Something I just noticed is that transformTableExpr takes a TableExpr\\n> > node and returns another TableExpr node. That's unlike what we do in\\n> > other places, where the node returned is of a different type than the\\n> > input node. I'm not real clear what happens if you try to re-transform\\n> > a node that was already transformed, but it seems worth thinking about.\\n>\\n> We're not 100% consistent on that --- there are cases such as RowExpr\\n> and CaseExpr where the same struct type is used for pre-parse-analysis\\n> and post-parse-analysis nodes. I think it's okay as long as the\\n> information content isn't markedly different, ie the transformation\\n> just consists of transforming all the sub-nodes.\\n>\\n> Being able to behave sanely on a re-transformation used to be an\\n> issue, but we no longer expect transformExpr to support that.\\n>\\n\\nI was not sure in this case - using new node was more clear for me -\\nsafeguard against some uninitialized or untransformed value. There in only\\nfew bytes memory more overhead.\\n\\nregards\\n\\nPavel\\n\\n\\n>\\n> regards, tom lane\\n>\\n\\n2016-11-21 21:16 GMT+01:00 Tom Lane <tgl@sss.pgh.pa.us>:Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n> Something I just noticed is that transformTableExpr takes a TableExpr\\n> node and returns another TableExpr node. That's unlike what we do in\\n> other places, where the node returned is of a different type than the\\n> input node. I'm not real clear what happens if you try to re-transform\\n> a node that was already transformed, but it seems worth thinking about.\\n\\nWe're not 100% consistent on that --- there are cases such as RowExpr\\nand CaseExpr where the same struct type is used for pre-parse-analysis\\nand post-parse-analysis nodes. I think it's okay as long as the\\ninformation content isn't markedly different, ie the transformation\\njust consists of transforming all the sub-nodes.\\n\\nBeing able to behave sanely on a re-transformation used to be an\\nissue, but we no longer expect transformExpr to support that.I was not sure in this case - using new node was more clear for me - safeguard against some uninitialized or untransformed value. There in only few bytes memory more overhead.regardsPavel \\n\\n regards, tom lane\",\n \"msg_date\": \"Tue, 22 Nov 2016 05:53:52 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"I found the whole TableExprGetTupleDesc() function a bit odd in\\nnodeFuncs.c, so I renamed it to ExecTypeFromTableExpr() and moved it to\\nexecTuples.c -- but only because that's where ExecTypeFromTL and others\\nalready live. I would have liked to move it to tupdesc.c instead, but\\nit requires knowledge of executor nodes, which is probably the reason\\nthat ExecTypeFromTL is in execTuples. I think we'd eat that bit of\\nugliness only because we're not the first. But anyway I quickly ran\\ninto another problem.\\n\\nI noticed that ExecTypeFromTableExpr is being called from the transform\\nphase, which is much earlier than the executor. I noticed because of\\nthe warning that the above movement added to nodeFuncs.c,\\nsrc/backend/nodes/nodeFuncs.c:509:5: warning: implicit declaration of function 'ExecTypeFromTableExpr' [-Wimplicit-function-declaration]\\n\\nso I thought, hm, is it okay to have parse analysis run an executor\\nfunction? (I suppose this is the reason you put it in nodeFuncs in the\\nfirst place). For fun, I tried this query under GDB, with a breakpoint\\non exprTypmod():\\n\\nSELECT X.* \\nFROM emp, \\nXMLTABLE ('//depts/dept/employee' passing doc \\n COLUMNS \\n empID INTEGER PATH '@id',\\n firstname int PATH 'name/first',\\n lastname VARCHAR(25) PATH 'name/last') AS X;\\n\\nand sure enough, the type is resolved during parse analysis:\\n\\nBreakpoint 1, exprTypmod (expr=expr@entry=0x1d23ad8)\\n at /pgsql/source/master/src/backend/nodes/nodeFuncs.c:283\\n283 if (!expr)\\n(gdb) print *expr\\n$2 = {type = T_TableExpr}\\n(gdb) bt\\n#0 exprTypmod (expr=expr@entry=0x1d23ad8)\\n at /pgsql/source/master/src/backend/nodes/nodeFuncs.c:283\\n#1 0x000000000080c500 in get_expr_result_type (expr=0x1d23ad8, \\n resultTypeId=0x7ffd482bfdb4, resultTupleDesc=0x7ffd482bfdb8)\\n at /pgsql/source/master/src/backend/utils/fmgr/funcapi.c:247\\n#2 0x000000000056de1b in expandRTE (rte=rte@entry=0x1d6b800, rtindex=2, \\n sublevels_up=0, location=location@entry=7, \\n include_dropped=include_dropped@entry=0 '\\\\000', \\n colnames=colnames@entry=0x7ffd482bfe10, colvars=0x7ffd482bfe18)\\n at /pgsql/source/master/src/backend/parser/parse_relation.c:2052\\n#3 0x000000000056e131 in expandRelAttrs (pstate=pstate@entry=0x1d238a8, \\n rte=rte@entry=0x1d6b800, rtindex=<optimized out>, \\n sublevels_up=<optimized out>, location=location@entry=7)\\n at /pgsql/source/master/src/backend/parser/parse_relation.c:2435\\n#4 0x000000000056fa64 in ExpandSingleTable (pstate=pstate@entry=0x1d238a8, \\n rte=rte@entry=0x1d6b800, location=7, \\n make_target_entry=make_target_entry@entry=1 '\\\\001')\\n at /pgsql/source/master/src/backend/parser/parse_target.c:1266\\n#5 0x000000000057135b in ExpandColumnRefStar (pstate=pstate@entry=0x1d238a8, \\n cref=0x1d22720, make_target_entry=make_target_entry@entry=1 '\\\\001')\\n at /pgsql/source/master/src/backend/parser/parse_target.c:1158\\n#6 0x00000000005716f9 in transformTargetList (pstate=0x1d238a8, \\n targetlist=<optimized out>, exprKind=EXPR_KIND_SELECT_TARGET)\\n\\nThis seems fine I guess, and it seems to say that we ought to move the\\ncode that generates the tupdesc to back parse analysis rather than\\nexecutor. Okay, fine. But let's find a better place than nodeFuncs.\\n\\nBut if I move the XMLTABLE() call to the target list instead, the type\\nis resolved at planner time:\\n\\nSELECT \\nXMLTABLE ('/dept/employee' passing $$<dept bldg=\\\"114\\\"> \\n <employee id=\\\"903\\\">\\n <name>\\n <first>Mary</first>\\n <last>Jones</last>\\n </name>\\n <office>415</office>\\n <phone>905-403-6112</phone>\\n <phone>647-504-4546</phone>\\n <salary currency=\\\"USD\\\">64000</salary>\\n </employee>\\n</dept>$$\\n COLUMNS \\n empID INTEGER PATH '@id',\\n firstname varchar(4) PATH 'name/first',\\n lastname VARCHAR(25) PATH 'name/last') AS X;\\n\\nBreakpoint 1, exprTypmod (expr=expr@entry=0x1d6bed8)\\n at /pgsql/source/master/src/backend/nodes/nodeFuncs.c:283\\n283 if (!expr)\\n(gdb) bt\\n#0 exprTypmod (expr=expr@entry=0x1d6bed8)\\n at /pgsql/source/master/src/backend/nodes/nodeFuncs.c:283\\n#1 0x0000000000654058 in set_pathtarget_cost_width (root=0x1d6bc68, \\n target=0x1d6c728)\\n at /pgsql/source/master/src/backend/optimizer/path/costsize.c:4729\\n#2 0x000000000066c197 in grouping_planner (root=0x1d6bc68, \\n inheritance_update=40 '(', inheritance_update@entry=0 '\\\\000', \\n tuple_fraction=0.01, tuple_fraction@entry=0)\\n at /pgsql/source/master/src/backend/optimizer/plan/planner.c:1745\\n#3 0x000000000066ef64 in subquery_planner (glob=glob@entry=0x1d6bbd0, \\n parse=parse@entry=0x1d23818, parent_root=parent_root@entry=0x0, \\n hasRecursion=hasRecursion@entry=0 '\\\\000', \\n tuple_fraction=tuple_fraction@entry=0)\\n at /pgsql/source/master/src/backend/optimizer/plan/planner.c:795\\n#4 0x000000000066fe5e in standard_planner (parse=0x1d23818, \\n cursorOptions=256, boundParams=<optimized out>)\\n at /pgsql/source/master/src/backend/optimizer/plan/planner.c:307\\n\\nThis is surprising, but I'm not sure it's wrong.\\n\\n-- \\n�lvaro Herrera https://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\": \"Tue, 22 Nov 2016 17:47:30 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-11-22 21:47 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> I found the whole TableExprGetTupleDesc() function a bit odd in\\n> nodeFuncs.c, so I renamed it to ExecTypeFromTableExpr() and moved it to\\n> execTuples.c -- but only because that's where ExecTypeFromTL and others\\n> already live. I would have liked to move it to tupdesc.c instead, but\\n> it requires knowledge of executor nodes, which is probably the reason\\n> that ExecTypeFromTL is in execTuples. I think we'd eat that bit of\\n> ugliness only because we're not the first. But anyway I quickly ran\\n> into another problem.\\n>\\n> I noticed that ExecTypeFromTableExpr is being called from the transform\\n> phase, which is much earlier than the executor. I noticed because of\\n> the warning that the above movement added to nodeFuncs.c,\\n> src/backend/nodes/nodeFuncs.c:509:5: warning: implicit declaration of\\n> function 'ExecTypeFromTableExpr' [-Wimplicit-function-declaration]\\n>\\n\\nThe tuple descriptor should not be serialized.\\n\\nWhen xmltable is called directly, then living tuple descriptor is used -\\ncreated in transform time. Another situation is when xmltable is used from\\nview, where transform time is skipped.\\n\\nOriginally I serialized generated type - but I had the problems with record\\ntypes - the current infrastructure expects serialization only real types.\\n\\nMy solution is a recheck of tuple descriptor in executor time. It is small\\noverhead - once per query - but it allows use xmltable from views without\\nnecessity to specify returned columns explicitly.\\n\\n\\n>\\n> so I thought, hm, is it okay to have parse analysis run an executor\\n> function? (I suppose this is the reason you put it in nodeFuncs in the\\n> first place). For fun, I tried this query under GDB, with a breakpoint\\n> on exprTypmod():\\n>\\n> SELECT X.*\\n> FROM emp,\\n> XMLTABLE ('//depts/dept/employee' passing doc\\n> COLUMNS\\n> empID INTEGER PATH '@id',\\n> firstname int PATH 'name/first',\\n> lastname VARCHAR(25) PATH 'name/last') AS X;\\n>\\n> and sure enough, the type is resolved during parse analysis:\\n>\\n> Breakpoint 1, exprTypmod (expr=expr@entry=0x1d23ad8)\\n> at /pgsql/source/master/src/backend/nodes/nodeFuncs.c:283\\n> 283 if (!expr)\\n> (gdb) print *expr\\n> $2 = {type = T_TableExpr}\\n> (gdb) bt\\n> #0 exprTypmod (expr=expr@entry=0x1d23ad8)\\n> at /pgsql/source/master/src/backend/nodes/nodeFuncs.c:283\\n> #1 0x000000000080c500 in get_expr_result_type (expr=0x1d23ad8,\\n> resultTypeId=0x7ffd482bfdb4, resultTupleDesc=0x7ffd482bfdb8)\\n> at /pgsql/source/master/src/backend/utils/fmgr/funcapi.c:247\\n> #2 0x000000000056de1b in expandRTE (rte=rte@entry=0x1d6b800, rtindex=2,\\n> sublevels_up=0, location=location@entry=7,\\n> include_dropped=include_dropped@entry=0 '\\\\000',\\n> colnames=colnames@entry=0x7ffd482bfe10, colvars=0x7ffd482bfe18)\\n> at /pgsql/source/master/src/backend/parser/parse_relation.c:2052\\n> #3 0x000000000056e131 in expandRelAttrs (pstate=pstate@entry=0x1d238a8,\\n> rte=rte@entry=0x1d6b800, rtindex=<optimized out>,\\n> sublevels_up=<optimized out>, location=location@entry=7)\\n> at /pgsql/source/master/src/backend/parser/parse_relation.c:2435\\n> #4 0x000000000056fa64 in ExpandSingleTable (pstate=pstate@entry=\\n> 0x1d238a8,\\n> rte=rte@entry=0x1d6b800, location=7,\\n> make_target_entry=make_target_entry@entry=1 '\\\\001')\\n> at /pgsql/source/master/src/backend/parser/parse_target.c:1266\\n> #5 0x000000000057135b in ExpandColumnRefStar (pstate=pstate@entry=\\n> 0x1d238a8,\\n> cref=0x1d22720, make_target_entry=make_target_entry@entry=1 '\\\\001')\\n> at /pgsql/source/master/src/backend/parser/parse_target.c:1158\\n> #6 0x00000000005716f9 in transformTargetList (pstate=0x1d238a8,\\n> targetlist=<optimized out>, exprKind=EXPR_KIND_SELECT_TARGET)\\n>\\n> This seems fine I guess, and it seems to say that we ought to move the\\n> code that generates the tupdesc to back parse analysis rather than\\n> executor. Okay, fine. But let's find a better place than nodeFuncs.\\n>\\n> But if I move the XMLTABLE() call to the target list instead, the type\\n> is resolved at planner time:\\n>\\n> SELECT\\n> XMLTABLE ('/dept/employee' passing $$<dept bldg=\\\"114\\\">\\n> <employee id=\\\"903\\\">\\n> <name>\\n> <first>Mary</first>\\n> <last>Jones</last>\\n> </name>\\n> <office>415</office>\\n> <phone>905-403-6112</phone>\\n> <phone>647-504-4546</phone>\\n> <salary currency=\\\"USD\\\">64000</salary>\\n> </employee>\\n> </dept>$$\\n> COLUMNS\\n> empID INTEGER PATH '@id',\\n> firstname varchar(4) PATH 'name/first',\\n> lastname VARCHAR(25) PATH 'name/last') AS X;\\n>\\n> Breakpoint 1, exprTypmod (expr=expr@entry=0x1d6bed8)\\n> at /pgsql/source/master/src/backend/nodes/nodeFuncs.c:283\\n> 283 if (!expr)\\n> (gdb) bt\\n> #0 exprTypmod (expr=expr@entry=0x1d6bed8)\\n> at /pgsql/source/master/src/backend/nodes/nodeFuncs.c:283\\n> #1 0x0000000000654058 in set_pathtarget_cost_width (root=0x1d6bc68,\\n> target=0x1d6c728)\\n> at /pgsql/source/master/src/backend/optimizer/path/costsize.c:4729\\n> #2 0x000000000066c197 in grouping_planner (root=0x1d6bc68,\\n> inheritance_update=40 '(', inheritance_update@entry=0 '\\\\000',\\n> tuple_fraction=0.01, tuple_fraction@entry=0)\\n> at /pgsql/source/master/src/backend/optimizer/plan/planner.c:1745\\n> #3 0x000000000066ef64 in subquery_planner (glob=glob@entry=0x1d6bbd0,\\n> parse=parse@entry=0x1d23818, parent_root=parent_root@entry=0x0,\\n> hasRecursion=hasRecursion@entry=0 '\\\\000',\\n> tuple_fraction=tuple_fraction@entry=0)\\n> at /pgsql/source/master/src/backend/optimizer/plan/planner.c:795\\n> #4 0x000000000066fe5e in standard_planner (parse=0x1d23818,\\n> cursorOptions=256, boundParams=<optimized out>)\\n> at /pgsql/source/master/src/backend/optimizer/plan/planner.c:307\\n>\\n> This is surprising, but I'm not sure it's wrong.\\n>\\n\\nThere are different processing for Set Returning nodes called from\\nparamlist and from tablelist. In last case the invokes exprTypmod early.\\n\\nThere is a different case, that you didn't check\\n\\nCREATE VIEW x AS SELECT xmltable(...)\\nCREATE VIEW x1 AS SELECT * FROM xmltable(...)\\n\\nclose session\\n\\nand in new session\\nSELECT * FROM x;\\nSELECT * FROM x1;\\n\\nRegards\\n\\nPavel\\n\\n\\n\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2016-11-22 21:47 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:I found the whole TableExprGetTupleDesc() function a bit odd in\\nnodeFuncs.c, so I renamed it to ExecTypeFromTableExpr() and moved it to\\nexecTuples.c -- but only because that's where ExecTypeFromTL and others\\nalready live. I would have liked to move it to tupdesc.c instead, but\\nit requires knowledge of executor nodes, which is probably the reason\\nthat ExecTypeFromTL is in execTuples. I think we'd eat that bit of\\nugliness only because we're not the first. But anyway I quickly ran\\ninto another problem.\\n\\nI noticed that ExecTypeFromTableExpr is being called from the transform\\nphase, which is much earlier than the executor. I noticed because of\\nthe warning that the above movement added to nodeFuncs.c,\\nsrc/backend/nodes/nodeFuncs.c:509:5: warning: implicit declaration of function 'ExecTypeFromTableExpr' [-Wimplicit-function-declaration]The tuple descriptor should not be serialized.When xmltable is called directly, then living tuple descriptor is used - created in transform time. Another situation is when xmltable is used from view, where transform time is skipped. Originally I serialized generated type - but I had the problems with record types - the current infrastructure expects serialization only real types.My solution is a recheck of tuple descriptor in executor time. It is small overhead - once per query - but it allows use xmltable from views without necessity to specify returned columns explicitly. \\n\\nso I thought, hm, is it okay to have parse analysis run an executor\\nfunction? (I suppose this is the reason you put it in nodeFuncs in the\\nfirst place). For fun, I tried this query under GDB, with a breakpoint\\non exprTypmod():\\n\\nSELECT X.*\\nFROM emp,\\nXMLTABLE ('//depts/dept/employee' passing doc\\n COLUMNS\\n empID INTEGER PATH '@id',\\n firstname int PATH 'name/first',\\n lastname VARCHAR(25) PATH 'name/last') AS X;\\n\\nand sure enough, the type is resolved during parse analysis:\\n\\nBreakpoint 1, exprTypmod (expr=expr@entry=0x1d23ad8)\\n at /pgsql/source/master/src/backend/nodes/nodeFuncs.c:283\\n283 if (!expr)\\n(gdb) print *expr\\n$2 = {type = T_TableExpr}\\n(gdb) bt\\n#0 exprTypmod (expr=expr@entry=0x1d23ad8)\\n at /pgsql/source/master/src/backend/nodes/nodeFuncs.c:283\\n#1 0x000000000080c500 in get_expr_result_type (expr=0x1d23ad8,\\n resultTypeId=0x7ffd482bfdb4, resultTupleDesc=0x7ffd482bfdb8)\\n at /pgsql/source/master/src/backend/utils/fmgr/funcapi.c:247\\n#2 0x000000000056de1b in expandRTE (rte=rte@entry=0x1d6b800, rtindex=2,\\n sublevels_up=0, location=location@entry=7,\\n include_dropped=include_dropped@entry=0 '\\\\000',\\n colnames=colnames@entry=0x7ffd482bfe10, colvars=0x7ffd482bfe18)\\n at /pgsql/source/master/src/backend/parser/parse_relation.c:2052\\n#3 0x000000000056e131 in expandRelAttrs (pstate=pstate@entry=0x1d238a8,\\n rte=rte@entry=0x1d6b800, rtindex=<optimized out>,\\n sublevels_up=<optimized out>, location=location@entry=7)\\n at /pgsql/source/master/src/backend/parser/parse_relation.c:2435\\n#4 0x000000000056fa64 in ExpandSingleTable (pstate=pstate@entry=0x1d238a8,\\n rte=rte@entry=0x1d6b800, location=7,\\n make_target_entry=make_target_entry@entry=1 '\\\\001')\\n at /pgsql/source/master/src/backend/parser/parse_target.c:1266\\n#5 0x000000000057135b in ExpandColumnRefStar (pstate=pstate@entry=0x1d238a8,\\n cref=0x1d22720, make_target_entry=make_target_entry@entry=1 '\\\\001')\\n at /pgsql/source/master/src/backend/parser/parse_target.c:1158\\n#6 0x00000000005716f9 in transformTargetList (pstate=0x1d238a8,\\n targetlist=<optimized out>, exprKind=EXPR_KIND_SELECT_TARGET)\\n\\nThis seems fine I guess, and it seems to say that we ought to move the\\ncode that generates the tupdesc to back parse analysis rather than\\nexecutor. Okay, fine. But let's find a better place than nodeFuncs.\\n\\nBut if I move the XMLTABLE() call to the target list instead, the type\\nis resolved at planner time:\\n\\nSELECT\\nXMLTABLE ('/dept/employee' passing $$<dept bldg=\\\"114\\\">\\n <employee id=\\\"903\\\">\\n <name>\\n <first>Mary</first>\\n <last>Jones</last>\\n </name>\\n <office>415</office>\\n <phone>905-403-6112</phone>\\n <phone>647-504-4546</phone>\\n <salary currency=\\\"USD\\\">64000</salary>\\n </employee>\\n</dept>$$\\n COLUMNS\\n empID INTEGER PATH '@id',\\n firstname varchar(4) PATH 'name/first',\\n lastname VARCHAR(25) PATH 'name/last') AS X;\\n\\nBreakpoint 1, exprTypmod (expr=expr@entry=0x1d6bed8)\\n at /pgsql/source/master/src/backend/nodes/nodeFuncs.c:283\\n283 if (!expr)\\n(gdb) bt\\n#0 exprTypmod (expr=expr@entry=0x1d6bed8)\\n at /pgsql/source/master/src/backend/nodes/nodeFuncs.c:283\\n#1 0x0000000000654058 in set_pathtarget_cost_width (root=0x1d6bc68,\\n target=0x1d6c728)\\n at /pgsql/source/master/src/backend/optimizer/path/costsize.c:4729\\n#2 0x000000000066c197 in grouping_planner (root=0x1d6bc68,\\n inheritance_update=40 '(', inheritance_update@entry=0 '\\\\000',\\n tuple_fraction=0.01, tuple_fraction@entry=0)\\n at /pgsql/source/master/src/backend/optimizer/plan/planner.c:1745\\n#3 0x000000000066ef64 in subquery_planner (glob=glob@entry=0x1d6bbd0,\\n parse=parse@entry=0x1d23818, parent_root=parent_root@entry=0x0,\\n hasRecursion=hasRecursion@entry=0 '\\\\000',\\n tuple_fraction=tuple_fraction@entry=0)\\n at /pgsql/source/master/src/backend/optimizer/plan/planner.c:795\\n#4 0x000000000066fe5e in standard_planner (parse=0x1d23818,\\n cursorOptions=256, boundParams=<optimized out>)\\n at /pgsql/source/master/src/backend/optimizer/plan/planner.c:307\\n\\nThis is surprising, but I'm not sure it's wrong.There are different processing for Set Returning nodes called from paramlist and from tablelist. In last case the invokes exprTypmod early.There is a different case, that you didn't checkCREATE VIEW x AS SELECT xmltable(...)CREATE VIEW x1 AS SELECT * FROM xmltable(...)close sessionand in new sessionSELECT * FROM x;SELECT * FROM x1;RegardsPavel \\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Wed, 23 Nov 2016 06:16:33 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"I tried to see if a following RTE was able to \\\"see\\\" the entries created by\\nXMLTABLE, and sure it can:\\n\\nSELECT X.*, generate_series \\nFROM emp, \\nXMLTABLE ('//depts/dept/employee' passing doc \\n COLUMNS \\n empID INTEGER PATH '@id',\\n firstname varchar(25) PATH 'name/first',\\n lastname VARCHAR(25) PATH 'name/last') AS X,\\ngenerate_series(900, empid);\\n\\n empid │ firstname │ lastname │ generate_series \\n───────┼───────────┼──────────┼─────────────────\\n 901 │ John │ Doe │ 900\\n 901 │ John │ Doe │ 901\\n 902 │ Peter │ Pan │ 900\\n 902 │ Peter │ Pan │ 901\\n 902 │ Peter │ Pan │ 902\\n 903 │ Mary │ Jones │ 900\\n 903 │ Mary │ Jones │ 901\\n 903 │ Mary │ Jones │ 902\\n 903 │ Mary │ Jones │ 903\\n(9 filas)\\n\\nCool.\\n\\nI'm still wondering how this works. I'll continue to explore the patch\\nin order to figure out.\\n\\n-- \\nÁlvaro Herrera https://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, 23 Nov 2016 13:54:48 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera wrote:\\n\\n> I'm still wondering how this works. I'll continue to explore the patch\\n> in order to figure out.\\n\\nAh, so it's already parsed as a \\\"range function\\\". That part looks good\\nto me.\\n\\n-- \\n�lvaro Herrera https://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, 23 Nov 2016 14:49:37 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Oh my, I just noticed we have a new xpath preprocessor in this patch\\ntoo. Where did this code come from -- did you write it all from\\nscratch?\\n\\n-- \\n�lvaro Herrera https://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, 23 Nov 2016 20:13:30 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Here's another version. Not there yet: need to move back the function\\nto create the tupdesc, as discussed. Not clear what's the best place,\\nhowever. I modified the grammar a bit (added the missing comma, removed\\nPATH as an unreserved keyword and just used IDENT, removed the \\\"Opt\\\"\\nversion for column options), and reworked the comments in the transform\\nphase (I tweaked the code here and there mostly to move things to nicer\\nplaces, but it's pretty much the same code).\\n\\nIn the new xpath_parser.c file I think we should tidy things up a bit.\\nFirst, it needs more commentary on what the entry function actually\\ndoes, in detail. Also, IMO that function should be at the top of the\\nfile, not at the bottom, followed by all its helpers. I would like some\\nmore clarity on the provenance of all this code, just to assess the\\nprobability of bugs; mostly as it's completely undocumented.\\n\\nI don't like the docs either. I think we should have a complete\\nreference to the syntax, followed by examples, rather than letting the\\nexamples drive the whole thing. I fixed the synopsis so that it's not\\none very long line.\\n\\nIf you use \\\"PATH '/'\\\" for a column, you get the text for all the entries\\nin the whole XML, rather than the text for the particular row being\\nprocessed. Isn't that rather weird, or to put it differently, completely\\nwrong? I didn't find a way to obtain the whole XML row when you have\\nthe COLUMNS option (which is what I was hoping for with the \\\"PATH '/'\\\").\\n\\n-- \\n�lvaro Herrera https://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, 23 Nov 2016 20:29:10 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Sorry, here's the patch. Power loss distracted me here.\\n\\nBy the way, the pgindent you did is slightly confused because you failed\\nto add the new struct types you define to typedefs.list.\\n\\nI have not looked at the new xml.c code at all, yet.\\n\\n-- \\n�lvaro Herrera https://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\": \"Wed, 23 Nov 2016 20:31:30 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 11/23/2016 06:31 PM, Alvaro Herrera wrote:\\n> Sorry, here's the patch. Power loss distracted me here.\\n>\\n> By the way, the pgindent you did is slightly confused because you failed\\n> to add the new struct types you define to typedefs.list.\\n\\n\\nTips on how to use pgindent for developers:\\n\\n<http://adpgtech.blogspot.com/2015/05/running-pgindent-on-non-core-code-or.html>\\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, 23 Nov 2016 18:44:16 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera wrote:\\n\\n> If you use \\\"PATH '/'\\\" for a column, you get the text for all the entries\\n> in the whole XML, rather than the text for the particular row being\\n> processed. Isn't that rather weird, or to put it differently, completely\\n> wrong? I didn't find a way to obtain the whole XML row when you have\\n> the COLUMNS option (which is what I was hoping for with the \\\"PATH '/'\\\").\\n\\nAh, apparently you need to use type XML for that column in order for\\nthis to happen. Example:\\n\\ninsert into emp values ($$\\n<depts >\\n <dept bldg=\\\"102\\\">\\n <employee id=\\\"905\\\">\\n <name>\\n <first>John</first>\\n <last>Doew</last>\\n </name> \\n <office>344</office> \\n <salary currency=\\\"USD\\\">55000</salary>\\n </employee>\\n\\n <employee id=\\\"908\\\">\\n <name>\\n <first>Peter</first>\\n <last>Panw</last>\\n </name>\\n <office>216</office>\\n <phone>905-416-5004</phone>\\n </employee>\\n </dept>\\n\\n <dept bldg=\\\"115\\\">\\n <employee id=\\\"909\\\">\\n <name>\\n <first>Mary</first>\\n <last>Jonesw</last>\\n </name>\\n <office>415</office>\\n <phone>905-403-6112</phone>\\n <phone>647-504-4546</phone>\\n <salary currency=\\\"USD\\\">64000</salary>\\n </employee>\\n </dept>\\n</depts>\\n$$);\\n\\nNote the weird salary_amount value here:\\n\\nSELECT x.*\\nFROM emp, \\nXMLTABLE ('//depts/dept/employee' passing doc \\n COLUMNS \\n i for ordinality,\\n empID int PATH '@id',\\n firstname varchar(25) PATH 'name/first' default 'FOOBAR',\\n lastname VARCHAR(25) PATH 'name/last',\\n salary xml path 'concat(salary/text(), salary/@currency)' default 'DONT KNOW', salary_amount xml path '/' )\\n WITH ORDINALITY\\n AS X (i, a, b, c) limit 1;\\n i │ a │ b │ c │ salary │ salary_amount │ ordinality \\n───┼─────┼──────┼──────┼──────────┼───────────────────────┼────────────\\n 1 │ 905 │ John │ Doew │ 55000USD │ ↵│ 1\\n │ │ │ │ │ ↵│ \\n │ │ │ │ │ ↵│ \\n │ │ │ │ │ John ↵│ \\n │ │ │ │ │ Doew ↵│ \\n │ │ │ │ │ ↵│ \\n │ │ │ │ │ 344 ↵│ \\n │ │ │ │ │ 55000 ↵│ \\n │ │ │ │ │ ↵│ \\n │ │ │ │ │ ↵│ \\n │ │ │ │ │ ↵│ \\n │ │ │ │ │ ↵│ \\n │ │ │ │ │ Peter ↵│ \\n │ │ │ │ │ Panw ↵│ \\n │ │ │ │ │ ↵│ \\n │ │ │ │ │ 216 ↵│ \\n │ │ │ │ │ 905-416-5004↵│ \\n │ │ │ │ │ ↵│ \\n │ │ │ │ │ ↵│ \\n │ │ │ │ │ ↵│ \\n │ │ │ │ │ ↵│ \\n │ │ │ │ │ ↵│ \\n │ │ │ │ │ ↵│ \\n │ │ │ │ │ Mary ↵│ \\n │ │ │ │ │ Jonesw ↵│ \\n │ │ │ │ │ ↵│ \\n │ │ │ │ │ 415 ↵│ \\n │ │ │ │ │ 905-403-6112↵│ \\n │ │ │ │ │ 647-504-4546↵│ \\n │ │ │ │ │ 64000 ↵│ \\n │ │ │ │ │ ↵│ \\n │ │ │ │ │ ↵│ \\n │ │ │ │ │ │ \\n(1 fila)\\n\\n\\nIf you declare salary_amount to be text instead, it doesn't happen anymore.\\nApparently if you put it in a namespace, it doesn't hapen either.\\n\\n-- \\nÁlvaro Herrera https://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, 24 Nov 2016 00:26:20 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2016-11-24 0:13 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Oh my, I just noticed we have a new xpath preprocessor in this patch\\n> too. Where did this code come from -- did you write it all from\\n> scratch?\\n>\\n\\nI wrote it from scratch - libxml2 has not any API for iteration over XPath\\nexpression (different than iteration over XPath expression result), and\\nwhat I have info, there will not be any new API in libxml2.\\n\\nThere are two purposes:\\n\\nSafe manipulation with XPath expression prefixes - ANSI SQL design\\nimplicitly expects some prefix, but it can be used manually. The prefix\\nshould not be used twice and in some situations, when it can breaks the\\nexpression.\\n\\nSecond goal is support default namespaces - when we needed parser for first\\ntask, then the enhancing for this task was not too much lines more.\\n\\nThis parser can be used for enhancing current XPath function - default\\nnamespaces are pretty nice, when you have to use namespaces.\\n\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\nHi2016-11-24 0:13 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Oh my, I just noticed we have a new xpath preprocessor in this patch\\ntoo. Where did this code come from -- did you write it all from\\nscratch?I wrote it from scratch - libxml2 has not any API for iteration over XPath expression (different than iteration over XPath expression result), and what I have info, there will not be any new API in libxml2.There are two purposes:Safe manipulation with XPath expression prefixes - ANSI SQL design implicitly expects some prefix, but it can be used manually. The prefix should not be used twice and in some situations, when it can breaks the expression.Second goal is support default namespaces - when we needed parser for first task, then the enhancing for this task was not too much lines more.This parser can be used for enhancing current XPath function - default namespaces are pretty nice, when you have to use namespaces. \\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Thu, 24 Nov 2016 05:52:46 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-11-24 0:29 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Here's another version. Not there yet: need to move back the function\\n> to create the tupdesc, as discussed. Not clear what's the best place,\\n> however. I modified the grammar a bit (added the missing comma, removed\\n> PATH as an unreserved keyword and just used IDENT, removed the \\\"Opt\\\"\\n> version for column options), and reworked the comments in the transform\\n> phase (I tweaked the code here and there mostly to move things to nicer\\n> places, but it's pretty much the same code).\\n>\\n> In the new xpath_parser.c file I think we should tidy things up a bit.\\n> First, it needs more commentary on what the entry function actually\\n> does, in detail. Also, IMO that function should be at the top of the\\n> file, not at the bottom, followed by all its helpers. I would like some\\n> more clarity on the provenance of all this code, just to assess the\\n> probability of bugs; mostly as it's completely undocumented.\\n>\\n> I don't like the docs either. I think we should have a complete\\n> reference to the syntax, followed by examples, rather than letting the\\n> examples drive the whole thing. I fixed the synopsis so that it's not\\n> one very long line.\\n>\\n> If you use \\\"PATH '/'\\\" for a column, you get the text for all the entries\\n> in the whole XML, rather than the text for the particular row being\\n> processed. Isn't that rather weird, or to put it differently, completely\\n> wrong? I didn't find a way to obtain the whole XML row when you have\\n> the COLUMNS option (which is what I was hoping for with the \\\"PATH '/'\\\").\\n>\\n\\nThis is a libxml2 behave\\n\\nPostprocessing only check result and try to push the result to expected\\ntypes.\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2016-11-24 0:29 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Here's another version. Not there yet: need to move back the function\\nto create the tupdesc, as discussed. Not clear what's the best place,\\nhowever. I modified the grammar a bit (added the missing comma, removed\\nPATH as an unreserved keyword and just used IDENT, removed the \\\"Opt\\\"\\nversion for column options), and reworked the comments in the transform\\nphase (I tweaked the code here and there mostly to move things to nicer\\nplaces, but it's pretty much the same code).\\n\\nIn the new xpath_parser.c file I think we should tidy things up a bit.\\nFirst, it needs more commentary on what the entry function actually\\ndoes, in detail. Also, IMO that function should be at the top of the\\nfile, not at the bottom, followed by all its helpers. I would like some\\nmore clarity on the provenance of all this code, just to assess the\\nprobability of bugs; mostly as it's completely undocumented.\\n\\nI don't like the docs either. I think we should have a complete\\nreference to the syntax, followed by examples, rather than letting the\\nexamples drive the whole thing. I fixed the synopsis so that it's not\\none very long line.\\n\\nIf you use \\\"PATH '/'\\\" for a column, you get the text for all the entries\\nin the whole XML, rather than the text for the particular row being\\nprocessed. Isn't that rather weird, or to put it differently, completely\\nwrong? I didn't find a way to obtain the whole XML row when you have\\nthe COLUMNS option (which is what I was hoping for with the \\\"PATH '/'\\\").This is a libxml2 behave Postprocessing only check result and try to push the result to expected types.RegardsPavel \\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Thu, 24 Nov 2016 06:04:23 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-11-24 5:52 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n\\n> Hi\\n>\\n> 2016-11-24 0:13 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>\\n>> Oh my, I just noticed we have a new xpath preprocessor in this patch\\n>> too. Where did this code come from -- did you write it all from\\n>> scratch?\\n>>\\n>\\n> I wrote it from scratch - libxml2 has not any API for iteration over XPath\\n> expression (different than iteration over XPath expression result), and\\n> what I have info, there will not be any new API in libxml2.\\n>\\n> There are two purposes:\\n>\\n> Safe manipulation with XPath expression prefixes - ANSI SQL design\\n> implicitly expects some prefix, but it can be used manually. The prefix\\n> should not be used twice and in some situations, when it can breaks the\\n> expression.\\n>\\n\\nImplicit prefix for column PATH expressions is \\\"./\\\". An user can use it\\nexplicitly.\\n\\nIn my initial patches, the manipulations with XPath expression was more\\ncomplex - now, it can be reduced - but then we lost default namespaces\\nsupport, what is nice feature, supported other providers.\\n\\n\\n\\n\\n>\\n> Second goal is support default namespaces - when we needed parser for\\n> first task, then the enhancing for this task was not too much lines more.\\n>\\n> This parser can be used for enhancing current XPath function - default\\n> namespaces are pretty nice, when you have to use namespaces.\\n>\\n>\\n>>\\n>> --\\n>> Álvaro Herrera https://www.2ndQuadrant.com/\\n>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>>\\n>\\n>\\n\\n2016-11-24 5:52 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:Hi2016-11-24 0:13 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Oh my, I just noticed we have a new xpath preprocessor in this patch\\ntoo. Where did this code come from -- did you write it all from\\nscratch?I wrote it from scratch - libxml2 has not any API for iteration over XPath expression (different than iteration over XPath expression result), and what I have info, there will not be any new API in libxml2.There are two purposes:Safe manipulation with XPath expression prefixes - ANSI SQL design implicitly expects some prefix, but it can be used manually. The prefix should not be used twice and in some situations, when it can breaks the expression.Implicit prefix for column PATH expressions is \\\"./\\\". An user can use it explicitly. In my initial patches, the manipulations with XPath expression was more complex - now, it can be reduced - but then we lost default namespaces support, what is nice feature, supported other providers. Second goal is support default namespaces - when we needed parser for first task, then the enhancing for this task was not too much lines more.This parser can be used for enhancing current XPath function - default namespaces are pretty nice, when you have to use namespaces. \\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Thu, 24 Nov 2016 06:18:31 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n> Hi\\n> \\n> 2016-11-24 0:13 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n> \\n> > Oh my, I just noticed we have a new xpath preprocessor in this patch\\n> > too. Where did this code come from -- did you write it all from\\n> > scratch?\\n> \\n> I wrote it from scratch - libxml2 has not any API for iteration over XPath\\n> expression (different than iteration over XPath expression result), and\\n> what I have info, there will not be any new API in libxml2.\\n\\nOkay, I agree that the default namespace stuff looks worthwhile in the\\nlong run. But I don't have enough time to review the xpath parser stuff\\nin the current commitfest, and I think it needs at the very least a lot\\nof additional code commentary.\\n\\nHowever I think the rest of it can reasonably go in -- I mean the SQL\\nparse of it, analysis, executor. Let me propose this: you split the\\npatch, leaving the xpath_parser.c stuff out and XMLNAMESPACES DEFAULT,\\nand we introduce just the TableExpr stuff plus the XMLTABLE function. I\\ncan commit that part in the current commitfest, and we leave the\\nxpath_parser plus associated features for the upcoming commitfest.\\n\\nDeal?\\n\\n-- \\n�lvaro Herrera https://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, 24 Nov 2016 14:29:46 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-11-24 18:29 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n> > Hi\\n> >\\n> > 2016-11-24 0:13 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n> >\\n> > > Oh my, I just noticed we have a new xpath preprocessor in this patch\\n> > > too. Where did this code come from -- did you write it all from\\n> > > scratch?\\n> >\\n> > I wrote it from scratch - libxml2 has not any API for iteration over\\n> XPath\\n> > expression (different than iteration over XPath expression result), and\\n> > what I have info, there will not be any new API in libxml2.\\n>\\n> Okay, I agree that the default namespace stuff looks worthwhile in the\\n> long run. But I don't have enough time to review the xpath parser stuff\\n> in the current commitfest, and I think it needs at the very least a lot\\n> of additional code commentary.\\n>\\n> However I think the rest of it can reasonably go in -- I mean the SQL\\n> parse of it, analysis, executor. Let me propose this: you split the\\n> patch, leaving the xpath_parser.c stuff out and XMLNAMESPACES DEFAULT,\\n> and we introduce just the TableExpr stuff plus the XMLTABLE function. I\\n> can commit that part in the current commitfest, and we leave the\\n> xpath_parser plus associated features for the upcoming commitfest.\\n\\n\\n> Deal?\\n>\\n\\nok\\n\\ncan me send your last work?\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2016-11-24 18:29 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\\n> Hi\\n>\\n> 2016-11-24 0:13 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>\\n> > Oh my, I just noticed we have a new xpath preprocessor in this patch\\n> > too. Where did this code come from -- did you write it all from\\n> > scratch?\\n>\\n> I wrote it from scratch - libxml2 has not any API for iteration over XPath\\n> expression (different than iteration over XPath expression result), and\\n> what I have info, there will not be any new API in libxml2.\\n\\nOkay, I agree that the default namespace stuff looks worthwhile in the\\nlong run. But I don't have enough time to review the xpath parser stuff\\nin the current commitfest, and I think it needs at the very least a lot\\nof additional code commentary.\\n\\nHowever I think the rest of it can reasonably go in -- I mean the SQL\\nparse of it, analysis, executor. Let me propose this: you split the\\npatch, leaving the xpath_parser.c stuff out and XMLNAMESPACES DEFAULT,\\nand we introduce just the TableExpr stuff plus the XMLTABLE function. I\\ncan commit that part in the current commitfest, and we leave the\\nxpath_parser plus associated features for the upcoming commitfest. \\n\\nDeal?okcan me send your last work?RegardsPavel \\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Thu, 24 Nov 2016 18:32:47 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n\\n> can me send your last work?\\n\\nSure, it's in the archives --\\nhttps://www.postgresql.org/message-id/20161123233130.oqf7jl6czehy5fiw@alvherre.pgsql\\n\\n-- \\n�lvaro Herrera https://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, 24 Nov 2016 14:41:19 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n> Pavel Stehule wrote:\\n>> 2016-11-24 0:13 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>>> Oh my, I just noticed we have a new xpath preprocessor in this patch\\n>>> too. Where did this code come from -- did you write it all from\\n>>> scratch?\\n\\n>> I wrote it from scratch - libxml2 has not any API for iteration over XPath\\n>> expression (different than iteration over XPath expression result), and\\n>> what I have info, there will not be any new API in libxml2.\\n\\n> Okay, I agree that the default namespace stuff looks worthwhile in the\\n> long run. But I don't have enough time to review the xpath parser stuff\\n> in the current commitfest, and I think it needs at the very least a lot\\n> of additional code commentary.\\n\\ncontrib/xml2 has always relied on libxslt for xpath functionality.\\nCan we do that here instead of writing, debugging, and documenting\\na pile of new code?\\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, 24 Nov 2016 12:51:52 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-11-24 18:51 GMT+01:00 Tom Lane <tgl@sss.pgh.pa.us>:\\n\\n> Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n> > Pavel Stehule wrote:\\n> >> 2016-11-24 0:13 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n> >>> Oh my, I just noticed we have a new xpath preprocessor in this patch\\n> >>> too. Where did this code come from -- did you write it all from\\n> >>> scratch?\\n>\\n> >> I wrote it from scratch - libxml2 has not any API for iteration over\\n> XPath\\n> >> expression (different than iteration over XPath expression result), and\\n> >> what I have info, there will not be any new API in libxml2.\\n>\\n> > Okay, I agree that the default namespace stuff looks worthwhile in the\\n> > long run. But I don't have enough time to review the xpath parser stuff\\n> > in the current commitfest, and I think it needs at the very least a lot\\n> > of additional code commentary.\\n>\\n> contrib/xml2 has always relied on libxslt for xpath functionality.\\n> Can we do that here instead of writing, debugging, and documenting\\n> a pile of new code?\\n>\\n\\nI am sorry - I don't see it. There is nothing complex manipulation with\\nXPath expressions.\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> regards, tom lane\\n>\\n\\n2016-11-24 18:51 GMT+01:00 Tom Lane <tgl@sss.pgh.pa.us>:Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n> Pavel Stehule wrote:\\n>> 2016-11-24 0:13 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>>> Oh my, I just noticed we have a new xpath preprocessor in this patch\\n>>> too. Where did this code come from -- did you write it all from\\n>>> scratch?\\n\\n>> I wrote it from scratch - libxml2 has not any API for iteration over XPath\\n>> expression (different than iteration over XPath expression result), and\\n>> what I have info, there will not be any new API in libxml2.\\n\\n> Okay, I agree that the default namespace stuff looks worthwhile in the\\n> long run. But I don't have enough time to review the xpath parser stuff\\n> in the current commitfest, and I think it needs at the very least a lot\\n> of additional code commentary.\\n\\ncontrib/xml2 has always relied on libxslt for xpath functionality.\\nCan we do that here instead of writing, debugging, and documenting\\na pile of new code?I am sorry - I don't see it. There is nothing complex manipulation with XPath expressions.RegardsPavel \\n\\n regards, tom lane\",\n \"msg_date\": \"Thu, 24 Nov 2016 19:31:55 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-11-24 18:41 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n>\\n> > can me send your last work?\\n>\\n> Sure, it's in the archives --\\n> https://www.postgresql.org/message-id/20161123233130.\\n> oqf7jl6czehy5fiw@alvherre.pgsql\\n\\n\\nHere is updated patch without default namespace support (and without XPath\\nexpression transformation).\\n\\nDue last changes in parser\\nhttps://github.com/postgres/postgres/commit/906bfcad7ba7cb3863fe0e2a7810be8e3cd84fbd\\nI had to use c_expr on other positions ( xmlnamespace definition).\\n\\nI don't think it is limit - in 99% there will be const literal.\\n\\nRegards\\n\\nPavel\\n\\n>\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL 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\": \"Fri, 25 Nov 2016 00:33:14 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 25, 2016 at 3:31 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n> 2016-11-24 18:51 GMT+01:00 Tom Lane <tgl@sss.pgh.pa.us>:\\n>> contrib/xml2 has always relied on libxslt for xpath functionality.\\n>> Can we do that here instead of writing, debugging, and documenting\\n>> a pile of new code?\\n>\\n> I am sorry - I don't see it. There is nothing complex manipulation with\\n> XPath expressions.\\n\\nYou are missing the point here, which is to make the implementation\\nfootprint as light as possible, especially if the added functionality\\nis already present in a dependency that Postgres can be linked to. OK,\\nlibxslt can only be linked with contrib/xml2/ now, but it would be at\\nleast worth looking at how much the current patch can be simplified\\nfor things like transformTableExpr or XmlTableGetValue by relying on\\nsome existing routines. Nit: I did not look at the patch in details,\\nbut I find the size of the latest version sent, 167kB, scary as it\\ncomplicates review and increases the likeliness of bugs.\\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, 25 Nov 2016 09:45:37 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Michael Paquier wrote:\\n\\n> Nit: I did not look at the patch in details,\\n> but I find the size of the latest version sent, 167kB, scary as it\\n> complicates review and increases the likeliness of bugs.\\n\\nHere's the stat. Note that removing the functionality as discussed\\nwould remove all of xpath_parser.c but I think the rest of it remains\\npretty much unchanged. So it's clearly a large patch, but there are\\nlarge docs and tests too, not just code.\\n\\n doc/src/sgml/func.sgml | 376 ++++++++++++++++++---\\n src/backend/executor/execQual.c | 335 +++++++++++++++++++\\n src/backend/executor/execTuples.c | 42 +++\\n src/backend/nodes/copyfuncs.c | 66 ++++\\n src/backend/nodes/equalfuncs.c | 51 +++\\n src/backend/nodes/nodeFuncs.c | 100 ++++++\\n src/backend/nodes/outfuncs.c | 51 +++\\n src/backend/nodes/readfuncs.c | 42 +++\\n src/backend/optimizer/util/clauses.c | 33 ++\\n src/backend/parser/gram.y | 181 ++++++++++-\\n src/backend/parser/parse_coerce.c | 33 +-\\n src/backend/parser/parse_expr.c | 182 +++++++++++\\n src/backend/parser/parse_target.c | 7 +\\n src/backend/utils/adt/Makefile | 2 +-\\n src/backend/utils/adt/ruleutils.c | 100 ++++++\\n src/backend/utils/adt/xml.c | 610 +++++++++++++++++++++++++++++++++++\\n src/backend/utils/adt/xpath_parser.c | 337 +++++++++++++++++++\\n src/backend/utils/fmgr/funcapi.c | 13 +\\n src/include/executor/executor.h | 1 +\\n src/include/executor/tableexpr.h | 69 ++++\\n src/include/funcapi.h | 1 -\\n src/include/nodes/execnodes.h | 31 ++\\n src/include/nodes/nodes.h | 4 +\\n src/include/nodes/parsenodes.h | 21 ++\\n src/include/nodes/primnodes.h | 40 +++\\n src/include/parser/kwlist.h | 3 +\\n src/include/parser/parse_coerce.h | 4 +\\n src/include/utils/xml.h | 2 +\\n src/include/utils/xpath_parser.h | 24 ++\\n src/test/regress/expected/xml.out | 415 ++++++++++++++++++++++++\\n src/test/regress/expected/xml_1.out | 323 +++++++++++++++++++\\n src/test/regress/expected/xml_2.out | 414 ++++++++++++++++++++++++\\n src/test/regress/sql/xml.sql | 170 ++++++++++\\n 33 files changed, 4019 insertions(+), 64 deletions(-)\\n\\n-- \\n�lvaro Herrera https://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, 24 Nov 2016 23:31:26 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-11-25 3:31 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Michael Paquier wrote:\\n>\\n> > Nit: I did not look at the patch in details,\\n> > but I find the size of the latest version sent, 167kB, scary as it\\n> > complicates review and increases the likeliness of bugs.\\n>\\n> Here's the stat. Note that removing the functionality as discussed\\n> would remove all of xpath_parser.c but I think the rest of it remains\\n> pretty much unchanged. So it's clearly a large patch, but there are\\n> large docs and tests too, not just code.\\n>\\n\\nyes, lot of is regress tests (expected part is 3x) - and XMLTABLE function\\nis not trivial.\\n\\nlot of code is mechanical - nodes related. The really complex part is only\\nin xml.c. There is one longer function only - the complexity is based on\\nmapping libxml2 result to PostgreSQL result (with catching exceptions due\\nreleasing libxml2 sources).\\n\\nThe all changes are well isolated - there is less risk to break some other.\\n\\n\\n>\\n> doc/src/sgml/func.sgml | 376 ++++++++++++++++++---\\n> src/backend/executor/execQual.c | 335 +++++++++++++++++++\\n> src/backend/executor/execTuples.c | 42 +++\\n> src/backend/nodes/copyfuncs.c | 66 ++++\\n> src/backend/nodes/equalfuncs.c | 51 +++\\n> src/backend/nodes/nodeFuncs.c | 100 ++++++\\n> src/backend/nodes/outfuncs.c | 51 +++\\n> src/backend/nodes/readfuncs.c | 42 +++\\n> src/backend/optimizer/util/clauses.c | 33 ++\\n> src/backend/parser/gram.y | 181 ++++++++++-\\n> src/backend/parser/parse_coerce.c | 33 +-\\n> src/backend/parser/parse_expr.c | 182 +++++++++++\\n> src/backend/parser/parse_target.c | 7 +\\n> src/backend/utils/adt/Makefile | 2 +-\\n> src/backend/utils/adt/ruleutils.c | 100 ++++++\\n> src/backend/utils/adt/xml.c | 610 ++++++++++++++++++++++++++++++\\n> +++++\\n> src/backend/utils/adt/xpath_parser.c | 337 +++++++++++++++++++\\n> src/backend/utils/fmgr/funcapi.c | 13 +\\n> src/include/executor/executor.h | 1 +\\n> src/include/executor/tableexpr.h | 69 ++++\\n> src/include/funcapi.h | 1 -\\n> src/include/nodes/execnodes.h | 31 ++\\n> src/include/nodes/nodes.h | 4 +\\n> src/include/nodes/parsenodes.h | 21 ++\\n> src/include/nodes/primnodes.h | 40 +++\\n> src/include/parser/kwlist.h | 3 +\\n> src/include/parser/parse_coerce.h | 4 +\\n> src/include/utils/xml.h | 2 +\\n> src/include/utils/xpath_parser.h | 24 ++\\n> src/test/regress/expected/xml.out | 415 ++++++++++++++++++++++++\\n> src/test/regress/expected/xml_1.out | 323 +++++++++++++++++++\\n> src/test/regress/expected/xml_2.out | 414 ++++++++++++++++++++++++\\n> src/test/regress/sql/xml.sql | 170 ++++++++++\\n> 33 files changed, 4019 insertions(+), 64 deletions(-)\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2016-11-25 3:31 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Michael Paquier wrote:\\n\\n> Nit: I did not look at the patch in details,\\n> but I find the size of the latest version sent, 167kB, scary as it\\n> complicates review and increases the likeliness of bugs.\\n\\nHere's the stat. Note that removing the functionality as discussed\\nwould remove all of xpath_parser.c but I think the rest of it remains\\npretty much unchanged. So it's clearly a large patch, but there are\\nlarge docs and tests too, not just code.yes, lot of is regress tests (expected part is 3x) - and XMLTABLE function is not trivial.lot of code is mechanical - nodes related. The really complex part is only in xml.c. There is one longer function only - the complexity is based on mapping libxml2 result to PostgreSQL result (with catching exceptions due releasing libxml2 sources).The all changes are well isolated - there is less risk to break some other. \\n\\n doc/src/sgml/func.sgml | 376 ++++++++++++++++++---\\n src/backend/executor/execQual.c | 335 +++++++++++++++++++\\n src/backend/executor/execTuples.c | 42 +++\\n src/backend/nodes/copyfuncs.c | 66 ++++\\n src/backend/nodes/equalfuncs.c | 51 +++\\n src/backend/nodes/nodeFuncs.c | 100 ++++++\\n src/backend/nodes/outfuncs.c | 51 +++\\n src/backend/nodes/readfuncs.c | 42 +++\\n src/backend/optimizer/util/clauses.c | 33 ++\\n src/backend/parser/gram.y | 181 ++++++++++-\\n src/backend/parser/parse_coerce.c | 33 +-\\n src/backend/parser/parse_expr.c | 182 +++++++++++\\n src/backend/parser/parse_target.c | 7 +\\n src/backend/utils/adt/Makefile | 2 +-\\n src/backend/utils/adt/ruleutils.c | 100 ++++++\\n src/backend/utils/adt/xml.c | 610 +++++++++++++++++++++++++++++++++++\\n src/backend/utils/adt/xpath_parser.c | 337 +++++++++++++++++++\\n src/backend/utils/fmgr/funcapi.c | 13 +\\n src/include/executor/executor.h | 1 +\\n src/include/executor/tableexpr.h | 69 ++++\\n src/include/funcapi.h | 1 -\\n src/include/nodes/execnodes.h | 31 ++\\n src/include/nodes/nodes.h | 4 +\\n src/include/nodes/parsenodes.h | 21 ++\\n src/include/nodes/primnodes.h | 40 +++\\n src/include/parser/kwlist.h | 3 +\\n src/include/parser/parse_coerce.h | 4 +\\n src/include/utils/xml.h | 2 +\\n src/include/utils/xpath_parser.h | 24 ++\\n src/test/regress/expected/xml.out | 415 ++++++++++++++++++++++++\\n src/test/regress/expected/xml_1.out | 323 +++++++++++++++++++\\n src/test/regress/expected/xml_2.out | 414 ++++++++++++++++++++++++\\n src/test/regress/sql/xml.sql | 170 ++++++++++\\n 33 files changed, 4019 insertions(+), 64 deletions(-)\\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Fri, 25 Nov 2016 07:44:12 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-11-25 7:44 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n\\n>\\n>\\n> 2016-11-25 3:31 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>\\n>> Michael Paquier wrote:\\n>>\\n>> > Nit: I did not look at the patch in details,\\n>> > but I find the size of the latest version sent, 167kB, scary as it\\n>> > complicates review and increases the likeliness of bugs.\\n>>\\n>> Here's the stat. Note that removing the functionality as discussed\\n>> would remove all of xpath_parser.c but I think the rest of it remains\\n>> pretty much unchanged. So it's clearly a large patch, but there are\\n>> large docs and tests too, not just code.\\n>>\\n>\\n> yes, lot of is regress tests (expected part is 3x) - and XMLTABLE function\\n> is not trivial.\\n>\\n\\nregress tests are about 50%\\n\\n\\n\\n>\\n> lot of code is mechanical - nodes related. The really complex part is only\\n> in xml.c. There is one longer function only - the complexity is based on\\n> mapping libxml2 result to PostgreSQL result (with catching exceptions due\\n> releasing libxml2 sources).\\n>\\n> The all changes are well isolated - there is less risk to break some\\n> other.\\n>\\n>\\n>>\\n>> doc/src/sgml/func.sgml | 376 ++++++++++++++++++---\\n>> src/backend/executor/execQual.c | 335 +++++++++++++++++++\\n>> src/backend/executor/execTuples.c | 42 +++\\n>> src/backend/nodes/copyfuncs.c | 66 ++++\\n>> src/backend/nodes/equalfuncs.c | 51 +++\\n>> src/backend/nodes/nodeFuncs.c | 100 ++++++\\n>> src/backend/nodes/outfuncs.c | 51 +++\\n>> src/backend/nodes/readfuncs.c | 42 +++\\n>> src/backend/optimizer/util/clauses.c | 33 ++\\n>> src/backend/parser/gram.y | 181 ++++++++++-\\n>> src/backend/parser/parse_coerce.c | 33 +-\\n>> src/backend/parser/parse_expr.c | 182 +++++++++++\\n>> src/backend/parser/parse_target.c | 7 +\\n>> src/backend/utils/adt/Makefile | 2 +-\\n>> src/backend/utils/adt/ruleutils.c | 100 ++++++\\n>> src/backend/utils/adt/xml.c | 610 ++++++++++++++++++++++++++++++\\n>> +++++\\n>> src/backend/utils/adt/xpath_parser.c | 337 +++++++++++++++++++\\n>> src/backend/utils/fmgr/funcapi.c | 13 +\\n>> src/include/executor/executor.h | 1 +\\n>> src/include/executor/tableexpr.h | 69 ++++\\n>> src/include/funcapi.h | 1 -\\n>> src/include/nodes/execnodes.h | 31 ++\\n>> src/include/nodes/nodes.h | 4 +\\n>> src/include/nodes/parsenodes.h | 21 ++\\n>> src/include/nodes/primnodes.h | 40 +++\\n>> src/include/parser/kwlist.h | 3 +\\n>> src/include/parser/parse_coerce.h | 4 +\\n>> src/include/utils/xml.h | 2 +\\n>> src/include/utils/xpath_parser.h | 24 ++\\n>> src/test/regress/expected/xml.out | 415 ++++++++++++++++++++++++\\n>> src/test/regress/expected/xml_1.out | 323 +++++++++++++++++++\\n>> src/test/regress/expected/xml_2.out | 414 ++++++++++++++++++++++++\\n>> src/test/regress/sql/xml.sql | 170 ++++++++++\\n>> 33 files changed, 4019 insertions(+), 64 deletions(-)\\n>>\\n>> --\\n>> Álvaro Herrera https://www.2ndQuadrant.com/\\n>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>>\\n>\\n>\\n\\n2016-11-25 7:44 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:2016-11-25 3:31 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Michael Paquier wrote:\\n\\n> Nit: I did not look at the patch in details,\\n> but I find the size of the latest version sent, 167kB, scary as it\\n> complicates review and increases the likeliness of bugs.\\n\\nHere's the stat. Note that removing the functionality as discussed\\nwould remove all of xpath_parser.c but I think the rest of it remains\\npretty much unchanged. So it's clearly a large patch, but there are\\nlarge docs and tests too, not just code.yes, lot of is regress tests (expected part is 3x) - and XMLTABLE function is not trivial.regress tests are about 50% lot of code is mechanical - nodes related. The really complex part is only in xml.c. There is one longer function only - the complexity is based on mapping libxml2 result to PostgreSQL result (with catching exceptions due releasing libxml2 sources).The all changes are well isolated - there is less risk to break some other. \\n\\n doc/src/sgml/func.sgml | 376 ++++++++++++++++++---\\n src/backend/executor/execQual.c | 335 +++++++++++++++++++\\n src/backend/executor/execTuples.c | 42 +++\\n src/backend/nodes/copyfuncs.c | 66 ++++\\n src/backend/nodes/equalfuncs.c | 51 +++\\n src/backend/nodes/nodeFuncs.c | 100 ++++++\\n src/backend/nodes/outfuncs.c | 51 +++\\n src/backend/nodes/readfuncs.c | 42 +++\\n src/backend/optimizer/util/clauses.c | 33 ++\\n src/backend/parser/gram.y | 181 ++++++++++-\\n src/backend/parser/parse_coerce.c | 33 +-\\n src/backend/parser/parse_expr.c | 182 +++++++++++\\n src/backend/parser/parse_target.c | 7 +\\n src/backend/utils/adt/Makefile | 2 +-\\n src/backend/utils/adt/ruleutils.c | 100 ++++++\\n src/backend/utils/adt/xml.c | 610 +++++++++++++++++++++++++++++++++++\\n src/backend/utils/adt/xpath_parser.c | 337 +++++++++++++++++++\\n src/backend/utils/fmgr/funcapi.c | 13 +\\n src/include/executor/executor.h | 1 +\\n src/include/executor/tableexpr.h | 69 ++++\\n src/include/funcapi.h | 1 -\\n src/include/nodes/execnodes.h | 31 ++\\n src/include/nodes/nodes.h | 4 +\\n src/include/nodes/parsenodes.h | 21 ++\\n src/include/nodes/primnodes.h | 40 +++\\n src/include/parser/kwlist.h | 3 +\\n src/include/parser/parse_coerce.h | 4 +\\n src/include/utils/xml.h | 2 +\\n src/include/utils/xpath_parser.h | 24 ++\\n src/test/regress/expected/xml.out | 415 ++++++++++++++++++++++++\\n src/test/regress/expected/xml_1.out | 323 +++++++++++++++++++\\n src/test/regress/expected/xml_2.out | 414 ++++++++++++++++++++++++\\n src/test/regress/sql/xml.sql | 170 ++++++++++\\n 33 files changed, 4019 insertions(+), 64 deletions(-)\\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Fri, 25 Nov 2016 07:46:09 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n\\n> Here is updated patch without default namespace support (and without XPath\\n> expression transformation).\\n> \\n> Due last changes in parser\\n> https://github.com/postgres/postgres/commit/906bfcad7ba7cb3863fe0e2a7810be8e3cd84fbd\\n> I had to use c_expr on other positions ( xmlnamespace definition).\\n> \\n> I don't think it is limit - in 99% there will be const literal.\\n\\nArgh. I can't avoid the feeling that I'm missing some parser trickery\\nhere. We have the XMLNAMESPACES keyword and the clause-terminating\\ncomma to protect these clauses, there must be a way to define this piece\\nof the grammar so that there's no conflict, without losing the freedom\\nin the expressions. But I don't see how. Now I agree that xml\\nnamespace definitions are going to be string literals in almost all\\ncases (or in extra sophisticated cases, column names) ... it's probably\\nbetter to spend the bison-fu in the document expression or the column\\noptions, or better yet the xmlexists_argument stuff. But I don't see\\npossibility of improvements in any of those places, so let's put it\\naside -- we can improve later, if need arises.\\n\\nIn any case, it looks like we can change c_expr to b_expr in a few\\nplaces, which is good because then operators work (in particular, unless\\nI misread the grammar, foo||bar doesn't work with c_expr and does work\\nwith b_expr, which seems the most useful in this case). Also, it makes\\nno sense to support (in the namespaces clause) DEFAULT a_expr if the\\nIDENT case uses only b_expr, so let's reduce both to just b_expr.\\n\\nWhile I'm looking at node definitions, I see a few things that could use\\nsome naming improvement. For example, \\\"expr\\\" for TableExpr is a bit\\nunexpressive. We could use \\\"document_expr\\\" there, perhaps. \\\"row_path\\\"\\nseems fixated on the XML case and the expression be path; let's use\\n\\\"row_expr\\\" there. And \\\"cols\\\" could be \\\"column_exprs\\\" perhaps. (All\\nthose renames cause fall-out in various node-related files, so let's\\nthink carefully to avoid renaming them multiple times.)\\n\\nIn primnodes, you kept the comment that says \\\"xpath\\\". Please update\\nthat to not-just-XML reality.\\n\\nPlease fix the comment in XmlTableAddNs; NULL is no longer a valid value.\\n\\nparse_expr.c has two unused variables; please remove them.\\n\\nThis test in ExecEvalTableExprProtected looks weird:\\n\\tif (i != tstate->for_ordinality_col - 1)\\nplease change to comparing \\\"i + 1\\\" (convert array index into attribute\\nnumber), and invert the boolean expression, leaving the for_ordinality\\ncase on top and the rest in the \\\"else\\\". That seems easier to read.\\nAlso, we customarily use post-increment (rownum++) instead of pre-incr.\\n\\nIn execQual.c I think it's neater to have ExecEvalTableExpr go before\\nits subroutine. Actually, I wonder whether it is really necessary to\\nhave a subroutine in the first place; you could just move the entire\\ncontents of that subroutine to within the PG_TRY block instead. The\\nonly thing you lose is one indentation level. I'm not sure about this\\none, but it's worth considering.\\n\\n-- \\n�lvaro Herrera https://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\": \"Mon, 28 Nov 2016 19:34:33 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-11-28 23:34 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n>\\n> > Here is updated patch without default namespace support (and without\\n> XPath\\n> > expression transformation).\\n> >\\n> > Due last changes in parser\\n> > https://github.com/postgres/postgres/commit/\\n> 906bfcad7ba7cb3863fe0e2a7810be8e3cd84fbd\\n> > I had to use c_expr on other positions ( xmlnamespace definition).\\n> >\\n> > I don't think it is limit - in 99% there will be const literal.\\n>\\n> Argh. I can't avoid the feeling that I'm missing some parser trickery\\n> here. We have the XMLNAMESPACES keyword and the clause-terminating\\n> comma to protect these clauses, there must be a way to define this piece\\n> of the grammar so that there's no conflict, without losing the freedom\\n> in the expressions. But I don't see how. Now I agree that xml\\n> namespace definitions are going to be string literals in almost all\\n> cases (or in extra sophisticated cases, column names) ... it's probably\\n> better to spend the bison-fu in the document expression or the column\\n> options, or better yet the xmlexists_argument stuff. But I don't see\\n> possibility of improvements in any of those places, so let's put it\\n> aside -- we can improve later, if need arises.\\n>\\n\\nThe problem is in unreserved keyword \\\"PASSING\\\" probably.\\n\\n\\n>\\n> In any case, it looks like we can change c_expr to b_expr in a few\\n> places, which is good because then operators work (in particular, unless\\n> I misread the grammar, foo||bar doesn't work with c_expr and does work\\n> with b_expr, which seems the most useful in this case). Also, it makes\\n> no sense to support (in the namespaces clause) DEFAULT a_expr if the\\n> IDENT case uses only b_expr, so let's reduce both to just b_expr.\\n>\\n\\nI changed all what was possible to b_expr.\\n\\n\\n>\\n> While I'm looking at node definitions, I see a few things that could use\\n> some naming improvement. For example, \\\"expr\\\" for TableExpr is a bit\\n> unexpressive. We could use \\\"document_expr\\\" there, perhaps. \\\"row_path\\\"\\n> seems fixated on the XML case and the expression be path; let's use\\n> \\\"row_expr\\\" there. And \\\"cols\\\" could be \\\"column_exprs\\\" perhaps. (All\\n> those renames cause fall-out in various node-related files, so let's\\n> think carefully to avoid renaming them multiple times.)\\n>\\n\\nColumns is not only expr - list - so I renamed it to \\\"columns\\\". Other\\nrenamed like you proposed\\n\\n\\n>\\n> In primnodes, you kept the comment that says \\\"xpath\\\". Please update\\n> that to not-just-XML reality.\\n>\\n\\nfixed\\n\\n\\n>\\n> Please fix the comment in XmlTableAddNs; NULL is no longer a valid value.\\n>\\n\\nfixed\\n\\n\\n>\\n> parse_expr.c has two unused variables; please remove them.\\n>\\n\\nfixed\\n\\n\\n>\\n> This test in ExecEvalTableExprProtected looks weird:\\n> if (i != tstate->for_ordinality_col - 1)\\n> please change to comparing \\\"i + 1\\\" (convert array index into attribute\\n> number), and invert the boolean expression, leaving the for_ordinality\\n> case on top and the rest in the \\\"else\\\". That seems easier to read.\\n> Also, we customarily use post-increment (rownum++) instead of pre-incr.\\n>\\n>\\nfiexed\\n\\n\\n> In execQual.c I think it's neater to have ExecEvalTableExpr go before\\n> its subroutine. Actually, I wonder whether it is really necessary to\\n> have a subroutine in the first place; you could just move the entire\\n> contents of that subroutine to within the PG_TRY block instead. The\\n> only thing you lose is one indentation level. I'm not sure about this\\n> one, but it's worth considering.\\n>\\n\\ndone\\n\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\nRegards\\n\\nPavel\\n\\n\\n-- \\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, 29 Nov 2016 22:36:18 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On 30 November 2016 at 05:36, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n\\n> The problem is in unreserved keyword \\\"PASSING\\\" probably.\\n\\nYeah, I think that's what I hit when trying to change it.\\n\\nCan't you just parenthesize the expression to use operators like ||\\netc? If so, not a big deal.\\n\\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, 30 Nov 2016 09:40:57 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-11-30 2:40 GMT+01:00 Craig Ringer <craig@2ndquadrant.com>:\\n\\n> On 30 November 2016 at 05:36, Pavel Stehule <pavel.stehule@gmail.com>\\n> wrote:\\n>\\n> > The problem is in unreserved keyword \\\"PASSING\\\" probably.\\n>\\n> Yeah, I think that's what I hit when trying to change it.\\n>\\n> Can't you just parenthesize the expression to use operators like ||\\n> etc? If so, not a big deal.\\n>\\n>\\n???\\n\\n\\n\\n>\\n>\\n>\\n> --\\n> Craig Ringer http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Training & Services\\n>\\n\\n2016-11-30 2:40 GMT+01:00 Craig Ringer <craig@2ndquadrant.com>:On 30 November 2016 at 05:36, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n\\n> The problem is in unreserved keyword \\\"PASSING\\\" probably.\\n\\nYeah, I think that's what I hit when trying to change it.\\n\\nCan't you just parenthesize the expression to use operators like ||\\netc? If so, not a big deal.\\n??? \\n\\n\\n\\n--\\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Wed, 30 Nov 2016 06:30:37 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n> 2016-11-30 2:40 GMT+01:00 Craig Ringer <craig@2ndquadrant.com>:\\n> \\n> > On 30 November 2016 at 05:36, Pavel Stehule <pavel.stehule@gmail.com>\\n> > wrote:\\n> >\\n> > > The problem is in unreserved keyword \\\"PASSING\\\" probably.\\n> >\\n> > Yeah, I think that's what I hit when trying to change it.\\n> >\\n> > Can't you just parenthesize the expression to use operators like ||\\n> > etc? If so, not a big deal.\\n> >\\n> ???\\n\\n\\\"'(' a_expr ')'\\\" is a c_expr; Craig suggests that we can just tell users\\nto manually add parens around any expressions that they want to use.\\nThat's not necessary most of the time since we've been able to use\\nb_expr in most places.\\n\\n-- \\n�lvaro Herrera https://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, 30 Nov 2016 09:38:04 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-11-30 13:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n> > 2016-11-30 2:40 GMT+01:00 Craig Ringer <craig@2ndquadrant.com>:\\n> >\\n> > > On 30 November 2016 at 05:36, Pavel Stehule <pavel.stehule@gmail.com>\\n> > > wrote:\\n> > >\\n> > > > The problem is in unreserved keyword \\\"PASSING\\\" probably.\\n> > >\\n> > > Yeah, I think that's what I hit when trying to change it.\\n> > >\\n> > > Can't you just parenthesize the expression to use operators like ||\\n> > > etc? If so, not a big deal.\\n> > >\\n> > ???\\n>\\n> \\\"'(' a_expr ')'\\\" is a c_expr; Craig suggests that we can just tell users\\n> to manually add parens around any expressions that they want to use.\\n> That's not necessary most of the time since we've been able to use\\n> b_expr in most places.\\n>\\n\\nNow I understand\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2016-11-30 13:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\\n> 2016-11-30 2:40 GMT+01:00 Craig Ringer <craig@2ndquadrant.com>:\\n>\\n> > On 30 November 2016 at 05:36, Pavel Stehule <pavel.stehule@gmail.com>\\n> > wrote:\\n> >\\n> > > The problem is in unreserved keyword \\\"PASSING\\\" probably.\\n> >\\n> > Yeah, I think that's what I hit when trying to change it.\\n> >\\n> > Can't you just parenthesize the expression to use operators like ||\\n> > etc? If so, not a big deal.\\n> >\\n> ???\\n\\n\\\"'(' a_expr ')'\\\" is a c_expr; Craig suggests that we can just tell users\\nto manually add parens around any expressions that they want to use.\\nThat's not necessary most of the time since we've been able to use\\nb_expr in most places.Now I understandRegardsPavel \\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Wed, 30 Nov 2016 14:53:03 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Dne 30. 11. 2016 14:53 napsal uživatel \\\"Pavel Stehule\\\" <\\npavel.stehule@gmail.com>:\\n>\\n>\\n>\\n> 2016-11-30 13:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>>\\n>> Pavel Stehule wrote:\\n>> > 2016-11-30 2:40 GMT+01:00 Craig Ringer <craig@2ndquadrant.com>:\\n>> >\\n>> > > On 30 November 2016 at 05:36, Pavel Stehule <pavel.stehule@gmail.com>\\n>> > > wrote:\\n>> > >\\n>> > > > The problem is in unreserved keyword \\\"PASSING\\\" probably.\\n>> > >\\n>> > > Yeah, I think that's what I hit when trying to change it.\\n>> > >\\n>> > > Can't you just parenthesize the expression to use operators like ||\\n>> > > etc? If so, not a big deal.\\n>> > >\\n>> > ???\\n>>\\n>> \\\"'(' a_expr ')'\\\" is a c_expr; Craig suggests that we can just tell users\\n>> to manually add parens around any expressions that they want to use.\\n>> That's not necessary most of the time since we've been able to use\\n>> b_expr in most places.\\n>\\n\\nstill there are one c_expr, but without new reserved word there are not\\nchange to reduce it.\\n\\n>\\n> Now I understand\\n>\\n> Regards\\n>\\n> Pavel\\n>\\n>>\\n>>\\n>> --\\n>> Álvaro Herrera https://www.2ndQuadrant.com/\\n>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n>\\n\\n\\nDne 30. 11. 2016 14:53 napsal uživatel \\\"Pavel Stehule\\\" <pavel.stehule@gmail.com>:\\n>\\n>\\n>\\n> 2016-11-30 13:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>>\\n>> Pavel Stehule wrote:\\n>> > 2016-11-30 2:40 GMT+01:00 Craig Ringer <craig@2ndquadrant.com>:\\n>> >\\n>> > > On 30 November 2016 at 05:36, Pavel Stehule <pavel.stehule@gmail.com>\\n>> > > wrote:\\n>> > >\\n>> > > > The problem is in unreserved keyword \\\"PASSING\\\" probably.\\n>> > >\\n>> > > Yeah, I think that's what I hit when trying to change it.\\n>> > >\\n>> > > Can't you just parenthesize the expression to use operators like ||\\n>> > > etc? If so, not a big deal.\\n>> > >\\n>> > ???\\n>>\\n>> \\\"'(' a_expr ')'\\\" is a c_expr; Craig suggests that we can just tell users\\n>> to manually add parens around any expressions that they want to use.\\n>> That's not necessary most of the time since we've been able to use\\n>> b_expr in most places.\\n>\\nstill there are one c_expr, but without new reserved word there are not change to reduce it.\\n>\\n> Now I understand\\n>\\n> Regards\\n>\\n> Pavel\\n> \\n>>\\n>>\\n>> --\\n>> Álvaro Herrera https://www.2ndQuadrant.com/\\n>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n>\",\n \"msg_date\": \"Wed, 30 Nov 2016 16:21:37 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 1, 2016 at 2:21 AM, Pavel Stehule <pavel.stehule@gmail.com>\\nwrote:\\n\\n> Dne 30. 11. 2016 14:53 napsal uživatel \\\"Pavel Stehule\\\" <\\n> pavel.stehule@gmail.com>:\\n> >\\n> >\\n> >\\n> > 2016-11-30 13:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n> >>\\n> >> Pavel Stehule wrote:\\n> >> > 2016-11-30 2:40 GMT+01:00 Craig Ringer <craig@2ndquadrant.com>:\\n> >> >\\n> >> > > On 30 November 2016 at 05:36, Pavel Stehule <\\n> pavel.stehule@gmail.com>\\n> >> > > wrote:\\n> >> > >\\n> >> > > > The problem is in unreserved keyword \\\"PASSING\\\" probably.\\n> >> > >\\n> >> > > Yeah, I think that's what I hit when trying to change it.\\n> >> > >\\n> >> > > Can't you just parenthesize the expression to use operators like ||\\n> >> > > etc? If so, not a big deal.\\n> >> > >\\n> >> > ???\\n> >>\\n> >> \\\"'(' a_expr ')'\\\" is a c_expr; Craig suggests that we can just tell users\\n> >> to manually add parens around any expressions that they want to use.\\n> >> That's not necessary most of the time since we've been able to use\\n> >> b_expr in most places.\\n> >\\n>\\n> still there are one c_expr, but without new reserved word there are not\\n> change to reduce it.\\n>\\n>\\n>\\nMoved to next CF with the same status (ready for committer).\\n\\nRegards,\\nHari Babu\\nFujitsu Australia\\n\\nOn Thu, Dec 1, 2016 at 2:21 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\nDne 30. 11. 2016 14:53 napsal uživatel \\\"Pavel Stehule\\\" <pavel.stehule@gmail.com>:\\n>\\n>\\n>\\n> 2016-11-30 13:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>>\\n>> Pavel Stehule wrote:\\n>> > 2016-11-30 2:40 GMT+01:00 Craig Ringer <craig@2ndquadrant.com>:\\n>> >\\n>> > > On 30 November 2016 at 05:36, Pavel Stehule <pavel.stehule@gmail.com>\\n>> > > wrote:\\n>> > >\\n>> > > > The problem is in unreserved keyword \\\"PASSING\\\" probably.\\n>> > >\\n>> > > Yeah, I think that's what I hit when trying to change it.\\n>> > >\\n>> > > Can't you just parenthesize the expression to use operators like ||\\n>> > > etc? If so, not a big deal.\\n>> > >\\n>> > ???\\n>>\\n>> \\\"'(' a_expr ')'\\\" is a c_expr; Craig suggests that we can just tell users\\n>> to manually add parens around any expressions that they want to use.\\n>> That's not necessary most of the time since we've been able to use\\n>> b_expr in most places.\\n>\\nstill there are one c_expr, but without new reserved word there are not change to reduce it.\\nMoved to next CF with the same status (ready for committer).Regards,Hari BabuFujitsu Australia\",\n \"msg_date\": \"Fri, 2 Dec 2016 17:30:51 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hm, you omitted tableexpr.h from the v15 patch ...\\n\\n-- \\n�lvaro Herrera https://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\": \"Fri, 2 Dec 2016 13:25:08 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-12-02 17:25 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Hm, you omitted tableexpr.h from the v15 patch ...\\n>\\n\\nI am sorry\\n\\nshould be ok now\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL 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\": \"Fri, 2 Dec 2016 20:22:00 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Here's version 17. I have made significant changes here.\\n\\n1. Restructure the execQual code. Instead of a PG_TRY wrapper, I have\\nsplit this code in three pieces; there's the main code with the PG_TRY\\nwrappers and is mainly in charge of the builderCxt pointer. In the\\nprevious coding there was a shim that examined builderCxt but was not\\nresponsible for setting it up, which was ugly. The second part is the\\n\\\"initializer\\\" which sets the row and column filters and does namespace\\nprocessing. The third part is the \\\"FetchRow\\\" logic. It seems to me\\nmuch cleaner this way.\\n\\n2. rename the \\\"builder\\\" stuff to use the \\\"routine\\\" terminology. This is\\nin line with what we do for other function-pointer-filled structs, such\\nas FdwRoutine, IndexAmRoutine etc. I also cleaned up the names a bit\\nmore.\\n\\n3. Added a magic number to the table builder context struct, so that we\\ncan barf appropriately. This is in line with PgXmlErrorContext --\\nmostly for future-proofing. I didn't test this too hard. Also, moved\\nthe XmlTableContext struct declaration nearer the top of the file, as is\\ncustomary. (We don't really need it that way, since the functions are\\nall declared taking void *, but it seems cleaner to me anyway).\\n\\n4. I added, edited, and fixed a large number of code comments.\\n\\nThis is looking much better now, but it still needs at least the\\nfollowing changes.\\n\\nFirst, we need to fix is the per_rowset_memcxt thingy. I think the way\\nit's currently being used is rather ugly; it looks to me like the memory\\ncontext does not belong into the XmlTableContext struct at all.\\nInstead, the executor code should keep the memcxt pointer in a state\\nstruct of its own, and it should be the executor's responsibility to\\nchange to the appropriate context before calling the table builder\\nfunctions. In particular, this means that the table context can no\\nlonger be a void * pointer; it needs to be a struct that's defined by\\nthe executor (probably a primnodes.h one). The void * pointer is\\nstashed inside that struct. Also, the \\\"routine\\\" pointer should not be\\npart of the void * struct, but of the executor's struct. So the\\nexecQual code can switch to the memory context, and destroy it\\nappropriately.\\n\\nSecond, we should make gram.y set a new \\\"function type\\\" value in the\\nTableExpr it creates, so that the downstream code (transformTableExpr,\\nExecInitExpr, ruleutils.c) really knows that the given function is\\nXmlTableExpr, instead of guessing just because it's the only implemented\\ncase. Probably this \\\"function type\\\" is an enum (currently with a single\\nvalue TableExprTypeXml or something like that) in primnodes.\\n\\nFinally, there's the pending task of renaming and moving\\nExecTypeFromTableExpr to some better place. Not sure that moving it\\nback to nodeFuncs is a nice idea. Looks to me like calling it from\\nExprTypmod is a rather ugly idea.\\n\\nHmm, ruleutils ... not sure what to think of this one.\\n\\nThe typedefs.list changes are just used to pgindent the affected code\\ncorrectly. It's not for commit.\\n\\n-- \\n�lvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\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, 2 Dec 2016 19:25:50 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2016-12-02 23:25 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Here's version 17. I have made significant changes here.\\n>\\n> 1. Restructure the execQual code. Instead of a PG_TRY wrapper, I have\\n> split this code in three pieces; there's the main code with the PG_TRY\\n> wrappers and is mainly in charge of the builderCxt pointer. In the\\n> previous coding there was a shim that examined builderCxt but was not\\n> responsible for setting it up, which was ugly. The second part is the\\n> \\\"initializer\\\" which sets the row and column filters and does namespace\\n> processing. The third part is the \\\"FetchRow\\\" logic. It seems to me\\n> much cleaner this way.\\n>\\n> 2. rename the \\\"builder\\\" stuff to use the \\\"routine\\\" terminology. This is\\n> in line with what we do for other function-pointer-filled structs, such\\n> as FdwRoutine, IndexAmRoutine etc. I also cleaned up the names a bit\\n> more.\\n>\\n> 3. Added a magic number to the table builder context struct, so that we\\n> can barf appropriately. This is in line with PgXmlErrorContext --\\n> mostly for future-proofing. I didn't test this too hard. Also, moved\\n> the XmlTableContext struct declaration nearer the top of the file, as is\\n> customary. (We don't really need it that way, since the functions are\\n> all declared taking void *, but it seems cleaner to me anyway).\\n>\\n> 4. I added, edited, and fixed a large number of code comments.\\n>\\n> This is looking much better now, but it still needs at least the\\n> following changes.\\n>\\n> First, we need to fix is the per_rowset_memcxt thingy. I think the way\\n> it's currently being used is rather ugly; it looks to me like the memory\\n> context does not belong into the XmlTableContext struct at all.\\n> Instead, the executor code should keep the memcxt pointer in a state\\n> struct of its own, and it should be the executor's responsibility to\\n> change to the appropriate context before calling the table builder\\n> functions. In particular, this means that the table context can no\\n> longer be a void * pointer; it needs to be a struct that's defined by\\n> the executor (probably a primnodes.h one). The void * pointer is\\n> stashed inside that struct. Also, the \\\"routine\\\" pointer should not be\\n> part of the void * struct, but of the executor's struct. So the\\n> execQual code can switch to the memory context, and destroy it\\n> appropriately.\\n>\\n> Second, we should make gram.y set a new \\\"function type\\\" value in the\\n> TableExpr it creates, so that the downstream code (transformTableExpr,\\n> ExecInitExpr, ruleutils.c) really knows that the given function is\\n> XmlTableExpr, instead of guessing just because it's the only implemented\\n> case. Probably this \\\"function type\\\" is an enum (currently with a single\\n> value TableExprTypeXml or something like that) in primnodes.\\n>\\n\\nIt has sense - I was not sure about it - because currently it is only one\\nvalue, you mentioned it.\\n\\n\\n>\\n> Finally, there's the pending task of renaming and moving\\n> ExecTypeFromTableExpr to some better place. Not sure that moving it\\n> back to nodeFuncs is a nice idea. Looks to me like calling it from\\n> ExprTypmod is a rather ugly idea.\\n>\\n\\nThe code is related to prim nodes - it is used more times than in executor.\\n\\n>\\n> Hmm, ruleutils ... not sure what to think of this one.\\n>\\n\\nit is little bit more complex - but it is related to complexity of XMLTABLE\\n\\n\\n>\\n> The typedefs.list changes are just used to pgindent the affected code\\n> correctly. It's not for commit.\\n>\\n\\nThe documentation is very precious. Nice\\n\\n+ /* XXX OK to do this? looks a bit out of place ... */\\n+ assign_record_type_typmod(typeInfo);\\n\\nI am thinking it is ok. It is tupdesc without fixed typid, typname used in\\nreturned value - you should to register this tupdesc in typcache.\\n\\nRegards\\n\\nPavel\\n\\n\\n\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\nHi2016-12-02 23:25 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Here's version 17. I have made significant changes here.\\n\\n1. Restructure the execQual code. Instead of a PG_TRY wrapper, I have\\nsplit this code in three pieces; there's the main code with the PG_TRY\\nwrappers and is mainly in charge of the builderCxt pointer. In the\\nprevious coding there was a shim that examined builderCxt but was not\\nresponsible for setting it up, which was ugly. The second part is the\\n\\\"initializer\\\" which sets the row and column filters and does namespace\\nprocessing. The third part is the \\\"FetchRow\\\" logic. It seems to me\\nmuch cleaner this way.\\n\\n2. rename the \\\"builder\\\" stuff to use the \\\"routine\\\" terminology. This is\\nin line with what we do for other function-pointer-filled structs, such\\nas FdwRoutine, IndexAmRoutine etc. I also cleaned up the names a bit\\nmore.\\n\\n3. Added a magic number to the table builder context struct, so that we\\ncan barf appropriately. This is in line with PgXmlErrorContext --\\nmostly for future-proofing. I didn't test this too hard. Also, moved\\nthe XmlTableContext struct declaration nearer the top of the file, as is\\ncustomary. (We don't really need it that way, since the functions are\\nall declared taking void *, but it seems cleaner to me anyway).\\n\\n4. I added, edited, and fixed a large number of code comments.\\n\\nThis is looking much better now, but it still needs at least the\\nfollowing changes.\\n\\nFirst, we need to fix is the per_rowset_memcxt thingy. I think the way\\nit's currently being used is rather ugly; it looks to me like the memory\\ncontext does not belong into the XmlTableContext struct at all.\\nInstead, the executor code should keep the memcxt pointer in a state\\nstruct of its own, and it should be the executor's responsibility to\\nchange to the appropriate context before calling the table builder\\nfunctions. In particular, this means that the table context can no\\nlonger be a void * pointer; it needs to be a struct that's defined by\\nthe executor (probably a primnodes.h one). The void * pointer is\\nstashed inside that struct. Also, the \\\"routine\\\" pointer should not be\\npart of the void * struct, but of the executor's struct. So the\\nexecQual code can switch to the memory context, and destroy it\\nappropriately.\\n\\nSecond, we should make gram.y set a new \\\"function type\\\" value in the\\nTableExpr it creates, so that the downstream code (transformTableExpr,\\nExecInitExpr, ruleutils.c) really knows that the given function is\\nXmlTableExpr, instead of guessing just because it's the only implemented\\ncase. Probably this \\\"function type\\\" is an enum (currently with a single\\nvalue TableExprTypeXml or something like that) in primnodes.It has sense - I was not sure about it - because currently it is only one value, you mentioned it. \\n\\nFinally, there's the pending task of renaming and moving\\nExecTypeFromTableExpr to some better place. Not sure that moving it\\nback to nodeFuncs is a nice idea. Looks to me like calling it from\\nExprTypmod is a rather ugly idea.The code is related to prim nodes - it is used more times than in executor.\\n\\nHmm, ruleutils ... not sure what to think of this one.it is little bit more complex - but it is related to complexity of XMLTABLE \\n\\nThe typedefs.list changes are just used to pgindent the affected code\\ncorrectly. It's not for commit.The documentation is very precious. Nice+ /* XXX OK to do this? looks a bit out of place ... */ + assign_record_type_typmod(typeInfo);I am thinking it is ok. It is tupdesc without fixed typid, typname used in returned value - you should to register this tupdesc in typcache.RegardsPavel \\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Sat, 3 Dec 2016 08:28:50 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n\\n> 2016-12-02 23:25 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> > This is looking much better now, but it still needs at least the\\n> > following changes.\\n> >\\n> > First, we need to fix is the per_rowset_memcxt thingy. I think the way\\n> > it's currently being used is rather ugly; it looks to me like the memory\\n> > context does not belong into the XmlTableContext struct at all.\\n> > Instead, the executor code should keep the memcxt pointer in a state\\n> > struct of its own, and it should be the executor's responsibility to\\n> > change to the appropriate context before calling the table builder\\n> > functions. In particular, this means that the table context can no\\n> > longer be a void * pointer; it needs to be a struct that's defined by\\n> > the executor (probably a primnodes.h one). The void * pointer is\\n> > stashed inside that struct. Also, the \\\"routine\\\" pointer should not be\\n> > part of the void * struct, but of the executor's struct. So the\\n> > execQual code can switch to the memory context, and destroy it\\n> > appropriately.\\n> >\\n> > Second, we should make gram.y set a new \\\"function type\\\" value in the\\n> > TableExpr it creates, so that the downstream code (transformTableExpr,\\n> > ExecInitExpr, ruleutils.c) really knows that the given function is\\n> > XmlTableExpr, instead of guessing just because it's the only implemented\\n> > case. Probably this \\\"function type\\\" is an enum (currently with a single\\n> > value TableExprTypeXml or something like that) in primnodes.\\n> \\n> It has sense - I was not sure about it - because currently it is only one\\n> value, you mentioned it.\\n\\nTrue. This is a minor point.\\n\\nAre you able to do the memory context change I describe?\\n\\n-- \\n�lvaro Herrera https://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\": \"Sat, 3 Dec 2016 12:03:58 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-12-03 16:03 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n>\\n> > 2016-12-02 23:25 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>\\n> > > This is looking much better now, but it still needs at least the\\n> > > following changes.\\n> > >\\n> > > First, we need to fix is the per_rowset_memcxt thingy. I think the way\\n> > > it's currently being used is rather ugly; it looks to me like the\\n> memory\\n> > > context does not belong into the XmlTableContext struct at all.\\n> > > Instead, the executor code should keep the memcxt pointer in a state\\n> > > struct of its own, and it should be the executor's responsibility to\\n> > > change to the appropriate context before calling the table builder\\n> > > functions. In particular, this means that the table context can no\\n> > > longer be a void * pointer; it needs to be a struct that's defined by\\n> > > the executor (probably a primnodes.h one). The void * pointer is\\n> > > stashed inside that struct. Also, the \\\"routine\\\" pointer should not be\\n> > > part of the void * struct, but of the executor's struct. So the\\n> > > execQual code can switch to the memory context, and destroy it\\n> > > appropriately.\\n> > >\\n> > > Second, we should make gram.y set a new \\\"function type\\\" value in the\\n> > > TableExpr it creates, so that the downstream code (transformTableExpr,\\n> > > ExecInitExpr, ruleutils.c) really knows that the given function is\\n> > > XmlTableExpr, instead of guessing just because it's the only\\n> implemented\\n> > > case. Probably this \\\"function type\\\" is an enum (currently with a\\n> single\\n> > > value TableExprTypeXml or something like that) in primnodes.\\n> >\\n> > It has sense - I was not sure about it - because currently it is only one\\n> > value, you mentioned it.\\n>\\n> True. This is a minor point.\\n>\\n> Are you able to do the memory context change I describe?\\n>\\n\\nI am not sure if I understand well to your ideas - please, check attached\\npatch.\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL 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\": \"Sun, 4 Dec 2016 23:00:17 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-12-04 23:00 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n\\n>\\n>\\n> 2016-12-03 16:03 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>\\n>> Pavel Stehule wrote:\\n>>\\n>> > 2016-12-02 23:25 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>>\\n>> > > This is looking much better now, but it still needs at least the\\n>> > > following changes.\\n>> > >\\n>> > > First, we need to fix is the per_rowset_memcxt thingy. I think the\\n>> way\\n>> > > it's currently being used is rather ugly; it looks to me like the\\n>> memory\\n>> > > context does not belong into the XmlTableContext struct at all.\\n>> > > Instead, the executor code should keep the memcxt pointer in a state\\n>> > > struct of its own, and it should be the executor's responsibility to\\n>> > > change to the appropriate context before calling the table builder\\n>> > > functions. In particular, this means that the table context can no\\n>> > > longer be a void * pointer; it needs to be a struct that's defined by\\n>> > > the executor (probably a primnodes.h one). The void * pointer is\\n>> > > stashed inside that struct. Also, the \\\"routine\\\" pointer should not be\\n>> > > part of the void * struct, but of the executor's struct. So the\\n>> > > execQual code can switch to the memory context, and destroy it\\n>> > > appropriately.\\n>> > >\\n>> > > Second, we should make gram.y set a new \\\"function type\\\" value in the\\n>> > > TableExpr it creates, so that the downstream code (transformTableExpr,\\n>> > > ExecInitExpr, ruleutils.c) really knows that the given function is\\n>> > > XmlTableExpr, instead of guessing just because it's the only\\n>> implemented\\n>> > > case. Probably this \\\"function type\\\" is an enum (currently with a\\n>> single\\n>> > > value TableExprTypeXml or something like that) in primnodes.\\n>> >\\n>> > It has sense - I was not sure about it - because currently it is only\\n>> one\\n>> > value, you mentioned it.\\n>>\\n>> True. This is a minor point.\\n>>\\n>> Are you able to do the memory context change I describe?\\n>>\\n>\\n> I am not sure if I understand well to your ideas - please, check attached\\n> patch.\\n>\\n\\nattached patch without your patch 0001\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> Regards\\n>\\n> Pavel\\n>\\n>\\n>>\\n>> --\\n>> Álvaro Herrera https://www.2ndQuadrant.com/\\n>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\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\": \"Sun, 4 Dec 2016 23:12:55 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n> 2016-12-04 23:00 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n\\n> > I am not sure if I understand well to your ideas - please, check attached\\n> > patch.\\n\\nThanks, that's what I meant, but I think you went a bit overboard\\ncreating new functions in execQual -- seems to me it would work just\\nfine to have the memory switches in the same function, rather than\\nhaving a number of separate functions just to change the context then\\ncall the method. Please remove these shim functions.\\n\\nAlso, you forgot to remove the now-unused per_rowset_memcxt struct member.\\n\\nAlso, please rename \\\"rc\\\" to something more meaningful -- maybe\\n\\\"rowcount\\\" is good enough. And \\\"doc\\\" would perhaps be better as\\n\\\"document\\\".\\n\\nI'm not completely sure the structs are really sensible yet. I may do\\nsome more changes tomorrow.\\n\\n-- \\n�lvaro Herrera https://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\": \"Sun, 4 Dec 2016 20:45:54 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-12-05 0:45 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n> > 2016-12-04 23:00 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n>\\n> > > I am not sure if I understand well to your ideas - please, check\\n> attached\\n> > > patch.\\n>\\n> Thanks, that's what I meant, but I think you went a bit overboard\\n> creating new functions in execQual -- seems to me it would work just\\n> fine to have the memory switches in the same function, rather than\\n> having a number of separate functions just to change the context then\\n> call the method. Please remove these shim functions.\\n>\\n\\ndone\\n\\n\\n>\\n> Also, you forgot to remove the now-unused per_rowset_memcxt struct member.\\n>\\n\\ndone\\n\\n\\n>\\n> Also, please rename \\\"rc\\\" to something more meaningful -- maybe\\n> \\\"rowcount\\\" is good enough. And \\\"doc\\\" would perhaps be better as\\n> \\\"document\\\".\\n>\\n\\ndone\\n\\nRegards\\n\\nPavel\\n\\n>\\n> I'm not completely sure the structs are really sensible yet. I may do\\n> some more changes tomorrow.\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL 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\": \"Mon, 5 Dec 2016 07:21:37 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Here's v21.\\n\\n* I changed the grammar by moving the NOT NULL to the column options,\\nand removing the IsNotNull production. It wasn't nice that \\\"NOT NULL\\nDEFAULT 0\\\" was not accepted, which it is with the new representation.\\n\\n* The tuple that's returned is natively a TupleTableSlot inside the\\ntable builder, not directly a HeapTuple. That stuff was ugly and wasn't\\nusing the proper abstraction anyway.\\n\\n* I changed the signatures of the methods so that they receive\\nTableExprState, and restructured the \\\"opaque\\\" data to be inside\\nTableExprState. Now we don't need to have things such as the tupdesc or\\nthe input functions be repeated in the opaque struct. Instead they\\nbelong to the TableExprState and the methods can read them from there.\\n\\nI managed to break the case with no COLUMNS. Probably related to the\\ntupdesc changes. It now crashes the regression test. Too tired to\\ndebug now; care to take a look? The other stuff seems to run fine,\\nthough of course the regression test crashes in the middle, so perhaps\\nthere are other problems.\\n\\n-- \\n�lvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\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, 7 Dec 2016 04:14:39 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-12-07 8:14 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Here's v21.\\n>\\n> * I changed the grammar by moving the NOT NULL to the column options,\\n> and removing the IsNotNull production. It wasn't nice that \\\"NOT NULL\\n> DEFAULT 0\\\" was not accepted, which it is with the new representation.\\n>\\n> * The tuple that's returned is natively a TupleTableSlot inside the\\n> table builder, not directly a HeapTuple. That stuff was ugly and wasn't\\n> using the proper abstraction anyway.\\n>\\n> * I changed the signatures of the methods so that they receive\\n> TableExprState, and restructured the \\\"opaque\\\" data to be inside\\n> TableExprState. Now we don't need to have things such as the tupdesc or\\n> the input functions be repeated in the opaque struct. Instead they\\n> belong to the TableExprState and the methods can read them from there.\\n>\\n> I managed to break the case with no COLUMNS. Probably related to the\\n> tupdesc changes. It now crashes the regression test. Too tired to\\n> debug now; care to take a look? The other stuff seems to run fine,\\n> though of course the regression test crashes in the middle, so perhaps\\n> there are other problems.\\n>\\n\\nI fixed two issues.\\n\\n1. there are not columns data when there are not any explicit column - fixed\\n\\n2. there was reverse setting in NOT NULL flag\\n\\nall tests passed now\\n\\nRegards\\n\\nPavel\\n\\n\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL 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\": \"Wed, 7 Dec 2016 14:47:50 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n\\n> I fixed two issues.\\n> \\n> 2. there was reverse setting in NOT NULL flag\\n\\nAh-hah, that was silly, thanks.\\n\\n> 1. there are not columns data when there are not any explicit column - fixed\\n\\nHmm. Now that I see how this works, by having the GetValue \\\"guess\\\" what\\nis going on and have a special case for it, I actually don't like it\\nvery much. It seems way too magical. I think we should do away with\\nthe \\\"if column is NULL\\\" case in GetValue, and instead inject a column\\nduring transformTableExpr if columns is NIL. This has implications on\\nExecInitExpr too, which currently checks for an empty column list -- it\\nwould no longer have to do so.\\n\\nMaybe this means we need an additional method, which would request \\\"the\\nexpr that returns the whole row\\\", so that transformExpr can work for\\nXmlTable (which I think would be something like \\\"./\\\") and the future\\nJsonTable stuff (I don't know how that one would work, but I assume it's\\nnot necessarily the same thing).\\n\\n-- \\n�lvaro Herrera https://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, 7 Dec 2016 14:34:59 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-12-07 18:34 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n>\\n> > I fixed two issues.\\n> >\\n> > 2. there was reverse setting in NOT NULL flag\\n>\\n> Ah-hah, that was silly, thanks.\\n>\\n> > 1. there are not columns data when there are not any explicit column -\\n> fixed\\n>\\n> Hmm. Now that I see how this works, by having the GetValue \\\"guess\\\" what\\n> is going on and have a special case for it, I actually don't like it\\n> very much. It seems way too magical. I think we should do away with\\n> the \\\"if column is NULL\\\" case in GetValue, and instead inject a column\\n> during transformTableExpr if columns is NIL. This has implications on\\n> ExecInitExpr too, which currently checks for an empty column list -- it\\n> would no longer have to do so.\\n>\\n\\nI prefer this way against second described. The implementation should be in\\ntable builder routines, not in executor.\\n\\nsending new update\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> Maybe this means we need an additional method, which would request \\\"the\\n> expr that returns the whole row\\\", so that transformExpr can work for\\n> XmlTable (which I think would be something like \\\"./\\\") and the future\\n> JsonTable stuff (I don't know how that one would work, but I assume it's\\n> not necessarily the same thing).\\n>\\n\\n\\n\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL 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\": \"Wed, 7 Dec 2016 20:37:23 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n> 2016-12-07 18:34 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> > Hmm. Now that I see how this works, by having the GetValue \\\"guess\\\" what\\n> > is going on and have a special case for it, I actually don't like it\\n> > very much. It seems way too magical. I think we should do away with\\n> > the \\\"if column is NULL\\\" case in GetValue, and instead inject a column\\n> > during transformTableExpr if columns is NIL. This has implications on\\n> > ExecInitExpr too, which currently checks for an empty column list -- it\\n> > would no longer have to do so.\\n> \\n> I prefer this way against second described. The implementation should be in\\n> table builder routines, not in executor.\\n\\nWell, given the way you have implemented it, I would prefer the original\\ntoo. But your v23 is not what I meant. Essentially what you do in v23\\nis to communicate the lack of COLUMNS clause in a different way --\\npreviously it was \\\"ncolumns = 0\\\", now it's \\\"is_auto_col=true\\\". It's\\nstill \\\"magic\\\". It's not an improvement.\\n\\nWhat I want to happen is that there is no magic at all; it's up to\\ntransformExpr to make sure that when COLUMNS is empty, one column\\nappears and it must not be a magic column that makes the xml.c code act\\ndifferently, but rather to xml.c it should appear that this is just a\\nnormal column that happens to return the entire row. If I say \\\"COLUMNS\\nfoo PATH '/'\\\" I should be able to obtain a similar behavior (except that\\nin the current code, if I ask for \\\"COLUMNS foo XML PATH '/'\\\" I don't get\\nXML at all but rather weird text where all tags have been stripped out,\\nwhich is very strange. I would expect the tags to be preserved if the\\noutput type is XML. Maybe the tag-stripping behavior should occur if\\nthe output type is some type of text.)\\n\\n\\nI still have to figure out how to fix the tupledesc thing. What we have\\nnow is not good.\\n\\n-- \\n�lvaro Herrera https://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, 7 Dec 2016 16:50:49 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-12-07 20:50 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n> > 2016-12-07 18:34 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>\\n> > > Hmm. Now that I see how this works, by having the GetValue \\\"guess\\\"\\n> what\\n> > > is going on and have a special case for it, I actually don't like it\\n> > > very much. It seems way too magical. I think we should do away with\\n> > > the \\\"if column is NULL\\\" case in GetValue, and instead inject a column\\n> > > during transformTableExpr if columns is NIL. This has implications on\\n> > > ExecInitExpr too, which currently checks for an empty column list -- it\\n> > > would no longer have to do so.\\n> >\\n> > I prefer this way against second described. The implementation should be\\n> in\\n> > table builder routines, not in executor.\\n>\\n> Well, given the way you have implemented it, I would prefer the original\\n> too. But your v23 is not what I meant. Essentially what you do in v23\\n> is to communicate the lack of COLUMNS clause in a different way --\\n> previously it was \\\"ncolumns = 0\\\", now it's \\\"is_auto_col=true\\\". It's\\n> still \\\"magic\\\". It's not an improvement.\\n>\\n\\nis_auto_col is used primary for asserting. The table builder has\\ninformation for decision in parameter path, when path is NULL.\\n\\nHard to say, if this info should be assigned to column or to table. In both\\nlocations has sense. But somewhere should be some flag.\\n\\n\\n>\\n> What I want to happen is that there is no magic at all; it's up to\\n> transformExpr to make sure that when COLUMNS is empty, one column\\n> appears and it must not be a magic column that makes the xml.c code act\\n> differently, but rather to xml.c it should appear that this is just a\\n> normal column that happens to return the entire row. If I say \\\"COLUMNS\\n> foo PATH '/'\\\" I should be able to obtain a similar behavior (except that\\n> in the current code, if I ask for \\\"COLUMNS foo XML PATH '/'\\\" I don't get\\n> XML at all but rather weird text where all tags have been stripped out,\\n> which is very strange. I would expect the tags to be preserved if the\\n> output type is XML. Maybe the tag-stripping behavior should occur if\\n> the output type is some type of text.)\\n>\\n\\nI am doing this. Just I using NULL for PATH.\\n\\n\\n>\\n>\\n> I still have to figure out how to fix the tupledesc thing. What we have\\n> now is not good.\\n>\\n\\ncannot be moved to nodefunc?\\n\\nPavel\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2016-12-07 20:50 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\\n> 2016-12-07 18:34 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> > Hmm. Now that I see how this works, by having the GetValue \\\"guess\\\" what\\n> > is going on and have a special case for it, I actually don't like it\\n> > very much. It seems way too magical. I think we should do away with\\n> > the \\\"if column is NULL\\\" case in GetValue, and instead inject a column\\n> > during transformTableExpr if columns is NIL. This has implications on\\n> > ExecInitExpr too, which currently checks for an empty column list -- it\\n> > would no longer have to do so.\\n>\\n> I prefer this way against second described. The implementation should be in\\n> table builder routines, not in executor.\\n\\nWell, given the way you have implemented it, I would prefer the original\\ntoo. But your v23 is not what I meant. Essentially what you do in v23\\nis to communicate the lack of COLUMNS clause in a different way --\\npreviously it was \\\"ncolumns = 0\\\", now it's \\\"is_auto_col=true\\\". It's\\nstill \\\"magic\\\". It's not an improvement.is_auto_col is used primary for asserting. The table builder has information for decision in parameter path, when path is NULL.Hard to say, if this info should be assigned to column or to table. In both locations has sense. But somewhere should be some flag. \\n\\nWhat I want to happen is that there is no magic at all; it's up to\\ntransformExpr to make sure that when COLUMNS is empty, one column\\nappears and it must not be a magic column that makes the xml.c code act\\ndifferently, but rather to xml.c it should appear that this is just a\\nnormal column that happens to return the entire row. If I say \\\"COLUMNS\\nfoo PATH '/'\\\" I should be able to obtain a similar behavior (except that\\nin the current code, if I ask for \\\"COLUMNS foo XML PATH '/'\\\" I don't get\\nXML at all but rather weird text where all tags have been stripped out,\\nwhich is very strange. I would expect the tags to be preserved if the\\noutput type is XML. Maybe the tag-stripping behavior should occur if\\nthe output type is some type of text.)I am doing this. Just I using NULL for PATH. \\n\\n\\nI still have to figure out how to fix the tupledesc thing. What we have\\nnow is not good.cannot be moved to nodefunc?Pavel \\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Wed, 7 Dec 2016 21:48:20 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-12-07 20:37 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n\\n>\\n>\\n> 2016-12-07 18:34 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>\\n>> Pavel Stehule wrote:\\n>>\\n>> > I fixed two issues.\\n>> >\\n>> > 2. there was reverse setting in NOT NULL flag\\n>>\\n>> Ah-hah, that was silly, thanks.\\n>>\\n>> > 1. there are not columns data when there are not any explicit column -\\n>> fixed\\n>>\\n>> Hmm. Now that I see how this works, by having the GetValue \\\"guess\\\" what\\n>> is going on and have a special case for it, I actually don't like it\\n>> very much. It seems way too magical. I think we should do away with\\n>> the \\\"if column is NULL\\\" case in GetValue, and instead inject a column\\n>> during transformTableExpr if columns is NIL. This has implications on\\n>> ExecInitExpr too, which currently checks for an empty column list -- it\\n>> would no longer have to do so.\\n>>\\n>\\n> I prefer this way against second described. The implementation should be\\n> in table builder routines, not in executor.\\n>\\n> sending new update\\n>\\n\\nnew update - no functional changes, just unbreaking after last changes in\\nmaster.\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> Regards\\n>\\n> Pavel\\n>\\n>\\n>>\\n>> Maybe this means we need an additional method, which would request \\\"the\\n>> expr that returns the whole row\\\", so that transformExpr can work for\\n>> XmlTable (which I think would be something like \\\"./\\\") and the future\\n>> JsonTable stuff (I don't know how that one would work, but I assume it's\\n>> not necessarily the same thing).\\n>>\\n>\\n>\\n>\\n>\\n>>\\n>> --\\n>> Álvaro Herrera https://www.2ndQuadrant.com/\\n>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\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\": \"Sun, 18 Dec 2016 16:27:54 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2016-12-18 16:27 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n\\n>\\n>\\n> 2016-12-07 20:37 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n>\\n>>\\n>>\\n>> 2016-12-07 18:34 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>>\\n>>> Pavel Stehule wrote:\\n>>>\\n>>> > I fixed two issues.\\n>>> >\\n>>> > 2. there was reverse setting in NOT NULL flag\\n>>>\\n>>> Ah-hah, that was silly, thanks.\\n>>>\\n>>> > 1. there are not columns data when there are not any explicit column -\\n>>> fixed\\n>>>\\n>>> Hmm. Now that I see how this works, by having the GetValue \\\"guess\\\" what\\n>>> is going on and have a special case for it, I actually don't like it\\n>>> very much. It seems way too magical. I think we should do away with\\n>>> the \\\"if column is NULL\\\" case in GetValue, and instead inject a column\\n>>> during transformTableExpr if columns is NIL. This has implications on\\n>>> ExecInitExpr too, which currently checks for an empty column list -- it\\n>>> would no longer have to do so.\\n>>>\\n>>\\n>> I prefer this way against second described. The implementation should be\\n>> in table builder routines, not in executor.\\n>>\\n>> sending new update\\n>>\\n>\\n> new update - no functional changes, just unbreaking after last changes in\\n> master.\\n>\\n\\nanother update - lot of cleaning\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> Regards\\n>\\n> Pavel\\n>\\n>\\n>>\\n>> Regards\\n>>\\n>> Pavel\\n>>\\n>>\\n>>>\\n>>> Maybe this means we need an additional method, which would request \\\"the\\n>>> expr that returns the whole row\\\", so that transformExpr can work for\\n>>> XmlTable (which I think would be something like \\\"./\\\") and the future\\n>>> JsonTable stuff (I don't know how that one would work, but I assume it's\\n>>> not necessarily the same thing).\\n>>>\\n>>\\n>>\\n>>\\n>>\\n>>>\\n>>> --\\n>>> Álvaro Herrera https://www.2ndQuadrant.com/\\n>>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>>>\\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\": \"Thu, 22 Dec 2016 17:11:47 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n\\n> another update - lot of cleaning\\n\\nAh, the tupledesc stuff in this one appears much more reasonable to me.\\n\\n-- \\n�lvaro Herrera https://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, 22 Dec 2016 13:27:17 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n\\n> another update - lot of cleaning\\n\\nThanks.\\n\\nThe more I look at this, the less I like using NameArgExpr for\\nnamespaces. It looks all wrong to me, and it causes ugly code all over.\\nMaybe I just need to look at it a bit longer.\\n\\n\\n\\n-- \\n�lvaro Herrera https://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, 11 Jan 2017 18:05:12 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera wrote:\\n\\n> The more I look at this, the less I like using NameArgExpr for\\n> namespaces. It looks all wrong to me, and it causes ugly code all over.\\n> Maybe I just need to look at it a bit longer.\\n\\nI think it'd be cleaner to use ResTarget for the namespaces, like\\nxml_attribute_el does, and split the names from actual exprs in the same\\nway. So things like ExecInitExpr become simpler because you just\\nrecurse to initialize the list, without having to examine each element\\nindividually. tabexprInitialize can just do forboth().\\n\\nThe main reason I don't like abusing NamedArgExpr is that the whole\\ncomment that explains it becomes a lie.\\n\\n-- \\n�lvaro Herrera https://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\": \"Wed, 11 Jan 2017 18:53:25 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2017-01-11 22:53 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Alvaro Herrera wrote:\\n>\\n> > The more I look at this, the less I like using NameArgExpr for\\n> > namespaces. It looks all wrong to me, and it causes ugly code all over.\\n> > Maybe I just need to look at it a bit longer.\\n>\\n> I think it'd be cleaner to use ResTarget for the namespaces, like\\n> xml_attribute_el does, and split the names from actual exprs in the same\\n> way. So things like ExecInitExpr become simpler because you just\\n> recurse to initialize the list, without having to examine each element\\n> individually. tabexprInitialize can just do forboth().\\n>\\n> The main reason I don't like abusing NamedArgExpr is that the whole\\n> comment that explains it becomes a lie.\\n>\\n\\nI used your proposed way based on Restarget\\n\\nUpdated patch attached.\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL 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, 12 Jan 2017 16:02:07 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n> \\n> I used your proposed way based on Restarget\\n\\nThanks. Some more tweaking to go yet before I consider this\\ncommittable, but it's much better now. Here's v28. I changed a few\\nthings:\\n\\n- make expression evaluation code more orthodox:\\n 1. avoid PG_TRY, use a ExprContext shutdown callback instead\\n 2. use a \\\"Fast\\\" evaluator, for calls past the first one\\n 3. don't look up fmgrinfos until execution time\\n 4. don't duplicate get_expr_result_type\\n- make parser accept DEFAULT namespace. Only xml implementation barfs.\\n (this means we lost the errposition pointer, but I don't really\\n care. We could fix it if we cared)\\n- clean up parse analysis code a little bit\\n- move decls/struct defs to better locations in source code\\n- remove leftover \\\"namespaces\\\" in TableExprState\\n- pgindent the whole mess.\\n\\nI don't like the xml.c code and the \\\"evalcols\\\" flag. That's next on my\\nlist to fix.\\n\\nI don't think to_xmlstr() is necessary, considering xml_text2xmlChar.\\nWe could just apply a cast of the source cstring to xmlChar.\\n\\n-- \\n�lvaro Herrera https://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\": \"Fri, 13 Jan 2017 17:32:08 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2017-01-13 21:32 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n> >\\n> > I used your proposed way based on Restarget\\n>\\n> Thanks. Some more tweaking to go yet before I consider this\\n> committable, but it's much better now. Here's v28. I changed a few\\n> things:\\n>\\n> - make expression evaluation code more orthodox:\\n> 1. avoid PG_TRY, use a ExprContext shutdown callback instead\\n> 2. use a \\\"Fast\\\" evaluator, for calls past the first one\\n> 3. don't look up fmgrinfos until execution time\\n> 4. don't duplicate get_expr_result_type\\n> - make parser accept DEFAULT namespace. Only xml implementation barfs.\\n> (this means we lost the errposition pointer, but I don't really\\n> care. We could fix it if we cared)\\n> - clean up parse analysis code a little bit\\n> - move decls/struct defs to better locations in source code\\n> - remove leftover \\\"namespaces\\\" in TableExprState\\n> - pgindent the whole mess.\\n>\\n>\\nI checked the changes and looks correct - although for some I had not\\ncourage :) - like dynamic change of exprstate->evalfunc\\n\\nI fixed test, and append forgotten header file\\n\\n\\n\\n\\n> I don't like the xml.c code and the \\\"evalcols\\\" flag. That's next on my\\n> list to fix.\\n>\\n\\nYou need some flag to specify if column paths are valid or not.\\n\\n\\n> I don't think to_xmlstr() is necessary, considering xml_text2xmlChar.\\n> We could just apply a cast of the source cstring to xmlChar.\\n>\\n\\nis it safe? For one byte encodings?\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL 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\": \"Sat, 14 Jan 2017 14:43:26 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-01-14 14:43 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n\\n> Hi\\n>\\n> 2017-01-13 21:32 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>\\n>> Pavel Stehule wrote:\\n>> >\\n>> > I used your proposed way based on Restarget\\n>>\\n>> Thanks. Some more tweaking to go yet before I consider this\\n>> committable, but it's much better now. Here's v28. I changed a few\\n>> things:\\n>>\\n>> - make expression evaluation code more orthodox:\\n>> 1. avoid PG_TRY, use a ExprContext shutdown callback instead\\n>> 2. use a \\\"Fast\\\" evaluator, for calls past the first one\\n>> 3. don't look up fmgrinfos until execution time\\n>> 4. don't duplicate get_expr_result_type\\n>> - make parser accept DEFAULT namespace. Only xml implementation barfs.\\n>> (this means we lost the errposition pointer, but I don't really\\n>> care. We could fix it if we cared)\\n>> - clean up parse analysis code a little bit\\n>> - move decls/struct defs to better locations in source code\\n>> - remove leftover \\\"namespaces\\\" in TableExprState\\n>> - pgindent the whole mess.\\n>>\\n>>\\n> I checked the changes and looks correct - although for some I had not\\n> courage :) - like dynamic change of exprstate->evalfunc\\n>\\n> I fixed test, and append forgotten header file\\n>\\n>\\n>\\n>\\n>> I don't like the xml.c code and the \\\"evalcols\\\" flag. That's next on my\\n>> list to fix.\\n>>\\n>\\n> You need some flag to specify if column paths are valid or not.\\n>\\n>\\n>> I don't think to_xmlstr() is necessary, considering xml_text2xmlChar.\\n>> We could just apply a cast of the source cstring to xmlChar.\\n>>\\n>\\n> is it safe? For one byte encodings?\\n>\\n\\nLooks so this patch breaks regression tests\\n\\nestoring database schemas in the new cluster\\n \\\\\\\"\\\\ cdefghijklmnopqrstuvwxyz{|}~\\n\\n regression\\n*failure*\\n\\nConsult the last few lines of \\\"pg_upgrade_dump_16387.log\\\" for\\nthe probable cause of the failure.\\nFailure, exiting\\n+ rm -rf /tmp/pg_upgrade_check-wSfzCh\\nMakefile:39: návod pro cíl „check“ selhal\\nmake[2]: *** [check] Chyba 1\\nmake[2]: Opouští se adresář „/home/pavel/src/postgresql/src/bin/pg_upgrade“\\n\\n\\npg_restore: [archiver (db)] Error while PROCESSING TOC:\\npg_restore: [archiver (db)] Error from TOC entry 496; 1259 47693 VIEW\\nxmltableview2 pavel\\npg_restore: [archiver (db)] could not execute query: ERROR: syntax error\\nat or near \\\"(\\\"\\nLINE 15: ...XMLTABLE(XMLNAMESPACES('http://x.y'::\\\"text\\\" AS zz)('/zz:rows...\\n\\nFixed in attached patch\\n\\n\\n\\n\\n> Regards\\n>\\n> Pavel\\n>\\n>\\n>>\\n>> --\\n>> Álvaro Herrera https://www.2ndQuadrant.com/\\n>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>>\\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\": \"Sun, 15 Jan 2017 07:29:22 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"I just realized that your new xml_xmlnodetostr is line-by-line identical\\nto the existing xml_xmlnodetoxmltype except for two or three lines.\\nI think that's wrong. I'm going to patch the existing function so that\\nthey can share code.\\n\\n-- \\n�lvaro Herrera https://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\": \"Mon, 16 Jan 2017 16:30:02 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Given\\nhttps://www.postgresql.org/message-id/20170116210019.a3glfwspg5lnfrnm@alap3.anarazel.de\\nwhich is going to heavily change how the executor works in this area, I\\nam returning this patch to you again. I would like a few rather minor\\nchanges:\\n\\n1. to_xmlstr can be replaced with calls to xmlCharStrdup.\\n2. don't need xml_xmlnodetostr either -- just use xml_xmlnodetoxmltype\\n (which returns text*) and extract the cstring from the varlena. It's\\n a bit more wasteful in terms of cycles, but I don't think we care.\\n If we do care, change the function so that it returns cstring, and\\n have the callers that want text wrap it in cstring_to_text.\\n3. have a new perValueCxt memcxt in TableExprState, child of buildercxt,\\n and switch to it just before GetValue() (reset it just before\\n switching). Then, don't worry about leaks in GetValue. This way,\\n the text* conversions et al don't matter.\\n\\nAfter that I think we're going to need to get this working on top of\\nAndres' changes. Which I'm afraid is going to be rather major surgery,\\nbut I haven't looked.\\n\\n-- \\n�lvaro Herrera https://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\": \"Mon, 16 Jan 2017 19:51:13 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"In case this still matters, I think GetValue should look more or less\\nlike this (untested):\\n\\n\\n/*\\n * Return the value for column number 'colnum' for the current row. If column\\n * -1 is requested, return representation of the whole row.\\n *\\n * This leaks memory, so be sure to reset often the context in which it's\\n * called.\\n */\\nstatic Datum\\nXmlTableGetValue(TableExprState *state, int colnum, bool *isnull)\\n{\\n#ifdef USE_LIBXML\\n\\tXmlTableBuilderData *xtCxt;\\n\\tDatum\\t\\tresult = (Datum) 0;\\n\\txmlNodePtr\\tcur;\\n\\tchar\\t *cstr = NULL;\\n\\tvolatile xmlXPathObjectPtr xpathobj;\\n\\n\\txtCxt = GetXmlTableBuilderPrivateData(state, \\\"XmlTableGetValue\\\");\\n\\n\\tAssert(xtCxt->xpathobj &&\\n\\t\\t xtCxt->xpathobj->type == XPATH_NODESET &&\\n\\t\\t xtCxt->xpathobj->nodesetval != NULL);\\n\\n\\t/* Propagate context related error context to libxml2 */\\n\\txmlSetStructuredErrorFunc((void *) xtCxt->xmlerrcxt, xml_errorHandler);\\n\\n\\tcur = xtCxt->xpathobj->nodesetval->nodeTab[xtCxt->row_count - 1];\\n\\tif (cur->type != XML_ELEMENT_NODE)\\n\\t\\telog(ERROR, \\\"unexpected xmlNode type\\\");\\n\\n\\t/* Handle whole row case the easy way. */\\n\\tif (colnum == -1)\\n\\t{\\n\\t\\ttext\\t *txt;\\n\\n\\t\\ttxt = xml_xmlnodetoxmltype(cur, xtCxt->xmlerrcxt);\\n\\t\\tresult = InputFunctionCall(&state->in_functions[0],\\n\\t\\t\\t\\t\\t\\t\\t\\t text_to_cstring(txt),\\n\\t\\t\\t\\t\\t\\t\\t\\t state->typioparams[0],\\n\\t\\t\\t\\t\\t\\t\\t\\t -1);\\n\\t\\t*isnull = false;\\n\\n\\t\\treturn result;\\n\\t}\\n\\n\\tAssert(xtCxt->xpathscomp[colnum] != NULL);\\n\\n\\txpathobj = NULL;\\n\\tPG_TRY();\\n\\t{\\n\\t\\tForm_pg_attribute attr;\\n\\n\\t\\tattr = state->resultSlot->tts_tupleDescriptor->attrs[colnum];\\n\\n\\t\\t/* Set current node as entry point for XPath evaluation */\\n\\t\\txmlXPathSetContextNode(cur, xtCxt->xpathcxt);\\n\\n\\t\\t/* Evaluate column path */\\n\\t\\txpathobj = xmlXPathCompiledEval(xtCxt->xpathscomp[colnum], xtCxt->xpathcxt);\\n\\t\\tif (xpathobj == NULL || xtCxt->xmlerrcxt->err_occurred)\\n\\t\\t\\txml_ereport(xtCxt->xmlerrcxt, ERROR, ERRCODE_INTERNAL_ERROR,\\n\\t\\t\\t\\t\\t\\t\\\"could not create XPath object\\\");\\n\\n\\t\\tif (xpathobj->type == XPATH_NODESET)\\n\\t\\t{\\n\\t\\t\\tint\\t\\t\\tcount;\\n\\t\\t\\tOid\\t\\t\\ttargettypid = attr->atttypid;\\n\\n\\t\\t\\tif (xpathobj->nodesetval != NULL)\\n\\t\\t\\t\\tcount = xpathobj->nodesetval->nodeNr;\\n\\n\\t\\t\\t/*\\n\\t\\t\\t * There are four possible cases, depending on the number of\\n\\t\\t\\t * nodes returned by the XPath expression and the type of the\\n\\t\\t\\t * target column: a) XPath returns no nodes. b) One node is\\n\\t\\t\\t * returned, and column is of type XML. c) One node, column type\\n\\t\\t\\t * other than XML. d) Multiple nodes are returned.\\n\\t\\t\\t */\\n\\t\\t\\tif (xpathobj->nodesetval == NULL)\\n\\t\\t\\t{\\n\\t\\t\\t\\t*isnull = true;\\n\\t\\t\\t}\\n\\t\\t\\telse if (count == 1 && targettypid == XMLOID)\\n\\t\\t\\t{\\n\\t\\t\\t\\ttextstr = xml_xmlnodetoxmltype(xpathobj->nodesetval->nodeTab[0],\\n\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t xtCxt->xmlerrcxt);\\n\\t\\t\\t\\tcstr = text_to_cstring(textstr);\\n\\t\\t\\t}\\n\\t\\t\\telse if (count == 1)\\n\\t\\t\\t{\\n\\t\\t\\t\\txmlChar *str;\\n\\n\\t\\t\\t\\tstr = xmlNodeListGetString(xtCxt->doc,\\n\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t xpathobj->nodesetval->nodeTab[0]->xmlChildrenNode,\\n\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t 1);\\n\\t\\t\\t\\tif (str)\\n\\t\\t\\t\\t{\\n\\t\\t\\t\\t\\tPG_TRY();\\n\\t\\t\\t\\t\\t{\\n\\t\\t\\t\\t\\t\\tcstr = pstrdup(str);\\n\\t\\t\\t\\t\\t}\\n\\t\\t\\t\\t\\tPG_CATCH();\\n\\t\\t\\t\\t\\t{\\n\\t\\t\\t\\t\\t\\txmlFree(str);\\n\\t\\t\\t\\t\\t\\tPG_RE_THROW();\\n\\t\\t\\t\\t\\t}\\n\\t\\t\\t\\t\\tPG_END_TRY();\\n\\t\\t\\t\\t\\txmlFree(str);\\n\\t\\t\\t\\t}\\n\\t\\t\\t\\telse\\n\\t\\t\\t\\t\\tcstr = pstrdup(\\\"\\\");\\n\\t\\t\\t}\\n\\t\\t\\telse\\n\\t\\t\\t{\\n\\t\\t\\t\\tStringInfoData buf;\\n\\t\\t\\t\\tint\\t\\t\\ti;\\n\\n\\t\\t\\t\\tAssert(count > 1);\\n\\n\\t\\t\\t\\t/*\\n\\t\\t\\t\\t * When evaluating the XPath expression returns multiple\\n\\t\\t\\t\\t * nodes, the result is the concatenation of them all.\\n\\t\\t\\t\\t * The target type must be XML.\\n\\t\\t\\t\\t */\\n\\t\\t\\t\\tif (targettypid != XMLOID)\\n\\t\\t\\t\\t\\tereport(ERROR,\\n\\t\\t\\t\\t\\t\\t\\t(errcode(ERRCODE_CARDINALITY_VIOLATION),\\n\\t\\t\\t\\t\\t\\t\\t errmsg(\\\"more than one value returned by column XPath expression\\\")));\\n\\n\\t\\t\\t\\tinitStringInfo(&buf);\\n\\t\\t\\t\\tfor (i = 0; i < count; i++)\\n\\t\\t\\t\\t\\t/* worth freeing the text here? Naahh ... */\\n\\t\\t\\t\\t\\tappendStringInfoText(&buf,\\n\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t xml_xmlnodetoxmltype(xpathobj->nodesetval->nodeTab[i],\\n\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t xtCxt->xmlerrcxt));\\n\\t\\t\\t\\tcstr = buf.data;\\n\\t\\t\\t}\\n\\t\\t}\\n\\t\\telse if (xpathobj->type == XPATH_STRING)\\n\\t\\t{\\n\\t\\t\\tcstr = (char *) xpathobj->stringval;\\n\\t\\t\\t*isnull = false;\\n\\t\\t}\\n\\t\\telse\\n\\t\\t\\telog(ERROR, \\\"unexpected XPath object type %u\\\", xpathobj->type);\\n\\n\\t\\t/*\\n\\t\\t * By here, either cstr contains the result value, or the isnull flag\\n\\t\\t * has been set.\\n\\t\\t */\\n\\t\\tAssert(cstr || *isnull);\\n\\n\\t\\tif (!*isnull)\\n\\t\\t\\tresult = InputFunctionCall(&state->in_functions[colnum],\\n\\t\\t\\t\\t\\t\\t\\t\\t\\t cstr,\\n\\t\\t\\t\\t\\t\\t\\t\\t\\t state->typioparams[colnum],\\n\\t\\t\\t\\t\\t\\t\\t\\t\\t attr->atttypmod);\\n\\t}\\n\\tPG_CATCH();\\n\\t{\\n\\t\\tif (xpathobj != NULL)\\n\\t\\t\\txmlXPathFreeObject(xpathobj);\\n\\t\\tPG_RE_THROW();\\n\\t}\\n\\tPG_END_TRY();\\n\\n\\tif (xpathobj)\\n\\t\\txmlXPathFreeObject(xpathobj);\\n\\n\\treturn result;\\n#else\\n\\tNO_XML_SUPPORT();\\n#endif /* not USE_LIBXML */\\n}\\n\\n-- \\n�lvaro Herrera https://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\": \"Mon, 16 Jan 2017 20:30:00 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-01-16 23:51 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Given\\n> https://www.postgresql.org/message-id/20170116210019.\\n> a3glfwspg5lnfrnm@alap3.anarazel.de\\n> which is going to heavily change how the executor works in this area, I\\n> am returning this patch to you again. I would like a few rather minor\\n> changes:\\n>\\n> 1. to_xmlstr can be replaced with calls to xmlCharStrdup.\\n> 2. don't need xml_xmlnodetostr either -- just use xml_xmlnodetoxmltype\\n> (which returns text*) and extract the cstring from the varlena. It's\\n> a bit more wasteful in terms of cycles, but I don't think we care.\\n> If we do care, change the function so that it returns cstring, and\\n> have the callers that want text wrap it in cstring_to_text.\\n> 3. have a new perValueCxt memcxt in TableExprState, child of buildercxt,\\n> and switch to it just before GetValue() (reset it just before\\n> switching). Then, don't worry about leaks in GetValue. This way,\\n> the text* conversions et al don't matter.\\n>\\n> After that I think we're going to need to get this working on top of\\n> Andres' changes. Which I'm afraid is going to be rather major surgery,\\n> but I haven't looked.\\n>\\n\\nI'll try to clean xml part first, and then I can reflect the SRF changes. I\\nam not sure if I understand to all your proposed changes here, I have to\\nlook there.\\n\\nRegards\\n\\nPavel\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2017-01-16 23:51 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Given\\nhttps://www.postgresql.org/message-id/20170116210019.a3glfwspg5lnfrnm@alap3.anarazel.de\\nwhich is going to heavily change how the executor works in this area, I\\nam returning this patch to you again. I would like a few rather minor\\nchanges:\\n\\n1. to_xmlstr can be replaced with calls to xmlCharStrdup.\\n2. don't need xml_xmlnodetostr either -- just use xml_xmlnodetoxmltype\\n (which returns text*) and extract the cstring from the varlena. It's\\n a bit more wasteful in terms of cycles, but I don't think we care.\\n If we do care, change the function so that it returns cstring, and\\n have the callers that want text wrap it in cstring_to_text.\\n3. have a new perValueCxt memcxt in TableExprState, child of buildercxt,\\n and switch to it just before GetValue() (reset it just before\\n switching). Then, don't worry about leaks in GetValue. This way,\\n the text* conversions et al don't matter.\\n\\nAfter that I think we're going to need to get this working on top of\\nAndres' changes. Which I'm afraid is going to be rather major surgery,\\nbut I haven't looked.I'll try to clean xml part first, and then I can reflect the SRF changes. I am not sure if I understand to all your proposed changes here, I have to look there.RegardsPavel \\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Tue, 17 Jan 2017 19:07:42 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2017-01-16 23:51 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Given\\n> https://www.postgresql.org/message-id/20170116210019.a3glfws\\n> pg5lnfrnm@alap3.anarazel.de\\n> which is going to heavily change how the executor works in this area, I\\n> am returning this patch to you again. I would like a few rather minor\\n> changes:\\n>\\n> 1. to_xmlstr can be replaced with calls to xmlCharStrdup.\\n>\\n\\nI checked this idea, and it doesn't look well - xmlCharStrdup created xml\\nstring in own memory - and it should be explicitly released with xmlFree().\\nIn this case is more practical using PostgreSQL memory context - because\\nthis memory is released safely in exception. I can rename this function to\\nmore conventional pg_xmlCharStrndup. This function can be used more time in\\ncurrent code.\\n\\n\\n> 2. don't need xml_xmlnodetostr either -- just use xml_xmlnodetoxmltype\\n> (which returns text*) and extract the cstring from the varlena. It's\\n> a bit more wasteful in terms of cycles, but I don't think we care.\\n> If we do care, change the function so that it returns cstring, and\\n> have the callers that want text wrap it in cstring_to_text.\\n>\\n\\ndone - it is related to not too often use case, and possible slowdown is\\nminimal\\n\\n\\n> 3. have a new perValueCxt memcxt in TableExprState, child of buildercxt,\\n> and switch to it just before GetValue() (reset it just before\\n> switching). Then, don't worry about leaks in GetValue. This way,\\n> the text* conversions et al don't matter.\\n>\\n\\ndone\\n\\n\\n>\\n> After that I think we're going to need to get this working on top of\\n> Andres' changes. Which I'm afraid is going to be rather major surgery,\\n> but I haven't looked.\\n>\\n\\nI am waiting on new commits in this area. This moment I have not idea what\\nwill be broken.\\n\\nattached updated patch with cleaned xml part\\n\\nRegards\\n\\nPavel\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL 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\": \"Wed, 18 Jan 2017 22:08:07 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\nNew update - rebase after yesterday changes.\\n\\nWhat you want to change?\\n\\nRegards\\n\\nPavel\\n\\n\\n-- \\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 Jan 2017 08:48:32 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n> Hi\\n> \\n> New update - rebase after yesterday changes.\\n> \\n> What you want to change?\\n\\nI think the problem might come from the still pending patch on that\\nseries, which Andres posted in\\nhttps://www.postgresql.org/message-id/20170118221154.aldebi7yyjvds5qa@alap3.anarazel.de\\nAs far as I understand, minor details of that patch might change before\\ncommit, but it is pretty much in definitive form.\\n\\n-- \\n�lvaro Herrera https://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, 19 Jan 2017 09:35:56 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-01-19 13:35 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n> > Hi\\n> >\\n> > New update - rebase after yesterday changes.\\n> >\\n> > What you want to change?\\n>\\n> I think the problem might come from the still pending patch on that\\n> series, which Andres posted in\\n> https://www.postgresql.org/message-id/20170118221154.\\n> aldebi7yyjvds5qa@alap3.anarazel.de\\n> As far as I understand, minor details of that patch might change before\\n> commit, but it is pretty much in definitive form.\\n>\\n\\nok, we have to wait - please, check XML part if it is good for you\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2017-01-19 13:35 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\\n> Hi\\n>\\n> New update - rebase after yesterday changes.\\n>\\n> What you want to change?\\n\\nI think the problem might come from the still pending patch on that\\nseries, which Andres posted in\\nhttps://www.postgresql.org/message-id/20170118221154.aldebi7yyjvds5qa@alap3.anarazel.de\\nAs far as I understand, minor details of that patch might change before\\ncommit, but it is pretty much in definitive form.ok, we have to wait - please, check XML part if it is good for youRegardsPavel \\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Thu, 19 Jan 2017 13:40:22 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2017-01-19 13:35 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n> > Hi\\n> >\\n> > New update - rebase after yesterday changes.\\n> >\\n> > What you want to change?\\n>\\n> I think the problem might come from the still pending patch on that\\n> series, which Andres posted in\\n> https://www.postgresql.org/message-id/20170118221154.\\n> aldebi7yyjvds5qa@alap3.anarazel.de\\n> As far as I understand, minor details of that patch might change before\\n> commit, but it is pretty much in definitive form.\\n>\\n\\nnew rebased update after these changes\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL 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\": \"Sat, 21 Jan 2017 09:30:17 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-01-21 9:30 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n\\n> Hi\\n>\\n> 2017-01-19 13:35 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>\\n>> Pavel Stehule wrote:\\n>> > Hi\\n>> >\\n>> > New update - rebase after yesterday changes.\\n>> >\\n>> > What you want to change?\\n>>\\n>> I think the problem might come from the still pending patch on that\\n>> series, which Andres posted in\\n>> https://www.postgresql.org/message-id/20170118221154.aldebi7\\n>> yyjvds5qa@alap3.anarazel.de\\n>> As far as I understand, minor details of that patch might change before\\n>> commit, but it is pretty much in definitive form.\\n>>\\n>\\n> new rebased update after these changes\\n>\\n\\nfix white spaces\\n\\npavel\\n\\n\\n>\\n> Regards\\n>\\n> Pavel\\n>\\n>\\n>>\\n>> --\\n>> Álvaro Herrera https://www.2ndQuadrant.com/\\n>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\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\": \"Sat, 21 Jan 2017 10:31:48 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2017-01-21 10:31 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n\\n>\\n>\\n> 2017-01-21 9:30 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n>\\n>> Hi\\n>>\\n>> 2017-01-19 13:35 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>>\\n>>> Pavel Stehule wrote:\\n>>> > Hi\\n>>> >\\n>>> > New update - rebase after yesterday changes.\\n>>> >\\n>>> > What you want to change?\\n>>>\\n>>> I think the problem might come from the still pending patch on that\\n>>> series, which Andres posted in\\n>>> https://www.postgresql.org/message-id/20170118221154.aldebi7\\n>>> yyjvds5qa@alap3.anarazel.de\\n>>> As far as I understand, minor details of that patch might change before\\n>>> commit, but it is pretty much in definitive form.\\n>>>\\n>>\\n>> new rebased update after these changes\\n>>\\n>\\n> fix white spaces\\n>\\n\\nfew fixes:\\n\\n* SELECT (xmltable(..)).* + regress tests\\n* compilation and regress tests without --with-libxml\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> pavel\\n>\\n>\\n>>\\n>> Regards\\n>>\\n>> Pavel\\n>>\\n>>\\n>>>\\n>>> --\\n>>> Álvaro Herrera https://www.2ndQuadrant.com/\\n>>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>>>\\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\": \"Sun, 22 Jan 2017 10:57:22 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n\\n> * SELECT (xmltable(..)).* + regress tests\\n> * compilation and regress tests without --with-libxml\\n\\nThanks. I just realized that this is doing more work than necessary --\\nI think it would be simpler to have tableexpr fill a tuplestore with the\\nresults, instead of just expecting function execution to apply\\nExecEvalExpr over and over to obtain the results. So evaluating a\\ntableexpr returns just the tuplestore, which function evaluation can\\nreturn as-is. That code doesn't use the value-per-call interface\\nanyway.\\n\\nI also realized that the expr context callback is not called if there's\\nan error, which leaves us without shutting down libxml properly. I\\nadded PG_TRY around the fetchrow calls, but I'm not sure that's correct\\neither, because there could be an error raised in other parts of the\\ncode, after we've already emitted a few rows (for example out of\\nmemory). I think the right way is to have PG_TRY around the execution\\nof the whole thing rather than just row at a time; and the tuplestore\\nmechanism helps us with that.\\n\\nI think it would be good to have a more complex test case in regress --\\nlet's say there is a table with some simple XML values, then we use\\nXMLFOREST (or maybe one of the table_to_xml functions) to generate a\\nlarge document, and then XMLTABLE uses that document as input document.\\n\\nPlease fix.\\n\\n-- \\n�lvaro Herrera https://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\": \"Tue, 24 Jan 2017 17:38:49 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2017-01-24 17:38:49 -0300, Alvaro Herrera wrote:\\n> +static Datum ExecEvalTableExpr(TableExprState *tstate, ExprContext *econtext,\\n> +\\t\\t\\t\\t bool *isnull);\\n> +static Datum ExecEvalTableExprFast(TableExprState *exprstate, ExprContext *econtext,\\n> +\\t\\t\\t\\t\\t bool *isNull);\\n> +static Datum tabexprFetchRow(TableExprState *tstate, ExprContext *econtext,\\n> +\\t\\t\\t\\tbool *isNull);\\n> +static void tabexprInitialize(TableExprState *tstate, ExprContext *econtext,\\n> +\\t\\t\\t\\t Datum doc);\\n> +static void ShutdownTableExpr(Datum arg);\\n\\nTo me this (and a lot of the other code) hints quite strongly that\\nexpression evalution is the wrong approach to implementing this. What\\nyou're essentially doing is building a vulcano style scan node. Even if\\nwe can this, we shouldn't double down on the bad decision to have these\\nmagic expressions that return multiple rows. There's historical reason\\nfor tSRFs, but we shouldn't add more weirdness like this.\\n\\nAndres\\n\\n\\n-- \\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 Jan 2017 14:37:44 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Andres Freund wrote:\\n> Hi,\\n> \\n> On 2017-01-24 17:38:49 -0300, Alvaro Herrera wrote:\\n> > +static Datum ExecEvalTableExpr(TableExprState *tstate, ExprContext *econtext,\\n> > +\\t\\t\\t\\t bool *isnull);\\n> > +static Datum ExecEvalTableExprFast(TableExprState *exprstate, ExprContext *econtext,\\n> > +\\t\\t\\t\\t\\t bool *isNull);\\n> > +static Datum tabexprFetchRow(TableExprState *tstate, ExprContext *econtext,\\n> > +\\t\\t\\t\\tbool *isNull);\\n> > +static void tabexprInitialize(TableExprState *tstate, ExprContext *econtext,\\n> > +\\t\\t\\t\\t Datum doc);\\n> > +static void ShutdownTableExpr(Datum arg);\\n> \\n> To me this (and a lot of the other code) hints quite strongly that\\n> expression evalution is the wrong approach to implementing this. What\\n> you're essentially doing is building a vulcano style scan node. Even if\\n> we can this, we shouldn't double down on the bad decision to have these\\n> magic expressions that return multiple rows. There's historical reason\\n> for tSRFs, but we shouldn't add more weirdness like this.\\n\\nThanks for giving it a look. I have long thought that this patch would\\nbe at odds with your overall executor work.\\n\\nXMLTABLE is specified by the standard to return multiple rows ... but\\nthen as far as my reading goes, it is only supposed to be supported in\\nthe range table (FROM clause) not in the target list. I wonder if\\nthis would end up better if we only tried to support it in RT. I asked\\nPavel to implement it like that a few weeks ago, but ...\\n\\n-- \\n�lvaro Herrera https://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\": \"Tue, 24 Jan 2017 21:32:56 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\\n> Andres Freund wrote:\\n> > Hi,\\n> > \\n> > On 2017-01-24 17:38:49 -0300, Alvaro Herrera wrote:\\n> > > +static Datum ExecEvalTableExpr(TableExprState *tstate, ExprContext *econtext,\\n> > > +\\t\\t\\t\\t bool *isnull);\\n> > > +static Datum ExecEvalTableExprFast(TableExprState *exprstate, ExprContext *econtext,\\n> > > +\\t\\t\\t\\t\\t bool *isNull);\\n> > > +static Datum tabexprFetchRow(TableExprState *tstate, ExprContext *econtext,\\n> > > +\\t\\t\\t\\tbool *isNull);\\n> > > +static void tabexprInitialize(TableExprState *tstate, ExprContext *econtext,\\n> > > +\\t\\t\\t\\t Datum doc);\\n> > > +static void ShutdownTableExpr(Datum arg);\\n> > \\n> > To me this (and a lot of the other code) hints quite strongly that\\n> > expression evalution is the wrong approach to implementing this. What\\n> > you're essentially doing is building a vulcano style scan node. Even if\\n> > we can this, we shouldn't double down on the bad decision to have these\\n> > magic expressions that return multiple rows. There's historical reason\\n> > for tSRFs, but we shouldn't add more weirdness like this.\\n> \\n> Thanks for giving it a look. I have long thought that this patch would\\n> be at odds with your overall executor work.\\n\\nNot fundamentally, but it makes it harder.\\n\\n\\n> XMLTABLE is specified by the standard to return multiple rows ... but\\n> then as far as my reading goes, it is only supposed to be supported in\\n> the range table (FROM clause) not in the target list. I wonder if\\n> this would end up better if we only tried to support it in RT. I asked\\n> Pavel to implement it like that a few weeks ago, but ...\\n\\nRight - it makes sense in the FROM list - but then it should be an\\nexecutor node, instead of some expression thingy.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n-- \\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 Jan 2017 16:35:11 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\\n>> XMLTABLE is specified by the standard to return multiple rows ... but\\n>> then as far as my reading goes, it is only supposed to be supported in\\n>> the range table (FROM clause) not in the target list. I wonder if\\n>> this would end up better if we only tried to support it in RT. I asked\\n>> Pavel to implement it like that a few weeks ago, but ...\\n\\n> Right - it makes sense in the FROM list - but then it should be an\\n> executor node, instead of some expression thingy.\\n\\n+1 --- we're out of the business of having simple expressions that\\nreturn rowsets.\\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, 24 Jan 2017 23:45:03 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2017-01-25 1:35 GMT+01:00 Andres Freund <andres@anarazel.de>:\\n\\n> On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\\n> > Andres Freund wrote:\\n> > > Hi,\\n> > >\\n> > > On 2017-01-24 17:38:49 -0300, Alvaro Herrera wrote:\\n> > > > +static Datum ExecEvalTableExpr(TableExprState *tstate, ExprContext\\n> *econtext,\\n> > > > + bool *isnull);\\n> > > > +static Datum ExecEvalTableExprFast(TableExprState *exprstate,\\n> ExprContext *econtext,\\n> > > > + bool *isNull);\\n> > > > +static Datum tabexprFetchRow(TableExprState *tstate, ExprContext\\n> *econtext,\\n> > > > + bool *isNull);\\n> > > > +static void tabexprInitialize(TableExprState *tstate, ExprContext\\n> *econtext,\\n> > > > + Datum doc);\\n> > > > +static void ShutdownTableExpr(Datum arg);\\n> > >\\n> > > To me this (and a lot of the other code) hints quite strongly that\\n> > > expression evalution is the wrong approach to implementing this. What\\n> > > you're essentially doing is building a vulcano style scan node. Even\\n> if\\n> > > we can this, we shouldn't double down on the bad decision to have these\\n> > > magic expressions that return multiple rows. There's historical reason\\n> > > for tSRFs, but we shouldn't add more weirdness like this.\\n> >\\n> > Thanks for giving it a look. I have long thought that this patch would\\n> > be at odds with your overall executor work.\\n>\\n> Not fundamentally, but it makes it harder.\\n>\\n\\nIf you plan to hold support SRFin target list, then nothing is different.\\nIn last patch is executed under nodeProjectSet.\\n\\n\\n>\\n>\\n> > XMLTABLE is specified by the standard to return multiple rows ... but\\n> > then as far as my reading goes, it is only supposed to be supported in\\n> > the range table (FROM clause) not in the target list. I wonder if\\n> > this would end up better if we only tried to support it in RT. I asked\\n> > Pavel to implement it like that a few weeks ago, but ...\\n>\\n> Right - it makes sense in the FROM list - but then it should be an\\n> executor node, instead of some expression thingy.\\n>\\n\\nThe XMLTABLE function is from user perspective, from implementation\\nperspective a form of SRF function. I use own executor node, because fcinfo\\nis complex already and not too enough to hold all information about result\\ncolumns.\\n\\nThe implementation as RT doesn't reduce code - it is just moving to\\ndifferent file.\\n\\nI'll try to explain my motivation. Please, check it and correct me if I am\\nwrong. I don't keep on my implementation - just try to implement XMLTABLE\\nbe consistent with another behave and be used all time without any\\nsurprise.\\n\\n1. Any function that produces a content can be used in target list. We\\nsupport SRF in target list and in FROM part. Why XMLTABLE should be a\\nexception?\\n\\n2. In standard the XMLTABLE is placed only on FROM part - but standard\\ndoesn't need to solve my question - there are not SRF functions allowed in\\ntargetlist.\\n\\nIf there be a common decision so this inconsistency (in behave of this kind\\nof functions) is expected, required - then I have not a problem to remove\\nthis support from XMLTABLE.\\n\\nIn this moment I don't see a technical reason for this step - with last\\nAndres changes the support of XMLTABLE in target list needs less than 40\\nlines and there is not any special path for XMLTABLE only. Andres write\\nsupport for SRF functions and SRF operator. TableExpr is third category.\\n\\nRegards\\n\\nPavel\\n\\n\\n> Greetings,\\n>\\n> Andres Freund\\n>\\n\\nHi2017-01-25 1:35 GMT+01:00 Andres Freund <andres@anarazel.de>:On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\\n> Andres Freund wrote:\\n> > Hi,\\n> >\\n> > On 2017-01-24 17:38:49 -0300, Alvaro Herrera wrote:\\n> > > +static Datum ExecEvalTableExpr(TableExprState *tstate, ExprContext *econtext,\\n> > > + bool *isnull);\\n> > > +static Datum ExecEvalTableExprFast(TableExprState *exprstate, ExprContext *econtext,\\n> > > + bool *isNull);\\n> > > +static Datum tabexprFetchRow(TableExprState *tstate, ExprContext *econtext,\\n> > > + bool *isNull);\\n> > > +static void tabexprInitialize(TableExprState *tstate, ExprContext *econtext,\\n> > > + Datum doc);\\n> > > +static void ShutdownTableExpr(Datum arg);\\n> >\\n> > To me this (and a lot of the other code) hints quite strongly that\\n> > expression evalution is the wrong approach to implementing this. What\\n> > you're essentially doing is building a vulcano style scan node. Even if\\n> > we can this, we shouldn't double down on the bad decision to have these\\n> > magic expressions that return multiple rows. There's historical reason\\n> > for tSRFs, but we shouldn't add more weirdness like this.\\n>\\n> Thanks for giving it a look. I have long thought that this patch would\\n> be at odds with your overall executor work.\\n\\nNot fundamentally, but it makes it harder.If you plan to hold support SRFin target list, then nothing is different. In last patch is executed under nodeProjectSet. \\n\\n\\n> XMLTABLE is specified by the standard to return multiple rows ... but\\n> then as far as my reading goes, it is only supposed to be supported in\\n> the range table (FROM clause) not in the target list. I wonder if\\n> this would end up better if we only tried to support it in RT. I asked\\n> Pavel to implement it like that a few weeks ago, but ...\\n\\nRight - it makes sense in the FROM list - but then it should be an\\nexecutor node, instead of some expression thingy.The XMLTABLE function is from user perspective, from implementation perspective a form of SRF function. I use own executor node, because fcinfo is complex already and not too enough to hold all information about result columns. The implementation as RT doesn't reduce code - it is just moving to different file. I'll try to explain my motivation. Please, check it and correct me if I am wrong. I don't keep on my implementation - just try to implement XMLTABLE be consistent with another behave and be used all time without any surprise. 1. Any function that produces a content can be used in target list. We support SRF in target list and in FROM part. Why XMLTABLE should be a exception?2. In standard the XMLTABLE is placed only on FROM part - but standard doesn't need to solve my question - there are not SRF functions allowed in targetlist.If there be a common decision so this inconsistency (in behave of this kind of functions) is expected, required - then I have not a problem to remove this support from XMLTABLE.In this moment I don't see a technical reason for this step - with last Andres changes the support of XMLTABLE in target list needs less than 40 lines and there is not any special path for XMLTABLE only. Andres write support for SRF functions and SRF operator. TableExpr is third category.RegardsPavel\\n\\nGreetings,\\n\\nAndres Freund\",\n \"msg_date\": \"Wed, 25 Jan 2017 05:45:24 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-01-25 5:45 GMT+01:00 Tom Lane <tgl@sss.pgh.pa.us>:\\n\\n> Andres Freund <andres@anarazel.de> writes:\\n> > On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\\n> >> XMLTABLE is specified by the standard to return multiple rows ... but\\n> >> then as far as my reading goes, it is only supposed to be supported in\\n> >> the range table (FROM clause) not in the target list. I wonder if\\n> >> this would end up better if we only tried to support it in RT. I asked\\n> >> Pavel to implement it like that a few weeks ago, but ...\\n>\\n> > Right - it makes sense in the FROM list - but then it should be an\\n> > executor node, instead of some expression thingy.\\n>\\n> +1 --- we're out of the business of having simple expressions that\\n> return rowsets.\\n>\\n\\nIf we do decision so this kind of function will have different behave than\\nother SRF functions, then I remove support for this.\\n\\nThere are not technical reasons (maybe I don't see it) - last Andres\\nchanges do well support for my code.\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> regards, tom lane\\n>\\n\\n2017-01-25 5:45 GMT+01:00 Tom Lane <tgl@sss.pgh.pa.us>:Andres Freund <andres@anarazel.de> writes:\\n> On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\\n>> XMLTABLE is specified by the standard to return multiple rows ... but\\n>> then as far as my reading goes, it is only supposed to be supported in\\n>> the range table (FROM clause) not in the target list. I wonder if\\n>> this would end up better if we only tried to support it in RT. I asked\\n>> Pavel to implement it like that a few weeks ago, but ...\\n\\n> Right - it makes sense in the FROM list - but then it should be an\\n> executor node, instead of some expression thingy.\\n\\n+1 --- we're out of the business of having simple expressions that\\nreturn rowsets.If we do decision so this kind of function will have different behave than other SRF functions, then I remove support for this.There are not technical reasons (maybe I don't see it) - last Andres changes do well support for my code.RegardsPavel \\n\\n regards, tom lane\",\n \"msg_date\": \"Wed, 25 Jan 2017 05:48:58 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-01-24 21:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n>\\n> > * SELECT (xmltable(..)).* + regress tests\\n> > * compilation and regress tests without --with-libxml\\n>\\n> Thanks. I just realized that this is doing more work than necessary --\\n>\\n\\n?? I don't understand?\\n\\n\\n> I think it would be simpler to have tableexpr fill a tuplestore with the\\n> results, instead of just expecting function execution to apply\\n> ExecEvalExpr over and over to obtain the results. So evaluating a\\n> tableexpr returns just the tuplestore, which function evaluation can\\n> return as-is. That code doesn't use the value-per-call interface\\n> anyway.\\n>\\n\\nok\\n\\n\\n> I also realized that the expr context callback is not called if there's\\n> an error, which leaves us without shutting down libxml properly. I\\n> added PG_TRY around the fetchrow calls, but I'm not sure that's correct\\n> either, because there could be an error raised in other parts of the\\n> code, after we've already emitted a few rows (for example out of\\n> memory). I think the right way is to have PG_TRY around the execution\\n> of the whole thing rather than just row at a time; and the tuplestore\\n> mechanism helps us with that.\\n>\\n\\n\\nok.\\n\\n\\n\\n>\\n> I think it would be good to have a more complex test case in regress --\\n> let's say there is a table with some simple XML values, then we use\\n> XMLFOREST (or maybe one of the table_to_xml functions) to generate a\\n> large document, and then XMLTABLE uses that document as input document.\\n>\\n> Please fix.\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2017-01-24 21:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\\n\\n> * SELECT (xmltable(..)).* + regress tests\\n> * compilation and regress tests without --with-libxml\\n\\nThanks. I just realized that this is doing more work than necessary --?? I don't understand? \\nI think it would be simpler to have tableexpr fill a tuplestore with the\\nresults, instead of just expecting function execution to apply\\nExecEvalExpr over and over to obtain the results. So evaluating a\\ntableexpr returns just the tuplestore, which function evaluation can\\nreturn as-is. That code doesn't use the value-per-call interface\\nanyway.ok\\n\\nI also realized that the expr context callback is not called if there's\\nan error, which leaves us without shutting down libxml properly. I\\nadded PG_TRY around the fetchrow calls, but I'm not sure that's correct\\neither, because there could be an error raised in other parts of the\\ncode, after we've already emitted a few rows (for example out of\\nmemory). I think the right way is to have PG_TRY around the execution\\nof the whole thing rather than just row at a time; and the tuplestore\\nmechanism helps us with that.ok. \\n\\nI think it would be good to have a more complex test case in regress --\\nlet's say there is a table with some simple XML values, then we use\\nXMLFOREST (or maybe one of the table_to_xml functions) to generate a\\nlarge document, and then XMLTABLE uses that document as input document.\\n\\nPlease fix.\\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Wed, 25 Jan 2017 09:26:27 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Tom Lane wrote:\\n> Andres Freund <andres@anarazel.de> writes:\\n> > On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\\n> >> XMLTABLE is specified by the standard to return multiple rows ... but\\n> >> then as far as my reading goes, it is only supposed to be supported in\\n> >> the range table (FROM clause) not in the target list. I wonder if\\n> >> this would end up better if we only tried to support it in RT. I asked\\n> >> Pavel to implement it like that a few weeks ago, but ...\\n> \\n> > Right - it makes sense in the FROM list - but then it should be an\\n> > executor node, instead of some expression thingy.\\n> \\n> +1 --- we're out of the business of having simple expressions that\\n> return rowsets.\\n\\nWell, that's it. I'm not committing this patch against two other\\ncommitters' opinion, plus I was already on the fence about the\\nimplementation anyway. I think you should just go with the flow and\\nimplement this by creating nodeTableexprscan.c. It's not even\\ndifficult.\\n\\n-- \\n�lvaro Herrera https://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, 25 Jan 2017 11:07:47 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2017-01-25 05:45:24 +0100, Pavel Stehule wrote:\\n> 2017-01-25 1:35 GMT+01:00 Andres Freund <andres@anarazel.de>:\\n> \\n> > On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\\n> > > Andres Freund wrote:\\n> > > > Hi,\\n> > > >\\n> > > > On 2017-01-24 17:38:49 -0300, Alvaro Herrera wrote:\\n> > > > > +static Datum ExecEvalTableExpr(TableExprState *tstate, ExprContext\\n> > *econtext,\\n> > > > > + bool *isnull);\\n> > > > > +static Datum ExecEvalTableExprFast(TableExprState *exprstate,\\n> > ExprContext *econtext,\\n> > > > > + bool *isNull);\\n> > > > > +static Datum tabexprFetchRow(TableExprState *tstate, ExprContext\\n> > *econtext,\\n> > > > > + bool *isNull);\\n> > > > > +static void tabexprInitialize(TableExprState *tstate, ExprContext\\n> > *econtext,\\n> > > > > + Datum doc);\\n> > > > > +static void ShutdownTableExpr(Datum arg);\\n> > > >\\n> > > > To me this (and a lot of the other code) hints quite strongly that\\n> > > > expression evalution is the wrong approach to implementing this. What\\n> > > > you're essentially doing is building a vulcano style scan node. Even\\n> > if\\n> > > > we can this, we shouldn't double down on the bad decision to have these\\n> > > > magic expressions that return multiple rows. There's historical reason\\n> > > > for tSRFs, but we shouldn't add more weirdness like this.\\n> > >\\n> > > Thanks for giving it a look. I have long thought that this patch would\\n> > > be at odds with your overall executor work.\\n> >\\n> > Not fundamentally, but it makes it harder.\\n> >\\n> \\n> If you plan to hold support SRFin target list, then nothing is different.\\n> In last patch is executed under nodeProjectSet.\\n\\nIt is, because we suddenly need to call different functions - and I'm\\nrevamping most of execQual to have an opcode dispatch based execution\\nmodel (which then also can be JITed).\\n\\n\\n> > > XMLTABLE is specified by the standard to return multiple rows ... but\\n> > > then as far as my reading goes, it is only supposed to be supported in\\n> > > the range table (FROM clause) not in the target list. I wonder if\\n> > > this would end up better if we only tried to support it in RT. I asked\\n> > > Pavel to implement it like that a few weeks ago, but ...\\n> >\\n> > Right - it makes sense in the FROM list - but then it should be an\\n> > executor node, instead of some expression thingy.\\n> >\\n> \\n> The XMLTABLE function is from user perspective, from implementation\\n> perspective a form of SRF function. I use own executor node, because fcinfo\\n> is complex already and not too enough to hold all information about result\\n> columns.\\n\\n\\n> The implementation as RT doesn't reduce code - it is just moving to\\n> different file.\\n\\nYou're introducing a wholly separate callback system (TableExprRoutine)\\nfor the new functionality. And that stuff is excruciatingly close to\\nstuff that the normal executor already knows how to do.\\n\\n\\n\\n> I'll try to explain my motivation. Please, check it and correct me if I am\\n> wrong. I don't keep on my implementation - just try to implement XMLTABLE\\n> be consistent with another behave and be used all time without any\\n> surprise.\\n> \\n> 1. Any function that produces a content can be used in target list. We\\n> support SRF in target list and in FROM part. Why XMLTABLE should be a\\n> exception?\\n\\ntargetlist SRFs were a big mistake. They cause a fair number of problems\\ncode-wise. They permeated for a long while into bits of both planner and\\nexecutor, where they really shouldn't belong. Even after the recent\\nchanges there's a fair amount of uglyness associated with them. We\\ncan't remove tSRFs for backward compatibility reasons, but that's not\\ntrue for XMLTABLE\\n\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n-- \\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, 25 Jan 2017 12:31:28 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"> > >\\n> >\\n> > If you plan to hold support SRFin target list, then nothing is different.\\n> > In last patch is executed under nodeProjectSet.\\n>\\n> It is, because we suddenly need to call different functions - and I'm\\n> revamping most of execQual to have an opcode dispatch based execution\\n> model (which then also can be JITed).\\n>\\n\\n> > > > XMLTABLE is specified by the standard to return multiple rows ... but\\n> > > > then as far as my reading goes, it is only supposed to be supported\\n> in\\n> > > > the range table (FROM clause) not in the target list. I wonder if\\n> > > > this would end up better if we only tried to support it in RT. I\\n> asked\\n> > > > Pavel to implement it like that a few weeks ago, but ...\\n> > >\\n> > > Right - it makes sense in the FROM list - but then it should be an\\n> > > executor node, instead of some expression thingy.\\n> > >\\n> >\\n> > The XMLTABLE function is from user perspective, from implementation\\n> > perspective a form of SRF function. I use own executor node, because\\n> fcinfo\\n> > is complex already and not too enough to hold all information about\\n> result\\n> > columns.\\n>\\n>\\n> > The implementation as RT doesn't reduce code - it is just moving to\\n> > different file.\\n>\\n> You're introducing a wholly separate callback system (TableExprRoutine)\\n> for the new functionality. And that stuff is excruciatingly close to\\n> stuff that the normal executor already knows how to do.\\n>\\n\\nThese callbacks are related to isolation TableExpr infrastructure and\\nTableExpr implementation - This design is prepared for reusing for\\nJSON_TABLE function.\\n\\nAny placing of TableExpr code should not impact this callback system (Or I\\nam absolutely out and executor is able do some work what is hidden to me).\\n\\n\\n>\\n>\\n>\\n> > I'll try to explain my motivation. Please, check it and correct me if I\\n> am\\n> > wrong. I don't keep on my implementation - just try to implement XMLTABLE\\n> > be consistent with another behave and be used all time without any\\n> > surprise.\\n> >\\n> > 1. Any function that produces a content can be used in target list. We\\n> > support SRF in target list and in FROM part. Why XMLTABLE should be a\\n> > exception?\\n>\\n> targetlist SRFs were a big mistake. They cause a fair number of problems\\n> code-wise. They permeated for a long while into bits of both planner and\\n> executor, where they really shouldn't belong. Even after the recent\\n> changes there's a fair amount of uglyness associated with them. We\\n> can't remove tSRFs for backward compatibility reasons, but that's not\\n> true for XMLTABLE\\n>\\n>\\n>\\nok\\n\\nI afraid when I cannot to reuse a SRF infrastructure, I have to reimplement\\nit partially :( - mainly for usage in \\\"ROWS FROM ()\\\"\\n\\n\\n\\nGreetings,\\n>\\n> Andres Freund\\n>\\n\\n\\n> >\\n>\\n> If you plan to hold support SRFin target list, then nothing is different.\\n> In last patch is executed under nodeProjectSet.\\n\\nIt is, because we suddenly need to call different functions - and I'm\\nrevamping most of execQual to have an opcode dispatch based execution\\nmodel (which then also can be JITed). \\n\\n> > > XMLTABLE is specified by the standard to return multiple rows ... but\\n> > > then as far as my reading goes, it is only supposed to be supported in\\n> > > the range table (FROM clause) not in the target list. I wonder if\\n> > > this would end up better if we only tried to support it in RT. I asked\\n> > > Pavel to implement it like that a few weeks ago, but ...\\n> >\\n> > Right - it makes sense in the FROM list - but then it should be an\\n> > executor node, instead of some expression thingy.\\n> >\\n>\\n> The XMLTABLE function is from user perspective, from implementation\\n> perspective a form of SRF function. I use own executor node, because fcinfo\\n> is complex already and not too enough to hold all information about result\\n> columns.\\n\\n\\n> The implementation as RT doesn't reduce code - it is just moving to\\n> different file.\\n\\nYou're introducing a wholly separate callback system (TableExprRoutine)\\nfor the new functionality. And that stuff is excruciatingly close to\\nstuff that the normal executor already knows how to do.These callbacks are related to isolation TableExpr infrastructure and TableExpr implementation - This design is prepared for reusing for JSON_TABLE function.Any placing of TableExpr code should not impact this callback system (Or I am absolutely out and executor is able do some work what is hidden to me). \\n\\n\\n\\n> I'll try to explain my motivation. Please, check it and correct me if I am\\n> wrong. I don't keep on my implementation - just try to implement XMLTABLE\\n> be consistent with another behave and be used all time without any\\n> surprise.\\n>\\n> 1. Any function that produces a content can be used in target list. We\\n> support SRF in target list and in FROM part. Why XMLTABLE should be a\\n> exception?\\n\\ntargetlist SRFs were a big mistake. They cause a fair number of problems\\ncode-wise. They permeated for a long while into bits of both planner and\\nexecutor, where they really shouldn't belong. Even after the recent\\nchanges there's a fair amount of uglyness associated with them. We\\ncan't remove tSRFs for backward compatibility reasons, but that's not\\ntrue for XMLTABLE\\n\\nokI afraid when I cannot to reuse a SRF infrastructure, I have to reimplement it partially :( - mainly for usage in \\\"ROWS FROM ()\\\" \\nGreetings,\\n\\nAndres Freund\",\n \"msg_date\": \"Wed, 25 Jan 2017 22:38:25 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\n> > > I'll try to explain my motivation. Please, check it and correct me if I\\n> > am\\n> > > wrong. I don't keep on my implementation - just try to implement XMLTABLE\\n> > > be consistent with another behave and be used all time without any\\n> > > surprise.\\n> > >\\n> > > 1. Any function that produces a content can be used in target list. We\\n> > > support SRF in target list and in FROM part. Why XMLTABLE should be a\\n> > > exception?\\n> >\\n> > targetlist SRFs were a big mistake. They cause a fair number of problems\\n> > code-wise. They permeated for a long while into bits of both planner and\\n> > executor, where they really shouldn't belong. Even after the recent\\n> > changes there's a fair amount of uglyness associated with them. We\\n> > can't remove tSRFs for backward compatibility reasons, but that's not\\n> > true for XMLTABLE\\n> >\\n> >\\n> >\\n> ok\\n> \\n> I afraid when I cannot to reuse a SRF infrastructure, I have to reimplement\\n> it partially :( - mainly for usage in \\\"ROWS FROM ()\\\"\\n\\nHuh?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n-- \\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, 25 Jan 2017 13:40:22 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-01-25 22:40 GMT+01:00 Andres Freund <andres@anarazel.de>:\\n\\n> Hi,\\n>\\n> > > > I'll try to explain my motivation. Please, check it and correct me\\n> if I\\n> > > am\\n> > > > wrong. I don't keep on my implementation - just try to implement\\n> XMLTABLE\\n> > > > be consistent with another behave and be used all time without any\\n> > > > surprise.\\n> > > >\\n> > > > 1. Any function that produces a content can be used in target list.\\n> We\\n> > > > support SRF in target list and in FROM part. Why XMLTABLE should be a\\n> > > > exception?\\n> > >\\n> > > targetlist SRFs were a big mistake. They cause a fair number of\\n> problems\\n> > > code-wise. They permeated for a long while into bits of both planner\\n> and\\n> > > executor, where they really shouldn't belong. Even after the recent\\n> > > changes there's a fair amount of uglyness associated with them. We\\n> > > can't remove tSRFs for backward compatibility reasons, but that's not\\n> > > true for XMLTABLE\\n> > >\\n> > >\\n> > >\\n> > ok\\n> >\\n> > I afraid when I cannot to reuse a SRF infrastructure, I have to\\n> reimplement\\n> > it partially :( - mainly for usage in \\\"ROWS FROM ()\\\"\\n>\\n\\nThe TableExpr implementation is based on SRF now. You and Alvaro propose\\nindependent implementation like generic executor node. I am sceptic so\\nFunctionScan supports reading from generic executor node.\\n\\nRegards\\n\\nPavel\\n\\n\\n> Huh?\\n>\\n> Greetings,\\n>\\n> Andres Freund\\n>\\n\\n2017-01-25 22:40 GMT+01:00 Andres Freund <andres@anarazel.de>:Hi,\\n\\n> > > I'll try to explain my motivation. Please, check it and correct me if I\\n> > am\\n> > > wrong. I don't keep on my implementation - just try to implement XMLTABLE\\n> > > be consistent with another behave and be used all time without any\\n> > > surprise.\\n> > >\\n> > > 1. Any function that produces a content can be used in target list. We\\n> > > support SRF in target list and in FROM part. Why XMLTABLE should be a\\n> > > exception?\\n> >\\n> > targetlist SRFs were a big mistake. They cause a fair number of problems\\n> > code-wise. They permeated for a long while into bits of both planner and\\n> > executor, where they really shouldn't belong. Even after the recent\\n> > changes there's a fair amount of uglyness associated with them. We\\n> > can't remove tSRFs for backward compatibility reasons, but that's not\\n> > true for XMLTABLE\\n> >\\n> >\\n> >\\n> ok\\n>\\n> I afraid when I cannot to reuse a SRF infrastructure, I have to reimplement\\n> it partially :( - mainly for usage in \\\"ROWS FROM ()\\\"The TableExpr implementation is based on SRF now. You and Alvaro propose independent implementation like generic executor node. I am sceptic so FunctionScan supports reading from generic executor node.RegardsPavel\\n\\nHuh?\\n\\nGreetings,\\n\\nAndres Freund\",\n \"msg_date\": \"Wed, 25 Jan 2017 22:51:37 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On 2017-01-25 22:51:37 +0100, Pavel Stehule wrote:\\n> 2017-01-25 22:40 GMT+01:00 Andres Freund <andres@anarazel.de>:\\n> > > I afraid when I cannot to reuse a SRF infrastructure, I have to\\n> > reimplement\\n> > > it partially :( - mainly for usage in \\\"ROWS FROM ()\\\"\\n> >\\n> \\n> The TableExpr implementation is based on SRF now. You and Alvaro propose\\n> independent implementation like generic executor node. I am sceptic so\\n> FunctionScan supports reading from generic executor node.\\n\\nWhy would it need to?\\n\\n\\n-- \\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, 25 Jan 2017 14:33:56 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-01-25 23:33 GMT+01:00 Andres Freund <andres@anarazel.de>:\\n\\n> On 2017-01-25 22:51:37 +0100, Pavel Stehule wrote:\\n> > 2017-01-25 22:40 GMT+01:00 Andres Freund <andres@anarazel.de>:\\n> > > > I afraid when I cannot to reuse a SRF infrastructure, I have to\\n> > > reimplement\\n> > > > it partially :( - mainly for usage in \\\"ROWS FROM ()\\\"\\n> > >\\n> >\\n> > The TableExpr implementation is based on SRF now. You and Alvaro propose\\n> > independent implementation like generic executor node. I am sceptic so\\n> > FunctionScan supports reading from generic executor node.\\n>\\n> Why would it need to?\\n>\\n\\nSimply - due consistency with any other functions that can returns rows.\\n\\nMaybe I don't understand to Alvaro proposal well - I have a XMLTABLE\\nfunction - TableExpr that looks like SRF function, has similar behave -\\nreturns more rows, but should be significantly different implemented, and\\nshould to have different limits - should not be used there and there ... It\\nis hard to see consistency there for me.\\n\\nRegards\\n\\nPavel\\n\\n2017-01-25 23:33 GMT+01:00 Andres Freund <andres@anarazel.de>:On 2017-01-25 22:51:37 +0100, Pavel Stehule wrote:\\n> 2017-01-25 22:40 GMT+01:00 Andres Freund <andres@anarazel.de>:\\n> > > I afraid when I cannot to reuse a SRF infrastructure, I have to\\n> > reimplement\\n> > > it partially :( - mainly for usage in \\\"ROWS FROM ()\\\"\\n> >\\n>\\n> The TableExpr implementation is based on SRF now. You and Alvaro propose\\n> independent implementation like generic executor node. I am sceptic so\\n> FunctionScan supports reading from generic executor node.\\n\\nWhy would it need to?Simply - due consistency with any other functions that can returns rows.Maybe I don't understand to Alvaro proposal well - I have a XMLTABLE function - TableExpr that looks like SRF function, has similar behave - returns more rows, but should be significantly different implemented, and should to have different limits - should not be used there and there ... It is hard to see consistency there for me.RegardsPavel\",\n \"msg_date\": \"Wed, 25 Jan 2017 23:43:22 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-01-25 15:07 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Tom Lane wrote:\\n> > Andres Freund <andres@anarazel.de> writes:\\n> > > On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\\n> > >> XMLTABLE is specified by the standard to return multiple rows ... but\\n> > >> then as far as my reading goes, it is only supposed to be supported in\\n> > >> the range table (FROM clause) not in the target list. I wonder if\\n> > >> this would end up better if we only tried to support it in RT. I\\n> asked\\n> > >> Pavel to implement it like that a few weeks ago, but ...\\n> >\\n> > > Right - it makes sense in the FROM list - but then it should be an\\n> > > executor node, instead of some expression thingy.\\n> >\\n> > +1 --- we're out of the business of having simple expressions that\\n> > return rowsets.\\n>\\n> Well, that's it. I'm not committing this patch against two other\\n> committers' opinion, plus I was already on the fence about the\\n> implementation anyway. I think you should just go with the flow and\\n> implement this by creating nodeTableexprscan.c. It's not even\\n> difficult.\\n>\\n\\nI am playing with this and the patch looks about 15kB longer - just due\\nimplementation basic scan functionality - and I didn't touch a planner.\\n\\nI am not happy from this - still I have a feeling so I try to reimplement\\nreduced SRF.\\n\\nRegards\\n\\nPavel\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2017-01-25 15:07 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Tom Lane wrote:\\n> Andres Freund <andres@anarazel.de> writes:\\n> > On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\\n> >> XMLTABLE is specified by the standard to return multiple rows ... but\\n> >> then as far as my reading goes, it is only supposed to be supported in\\n> >> the range table (FROM clause) not in the target list. I wonder if\\n> >> this would end up better if we only tried to support it in RT. I asked\\n> >> Pavel to implement it like that a few weeks ago, but ...\\n>\\n> > Right - it makes sense in the FROM list - but then it should be an\\n> > executor node, instead of some expression thingy.\\n>\\n> +1 --- we're out of the business of having simple expressions that\\n> return rowsets.\\n\\nWell, that's it. I'm not committing this patch against two other\\ncommitters' opinion, plus I was already on the fence about the\\nimplementation anyway. I think you should just go with the flow and\\nimplement this by creating nodeTableexprscan.c. It's not even\\ndifficult.I am playing with this and the patch looks about 15kB longer - just due implementation basic scan functionality - and I didn't touch a planner. I am not happy from this - still I have a feeling so I try to reimplement reduced SRF. RegardsPavel\\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Thu, 26 Jan 2017 01:51:00 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2017-01-25 15:07 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Tom Lane wrote:\\n> > Andres Freund <andres@anarazel.de> writes:\\n> > > On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:\\n> > >> XMLTABLE is specified by the standard to return multiple rows ... but\\n> > >> then as far as my reading goes, it is only supposed to be supported in\\n> > >> the range table (FROM clause) not in the target list. I wonder if\\n> > >> this would end up better if we only tried to support it in RT. I\\n> asked\\n> > >> Pavel to implement it like that a few weeks ago, but ...\\n> >\\n> > > Right - it makes sense in the FROM list - but then it should be an\\n> > > executor node, instead of some expression thingy.\\n> >\\n> > +1 --- we're out of the business of having simple expressions that\\n> > return rowsets.\\n>\\n> Well, that's it. I'm not committing this patch against two other\\n> committers' opinion, plus I was already on the fence about the\\n> implementation anyway. I think you should just go with the flow and\\n> implement this by creating nodeTableexprscan.c. It's not even\\n> difficult.\\n>\\n\\nI am sending new version - it is based on own executor scan node and\\ntuplestore.\\n\\nSome now obsolete regress tests removed, some new added.\\n\\nThe executor code (memory context usage) should be cleaned little bit - but\\nother code should be good.\\n\\nRegards\\n\\nPavel\\n\\n\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL 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\": \"Mon, 30 Jan 2017 20:22:04 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n\\n> I am sending new version - it is based on own executor scan node and\\n> tuplestore.\\n> \\n> Some now obsolete regress tests removed, some new added.\\n> \\n> The executor code (memory context usage) should be cleaned little bit - but\\n> other code should be good.\\n\\nI think you forgot nodeTableFuncscan.c.\\n\\n-- \\n�lvaro Herrera https://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\": \"Mon, 30 Jan 2017 16:38:59 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2017-01-30 20:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n>\\n> > I am sending new version - it is based on own executor scan node and\\n> > tuplestore.\\n> >\\n> > Some now obsolete regress tests removed, some new added.\\n> >\\n> > The executor code (memory context usage) should be cleaned little bit -\\n> but\\n> > other code should be good.\\n>\\n> I think you forgot nodeTableFuncscan.c.\\n>\\n\\ntrue, I am sorry\\n\\nattached\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL 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\": \"Mon, 30 Jan 2017 21:18:53 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 31, 2017 at 5:18 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n> true, I am sorry\\n\\nLast status is a new patch and no reviews. On top of that this thread\\nis quite active. So moved to next CF. Pavel, please be careful about\\nthe status of the patch on the CF app, it was set to \\\"waiting on\\nauthor\\\"...\\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\": \"Tue, 31 Jan 2017 15:33:05 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-01-24 21:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n>\\n> > * SELECT (xmltable(..)).* + regress tests\\n> > * compilation and regress tests without --with-libxml\\n>\\n> Thanks. I just realized that this is doing more work than necessary --\\n> I think it would be simpler to have tableexpr fill a tuplestore with the\\n> results, instead of just expecting function execution to apply\\n> ExecEvalExpr over and over to obtain the results. So evaluating a\\n> tableexpr returns just the tuplestore, which function evaluation can\\n> return as-is. That code doesn't use the value-per-call interface\\n> anyway.\\n>\\n> I also realized that the expr context callback is not called if there's\\n> an error, which leaves us without shutting down libxml properly. I\\n> added PG_TRY around the fetchrow calls, but I'm not sure that's correct\\n> either, because there could be an error raised in other parts of the\\n> code, after we've already emitted a few rows (for example out of\\n> memory). I think the right way is to have PG_TRY around the execution\\n> of the whole thing rather than just row at a time; and the tuplestore\\n> mechanism helps us with that.\\n>\\n> I think it would be good to have a more complex test case in regress --\\n> let's say there is a table with some simple XML values, then we use\\n> XMLFOREST (or maybe one of the table_to_xml functions) to generate a\\n> large document, and then XMLTABLE uses that document as input document.\\n>\\n\\nI have a 16K lines long real XML 6.MB. Probably we would not to append it\\nto regress tests.\\n\\nIt is really fast - original customer implementation 20min, nested our\\nxpath implementation 10 sec, PLPython xml reader 5 sec, xmltable 400ms\\n\\nI have a plan to create tests based on pg_proc and CTE - if all works, then\\nthe query must be empty\\n\\nwith x as (select proname, proowner, procost, pronargs,\\narray_to_string(proargnames,',') as proargnames,\\narray_to_string(proargtypes,',') as proargtypes from pg_proc), y as (select\\nxmlelement(name proc, xmlforest(proname, proowner, procost, pronargs,\\nproargnames, proargtypes)) as proc from x), z as (select xmltable.* from y,\\nlateral xmltable('/proc' passing proc columns proname name, proowner oid,\\nprocost float, pronargs int, proargnames text, proargtypes text)) select *\\nfrom z except select * from x;\\n\\n\\n\\n>\\n> Please fix.\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2017-01-24 21:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\\n\\n> * SELECT (xmltable(..)).* + regress tests\\n> * compilation and regress tests without --with-libxml\\n\\nThanks. I just realized that this is doing more work than necessary --\\nI think it would be simpler to have tableexpr fill a tuplestore with the\\nresults, instead of just expecting function execution to apply\\nExecEvalExpr over and over to obtain the results. So evaluating a\\ntableexpr returns just the tuplestore, which function evaluation can\\nreturn as-is. That code doesn't use the value-per-call interface\\nanyway.\\n\\nI also realized that the expr context callback is not called if there's\\nan error, which leaves us without shutting down libxml properly. I\\nadded PG_TRY around the fetchrow calls, but I'm not sure that's correct\\neither, because there could be an error raised in other parts of the\\ncode, after we've already emitted a few rows (for example out of\\nmemory). I think the right way is to have PG_TRY around the execution\\nof the whole thing rather than just row at a time; and the tuplestore\\nmechanism helps us with that.\\n\\nI think it would be good to have a more complex test case in regress --\\nlet's say there is a table with some simple XML values, then we use\\nXMLFOREST (or maybe one of the table_to_xml functions) to generate a\\nlarge document, and then XMLTABLE uses that document as input document.I have a 16K lines long real XML 6.MB. Probably we would not to append it to regress tests.It is really fast - original customer implementation 20min, nested our xpath implementation 10 sec, PLPython xml reader 5 sec, xmltable 400msI have a plan to create tests based on pg_proc and CTE - if all works, then the query must be emptywith x as (select proname, proowner, procost, pronargs, array_to_string(proargnames,',') as proargnames, array_to_string(proargtypes,',') as proargtypes from pg_proc), y as (select xmlelement(name proc, xmlforest(proname, proowner, procost, pronargs, proargnames, proargtypes)) as proc from x), z as (select xmltable.* from y, lateral xmltable('/proc' passing proc columns proname name, proowner oid, procost float, pronargs int, proargnames text, proargtypes text)) select * from z except select * from x; \\n\\nPlease fix.\\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Tue, 31 Jan 2017 11:26:32 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n> 2017-01-24 21:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> > I think it would be good to have a more complex test case in regress --\\n> > let's say there is a table with some simple XML values, then we use\\n> > XMLFOREST (or maybe one of the table_to_xml functions) to generate a\\n> > large document, and then XMLTABLE uses that document as input document.\\n> \\n> I have a 16K lines long real XML 6.MB. Probably we would not to append it\\n> to regress tests.\\n> \\n> It is really fast - original customer implementation 20min, nested our\\n> xpath implementation 10 sec, PLPython xml reader 5 sec, xmltable 400ms\\n\\nThat's great numbers, kudos for the hard work here. That will make for\\na nice headline in pg10 PR materials. But what I was getting at is that\\nI would like to exercise a bit more of the expression handling in\\nxmltable execution, to make sure it doesn't handle just string literals.\\n\\n> I have a plan to create tests based on pg_proc and CTE - if all works, then\\n> the query must be empty\\n> \\n> with x as (select proname, proowner, procost, pronargs,\\n> array_to_string(proargnames,',') as proargnames,\\n> array_to_string(proargtypes,',') as proargtypes from pg_proc), y as (select\\n> xmlelement(name proc, xmlforest(proname, proowner, procost, pronargs,\\n> proargnames, proargtypes)) as proc from x), z as (select xmltable.* from y,\\n> lateral xmltable('/proc' passing proc columns proname name, proowner oid,\\n> procost float, pronargs int, proargnames text, proargtypes text)) select *\\n> from z except select * from x;\\n\\nNice one :-)\\n\\n-- \\n�lvaro Herrera https://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\": \"Tue, 31 Jan 2017 10:57:07 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-01-31 14:57 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n> > 2017-01-24 21:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>\\n> > > I think it would be good to have a more complex test case in regress --\\n> > > let's say there is a table with some simple XML values, then we use\\n> > > XMLFOREST (or maybe one of the table_to_xml functions) to generate a\\n> > > large document, and then XMLTABLE uses that document as input document.\\n> >\\n> > I have a 16K lines long real XML 6.MB. Probably we would not to append it\\n> > to regress tests.\\n> >\\n> > It is really fast - original customer implementation 20min, nested our\\n> > xpath implementation 10 sec, PLPython xml reader 5 sec, xmltable 400ms\\n>\\n> That's great numbers, kudos for the hard work here. That will make for\\n> a nice headline in pg10 PR materials. But what I was getting at is that\\n> I would like to exercise a bit more of the expression handling in\\n> xmltable execution, to make sure it doesn't handle just string literals.\\n>\\n\\nI'll try to write some more dynamic examples.\\n\\n\\n>\\n> > I have a plan to create tests based on pg_proc and CTE - if all works,\\n> then\\n> > the query must be empty\\n> >\\n> > with x as (select proname, proowner, procost, pronargs,\\n> > array_to_string(proargnames,',') as proargnames,\\n> > array_to_string(proargtypes,',') as proargtypes from pg_proc), y as\\n> (select\\n> > xmlelement(name proc, xmlforest(proname, proowner, procost, pronargs,\\n> > proargnames, proargtypes)) as proc from x), z as (select xmltable.* from\\n> y,\\n> > lateral xmltable('/proc' passing proc columns proname name, proowner oid,\\n> > procost float, pronargs int, proargnames text, proargtypes text)) select\\n> *\\n> > from z except select * from x;\\n>\\n> Nice one :-)\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2017-01-31 14:57 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\\n> 2017-01-24 21:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> > I think it would be good to have a more complex test case in regress --\\n> > let's say there is a table with some simple XML values, then we use\\n> > XMLFOREST (or maybe one of the table_to_xml functions) to generate a\\n> > large document, and then XMLTABLE uses that document as input document.\\n>\\n> I have a 16K lines long real XML 6.MB. Probably we would not to append it\\n> to regress tests.\\n>\\n> It is really fast - original customer implementation 20min, nested our\\n> xpath implementation 10 sec, PLPython xml reader 5 sec, xmltable 400ms\\n\\nThat's great numbers, kudos for the hard work here. That will make for\\na nice headline in pg10 PR materials. But what I was getting at is that\\nI would like to exercise a bit more of the expression handling in\\nxmltable execution, to make sure it doesn't handle just string literals.I'll try to write some more dynamic examples. \\n\\n> I have a plan to create tests based on pg_proc and CTE - if all works, then\\n> the query must be empty\\n>\\n> with x as (select proname, proowner, procost, pronargs,\\n> array_to_string(proargnames,',') as proargnames,\\n> array_to_string(proargtypes,',') as proargtypes from pg_proc), y as (select\\n> xmlelement(name proc, xmlforest(proname, proowner, procost, pronargs,\\n> proargnames, proargtypes)) as proc from x), z as (select xmltable.* from y,\\n> lateral xmltable('/proc' passing proc columns proname name, proowner oid,\\n> procost float, pronargs int, proargnames text, proargtypes text)) select *\\n> from z except select * from x;\\n\\nNice one :-)\\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Tue, 31 Jan 2017 16:32:57 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2017-01-31 14:57 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n> > 2017-01-24 21:38 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>\\n> > > I think it would be good to have a more complex test case in regress --\\n> > > let's say there is a table with some simple XML values, then we use\\n> > > XMLFOREST (or maybe one of the table_to_xml functions) to generate a\\n> > > large document, and then XMLTABLE uses that document as input document.\\n> >\\n> > I have a 16K lines long real XML 6.MB. Probably we would not to append it\\n> > to regress tests.\\n> >\\n> > It is really fast - original customer implementation 20min, nested our\\n> > xpath implementation 10 sec, PLPython xml reader 5 sec, xmltable 400ms\\n>\\n> That's great numbers, kudos for the hard work here. That will make for\\n> a nice headline in pg10 PR materials. But what I was getting at is that\\n> I would like to exercise a bit more of the expression handling in\\n> xmltable execution, to make sure it doesn't handle just string literals.\\n>\\n\\ndone\\n\\n\\n>\\n> > I have a plan to create tests based on pg_proc and CTE - if all works,\\n> then\\n> > the query must be empty\\n> >\\n> > with x as (select proname, proowner, procost, pronargs,\\n> > array_to_string(proargnames,',') as proargnames,\\n> > array_to_string(proargtypes,',') as proargtypes from pg_proc), y as\\n> (select\\n> > xmlelement(name proc, xmlforest(proname, proowner, procost, pronargs,\\n> > proargnames, proargtypes)) as proc from x), z as (select xmltable.* from\\n> y,\\n> > lateral xmltable('/proc' passing proc columns proname name, proowner oid,\\n> > procost float, pronargs int, proargnames text, proargtypes text)) select\\n> *\\n> > from z except select * from x;\\n>\\n> Nice one :-)\\n>\\n\\nplease see attached patch\\n\\n* enhanced regress tests\\n* clean memory context work\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL 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\": \"Tue, 31 Jan 2017 22:20:49 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n\\n> please see attached patch\\n>\\n> * enhanced regress tests\\n> * clean memory context work\\n>\\n\\nnew update\\n\\nfix a bug in string compare\\nfix some typo and obsolete comments\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> Regards\\n>\\n> Pavel\\n>\\n>\\n>>\\n>> --\\n>> Álvaro Herrera https://www.2ndQuadrant.com/\\n>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\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\": \"Thu, 16 Feb 2017 06:38:04 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2017-02-16 6:38 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n\\n> Hi\\n>\\n>\\n>> please see attached patch\\n>>\\n>> * enhanced regress tests\\n>> * clean memory context work\\n>>\\n>\\n> new update\\n>\\n> fix a bug in string compare\\n> fix some typo and obsolete comments\\n>\\n> Regards\\n>\\n\\nsome minor but interesting fix.\\n\\nI found so some xml values imported via recv function can have\\ninconsistency between header encoding and used encoding. Internally the\\nheader encoding is removed - see xml_out function.\\n\\nSo now, when I have to prepare data for libxml2, I don't do direct cast,\\nbut I use xml_out_internal instead. Maybe this technique should be used\\nelsewhere? Same issue I see on xpath function.\\n\\nSolved issue is not too often probably - the some different encoding than\\nutf8 should be used in XML document and XML document should be loaded with\\nrecv function.\\n\\nRegards\\n\\nPavel\\n\\n\\n\\n>\\n> Pavel\\n>\\n>\\n>>\\n>> Regards\\n>>\\n>> Pavel\\n>>\\n>>\\n>>>\\n>>> --\\n>>> Álvaro Herrera https://www.2ndQuadrant.com/\\n>>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>>>\\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 Feb 2017 19:12:55 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"\\nI've been giving this a look. I started by tweaking the docs once\\nagain, and while verifying that the example works as expected, I\\nreplayed what I have in sgml:\\n\\n... begin SGML paste ...\\n <para>\\n For example, given the following XML document:\\n <screen><![CDATA[\\n<ROWS>\\n <ROW id=\\\"1\\\">\\n <COUNTRY_ID>AU</COUNTRY_ID>\\n <COUNTRY_NAME>Australia</COUNTRY_NAME>\\n </ROW>\\n <ROW id=\\\"5\\\">\\n <COUNTRY_ID>JP</COUNTRY_ID>\\n <COUNTRY_NAME>Japan</COUNTRY_NAME>\\n <PREMIER_NAME>Sinzo Abe</PREMIER_NAME>\\n </ROW>\\n <ROW id=\\\"6\\\">\\n <COUNTRY_ID>SG</COUNTRY_ID>\\n <COUNTRY_NAME>Singapore</COUNTRY_NAME>\\n <SIZE unit=\\\"km\\\">791</SIZE>\\n </ROW>\\n</ROWS>\\n]]></screen>\\n\\n the following query produces the result shown below:\\n\\n<screen><![CDATA[\\nSELECT xmltable.*\\n FROM (SELECT data FROM xmldata) x,\\n LATERAL xmltable('//ROWS/ROW'\\n PASSING data\\n COLUMNS id int PATH '@id',\\n ordinality FOR ORDINALITY,\\n country_name text PATH 'COUNTRY_NAME',\\n country_id text PATH 'COUNTRY_ID',\\n size float PATH 'SIZE[@unit = \\\"km\\\"]/text()',\\n unit text PATH 'SIZE/@unit',\\n premier_name text PATH 'PREMIER_NAME' DEFAULT 'not specified');\\n... end SGML paste ...\\n\\n\\nBut the query doesn't actually return a table, but instead it fails with\\nthis error:\\n ERROR: invalid input syntax for type double precision: \\\"\\\"\\nThis is because of the \\\"size\\\" column (if I remove SIZE from the COLUMNS\\nclause, the query returns correctly). Apparently, for the rows where\\nSIZE is not given, we try to inssert an empty string instead of a NULL\\nvalue, which is what I expected.\\n\\nI'm using your v44 code, but trimmed both the XML document used in SGML\\nas well as modified the query slightly to show additional features. But\\nthose changes should not cause the above error ...\\n\\n-- \\n�lvaro Herrera https://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, 1 Mar 2017 21:12:45 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2017-03-02 1:12 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n>\\n> I've been giving this a look. I started by tweaking the docs once\\n> again, and while verifying that the example works as expected, I\\n> replayed what I have in sgml:\\n>\\n> ... begin SGML paste ...\\n> <para>\\n> For example, given the following XML document:\\n> <screen><![CDATA[\\n> <ROWS>\\n> <ROW id=\\\"1\\\">\\n> <COUNTRY_ID>AU</COUNTRY_ID>\\n> <COUNTRY_NAME>Australia</COUNTRY_NAME>\\n> </ROW>\\n> <ROW id=\\\"5\\\">\\n> <COUNTRY_ID>JP</COUNTRY_ID>\\n> <COUNTRY_NAME>Japan</COUNTRY_NAME>\\n> <PREMIER_NAME>Sinzo Abe</PREMIER_NAME>\\n> </ROW>\\n> <ROW id=\\\"6\\\">\\n> <COUNTRY_ID>SG</COUNTRY_ID>\\n> <COUNTRY_NAME>Singapore</COUNTRY_NAME>\\n> <SIZE unit=\\\"km\\\">791</SIZE>\\n> </ROW>\\n> </ROWS>\\n> ]]></screen>\\n>\\n> the following query produces the result shown below:\\n>\\n> <screen><![CDATA[\\n> SELECT xmltable.*\\n> FROM (SELECT data FROM xmldata) x,\\n> LATERAL xmltable('//ROWS/ROW'\\n> PASSING data\\n> COLUMNS id int PATH '@id',\\n> ordinality FOR ORDINALITY,\\n> country_name text PATH 'COUNTRY_NAME',\\n> country_id text PATH 'COUNTRY_ID',\\n> size float PATH 'SIZE[@unit =\\n> \\\"km\\\"]/text()',\\n> unit text PATH 'SIZE/@unit',\\n> premier_name text PATH 'PREMIER_NAME'\\n> DEFAULT 'not specified');\\n> ... end SGML paste ...\\n>\\n>\\n> But the query doesn't actually return a table, but instead it fails with\\n> this error:\\n> ERROR: invalid input syntax for type double precision: \\\"\\\"\\n> This is because of the \\\"size\\\" column (if I remove SIZE from the COLUMNS\\n> clause, the query returns correctly). Apparently, for the rows where\\n> SIZE is not given, we try to inssert an empty string instead of a NULL\\n> value, which is what I expected.\\n>\\n> I'm using your v44 code, but trimmed both the XML document used in SGML\\n> as well as modified the query slightly to show additional features. But\\n> those changes should not cause the above error ...\\n>\\n\\nThe example in doc is obsolete. Following example works without problems.\\n\\n SELECT xmltable.*\\n\\n FROM (SELECT data FROM xmldata) x,\\n LATERAL xmltable('//ROWS/ROW'\\n PASSING data\\n COLUMNS id int PATH '@id',\\n ordinality FOR ORDINALITY,\\n country_name text PATH 'COUNTRY_NAME',\\n country_id text PATH 'COUNTRY_ID',\\n size float PATH 'SIZE[@unit = \\\"km\\\"]',\\n unit text PATH 'SIZE/@unit',\\n premier_name text PATH 'PREMIER_NAME'\\nDEFAULT 'not specified');\\n\\n\\nIt is related to older variants of this patch, where I explicitly mapped\\nempty strings to NULL.\\n\\nNow, I don't do it - I use libxml2 result with following mapping\\n\\nNo tag ... NULL\\nempty tag ... empty string\\n\\nImportant question is about mapping empty tags to Postgres. I prefer\\ncurrent behave, because I have a possibility to differ between these states\\non application level. If we returns NULL for empty tag, then there will not\\nbe possible detect if XML has tag (although empty) or not. The change is\\nsimple - just one row - but I am thinking so current behave is better.\\nThere is possible risk of using /text() somewhere - it enforce a empty tag\\nwith all negative impacts.\\n\\nI prefer to fix doc in conformance with regress tests and append note about\\nmapping these corner cases from XML to relations.\\n\\nWhat do you think about it?\\n\\nRegards\\n\\nPavel\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\nHi2017-03-02 1:12 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\nI've been giving this a look. I started by tweaking the docs once\\nagain, and while verifying that the example works as expected, I\\nreplayed what I have in sgml:\\n\\n... begin SGML paste ...\\n <para>\\n For example, given the following XML document:\\n <screen><![CDATA[\\n<ROWS>\\n <ROW id=\\\"1\\\">\\n <COUNTRY_ID>AU</COUNTRY_ID>\\n <COUNTRY_NAME>Australia</COUNTRY_NAME>\\n </ROW>\\n <ROW id=\\\"5\\\">\\n <COUNTRY_ID>JP</COUNTRY_ID>\\n <COUNTRY_NAME>Japan</COUNTRY_NAME>\\n <PREMIER_NAME>Sinzo Abe</PREMIER_NAME>\\n </ROW>\\n <ROW id=\\\"6\\\">\\n <COUNTRY_ID>SG</COUNTRY_ID>\\n <COUNTRY_NAME>Singapore</COUNTRY_NAME>\\n <SIZE unit=\\\"km\\\">791</SIZE>\\n </ROW>\\n</ROWS>\\n]]></screen>\\n\\n the following query produces the result shown below:\\n\\n<screen><![CDATA[\\nSELECT xmltable.*\\n FROM (SELECT data FROM xmldata) x,\\n LATERAL xmltable('//ROWS/ROW'\\n PASSING data\\n COLUMNS id int PATH '@id',\\n ordinality FOR ORDINALITY,\\n country_name text PATH 'COUNTRY_NAME',\\n country_id text PATH 'COUNTRY_ID',\\n size float PATH 'SIZE[@unit = \\\"km\\\"]/text()',\\n unit text PATH 'SIZE/@unit',\\n premier_name text PATH 'PREMIER_NAME' DEFAULT 'not specified');\\n... end SGML paste ...\\n\\n\\nBut the query doesn't actually return a table, but instead it fails with\\nthis error:\\n ERROR: invalid input syntax for type double precision: \\\"\\\"\\nThis is because of the \\\"size\\\" column (if I remove SIZE from the COLUMNS\\nclause, the query returns correctly). Apparently, for the rows where\\nSIZE is not given, we try to inssert an empty string instead of a NULL\\nvalue, which is what I expected.\\n\\nI'm using your v44 code, but trimmed both the XML document used in SGML\\nas well as modified the query slightly to show additional features. But\\nthose changes should not cause the above error ...The example in doc is obsolete. Following example works without problems. SELECT xmltable.* FROM (SELECT data FROM xmldata) x,\\n LATERAL xmltable('//ROWS/ROW'\\n PASSING data\\n COLUMNS id int PATH '@id',\\n ordinality FOR ORDINALITY,\\n country_name text PATH 'COUNTRY_NAME',\\n country_id text PATH 'COUNTRY_ID',\\n size float PATH 'SIZE[@unit = \\\"km\\\"]',\\n unit text PATH 'SIZE/@unit',\\n premier_name text PATH 'PREMIER_NAME' DEFAULT 'not specified');It is related to older variants of this patch, where I explicitly mapped empty strings to NULL. Now, I don't do it - I use libxml2 result with following mappingNo tag ... NULLempty tag ... empty stringImportant question is about mapping empty tags to Postgres. I prefer current behave, because I have a possibility to differ between these states on application level. If we returns NULL for empty tag, then there will not be possible detect if XML has tag (although empty) or not. The change is simple - just one row - but I am thinking so current behave is better. There is possible risk of using /text() somewhere - it enforce a empty tag with all negative impacts.I prefer to fix doc in conformance with regress tests and append note about mapping these corner cases from XML to relations.What do you think about it?RegardsPavel\\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Thu, 2 Mar 2017 08:04:14 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-03-02 8:04 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n\\n> Hi\\n>\\n> 2017-03-02 1:12 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>\\n>>\\n>> I've been giving this a look. I started by tweaking the docs once\\n>> again, and while verifying that the example works as expected, I\\n>> replayed what I have in sgml:\\n>>\\n>> ... begin SGML paste ...\\n>> <para>\\n>> For example, given the following XML document:\\n>> <screen><![CDATA[\\n>> <ROWS>\\n>> <ROW id=\\\"1\\\">\\n>> <COUNTRY_ID>AU</COUNTRY_ID>\\n>> <COUNTRY_NAME>Australia</COUNTRY_NAME>\\n>> </ROW>\\n>> <ROW id=\\\"5\\\">\\n>> <COUNTRY_ID>JP</COUNTRY_ID>\\n>> <COUNTRY_NAME>Japan</COUNTRY_NAME>\\n>> <PREMIER_NAME>Sinzo Abe</PREMIER_NAME>\\n>> </ROW>\\n>> <ROW id=\\\"6\\\">\\n>> <COUNTRY_ID>SG</COUNTRY_ID>\\n>> <COUNTRY_NAME>Singapore</COUNTRY_NAME>\\n>> <SIZE unit=\\\"km\\\">791</SIZE>\\n>> </ROW>\\n>> </ROWS>\\n>> ]]></screen>\\n>>\\n>> the following query produces the result shown below:\\n>>\\n>> <screen><![CDATA[\\n>> SELECT xmltable.*\\n>> FROM (SELECT data FROM xmldata) x,\\n>> LATERAL xmltable('//ROWS/ROW'\\n>> PASSING data\\n>> COLUMNS id int PATH '@id',\\n>> ordinality FOR ORDINALITY,\\n>> country_name text PATH 'COUNTRY_NAME',\\n>> country_id text PATH 'COUNTRY_ID',\\n>> size float PATH 'SIZE[@unit =\\n>> \\\"km\\\"]/text()',\\n>> unit text PATH 'SIZE/@unit',\\n>> premier_name text PATH 'PREMIER_NAME'\\n>> DEFAULT 'not specified');\\n>> ... end SGML paste ...\\n>>\\n>>\\n>> But the query doesn't actually return a table, but instead it fails with\\n>> this error:\\n>> ERROR: invalid input syntax for type double precision: \\\"\\\"\\n>> This is because of the \\\"size\\\" column (if I remove SIZE from the COLUMNS\\n>> clause, the query returns correctly). Apparently, for the rows where\\n>> SIZE is not given, we try to inssert an empty string instead of a NULL\\n>> value, which is what I expected.\\n>>\\n>> I'm using your v44 code, but trimmed both the XML document used in SGML\\n>> as well as modified the query slightly to show additional features. But\\n>> those changes should not cause the above error ...\\n>>\\n>\\n> The example in doc is obsolete. Following example works without problems.\\n>\\n> SELECT xmltable.*\\n>\\n> FROM (SELECT data FROM xmldata) x,\\n> LATERAL xmltable('//ROWS/ROW'\\n> PASSING data\\n> COLUMNS id int PATH '@id',\\n> ordinality FOR ORDINALITY,\\n> country_name text PATH 'COUNTRY_NAME',\\n> country_id text PATH 'COUNTRY_ID',\\n> size float PATH 'SIZE[@unit = \\\"km\\\"]',\\n> unit text PATH 'SIZE/@unit',\\n> premier_name text PATH 'PREMIER_NAME' DEFAULT 'not specified');\\n>\\n>\\n> It is related to older variants of this patch, where I explicitly mapped\\n> empty strings to NULL.\\n>\\n> Now, I don't do it - I use libxml2 result with following mapping\\n>\\n> No tag ... NULL\\n> empty tag ... empty string\\n>\\n> Important question is about mapping empty tags to Postgres. I prefer\\n> current behave, because I have a possibility to differ between these states\\n> on application level. If we returns NULL for empty tag, then there will not\\n> be possible detect if XML has tag (although empty) or not. The change is\\n> simple - just one row - but I am thinking so current behave is better.\\n> There is possible risk of using /text() somewhere - it enforce a empty tag\\n> with all negative impacts.\\n>\\n> I prefer to fix doc in conformance with regress tests and append note\\n> about mapping these corner cases from XML to relations.\\n>\\n> What do you think about it?\\n>\\n\\nIt is documented already\\n\\n\\\"If the <literal>PATH</> matches an empty tag the result is an empty string\\\"\\n\\nAttached new patch\\n\\ncleaned documentation\\nregress tests is more robust\\nappended comment in src related to generating empty string for empty tag\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> Regards\\n>\\n> Pavel\\n>\\n>\\n>\\n>\\n>\\n>\\n>\\n>\\n>\\n>> --\\n>> Álvaro Herrera https://www.2ndQuadrant.com/\\n>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\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\": \"Thu, 2 Mar 2017 09:13:02 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n\\n> It is documented already\\n> \\n> \\\"If the <literal>PATH</> matches an empty tag the result is an empty string\\\"\\n\\nHmm, okay. But what we have here is not an empty tag, but a tag that is\\ncompletely missing. I don't think those two cases should be treated in\\nthe same way ...\\n\\n> Attached new patch\\n> \\n> cleaned documentation\\n> regress tests is more robust\\n> appended comment in src related to generating empty string for empty tag\\n\\nThanks, I incorporated those changes. Here's v46. I rewrote the\\ndocumentation, and fixed a couple of incorrectly copied&pasted comments\\nin the new executor code; I think that one looks good. In the future we\\ncould rewrite it to avoid the need for a tuplestore, but I think the\\ncurrent approach is good enough for a pg10 implementation.\\n\\nBarring serious problems, I intend to commit this later today.\\n\\n-- \\n�lvaro Herrera https://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, 2 Mar 2017 14:13:53 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Dne 2. 3. 2017 18:14 napsal uživatel \\\"Alvaro Herrera\\\" <\\nalvherre@2ndquadrant.com>:\\n\\nPavel Stehule wrote:\\n\\n> It is documented already\\n>\\n> \\\"If the <literal>PATH</> matches an empty tag the result is an empty\\nstring\\\"\\n\\nHmm, okay. But what we have here is not an empty tag, but a tag that is\\ncompletely missing. I don't think those two cases should be treated in\\nthe same way ...\\n\\n\\nthis information is not propagated from libxml2.\\n\\n\\n> Attached new patch\\n>\\n> cleaned documentation\\n> regress tests is more robust\\n> appended comment in src related to generating empty string for empty tag\\n\\nThanks, I incorporated those changes. Here's v46. I rewrote the\\ndocumentation, and fixed a couple of incorrectly copied&pasted comments\\nin the new executor code; I think that one looks good. In the future we\\ncould rewrite it to avoid the need for a tuplestore, but I think the\\ncurrent approach is good enough for a pg10 implementation.\\n\\nBarring serious problems, I intend to commit this later today.\\n\\n\\n\\nthank you very much\\n\\nregards\\n\\nPavel\\n\\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\nDne 2. 3. 2017 18:14 napsal uživatel \\\"Alvaro Herrera\\\" <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\\n\\n> It is documented already\\n>\\n> \\\"If the <literal>PATH</> matches an empty tag the result is an empty string\\\"\\n\\nHmm, okay. But what we have here is not an empty tag, but a tag that is\\ncompletely missing. I don't think those two cases should be treated in\\nthe same way ...this information is not propagated from libxml2.\\n\\n> Attached new patch\\n>\\n> cleaned documentation\\n> regress tests is more robust\\n> appended comment in src related to generating empty string for empty tag\\n\\nThanks, I incorporated those changes. Here's v46. I rewrote the\\ndocumentation, and fixed a couple of incorrectly copied&pasted comments\\nin the new executor code; I think that one looks good. In the future we\\ncould rewrite it to avoid the need for a tuplestore, but I think the\\ncurrent approach is good enough for a pg10 implementation.\\n\\nBarring serious problems, I intend to commit this later today.\\nthank you very muchregardsPavel\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Thu, 2 Mar 2017 18:46:23 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"So in the old (non-executor-node) implementation, you could attach WITH\\nORDINALITY to the xmltable expression and it would count the output\\nrows, regardless of which XML document it comes from. With the new\\nimplementation, the grammar no longer accepts it. To count output rows,\\nyou still need to use row_number(). Maybe this is okay. This is the\\nexample from the docs, and I add another XML document with two more rows\\nfor xmltable. Look at the three numbering columns ...\\n\\nCREATE TABLE xmldata AS SELECT\\nxml $$\\n<ROWS> \\n <ROW id=\\\"1\\\">\\n <COUNTRY_ID>AU</COUNTRY_ID>\\n <COUNTRY_NAME>Australia</COUNTRY_NAME>\\n </ROW> \\n <ROW id=\\\"5\\\"> \\n <COUNTRY_ID>JP</COUNTRY_ID>\\n <COUNTRY_NAME>Japan</COUNTRY_NAME>\\n <PREMIER_NAME>Shinzo Abe</PREMIER_NAME>\\n <SIZE unit=\\\"sq_mi\\\">145935</SIZE>\\n </ROW>\\n <ROW id=\\\"6\\\">\\n <COUNTRY_ID>SG</COUNTRY_ID>\\n <COUNTRY_NAME>Singapore</COUNTRY_NAME>\\n <SIZE unit=\\\"sq_km\\\">697</SIZE>\\n </ROW>\\n</ROWS>\\n$$ AS data;\\n\\n insert into xmldata values ($$\\n <ROWS><ROW id=\\\"2\\\"><COUNTRY_ID>CL</COUNTRY_ID><COUNTRY_NAME>Chile</COUNTRY_NAME></ROW>\\n <ROW id=\\\"3\\\"><COUNTRY_ID>AR</COUNTRY_ID><COUNTRY_NAME>Argentina</COUNTRY_NAME></ROW></ROWS>$$);\\n\\nSELECT ROW_NUMBER() OVER (), xmltable.*\\n FROM xmldata, \\n XMLTABLE('//ROWS/ROW' \\n PASSING data\\n COLUMNS id int PATH '@id',\\n ordinality FOR ORDINALITY,\\n \\\"COUNTRY_NAME\\\" text,\\n country_id text PATH 'COUNTRY_ID',\\n size_sq_km float PATH 'SIZE[@unit = \\\"sq_km\\\"]',\\n size_other text PATH\\n 'concat(SIZE[@unit!=\\\"sq_km\\\"], \\\" \\\", SIZE[@unit!=\\\"sq_km\\\"]/@unit)',\\n premier_name text PATH 'PREMIER_NAME' DEFAULT 'not specified')\\n;\\n\\n row_number │ id │ ordinality │ COUNTRY_NAME │ country_id │ size_sq_km │ size_other │ premier_name \\n────────────┼────┼────────────┼──────────────┼────────────┼────────────┼──────────────┼───────────────\\n 1 │ 1 │ 1 │ Australia │ AU │ │ │ not specified\\n 2 │ 5 │ 2 │ Japan │ JP │ │ 145935 sq_mi │ Shinzo Abe\\n 3 │ 6 │ 3 │ Singapore │ SG │ 697 │ │ not specified\\n 4 │ 2 │ 1 │ Chile │ CL │ │ │ not specified\\n 5 │ 3 │ 2 │ Argentina │ AR │ │ │ not specified\\n\\n\\n-- \\nÁlvaro Herrera https://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, 2 Mar 2017 15:32:05 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-03-02 19:32 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> So in the old (non-executor-node) implementation, you could attach WITH\\n> ORDINALITY to the xmltable expression and it would count the output\\n> rows, regardless of which XML document it comes from. With the new\\n> implementation, the grammar no longer accepts it. To count output rows,\\n> you still need to use row_number(). Maybe this is okay. This is the\\n> example from the docs, and I add another XML document with two more rows\\n> for xmltable. Look at the three numbering columns ...\\n>\\n\\nIt is expected - now tablefunc are not special case of SRF, so it lost all\\nSRF functionality. It is not critical lost - it supports internally FOR\\nORDINALITY column, and classic ROW_NUMBER can be used. It can be enhanced\\nto support WITH ORDINALITY in future, but I have not any use case for it.\\n\\nRegards\\n\\nPavel\\n\\n\\n\\n>\\n> CREATE TABLE xmldata AS SELECT\\n> xml $$\\n> <ROWS>\\n> <ROW id=\\\"1\\\">\\n> <COUNTRY_ID>AU</COUNTRY_ID>\\n> <COUNTRY_NAME>Australia</COUNTRY_NAME>\\n> </ROW>\\n> <ROW id=\\\"5\\\">\\n> <COUNTRY_ID>JP</COUNTRY_ID>\\n> <COUNTRY_NAME>Japan</COUNTRY_NAME>\\n> <PREMIER_NAME>Shinzo Abe</PREMIER_NAME>\\n> <SIZE unit=\\\"sq_mi\\\">145935</SIZE>\\n> </ROW>\\n> <ROW id=\\\"6\\\">\\n> <COUNTRY_ID>SG</COUNTRY_ID>\\n> <COUNTRY_NAME>Singapore</COUNTRY_NAME>\\n> <SIZE unit=\\\"sq_km\\\">697</SIZE>\\n> </ROW>\\n> </ROWS>\\n> $$ AS data;\\n>\\n> insert into xmldata values ($$\\n> <ROWS><ROW id=\\\"2\\\"><COUNTRY_ID>CL</COUNTRY_ID><COUNTRY_NAME>\\n> Chile</COUNTRY_NAME></ROW>\\n> <ROW id=\\\"3\\\"><COUNTRY_ID>AR</COUNTRY_ID><COUNTRY_NAME>\\n> Argentina</COUNTRY_NAME></ROW></ROWS>$$);\\n>\\n> SELECT ROW_NUMBER() OVER (), xmltable.*\\n> FROM xmldata,\\n> XMLTABLE('//ROWS/ROW'\\n> PASSING data\\n> COLUMNS id int PATH '@id',\\n> ordinality FOR ORDINALITY,\\n> \\\"COUNTRY_NAME\\\" text,\\n> country_id text PATH 'COUNTRY_ID',\\n> size_sq_km float PATH 'SIZE[@unit = \\\"sq_km\\\"]',\\n> size_other text PATH\\n> 'concat(SIZE[@unit!=\\\"sq_km\\\"], \\\" \\\",\\n> SIZE[@unit!=\\\"sq_km\\\"]/@unit)',\\n> premier_name text PATH 'PREMIER_NAME' DEFAULT 'not\\n> specified')\\n> ;\\n>\\n> row_number │ id │ ordinality │ COUNTRY_NAME │ country_id │ size_sq_km │\\n> size_other │ premier_name\\n> ────────────┼────┼────────────┼──────────────┼────────────┼─\\n> ───────────┼──────────────┼───────────────\\n> 1 │ 1 │ 1 │ Australia │ AU │ │\\n> │ not specified\\n> 2 │ 5 │ 2 │ Japan │ JP │ │\\n> 145935 sq_mi │ Shinzo Abe\\n> 3 │ 6 │ 3 │ Singapore │ SG │ 697 │\\n> │ not specified\\n> 4 │ 2 │ 1 │ Chile │ CL │ │\\n> │ not specified\\n> 5 │ 3 │ 2 │ Argentina │ AR │ │\\n> │ not specified\\n>\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2017-03-02 19:32 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:So in the old (non-executor-node) implementation, you could attach WITH\\r\\nORDINALITY to the xmltable expression and it would count the output\\r\\nrows, regardless of which XML document it comes from. With the new\\r\\nimplementation, the grammar no longer accepts it. To count output rows,\\r\\nyou still need to use row_number(). Maybe this is okay. This is the\\r\\nexample from the docs, and I add another XML document with two more rows\\r\\nfor xmltable. Look at the three numbering columns ...It is expected - now tablefunc are not special case of SRF, so it lost all SRF functionality. It is not critical lost - it supports internally FOR ORDINALITY column, and classic ROW_NUMBER can be used. It can be enhanced to support WITH ORDINALITY in future, but I have not any use case for it.RegardsPavel \\n\\r\\nCREATE TABLE xmldata AS SELECT\\r\\nxml $$\\n<ROWS>\\r\\n <ROW id=\\\"1\\\">\\r\\n <COUNTRY_ID>AU</COUNTRY_ID>\\r\\n <COUNTRY_NAME>Australia</COUNTRY_NAME>\\r\\n </ROW>\\r\\n <ROW id=\\\"5\\\">\\r\\n <COUNTRY_ID>JP</COUNTRY_ID>\\r\\n <COUNTRY_NAME>Japan</COUNTRY_NAME>\\n <PREMIER_NAME>Shinzo Abe</PREMIER_NAME>\\r\\n <SIZE unit=\\\"sq_mi\\\">145935</SIZE>\\n </ROW>\\r\\n <ROW id=\\\"6\\\">\\r\\n <COUNTRY_ID>SG</COUNTRY_ID>\\r\\n <COUNTRY_NAME>Singapore</COUNTRY_NAME>\\n <SIZE unit=\\\"sq_km\\\">697</SIZE>\\r\\n </ROW>\\r\\n</ROWS>\\r\\n$$ AS data;\\n\\r\\n insert into xmldata values ($$\\r\\n <ROWS><ROW id=\\\"2\\\"><COUNTRY_ID>CL</COUNTRY_ID><COUNTRY_NAME>Chile</COUNTRY_NAME></ROW>\\r\\n <ROW id=\\\"3\\\"><COUNTRY_ID>AR</COUNTRY_ID><COUNTRY_NAME>Argentina</COUNTRY_NAME></ROW></ROWS>$$);\\n\\r\\nSELECT ROW_NUMBER() OVER (), xmltable.*\\r\\n FROM xmldata,\\r\\n XMLTABLE('//ROWS/ROW'\\n PASSING data\\r\\n COLUMNS id int PATH '@id',\\r\\n ordinality FOR ORDINALITY,\\n \\\"COUNTRY_NAME\\\" text,\\n country_id text PATH 'COUNTRY_ID',\\n size_sq_km float PATH 'SIZE[@unit = \\\"sq_km\\\"]',\\r\\n size_other text PATH\\r\\n 'concat(SIZE[@unit!=\\\"sq_km\\\"], \\\" \\\", SIZE[@unit!=\\\"sq_km\\\"]/@unit)',\\n premier_name text PATH 'PREMIER_NAME' DEFAULT 'not specified')\\r\\n;\\n\\n row_number │ id │ ordinality │ COUNTRY_NAME │ country_id │ size_sq_km │ size_other │ premier_name\\r\\n────────────┼────┼────────────┼──────────────┼────────────┼────────────┼──────────────┼───────────────\\r\\n 1 │ 1 │ 1 │ Australia │ AU │ │ │ not specified\\r\\n 2 │ 5 │ 2 │ Japan │ JP │ │ 145935 sq_mi │ Shinzo Abe\\r\\n 3 │ 6 │ 3 │ Singapore │ SG │ 697 │ │ not specified\\r\\n 4 │ 2 │ 1 │ Chile │ CL │ │ │ not specified\\r\\n 5 │ 3 │ 2 │ Argentina │ AR │ │ │ not specified\\n\\n\\r\\n--\\r\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\r\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Thu, 2 Mar 2017 20:17:12 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n> 2017-03-02 19:32 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n> \\n> > So in the old (non-executor-node) implementation, you could attach WITH\\n> > ORDINALITY to the xmltable expression and it would count the output\\n> > rows, regardless of which XML document it comes from. With the new\\n> > implementation, the grammar no longer accepts it. To count output rows,\\n> > you still need to use row_number(). Maybe this is okay. This is the\\n> > example from the docs, and I add another XML document with two more rows\\n> > for xmltable. Look at the three numbering columns ...\\n> >\\n> \\n> It is expected - now tablefunc are not special case of SRF, so it lost all\\n> SRF functionality. It is not critical lost - it supports internally FOR\\n> ORDINALITY column, and classic ROW_NUMBER can be used. It can be enhanced\\n> to support WITH ORDINALITY in future, but I have not any use case for it.\\n\\nFine.\\n\\nAfter looking at the new executor code a bit, I noticed that we don't\\nneed the resultSlot anymore; we can use the ss_ScanTupleSlot instead.\\nBecause resultSlot was being used in the xml.c code (which already\\nappeared a bit dubious to me), I changed the interface so that instead\\nthe things that it read from it are passed as parameters -- namely, in\\nInitBuilder we pass natts, and in GetValue we pass typid and typmod.\\n\\nSecondly, I noticed we have the FetchRow routine produce a minimal\\ntuple, put it in a slot; then its caller takes the slot and put the\\ntuple in the tuplestore. This is pointless; we can just have FetchRow\\nput the tuple in the tuplestore directly and not bother with any slot\\nmanipulations there. This simplifies the code a bit.\\n\\nHere's v47 with those changes.\\n\\n-- \\n�lvaro Herrera https://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, 2 Mar 2017 18:35:44 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-03-02 22:35 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n> > 2017-03-02 19:32 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n> >\\n> > > So in the old (non-executor-node) implementation, you could attach WITH\\n> > > ORDINALITY to the xmltable expression and it would count the output\\n> > > rows, regardless of which XML document it comes from. With the new\\n> > > implementation, the grammar no longer accepts it. To count output\\n> rows,\\n> > > you still need to use row_number(). Maybe this is okay. This is the\\n> > > example from the docs, and I add another XML document with two more\\n> rows\\n> > > for xmltable. Look at the three numbering columns ...\\n> > >\\n> >\\n> > It is expected - now tablefunc are not special case of SRF, so it lost\\n> all\\n> > SRF functionality. It is not critical lost - it supports internally FOR\\n> > ORDINALITY column, and classic ROW_NUMBER can be used. It can be enhanced\\n> > to support WITH ORDINALITY in future, but I have not any use case for it.\\n>\\n> Fine.\\n>\\n> After looking at the new executor code a bit, I noticed that we don't\\n> need the resultSlot anymore; we can use the ss_ScanTupleSlot instead.\\n> Because resultSlot was being used in the xml.c code (which already\\n> appeared a bit dubious to me), I changed the interface so that instead\\n> the things that it read from it are passed as parameters -- namely, in\\n> InitBuilder we pass natts, and in GetValue we pass typid and typmod.\\n>\\n\\nI had similar feeling\\n\\n>\\n> Secondly, I noticed we have the FetchRow routine produce a minimal\\n> tuple, put it in a slot; then its caller takes the slot and put the\\n> tuple in the tuplestore. This is pointless; we can just have FetchRow\\n> put the tuple in the tuplestore directly and not bother with any slot\\n> manipulations there. This simplifies the code a bit.\\n>\\n>\\nhas sense\\n\\nattached update with fixed tests\\n\\nRegards\\n\\nPavel\\n\\n\\n\\n> Here's v47 with those changes.\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL 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\": \"Fri, 3 Mar 2017 10:28:54 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n\\n> attached update with fixed tests\\n\\nHeh, I noticed that you removed the libxml \\\"context\\\" lines that\\ndifferentiate xml.out from xml_2.out when doing this. My implementation\\nemits those lines, so it was failing for me. I restored them.\\n\\nI also changed a few things to avoid copying into TableFuncScanState\\nthings that come from the TableFunc itself, since the executor state\\nnode can grab them from the plan node. Let's do that. So instead of\\n\\\"evalcols\\\" the code now checks that the column list is empty; and also,\\nread the ordinality column number from the plan node.\\n\\nI have to bounce this back to you one more time, hopefully the last one\\nI hope. Two things:\\n\\n1. Please verify that pg_stat_statements behaves correctly. The patch\\ndoesn't have changes to contrib/ so without testing I'm guessing that it\\ndoesn't work. I think something very simple should do.\\n\\n2. As I've complained many times, I find the way we manage an empty\\nCOLUMNS clause pretty bad. The standard doesn't require that syntax\\n(COLUMNS is required), and I don't like the implementation, so why not\\nprovide the feature in a different way? My proposal is to change the\\ncolumn options in gram.y to be something like this:\\n\\nxmltable_column_option_el:\\n\\t\\t\\tIDENT b_expr\\n\\t\\t\\t\\t{ $$ = makeDefElem($1, $2, @1); }\\n\\t\\t\\t| DEFAULT b_expr\\n\\t\\t\\t\\t{ $$ = makeDefElem(\\\"default\\\", $2, @1); }\\n\\t\\t\\t| FULL VALUE_P\\n\\t\\t\\t\\t{ $$ = makeDefElem(\\\"full_value\\\", NULL, @1); }\\n\\t\\t\\t| NOT NULL_P\\n\\t\\t\\t\\t{ $$ = makeDefElem(\\\"is_not_null\\\", (Node *) makeInteger(true), @1); }\\n\\t\\t\\t| NULL_P\\n\\t\\t\\t\\t{ $$ = makeDefElem(\\\"is_not_null\\\", (Node *) makeInteger(false), @1); }\\n\\t\\t;\\n\\nNote the FULL VALUE. Then we can process it like\\n\\n\\t\\t\\telse if (strcmp(defel->defname, \\\"full_value\\\") == 0)\\n\\t\\t\\t{\\n\\t\\t\\t\\tif (fc->colexpr != NULL)\\n\\t\\t\\t\\t\\tereport(ERROR,\\n\\t\\t\\t\\t\\t\\t\\t(errcode(ERRCODE_SYNTAX_ERROR),\\n\\t\\t\\t\\t\\t\\t\\t errmsg(\\\"FULL ROW may not be specified together with PATH\\\"),\\n\\t\\t\\t\\t\\t\\t\\t parser_errposition(defel->location)));\\n\\t\\t\\t\\tfc->full_row = true;\\n\\t\\t\\t}\\n\\nSo if you want the full XML value of the row, you have to specify it,\\n\\n .. XMLTABLE ( ... COLUMNS ..., whole_row xml FULL VALUE, ... )\\n\\nThis has the extra feature that you can add, say, an ORDINALITY column\\ntogether with the XML value, something that you cannot do with the\\ncurrent implementation.\\n\\nIt doesn't have to be FULL VALUE, but I couldn't think of anything\\nbetter. (I didn't want to add any new keywords for this.) If you have\\na better idea, let's discuss.\\n\\nCode-wise, this completely removes the \\\"else\\\" block in transformRangeTableFunc\\nwhich I marked with an XXX comment. That's a good thing -- let's get\\nrid of that. Also, it should remove the need for the separate \\\"if\\n!columns\\\" case in tfuncLoadRows. All those cases would become part of\\nthe normal code path instead of special cases. I think\\nXmlTableSetColumnFilter doesn't need any change (we just don't call if\\nfor the FULL VALUE row); and XmlTableGetValue needs a special case that\\nif the column filter is NULL (i.e. SetColumnFilter wasn't called for\\nthat column) then return the whole row.\\n\\n\\nOf course, this opens an implementation issue: how do you annotate\\nthings from parse analysis till execution? The current TableFunc\\nstructure doesn't help, because there are only lists of column names and\\nexpressions; and we can't use the case of a NULL colexpr, because that\\ncase is already used by the column filter being the column name (a\\nfeature required by the standard). A simple way would be to have a new\\n\\\"colno\\\" struct member, to store a column number for the column marked\\nFULL VALUE (just like ordinalitycol). This means you can't have more\\nthan one of those FULL VALUE columns, but that seems okay.\\n\\n\\n(Of course, this means that the two cases that have no COLUMNS in the\\n\\\"xmltable\\\" production in gram.y should go away).\\n\\n-- \\n�lvaro Herrera https://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\": \"Fri, 3 Mar 2017 15:15:11 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-03-03 19:15 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n>\\n> > attached update with fixed tests\\n>\\n> Heh, I noticed that you removed the libxml \\\"context\\\" lines that\\n> differentiate xml.out from xml_2.out when doing this. My implementation\\n> emits those lines, so it was failing for me. I restored them.\\n>\\n> I also changed a few things to avoid copying into TableFuncScanState\\n> things that come from the TableFunc itself, since the executor state\\n> node can grab them from the plan node. Let's do that. So instead of\\n> \\\"evalcols\\\" the code now checks that the column list is empty; and also,\\n> read the ordinality column number from the plan node.\\n>\\n> I have to bounce this back to you one more time, hopefully the last one\\n> I hope. Two things:\\n>\\n> 1. Please verify that pg_stat_statements behaves correctly. The patch\\n> doesn't have changes to contrib/ so without testing I'm guessing that it\\n> doesn't work. I think something very simple should do.\\n>\\n> 2. As I've complained many times, I find the way we manage an empty\\n> COLUMNS clause pretty bad. The standard doesn't require that syntax\\n> (COLUMNS is required), and I don't like the implementation, so why not\\n> provide the feature in a different way? My proposal is to change the\\n> column options in gram.y to be something like this:\\n\\n\\nThe clause COLUMNS is optional on Oracle and DB2\\n\\nSo I prefer a Oracle, DB2 design. If you are strongly against it, then we\\ncan remove it to be ANSI/SQL only.\\n\\nI am don't see an good idea to introduce third syntax.\\n\\n\\n> xmltable_column_option_el:\\n> IDENT b_expr\\n> { $$ = makeDefElem($1, $2, @1); }\\n> | DEFAULT b_expr\\n> { $$ = makeDefElem(\\\"default\\\", $2, @1); }\\n> | FULL VALUE_P\\n> { $$ = makeDefElem(\\\"full_value\\\", NULL,\\n> @1); }\\n> | NOT NULL_P\\n> { $$ = makeDefElem(\\\"is_not_null\\\", (Node *)\\n> makeInteger(true), @1); }\\n> | NULL_P\\n> { $$ = makeDefElem(\\\"is_not_null\\\", (Node *)\\n> makeInteger(false), @1); }\\n> ;\\n>\\n> Note the FULL VALUE. Then we can process it like\\n>\\n> else if (strcmp(defel->defname, \\\"full_value\\\") == 0)\\n> {\\n> if (fc->colexpr != NULL)\\n> ereport(ERROR,\\n>\\n> (errcode(ERRCODE_SYNTAX_ERROR),\\n> errmsg(\\\"FULL ROW\\n> may not be specified together with PATH\\\"),\\n>\\n> parser_errposition(defel->location)));\\n> fc->full_row = true;\\n> }\\n>\\n> So if you want the full XML value of the row, you have to specify it,\\n>\\n> .. XMLTABLE ( ... COLUMNS ..., whole_row xml FULL VALUE, ... )\\n>\\n> This has the extra feature that you can add, say, an ORDINALITY column\\n> together with the XML value, something that you cannot do with the\\n> current implementation.\\n>\\n> It doesn't have to be FULL VALUE, but I couldn't think of anything\\n> better. (I didn't want to add any new keywords for this.) If you have\\n> a better idea, let's discuss.\\n>\\n\\nI don't see a introduction own syntax as necessary solution here - use\\nOracle, DB2 compatible syntax, or ANSI.\\n\\nIt is partially corner case - the benefit of this case is almost bigger\\ncompatibility with mentioned databases.\\n\\n\\n>\\n> Code-wise, this completely removes the \\\"else\\\" block in\\n> transformRangeTableFunc\\n> which I marked with an XXX comment. That's a good thing -- let's get\\n> rid of that. Also, it should remove the need for the separate \\\"if\\n> !columns\\\" case in tfuncLoadRows. All those cases would become part of\\n> the normal code path instead of special cases. I think\\n> XmlTableSetColumnFilter doesn't need any change (we just don't call if\\n> for the FULL VALUE row); and XmlTableGetValue needs a special case that\\n> if the column filter is NULL (i.e. SetColumnFilter wasn't called for\\n> that column) then return the whole row.\\n>\\n>\\n> Of course, this opens an implementation issue: how do you annotate\\n> things from parse analysis till execution? The current TableFunc\\n> structure doesn't help, because there are only lists of column names and\\n> expressions; and we can't use the case of a NULL colexpr, because that\\n> case is already used by the column filter being the column name (a\\n> feature required by the standard). A simple way would be to have a new\\n> \\\"colno\\\" struct member, to store a column number for the column marked\\n> FULL VALUE (just like ordinalitycol). This means you can't have more\\n> than one of those FULL VALUE columns, but that seems okay.\\n>\\n>\\n> (Of course, this means that the two cases that have no COLUMNS in the\\n> \\\"xmltable\\\" production in gram.y should go away).\\n>\\n\\nYou are commiter, and you should to decide - as first I prefer current\\nstate, as second a remove this part - it should be good for you too,\\nbecause code that you don't like will be left.\\n\\nI dislike introduce new syntax - this case is not too important for this.\\n\\nRegards\\n\\nPavel\\n\\n\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2017-03-03 19:15 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\\n\\n> attached update with fixed tests\\n\\nHeh, I noticed that you removed the libxml \\\"context\\\" lines that\\ndifferentiate xml.out from xml_2.out when doing this. My implementation\\nemits those lines, so it was failing for me. I restored them.\\n\\nI also changed a few things to avoid copying into TableFuncScanState\\nthings that come from the TableFunc itself, since the executor state\\nnode can grab them from the plan node. Let's do that. So instead of\\n\\\"evalcols\\\" the code now checks that the column list is empty; and also,\\nread the ordinality column number from the plan node.\\n\\nI have to bounce this back to you one more time, hopefully the last one\\nI hope. Two things:\\n\\n1. Please verify that pg_stat_statements behaves correctly. The patch\\ndoesn't have changes to contrib/ so without testing I'm guessing that it\\ndoesn't work. I think something very simple should do.\\n\\n2. As I've complained many times, I find the way we manage an empty\\nCOLUMNS clause pretty bad. The standard doesn't require that syntax\\n(COLUMNS is required), and I don't like the implementation, so why not\\nprovide the feature in a different way? My proposal is to change the\\ncolumn options in gram.y to be something like this:The clause COLUMNS is optional on Oracle and DB2 So I prefer a Oracle, DB2 design. If you are strongly against it, then we can remove it to be ANSI/SQL only.I am don't see an good idea to introduce third syntax.\\n\\nxmltable_column_option_el:\\n IDENT b_expr\\n { $$ = makeDefElem($1, $2, @1); }\\n | DEFAULT b_expr\\n { $$ = makeDefElem(\\\"default\\\", $2, @1); }\\n | FULL VALUE_P\\n { $$ = makeDefElem(\\\"full_value\\\", NULL, @1); }\\n | NOT NULL_P\\n { $$ = makeDefElem(\\\"is_not_null\\\", (Node *) makeInteger(true), @1); }\\n | NULL_P\\n { $$ = makeDefElem(\\\"is_not_null\\\", (Node *) makeInteger(false), @1); }\\n ;\\n\\nNote the FULL VALUE. Then we can process it like\\n\\n else if (strcmp(defel->defname, \\\"full_value\\\") == 0)\\n {\\n if (fc->colexpr != NULL)\\n ereport(ERROR,\\n (errcode(ERRCODE_SYNTAX_ERROR),\\n errmsg(\\\"FULL ROW may not be specified together with PATH\\\"),\\n parser_errposition(defel->location)));\\n fc->full_row = true;\\n }\\n\\nSo if you want the full XML value of the row, you have to specify it,\\n\\n .. XMLTABLE ( ... COLUMNS ..., whole_row xml FULL VALUE, ... )\\n\\nThis has the extra feature that you can add, say, an ORDINALITY column\\ntogether with the XML value, something that you cannot do with the\\ncurrent implementation.\\n\\nIt doesn't have to be FULL VALUE, but I couldn't think of anything\\nbetter. (I didn't want to add any new keywords for this.) If you have\\na better idea, let's discuss.I don't see a introduction own syntax as necessary solution here - use Oracle, DB2 compatible syntax, or ANSI. It is partially corner case - the benefit of this case is almost bigger compatibility with mentioned databases. \\n\\nCode-wise, this completely removes the \\\"else\\\" block in transformRangeTableFunc\\nwhich I marked with an XXX comment. That's a good thing -- let's get\\nrid of that. Also, it should remove the need for the separate \\\"if\\n!columns\\\" case in tfuncLoadRows. All those cases would become part of\\nthe normal code path instead of special cases. I think\\nXmlTableSetColumnFilter doesn't need any change (we just don't call if\\nfor the FULL VALUE row); and XmlTableGetValue needs a special case that\\nif the column filter is NULL (i.e. SetColumnFilter wasn't called for\\nthat column) then return the whole row.\\n\\n\\nOf course, this opens an implementation issue: how do you annotate\\nthings from parse analysis till execution? The current TableFunc\\nstructure doesn't help, because there are only lists of column names and\\nexpressions; and we can't use the case of a NULL colexpr, because that\\ncase is already used by the column filter being the column name (a\\nfeature required by the standard). A simple way would be to have a new\\n\\\"colno\\\" struct member, to store a column number for the column marked\\nFULL VALUE (just like ordinalitycol). This means you can't have more\\nthan one of those FULL VALUE columns, but that seems okay.\\n\\n\\n(Of course, this means that the two cases that have no COLUMNS in the\\n\\\"xmltable\\\" production in gram.y should go away).You are commiter, and you should to decide - as first I prefer current state, as second a remove this part - it should be good for you too, because code that you don't like will be left.I dislike introduce new syntax - this case is not too important for this.RegardsPavel\\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Fri, 3 Mar 2017 19:42:11 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-03-03 19:42 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:\\n\\n>\\n>\\n> 2017-03-03 19:15 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>\\n>> Pavel Stehule wrote:\\n>>\\n>> > attached update with fixed tests\\n>>\\n>> Heh, I noticed that you removed the libxml \\\"context\\\" lines that\\n>> differentiate xml.out from xml_2.out when doing this. My implementation\\n>> emits those lines, so it was failing for me. I restored them.\\n>>\\n>> I also changed a few things to avoid copying into TableFuncScanState\\n>> things that come from the TableFunc itself, since the executor state\\n>> node can grab them from the plan node. Let's do that. So instead of\\n>> \\\"evalcols\\\" the code now checks that the column list is empty; and also,\\n>> read the ordinality column number from the plan node.\\n>>\\n>> I have to bounce this back to you one more time, hopefully the last one\\n>> I hope. Two things:\\n>>\\n>> 1. Please verify that pg_stat_statements behaves correctly. The patch\\n>> doesn't have changes to contrib/ so without testing I'm guessing that it\\n>> doesn't work. I think something very simple should do.\\n>>\\n>> 2. As I've complained many times, I find the way we manage an empty\\n>> COLUMNS clause pretty bad. The standard doesn't require that syntax\\n>> (COLUMNS is required), and I don't like the implementation, so why not\\n>> provide the feature in a different way? My proposal is to change the\\n>> column options in gram.y to be something like this:\\n>\\n>\\n> The clause COLUMNS is optional on Oracle and DB2\\n>\\n> So I prefer a Oracle, DB2 design. If you are strongly against it, then we\\n> can remove it to be ANSI/SQL only.\\n>\\n> I am don't see an good idea to introduce third syntax.\\n>\\n>\\n>> xmltable_column_option_el:\\n>> IDENT b_expr\\n>> { $$ = makeDefElem($1, $2, @1); }\\n>> | DEFAULT b_expr\\n>> { $$ = makeDefElem(\\\"default\\\", $2, @1); }\\n>> | FULL VALUE_P\\n>> { $$ = makeDefElem(\\\"full_value\\\", NULL,\\n>> @1); }\\n>> | NOT NULL_P\\n>> { $$ = makeDefElem(\\\"is_not_null\\\", (Node\\n>> *) makeInteger(true), @1); }\\n>> | NULL_P\\n>> { $$ = makeDefElem(\\\"is_not_null\\\", (Node\\n>> *) makeInteger(false), @1); }\\n>> ;\\n>>\\n>> Note the FULL VALUE. Then we can process it like\\n>>\\n>> else if (strcmp(defel->defname, \\\"full_value\\\") ==\\n>> 0)\\n>> {\\n>> if (fc->colexpr != NULL)\\n>> ereport(ERROR,\\n>>\\n>> (errcode(ERRCODE_SYNTAX_ERROR),\\n>> errmsg(\\\"FULL ROW\\n>> may not be specified together with PATH\\\"),\\n>>\\n>> parser_errposition(defel->location)));\\n>> fc->full_row = true;\\n>> }\\n>>\\n>> So if you want the full XML value of the row, you have to specify it,\\n>>\\n>> .. XMLTABLE ( ... COLUMNS ..., whole_row xml FULL VALUE, ... )\\n>>\\n>> This has the extra feature that you can add, say, an ORDINALITY column\\n>> together with the XML value, something that you cannot do with the\\n>> current implementation.\\n>>\\n>> It doesn't have to be FULL VALUE, but I couldn't think of anything\\n>> better. (I didn't want to add any new keywords for this.) If you have\\n>> a better idea, let's discuss.\\n>>\\n>\\n> I don't see a introduction own syntax as necessary solution here - use\\n> Oracle, DB2 compatible syntax, or ANSI.\\n>\\n> It is partially corner case - the benefit of this case is almost bigger\\n> compatibility with mentioned databases.\\n>\\n>\\n>>\\n>> Code-wise, this completely removes the \\\"else\\\" block in\\n>> transformRangeTableFunc\\n>> which I marked with an XXX comment. That's a good thing -- let's get\\n>> rid of that. Also, it should remove the need for the separate \\\"if\\n>> !columns\\\" case in tfuncLoadRows. All those cases would become part of\\n>> the normal code path instead of special cases. I think\\n>> XmlTableSetColumnFilter doesn't need any change (we just don't call if\\n>> for the FULL VALUE row); and XmlTableGetValue needs a special case that\\n>> if the column filter is NULL (i.e. SetColumnFilter wasn't called for\\n>> that column) then return the whole row.\\n>>\\n>>\\n>> Of course, this opens an implementation issue: how do you annotate\\n>> things from parse analysis till execution? The current TableFunc\\n>> structure doesn't help, because there are only lists of column names and\\n>> expressions; and we can't use the case of a NULL colexpr, because that\\n>> case is already used by the column filter being the column name (a\\n>> feature required by the standard). A simple way would be to have a new\\n>> \\\"colno\\\" struct member, to store a column number for the column marked\\n>> FULL VALUE (just like ordinalitycol). This means you can't have more\\n>> than one of those FULL VALUE columns, but that seems okay.\\n>>\\n>>\\n>> (Of course, this means that the two cases that have no COLUMNS in the\\n>> \\\"xmltable\\\" production in gram.y should go away).\\n>>\\n>\\n> You are commiter, and you should to decide - as first I prefer current\\n> state, as second a remove this part - it should be good for you too,\\n> because code that you don't like will be left.\\n>\\n> I dislike introduce new syntax - this case is not too important for this.\\n>\\n>\\nI am able to prepare reduced version if we do a agreement\\n\\nRegards\\n\\nPavel\\n\\n\\n> Regards\\n>\\n> Pavel\\n>\\n>\\n>> --\\n>> Álvaro Herrera https://www.2ndQuadrant.com/\\n>> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>>\\n>\\n>\\n\\n2017-03-03 19:42 GMT+01:00 Pavel Stehule <pavel.stehule@gmail.com>:2017-03-03 19:15 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\\n\\n> attached update with fixed tests\\n\\nHeh, I noticed that you removed the libxml \\\"context\\\" lines that\\ndifferentiate xml.out from xml_2.out when doing this. My implementation\\nemits those lines, so it was failing for me. I restored them.\\n\\nI also changed a few things to avoid copying into TableFuncScanState\\nthings that come from the TableFunc itself, since the executor state\\nnode can grab them from the plan node. Let's do that. So instead of\\n\\\"evalcols\\\" the code now checks that the column list is empty; and also,\\nread the ordinality column number from the plan node.\\n\\nI have to bounce this back to you one more time, hopefully the last one\\nI hope. Two things:\\n\\n1. Please verify that pg_stat_statements behaves correctly. The patch\\ndoesn't have changes to contrib/ so without testing I'm guessing that it\\ndoesn't work. I think something very simple should do.\\n\\n2. As I've complained many times, I find the way we manage an empty\\nCOLUMNS clause pretty bad. The standard doesn't require that syntax\\n(COLUMNS is required), and I don't like the implementation, so why not\\nprovide the feature in a different way? My proposal is to change the\\ncolumn options in gram.y to be something like this:The clause COLUMNS is optional on Oracle and DB2 So I prefer a Oracle, DB2 design. If you are strongly against it, then we can remove it to be ANSI/SQL only.I am don't see an good idea to introduce third syntax.\\n\\nxmltable_column_option_el:\\n IDENT b_expr\\n { $$ = makeDefElem($1, $2, @1); }\\n | DEFAULT b_expr\\n { $$ = makeDefElem(\\\"default\\\", $2, @1); }\\n | FULL VALUE_P\\n { $$ = makeDefElem(\\\"full_value\\\", NULL, @1); }\\n | NOT NULL_P\\n { $$ = makeDefElem(\\\"is_not_null\\\", (Node *) makeInteger(true), @1); }\\n | NULL_P\\n { $$ = makeDefElem(\\\"is_not_null\\\", (Node *) makeInteger(false), @1); }\\n ;\\n\\nNote the FULL VALUE. Then we can process it like\\n\\n else if (strcmp(defel->defname, \\\"full_value\\\") == 0)\\n {\\n if (fc->colexpr != NULL)\\n ereport(ERROR,\\n (errcode(ERRCODE_SYNTAX_ERROR),\\n errmsg(\\\"FULL ROW may not be specified together with PATH\\\"),\\n parser_errposition(defel->location)));\\n fc->full_row = true;\\n }\\n\\nSo if you want the full XML value of the row, you have to specify it,\\n\\n .. XMLTABLE ( ... COLUMNS ..., whole_row xml FULL VALUE, ... )\\n\\nThis has the extra feature that you can add, say, an ORDINALITY column\\ntogether with the XML value, something that you cannot do with the\\ncurrent implementation.\\n\\nIt doesn't have to be FULL VALUE, but I couldn't think of anything\\nbetter. (I didn't want to add any new keywords for this.) If you have\\na better idea, let's discuss.I don't see a introduction own syntax as necessary solution here - use Oracle, DB2 compatible syntax, or ANSI. It is partially corner case - the benefit of this case is almost bigger compatibility with mentioned databases. \\n\\nCode-wise, this completely removes the \\\"else\\\" block in transformRangeTableFunc\\nwhich I marked with an XXX comment. That's a good thing -- let's get\\nrid of that. Also, it should remove the need for the separate \\\"if\\n!columns\\\" case in tfuncLoadRows. All those cases would become part of\\nthe normal code path instead of special cases. I think\\nXmlTableSetColumnFilter doesn't need any change (we just don't call if\\nfor the FULL VALUE row); and XmlTableGetValue needs a special case that\\nif the column filter is NULL (i.e. SetColumnFilter wasn't called for\\nthat column) then return the whole row.\\n\\n\\nOf course, this opens an implementation issue: how do you annotate\\nthings from parse analysis till execution? The current TableFunc\\nstructure doesn't help, because there are only lists of column names and\\nexpressions; and we can't use the case of a NULL colexpr, because that\\ncase is already used by the column filter being the column name (a\\nfeature required by the standard). A simple way would be to have a new\\n\\\"colno\\\" struct member, to store a column number for the column marked\\nFULL VALUE (just like ordinalitycol). This means you can't have more\\nthan one of those FULL VALUE columns, but that seems okay.\\n\\n\\n(Of course, this means that the two cases that have no COLUMNS in the\\n\\\"xmltable\\\" production in gram.y should go away).You are commiter, and you should to decide - as first I prefer current state, as second a remove this part - it should be good for you too, because code that you don't like will be left.I dislike introduce new syntax - this case is not too important for this.I am able to prepare reduced version if we do a agreement RegardsPavel RegardsPavel\\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Fri, 3 Mar 2017 20:23:03 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n> 2017-03-03 19:15 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> > 2. As I've complained many times, I find the way we manage an empty\\n> > COLUMNS clause pretty bad. The standard doesn't require that syntax\\n> > (COLUMNS is required), and I don't like the implementation, so why not\\n> > provide the feature in a different way? My proposal is to change the\\n> > column options in gram.y to be something like this:\\n> \\n> The clause COLUMNS is optional on Oracle and DB2\\n> \\n> So I prefer a Oracle, DB2 design. If you are strongly against it, then we\\n> can remove it to be ANSI/SQL only.\\n> \\n> I am don't see an good idea to introduce third syntax.\\n\\nOK. I think trying to be syntax compatible with DB2 or Oracle is a lost\\ncause, because the syntax used in the XPath expressions seems different\\n-- I think Oracle uses XQuery (which we don't support) and DB2 uses ...\\nnot sure what it is, but it doesn't work in our implementation\\n(stuff like '$d/employees/emp' in the row expression.)\\n\\nIn existing applications using those Oracle/DB2, is it common to omit\\nthe COLUMNS clause? I searched for \\\"xmltable oracle\\\" and had a look at\\nthe first few hits outside of the oracle docs:\\nhttp://viralpatel.net/blogs/oracle-xmltable-tutorial/\\nhttp://www.dba-oracle.com/t_xmltable.htm\\nhttp://stackoverflow.com/questions/12690868/how-to-use-xmltable-in-oracle\\nhttps://asktom.oracle.com/pls/asktom/f?p=100:11:0::::P11_QUESTION_ID:9533111800346252295\\nhttp://stackoverflow.com/questions/1222570/what-is-an-xmltable\\nhttps://community.oracle.com/thread/3955198\\n\\nNot a single one of these omit the COLUMNS clause (though the second one\\nmentions that the clause can be omitted).\\n\\nI also looked at a few samples with DB2 -- same thing; it is possible,\\nbut is it common?\\n\\nAnyway, I noticed that \\\"xml PATH '.'\\\" can be used to obtain the full XML\\nof the row, which I think is the feature I wanted, so I think we're\\ncovered and we can omit the case with no COLUMNS, since we already have\\nthe feature in another way. No need to implement anything further, and\\nwe can rip out the special case I don't like. Example:\\n\\nCREATE TABLE EMPLOYEES\\n(\\n id integer,\\n data XML\\n);\\nINSERT INTO EMPLOYEES\\n VALUES (1, '<Employees>\\n <Employee emplid=\\\"1111\\\" type=\\\"admin\\\">\\n <firstname>John</firstname>\\n <lastname>Watson</lastname>\\n <age>30</age>\\n <email>johnwatson@sh.com</email>\\n </Employee>\\n <Employee emplid=\\\"2222\\\" type=\\\"admin\\\">\\n <firstname>Sherlock</firstname>\\n <lastname>Homes</lastname>\\n <age>32</age>\\n <email>sherlock@sh.com</email>\\n </Employee>\\n <Employee emplid=\\\"3333\\\" type=\\\"user\\\">\\n <firstname>Jim</firstname>\\n <lastname>Moriarty</lastname>\\n <age>52</age>\\n <email>jim@sh.com</email>\\n </Employee>\\n <Employee emplid=\\\"4444\\\" type=\\\"user\\\">\\n <firstname>Mycroft</firstname>\\n <lastname>Holmes</lastname>\\n <age>41</age>\\n <email>mycroft@sh.com</email>\\n </Employee>\\n</Employees>');\\n\\nThis is with COLUMNS omitted:\\n\\nalvherre=# select xmltable.* from employees, xmltable('/Employees/Employee' passing data);\\n xmltable \\n──────────────────────────────────────────\\n <Employee emplid=\\\"1111\\\" type=\\\"admin\\\"> ↵\\n <firstname>John</firstname> ↵\\n <lastname>Watson</lastname> ↵\\n <age>30</age> ↵\\n <email>johnwatson@sh.com</email>↵\\n </Employee>\\n <Employee emplid=\\\"2222\\\" type=\\\"admin\\\"> ↵\\n <firstname>Sherlock</firstname> ↵\\n <lastname>Homes</lastname> ↵\\n <age>32</age> ↵\\n <email>sherlock@sh.com</email> ↵\\n </Employee>\\n <Employee emplid=\\\"3333\\\" type=\\\"user\\\"> ↵\\n <firstname>Jim</firstname> ↵\\n <lastname>Moriarty</lastname> ↵\\n <age>52</age> ↵\\n <email>jim@sh.com</email> ↵\\n </Employee>\\n <Employee emplid=\\\"4444\\\" type=\\\"user\\\"> ↵\\n <firstname>Mycroft</firstname> ↵\\n <lastname>Holmes</lastname> ↵\\n <age>41</age> ↵\\n <email>mycroft@sh.com</email> ↵\\n </Employee>\\n\\nand this is what you get with \\\"xml PATH '.'\\\" (I threw in ORDINALITY just\\nfor fun):\\n\\nalvherre=# select xmltable.* from employees, xmltable('/Employees/Employee' passing data columns row_number for ordinality, emp xml path '.');\\n row_number │ emp \\n────────────┼──────────────────────────────────────────\\n 1 │ <Employee emplid=\\\"1111\\\" type=\\\"admin\\\"> ↵\\n │ <firstname>John</firstname> ↵\\n │ <lastname>Watson</lastname> ↵\\n │ <age>30</age> ↵\\n │ <email>johnwatson@sh.com</email>↵\\n │ </Employee>\\n 2 │ <Employee emplid=\\\"2222\\\" type=\\\"admin\\\"> ↵\\n │ <firstname>Sherlock</firstname> ↵\\n │ <lastname>Homes</lastname> ↵\\n │ <age>32</age> ↵\\n │ <email>sherlock@sh.com</email> ↵\\n │ </Employee>\\n 3 │ <Employee emplid=\\\"3333\\\" type=\\\"user\\\"> ↵\\n │ <firstname>Jim</firstname> ↵\\n │ <lastname>Moriarty</lastname> ↵\\n │ <age>52</age> ↵\\n │ <email>jim@sh.com</email> ↵\\n │ </Employee>\\n 4 │ <Employee emplid=\\\"4444\\\" type=\\\"user\\\"> ↵\\n │ <firstname>Mycroft</firstname> ↵\\n │ <lastname>Holmes</lastname> ↵\\n │ <age>41</age> ↵\\n │ <email>mycroft@sh.com</email> ↵\\n │ </Employee>\\n\\n-- \\nÁlvaro Herrera https://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\": \"Fri, 3 Mar 2017 17:04:28 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-03-03 21:04 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n> > 2017-03-03 19:15 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>\\n> > > 2. As I've complained many times, I find the way we manage an empty\\n> > > COLUMNS clause pretty bad. The standard doesn't require that syntax\\n> > > (COLUMNS is required), and I don't like the implementation, so why not\\n> > > provide the feature in a different way? My proposal is to change the\\n> > > column options in gram.y to be something like this:\\n> >\\n> > The clause COLUMNS is optional on Oracle and DB2\\n> >\\n> > So I prefer a Oracle, DB2 design. If you are strongly against it, then we\\n> > can remove it to be ANSI/SQL only.\\n> >\\n> > I am don't see an good idea to introduce third syntax.\\n>\\n> OK. I think trying to be syntax compatible with DB2 or Oracle is a lost\\n> cause, because the syntax used in the XPath expressions seems different\\n> -- I think Oracle uses XQuery (which we don't support) and DB2 uses ...\\n> not sure what it is, but it doesn't work in our implementation\\n> (stuff like '$d/employees/emp' in the row expression.)\\n>\\n\\n100% compatibility is not possible - but XPath is subset of XQuery and in\\nreality - the full XQuery examples of XMLTABLE is not often.\\n\\nAlmost all examples of usage XMLTABLE, what I found in blogs, uses XPath\\nonly\\n\\n\\n>\\n> In existing applications using those Oracle/DB2, is it common to omit\\n> the COLUMNS clause? I searched for \\\"xmltable oracle\\\" and had a look at\\n> the first few hits outside of the oracle docs:\\n> http://viralpatel.net/blogs/oracle-xmltable-tutorial/\\n> http://www.dba-oracle.com/t_xmltable.htm\\n> http://stackoverflow.com/questions/12690868/how-to-use-xmltable-in-oracle\\n> https://asktom.oracle.com/pls/asktom/f?p=100:11:0::::P11_QUESTION_ID:\\n> 9533111800346252295\\n> http://stackoverflow.com/questions/1222570/what-is-an-xmltable\\n> https://community.oracle.com/thread/3955198\\n>\\n> Not a single one of these omit the COLUMNS clause (though the second one\\n> mentions that the clause can be omitted).\\n>\\n> I also looked at a few samples with DB2 -- same thing; it is possible,\\n> but is it common?\\n>\\n\\nI don't think so it is common - it is corner case - and I can live without\\nit well\\n\\n\\n>\\n> Anyway, I noticed that \\\"xml PATH '.'\\\" can be used to obtain the full XML\\n> of the row, which I think is the feature I wanted, so I think we're\\n> covered and we can omit the case with no COLUMNS, since we already have\\n> the feature in another way. No need to implement anything further, and\\n> we can rip out the special case I don't like. Example:\\n>\\n\\nyes,\\n\\n\\n>\\n> CREATE TABLE EMPLOYEES\\n> (\\n> id integer,\\n> data XML\\n> );\\n> INSERT INTO EMPLOYEES\\n> VALUES (1, '<Employees>\\n> <Employee emplid=\\\"1111\\\" type=\\\"admin\\\">\\n> <firstname>John</firstname>\\n> <lastname>Watson</lastname>\\n> <age>30</age>\\n> <email>johnwatson@sh.com</email>\\n> </Employee>\\n> <Employee emplid=\\\"2222\\\" type=\\\"admin\\\">\\n> <firstname>Sherlock</firstname>\\n> <lastname>Homes</lastname>\\n> <age>32</age>\\n> <email>sherlock@sh.com</email>\\n> </Employee>\\n> <Employee emplid=\\\"3333\\\" type=\\\"user\\\">\\n> <firstname>Jim</firstname>\\n> <lastname>Moriarty</lastname>\\n> <age>52</age>\\n> <email>jim@sh.com</email>\\n> </Employee>\\n> <Employee emplid=\\\"4444\\\" type=\\\"user\\\">\\n> <firstname>Mycroft</firstname>\\n> <lastname>Holmes</lastname>\\n> <age>41</age>\\n> <email>mycroft@sh.com</email>\\n> </Employee>\\n> </Employees>');\\n>\\n> This is with COLUMNS omitted:\\n>\\n> alvherre=# select xmltable.* from employees,\\n> xmltable('/Employees/Employee' passing data);\\n> xmltable\\n> ──────────────────────────────────────────\\n> <Employee emplid=\\\"1111\\\" type=\\\"admin\\\"> ↵\\n> <firstname>John</firstname> ↵\\n> <lastname>Watson</lastname> ↵\\n> <age>30</age> ↵\\n> <email>johnwatson@sh.com</email>↵\\n> </Employee>\\n> <Employee emplid=\\\"2222\\\" type=\\\"admin\\\"> ↵\\n> <firstname>Sherlock</firstname> ↵\\n> <lastname>Homes</lastname> ↵\\n> <age>32</age> ↵\\n> <email>sherlock@sh.com</email> ↵\\n> </Employee>\\n> <Employee emplid=\\\"3333\\\" type=\\\"user\\\"> ↵\\n> <firstname>Jim</firstname> ↵\\n> <lastname>Moriarty</lastname> ↵\\n> <age>52</age> ↵\\n> <email>jim@sh.com</email> ↵\\n> </Employee>\\n> <Employee emplid=\\\"4444\\\" type=\\\"user\\\"> ↵\\n> <firstname>Mycroft</firstname> ↵\\n> <lastname>Holmes</lastname> ↵\\n> <age>41</age> ↵\\n> <email>mycroft@sh.com</email> ↵\\n> </Employee>\\n>\\n> and this is what you get with \\\"xml PATH '.'\\\" (I threw in ORDINALITY just\\n> for fun):\\n>\\n> alvherre=# select xmltable.* from employees,\\n> xmltable('/Employees/Employee' passing data columns row_number for\\n> ordinality, emp xml path '.');\\n> row_number │ emp\\n> ────────────┼──────────────────────────────────────────\\n> 1 │ <Employee emplid=\\\"1111\\\" type=\\\"admin\\\"> ↵\\n> │ <firstname>John</firstname> ↵\\n> │ <lastname>Watson</lastname> ↵\\n> │ <age>30</age> ↵\\n> │ <email>johnwatson@sh.com</email>↵\\n> │ </Employee>\\n> 2 │ <Employee emplid=\\\"2222\\\" type=\\\"admin\\\"> ↵\\n> │ <firstname>Sherlock</firstname> ↵\\n> │ <lastname>Homes</lastname> ↵\\n> │ <age>32</age> ↵\\n> │ <email>sherlock@sh.com</email> ↵\\n> │ </Employee>\\n> 3 │ <Employee emplid=\\\"3333\\\" type=\\\"user\\\"> ↵\\n> │ <firstname>Jim</firstname> ↵\\n> │ <lastname>Moriarty</lastname> ↵\\n> │ <age>52</age> ↵\\n> │ <email>jim@sh.com</email> ↵\\n> │ </Employee>\\n> 4 │ <Employee emplid=\\\"4444\\\" type=\\\"user\\\"> ↵\\n> │ <firstname>Mycroft</firstname> ↵\\n> │ <lastname>Holmes</lastname> ↵\\n> │ <age>41</age> ↵\\n> │ <email>mycroft@sh.com</email> ↵\\n> │ </Employee>\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2017-03-03 21:04 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\\n> 2017-03-03 19:15 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> > 2. As I've complained many times, I find the way we manage an empty\\n> > COLUMNS clause pretty bad. The standard doesn't require that syntax\\n> > (COLUMNS is required), and I don't like the implementation, so why not\\n> > provide the feature in a different way? My proposal is to change the\\n> > column options in gram.y to be something like this:\\n>\\n> The clause COLUMNS is optional on Oracle and DB2\\n>\\n> So I prefer a Oracle, DB2 design. If you are strongly against it, then we\\n> can remove it to be ANSI/SQL only.\\n>\\n> I am don't see an good idea to introduce third syntax.\\n\\nOK. I think trying to be syntax compatible with DB2 or Oracle is a lost\\ncause, because the syntax used in the XPath expressions seems different\\n-- I think Oracle uses XQuery (which we don't support) and DB2 uses ...\\nnot sure what it is, but it doesn't work in our implementation\\n(stuff like '$d/employees/emp' in the row expression.)100% compatibility is not possible - but XPath is subset of XQuery and in reality - the full XQuery examples of XMLTABLE is not often.Almost all examples of usage XMLTABLE, what I found in blogs, uses XPath only \\n\\nIn existing applications using those Oracle/DB2, is it common to omit\\nthe COLUMNS clause? I searched for \\\"xmltable oracle\\\" and had a look at\\nthe first few hits outside of the oracle docs:\\nhttp://viralpatel.net/blogs/oracle-xmltable-tutorial/\\nhttp://www.dba-oracle.com/t_xmltable.htm\\nhttp://stackoverflow.com/questions/12690868/how-to-use-xmltable-in-oracle\\nhttps://asktom.oracle.com/pls/asktom/f?p=100:11:0::::P11_QUESTION_ID:9533111800346252295\\nhttp://stackoverflow.com/questions/1222570/what-is-an-xmltable\\nhttps://community.oracle.com/thread/3955198\\n\\nNot a single one of these omit the COLUMNS clause (though the second one\\nmentions that the clause can be omitted).\\n\\nI also looked at a few samples with DB2 -- same thing; it is possible,\\nbut is it common?I don't think so it is common - it is corner case - and I can live without it well \\n\\nAnyway, I noticed that \\\"xml PATH '.'\\\" can be used to obtain the full XML\\nof the row, which I think is the feature I wanted, so I think we're\\ncovered and we can omit the case with no COLUMNS, since we already have\\nthe feature in another way. No need to implement anything further, and\\nwe can rip out the special case I don't like. Example:yes, \\n\\nCREATE TABLE EMPLOYEES\\n(\\n id integer,\\n data XML\\n);\\nINSERT INTO EMPLOYEES\\n VALUES (1, '<Employees>\\n <Employee emplid=\\\"1111\\\" type=\\\"admin\\\">\\n <firstname>John</firstname>\\n <lastname>Watson</lastname>\\n <age>30</age>\\n <email>johnwatson@sh.com</email>\\n </Employee>\\n <Employee emplid=\\\"2222\\\" type=\\\"admin\\\">\\n <firstname>Sherlock</firstname>\\n <lastname>Homes</lastname>\\n <age>32</age>\\n <email>sherlock@sh.com</email>\\n </Employee>\\n <Employee emplid=\\\"3333\\\" type=\\\"user\\\">\\n <firstname>Jim</firstname>\\n <lastname>Moriarty</lastname>\\n <age>52</age>\\n <email>jim@sh.com</email>\\n </Employee>\\n <Employee emplid=\\\"4444\\\" type=\\\"user\\\">\\n <firstname>Mycroft</firstname>\\n <lastname>Holmes</lastname>\\n <age>41</age>\\n <email>mycroft@sh.com</email>\\n </Employee>\\n</Employees>');\\n\\nThis is with COLUMNS omitted:\\n\\nalvherre=# select xmltable.* from employees, xmltable('/Employees/Employee' passing data);\\n xmltable\\n──────────────────────────────────────────\\n <Employee emplid=\\\"1111\\\" type=\\\"admin\\\"> ↵\\n <firstname>John</firstname> ↵\\n <lastname>Watson</lastname> ↵\\n <age>30</age> ↵\\n <email>johnwatson@sh.com</email>↵\\n </Employee>\\n <Employee emplid=\\\"2222\\\" type=\\\"admin\\\"> ↵\\n <firstname>Sherlock</firstname> ↵\\n <lastname>Homes</lastname> ↵\\n <age>32</age> ↵\\n <email>sherlock@sh.com</email> ↵\\n </Employee>\\n <Employee emplid=\\\"3333\\\" type=\\\"user\\\"> ↵\\n <firstname>Jim</firstname> ↵\\n <lastname>Moriarty</lastname> ↵\\n <age>52</age> ↵\\n <email>jim@sh.com</email> ↵\\n </Employee>\\n <Employee emplid=\\\"4444\\\" type=\\\"user\\\"> ↵\\n <firstname>Mycroft</firstname> ↵\\n <lastname>Holmes</lastname> ↵\\n <age>41</age> ↵\\n <email>mycroft@sh.com</email> ↵\\n </Employee>\\n\\nand this is what you get with \\\"xml PATH '.'\\\" (I threw in ORDINALITY just\\nfor fun):\\n\\nalvherre=# select xmltable.* from employees, xmltable('/Employees/Employee' passing data columns row_number for ordinality, emp xml path '.');\\n row_number │ emp\\n────────────┼──────────────────────────────────────────\\n 1 │ <Employee emplid=\\\"1111\\\" type=\\\"admin\\\"> ↵\\n │ <firstname>John</firstname> ↵\\n │ <lastname>Watson</lastname> ↵\\n │ <age>30</age> ↵\\n │ <email>johnwatson@sh.com</email>↵\\n │ </Employee>\\n 2 │ <Employee emplid=\\\"2222\\\" type=\\\"admin\\\"> ↵\\n │ <firstname>Sherlock</firstname> ↵\\n │ <lastname>Homes</lastname> ↵\\n │ <age>32</age> ↵\\n │ <email>sherlock@sh.com</email> ↵\\n │ </Employee>\\n 3 │ <Employee emplid=\\\"3333\\\" type=\\\"user\\\"> ↵\\n │ <firstname>Jim</firstname> ↵\\n │ <lastname>Moriarty</lastname> ↵\\n │ <age>52</age> ↵\\n │ <email>jim@sh.com</email> ↵\\n │ </Employee>\\n 4 │ <Employee emplid=\\\"4444\\\" type=\\\"user\\\"> ↵\\n │ <firstname>Mycroft</firstname> ↵\\n │ <lastname>Holmes</lastname> ↵\\n │ <age>41</age> ↵\\n │ <email>mycroft@sh.com</email> ↵\\n │ </Employee>\\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Fri, 3 Mar 2017 22:41:57 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Hi\\n\\nI used your idea about special columns when COLUMNS are not explicitly\\ndefined.\\n\\nAll lines that you are dislike removed.\\n\\nNow, almost all code, related to this behave, is in next few lines.\\n\\n+ /*\\n+ * Use implicit column when it is necessary. The COLUMNS clause is\\noptional\\n+ * on Oracle and DB2. In this case a result is complete row of XML type.\\n+ */\\n+ if (rtf->columns == NIL)\\n+ {\\n+ RangeTableFuncCol *fc = makeNode(RangeTableFuncCol);\\n+ A_Const *n = makeNode(A_Const);\\n+\\n+ fc->colname = \\\"xmltable\\\";\\n+ fc->typeName = makeTypeNameFromOid(XMLOID, -1);\\n+ n->val.type = T_String;\\n+ n->val.val.str = \\\".\\\";\\n+ n->location = -1;\\n+\\n+ fc->colexpr = (Node *) n;\\n+ rtf->columns = list_make1(fc);\\n+ }\\n\\nall regress tests passing.\\n\\nRegards\\n\\nPavel\\n\\n\\n-- \\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 Mar 2017 10:32:18 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n> Hi\\n> \\n> I used your idea about special columns when COLUMNS are not explicitly\\n> defined.\\n> \\n> All lines that you are dislike removed.\\n\\nI just pushed XMLTABLE, after some additional changes. Please test it\\nthoroughly and report any problems.\\n\\nI didn't add the change you proposed here to keep COLUMNS optional;\\ninstead, I just made COLUMNS mandatory. I think what you propose here\\nis not entirely out of the question, but you left out ruleutils.c\\nsupport for it, so I decided to leave it aside for now so that I could\\nget this patch out of my plate once and for all. If you really want\\nthat feature, you can submit another patch for it and discuss with the\\nRMT whether it belongs in PG10 or not.\\n\\nSome changes I made:\\n* I added some pg_stat_statements support. It works fine for simple\\ntests, but deeper testing of it would be appreciated.\\n\\n* I removed the \\\"buildercxt\\\" memory context. It seemed mostly\\npointless, and I was disturbed by the MemoryContextResetOnly().\\nPer-value memory still uses the per-value memory context, but the rest\\nof the stuff is in the per-query context, which should be pretty much\\nthe same.\\n\\n* Desultory stylistic changes\\n\\n-- \\n�lvaro Herrera https://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, 8 Mar 2017 13:01:01 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-03-08 17:01 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n> > Hi\\n> >\\n> > I used your idea about special columns when COLUMNS are not explicitly\\n> > defined.\\n> >\\n> > All lines that you are dislike removed.\\n>\\n> I just pushed XMLTABLE, after some additional changes. Please test it\\n> thoroughly and report any problems.\\n>\\n\\nThank you\\n\\n>\\n> I didn't add the change you proposed here to keep COLUMNS optional;\\n> instead, I just made COLUMNS mandatory. I think what you propose here\\n> is not entirely out of the question, but you left out ruleutils.c\\n> support for it, so I decided to leave it aside for now so that I could\\n> get this patch out of my plate once and for all. If you really want\\n> that feature, you can submit another patch for it and discuss with the\\n> RMT whether it belongs in PG10 or not.\\n>\\n\\nIt is interesting feature - because it replaces XPATH function, but not\\nimportant enough.\\n\\nFor daily work the default schema support is much more interesting.\\n\\n\\n>\\n> Some changes I made:\\n> * I added some pg_stat_statements support. It works fine for simple\\n> tests, but deeper testing of it would be appreciated.\\n>\\n> * I removed the \\\"buildercxt\\\" memory context. It seemed mostly\\n> pointless, and I was disturbed by the MemoryContextResetOnly().\\n> Per-value memory still uses the per-value memory context, but the rest\\n> of the stuff is in the per-query context, which should be pretty much\\n> the same.\\n>\\n> * Desultory stylistic changes\\n>\\n\\nok\\n\\nRegards\\n\\nPavel\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2017-03-08 17:01 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\\n> Hi\\n>\\n> I used your idea about special columns when COLUMNS are not explicitly\\n> defined.\\n>\\n> All lines that you are dislike removed.\\n\\nI just pushed XMLTABLE, after some additional changes. Please test it\\nthoroughly and report any problems.Thank you \\n\\nI didn't add the change you proposed here to keep COLUMNS optional;\\ninstead, I just made COLUMNS mandatory. I think what you propose here\\nis not entirely out of the question, but you left out ruleutils.c\\nsupport for it, so I decided to leave it aside for now so that I could\\nget this patch out of my plate once and for all. If you really want\\nthat feature, you can submit another patch for it and discuss with the\\nRMT whether it belongs in PG10 or not.It is interesting feature - because it replaces XPATH function, but not important enough.For daily work the default schema support is much more interesting. \\n\\nSome changes I made:\\n* I added some pg_stat_statements support. It works fine for simple\\ntests, but deeper testing of it would be appreciated.\\n\\n* I removed the \\\"buildercxt\\\" memory context. It seemed mostly\\npointless, and I was disturbed by the MemoryContextResetOnly().\\nPer-value memory still uses the per-value memory context, but the rest\\nof the stuff is in the per-query context, which should be pretty much\\nthe same.\\n\\n* Desultory stylistic changesokRegardsPavel \\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Wed, 8 Mar 2017 17:10:53 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n> 2017-03-08 17:01 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> > I didn't add the change you proposed here to keep COLUMNS optional;\\n> > instead, I just made COLUMNS mandatory. I think what you propose here\\n> > is not entirely out of the question, but you left out ruleutils.c\\n> > support for it, so I decided to leave it aside for now so that I could\\n> > get this patch out of my plate once and for all. If you really want\\n> > that feature, you can submit another patch for it and discuss with the\\n> > RMT whether it belongs in PG10 or not.\\n> \\n> It is interesting feature - because it replaces XPATH function, but not\\n> important enough.\\n\\nOK.\\n\\n> For daily work the default schema support is much more interesting.\\n\\nLet's see that one, then. It was part of the original submission so\\ndepending on how the patch we looks can still cram it in. But other\\npatches have priority for me now.\\n\\n-- \\n�lvaro Herrera https://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, 8 Mar 2017 13:32:19 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"2017-03-08 17:32 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> Pavel Stehule wrote:\\n> > 2017-03-08 17:01 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n>\\n> > > I didn't add the change you proposed here to keep COLUMNS optional;\\n> > > instead, I just made COLUMNS mandatory. I think what you propose here\\n> > > is not entirely out of the question, but you left out ruleutils.c\\n> > > support for it, so I decided to leave it aside for now so that I could\\n> > > get this patch out of my plate once and for all. If you really want\\n> > > that feature, you can submit another patch for it and discuss with the\\n> > > RMT whether it belongs in PG10 or not.\\n> >\\n> > It is interesting feature - because it replaces XPATH function, but not\\n> > important enough.\\n>\\n> OK.\\n>\\n> > For daily work the default schema support is much more interesting.\\n>\\n> Let's see that one, then. It was part of the original submission so\\n> depending on how the patch we looks can still cram it in. But other\\n> patches have priority for me now.\\n>\\n\\nIt is theme for 11\\n\\nThank you very much\\n\\nPavel\\n\\n\\n>\\n> --\\n> Álvaro Herrera https://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\n2017-03-08 17:32 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:Pavel Stehule wrote:\\n> 2017-03-08 17:01 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> > I didn't add the change you proposed here to keep COLUMNS optional;\\n> > instead, I just made COLUMNS mandatory. I think what you propose here\\n> > is not entirely out of the question, but you left out ruleutils.c\\n> > support for it, so I decided to leave it aside for now so that I could\\n> > get this patch out of my plate once and for all. If you really want\\n> > that feature, you can submit another patch for it and discuss with the\\n> > RMT whether it belongs in PG10 or not.\\n>\\n> It is interesting feature - because it replaces XPATH function, but not\\n> important enough.\\n\\nOK.\\n\\n> For daily work the default schema support is much more interesting.\\n\\nLet's see that one, then. It was part of the original submission so\\ndepending on how the patch we looks can still cram it in. But other\\npatches have priority for me now.It is theme for 11Thank you very muchPavel \\n\\n--\\nÁlvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Wed, 8 Mar 2017 17:46:48 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: patch: function xmltable\"\n },\n {\n \"msg_contents\": \"Pavel Stehule wrote:\\n> 2017-03-08 17:32 GMT+01:00 Alvaro Herrera <alvherre@2ndquadrant.com>:\\n\\n> > > For daily work the default schema support is much more interesting.\\n> >\\n> > Let's see that one, then. It was part of the original submission so\\n> > depending on how the patch we looks can still cram it in. But other\\n> > patches have priority for me now.\\n> \\n> It is theme for 11\\n\\nAh, great.\\n\\n-- \\n�lvaro Herrera https://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, 8 Mar 2017 14:01:37 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: patch: function xmltable\"\n }\n]","isConversational":false}}},{"rowIdx":42986,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nThe postgres_fdw join pushdown in 9.6 is great, but it can't handle full \njoins on relations with restrictions. The reason for that is, \npostgres_fdw can't support deparsing subqueries when creating a remote \njoin query. So, by adding the deparsing logic to it, I removed that \nlimitation. Attached is a patch for that, which is based on the patch I \nposted before [1].\n\nAlso, by the deparsing logic, I improved the handling of PHVs so that \nPHVs are evaluated remotely if it's safe, as discussed in [2]. Here is \nan example from the regression test, which pushes down multiple levels \nof PHVs to the remote:\n\nEXPLAIN (VERBOSE, COSTS OFF)\nSELECT ss.*, ft2.c1 FROM ft2 LEFT JOIN (SELECT 13, q.a, ft2.c1 FROM \n(SELECT 13 FROM ft1 WHERE c1 = 13) q(a) RIGHT JOIN ft2 ON (q.a = ft2.c1) \nWHERE ft2.\\\nc1 BETWEEN 10 AND 15) ss(f1, f2, f3) ON (ft2.c1 = ss.f1) WHERE ft2.c1 \nBETWEEN 10 AND 15;\n \\\n QUERY PLAN \\\n\n-------------------------------------------------------------------------------------------------------------------------------------------------------\\\n-------------------------------------------------------------------------------------------------------------------------------------------------------\\\n---------------------------------------------------------------\n Foreign Scan\n Output: (13), (13), ft2_1.c1, ft2.c1\n Relations: (public.ft2) LEFT JOIN ((public.ft2) LEFT JOIN (public.ft1))\n Remote SQL: SELECT r1.\"C 1\", ss2.c1, ss2.c2, ss2.c3 FROM (\"S 1\".\"T \n1\" r1 LEFT JOIN (SELECT r5.\"C 1\", 13, ss1.c1 FROM (\"S 1\".\"T 1\" r5 LEFT \nJOIN (SELE\\\nCT 13 FROM \"S 1\".\"T 1\" WHERE ((\"C 1\" = 13))) ss1(c1) ON (((13 = r5.\"C \n1\")))) WHERE ((r5.\"C 1\" >= 10)) AND ((r5.\"C 1\" <= 15))) ss2(c1, c2, c3) \nON (((r1.\\\n\"C 1\" = 13)))) WHERE ((r1.\"C 1\" >= 10)) AND ((r1.\"C 1\" <= 15))\n(4 rows)\n\nAlso, in the same way as the PHV handling, I improved the handling of \nwhole-row reference (and system columns other than ctid), as proposed in \n[3]. We don't need the \"\"CASE WHEN ... IS NOT NULL THEN ROW(...) END\" \nconditions any more, as shown in the below example:\n\nEXPLAIN (VERBOSE, COSTS OFF)\nSELECT t1.ctid, t1, t2, t1.c1 FROM ft1 t1 JOIN ft2 t2 ON (t1.c1 = t2.c1) \nORDER BY t1.c3, t1.c1 OFFSET 100 LIMIT 10;\n \\\n QUERY PLAN \\\n\n-------------------------------------------------------------------------------------------------------------------------------------------------------\\\n-------------------------------------------------------------------------------------------------------------------------------------------------------\\\n-----------------------------------------------------------\n Limit\n Output: t1.ctid, t1.*, t2.*, t1.c1, t1.c3\n -> Foreign Scan\n Output: t1.ctid, t1.*, t2.*, t1.c1, t1.c3\n Relations: (public.ft1 t1) INNER JOIN (public.ft2 t2)\n Remote SQL: SELECT ss1.c1, ss1.c2, ss1.c3, ss1.c4, ss2.c1 \nFROM ((SELECT ctid, ROW(\"C 1\", c2, c3, c4, c5, c6, c7, c8), \"C 1\", c3 \nFROM \"S 1\".\"T \\\n1\") ss1(c1, c2, c3, c4) INNER JOIN (SELECT ROW(\"C 1\", c2, c3, c4, c5, \nc6, c7, c8), \"C 1\" FROM \"S 1\".\"T 1\") ss2(c1, c2) ON (TRUE)) WHERE \n((ss1.c3 = ss2.\\\nc2)) ORDER BY ss1.c4 ASC NULLS LAST, ss1.c3 ASC NULLS LAST\n(6 rows)\n\nI'll add this to the next CF. Comments are welcome!\n\nBest regards,\nEtsuro Fujita\n\n[1] https://www.postgresql.org/message-id/5710D7E2.7010302%40lab.ntt.co.jp\n[2] \nhttps://www.postgresql.org/message-id/b4549406-909f-7d15-dc34-499835a8f0b3%40lab.ntt.co.jp\n[3] \nhttps://www.postgresql.org/message-id/a1fa1c4c-bf96-8ea5-cff5-85b927298e73%40lab.ntt.co.jp\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, 19 Aug 2016 18:20:32 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Push down more full joins in postgres_fdw"},{"msg_contents":"Thanks Fujita-san for working on this. I took a quick look at the patch.\nHere are some quick comments.\n\n1. deparsePlaceHolderVar looks odd - each of the deparse* function is named\nas deparse + <name of the parser node the string would parse into>.\nPlaceHolderVar is not a parser node, so no string is going to be parsed as\na PlaceHolderVar. May be you want to name it as\ndeparseExerFromPlaceholderVar or something like that.\n\n2. You will need to check phlevelsup member while assessing whether a PHV\nis safe to push down.\n\n3. I think registerAlias stuff should happen really at the time of creating\npaths and should be stored in fpinfo. Without that it will be computed\nevery time we deparse the query. This means every time we try to EXPLAIN\nthe query at various levels in the join tree and once for the query to be\nsent to the foreign server.\n\n4. The changes related to retrieved_attrs look unrelated to the patch.\nEither your patch should use the current method of handling retrieved_attrs\nor there should be a separate patch for retrieved_attrs changes. May be you\nwant to take a look at the discussion in join pushdown thread as to why we\nassume retrieved_attrs to be non-NIL always.\n\n5. The blocks related to inner and outer relations in\ndeparseFromExprForRel() look same. We should probably separate that code\nout into a function and call it at two places.\n\n6.\n! if (is_placeholder)\n! errcontext(\"placeholder expression at position %d in select list\",\n! errpos->cur_attno);\nA user wouldn't know what a placeholder expression is, there is no such\nterm in the documentation. We have to device a better way to provide an\nerror context for an expression in general.\n\n\n\n\nOn Fri, Aug 19, 2016 at 2:50 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\nwrote:\n\n> Hi,\n>\n> The postgres_fdw join pushdown in 9.6 is great, but it can't handle full\n> joins on relations with restrictions. The reason for that is, postgres_fdw\n> can't support deparsing subqueries when creating a remote join query. So,\n> by adding the deparsing logic to it, I removed that limitation. Attached\n> is a patch for that, which is based on the patch I posted before [1].\n>\n> Also, by the deparsing logic, I improved the handling of PHVs so that PHVs\n> are evaluated remotely if it's safe, as discussed in [2]. Here is an\n> example from the regression test, which pushes down multiple levels of PHVs\n> to the remote:\n>\n> EXPLAIN (VERBOSE, COSTS OFF)\n> SELECT ss.*, ft2.c1 FROM ft2 LEFT JOIN (SELECT 13, q.a, ft2.c1 FROM\n> (SELECT 13 FROM ft1 WHERE c1 = 13) q(a) RIGHT JOIN ft2 ON (q.a = ft2.c1)\n> WHERE ft2.\\\n> c1 BETWEEN 10 AND 15) ss(f1, f2, f3) ON (ft2.c1 = ss.f1) WHERE ft2.c1\n> BETWEEN 10 AND 15;\n> \\\n> QUERY PLAN \\\n>\n> ------------------------------------------------------------\n> ------------------------------------------------------------\n> -------------------------------\\\n> ------------------------------------------------------------\n> ------------------------------------------------------------\n> -------------------------------\\\n> ---------------------------------------------------------------\n> Foreign Scan\n> Output: (13), (13), ft2_1.c1, ft2.c1\n> Relations: (public.ft2) LEFT JOIN ((public.ft2) LEFT JOIN (public.ft1))\n> Remote SQL: SELECT r1.\"C 1\", ss2.c1, ss2.c2, ss2.c3 FROM (\"S 1\".\"T 1\"\n> r1 LEFT JOIN (SELECT r5.\"C 1\", 13, ss1.c1 FROM (\"S 1\".\"T 1\" r5 LEFT JOIN\n> (SELE\\\n> CT 13 FROM \"S 1\".\"T 1\" WHERE ((\"C 1\" = 13))) ss1(c1) ON (((13 = r5.\"C\n> 1\")))) WHERE ((r5.\"C 1\" >= 10)) AND ((r5.\"C 1\" <= 15))) ss2(c1, c2, c3) ON\n> (((r1.\\\n> \"C 1\" = 13)))) WHERE ((r1.\"C 1\" >= 10)) AND ((r1.\"C 1\" <= 15))\n> (4 rows)\n>\n> Also, in the same way as the PHV handling, I improved the handling of\n> whole-row reference (and system columns other than ctid), as proposed in\n> [3]. We don't need the \"\"CASE WHEN ... IS NOT NULL THEN ROW(...) END\"\n> conditions any more, as shown in the below example:\n>\n> EXPLAIN (VERBOSE, COSTS OFF)\n> SELECT t1.ctid, t1, t2, t1.c1 FROM ft1 t1 JOIN ft2 t2 ON (t1.c1 = t2.c1)\n> ORDER BY t1.c3, t1.c1 OFFSET 100 LIMIT 10;\n> \\\n> QUERY PLAN \\\n>\n> ------------------------------------------------------------\n> ------------------------------------------------------------\n> -------------------------------\\\n> ------------------------------------------------------------\n> ------------------------------------------------------------\n> -------------------------------\\\n> -----------------------------------------------------------\n> Limit\n> Output: t1.ctid, t1.*, t2.*, t1.c1, t1.c3\n> -> Foreign Scan\n> Output: t1.ctid, t1.*, t2.*, t1.c1, t1.c3\n> Relations: (public.ft1 t1) INNER JOIN (public.ft2 t2)\n> Remote SQL: SELECT ss1.c1, ss1.c2, ss1.c3, ss1.c4, ss2.c1 FROM\n> ((SELECT ctid, ROW(\"C 1\", c2, c3, c4, c5, c6, c7, c8), \"C 1\", c3 FROM \"S\n> 1\".\"T \\\n> 1\") ss1(c1, c2, c3, c4) INNER JOIN (SELECT ROW(\"C 1\", c2, c3, c4, c5, c6,\n> c7, c8), \"C 1\" FROM \"S 1\".\"T 1\") ss2(c1, c2) ON (TRUE)) WHERE ((ss1.c3 =\n> ss2.\\\n> c2)) ORDER BY ss1.c4 ASC NULLS LAST, ss1.c3 ASC NULLS LAST\n> (6 rows)\n>\n> I'll add this to the next CF. Comments are welcome!\n>\n> Best regards,\n> Etsuro Fujita\n>\n> [1] https://www.postgresql.org/message-id/5710D7E2.7010302%40lab.ntt.co.jp\n> [2] https://www.postgresql.org/message-id/b4549406-909f-7d15-dc3\n> 4-499835a8f0b3%40lab.ntt.co.jp\n> [3] https://www.postgresql.org/message-id/a1fa1c4c-bf96-8ea5-cff\n> 5-85b927298e73%40lab.ntt.co.jp\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>\n\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database Company\n\nThanks Fujita-san for working on this. I took a quick look at the patch. Here are some quick comments.1. deparsePlaceHolderVar looks odd - each of the deparse* function is named as deparse + <name of the parser node the string would parse into>. PlaceHolderVar is not a parser node, so no string is going to be parsed as a PlaceHolderVar. May be you want to name it as deparseExerFromPlaceholderVar or something like that.2. You will need to check phlevelsup member while assessing whether a PHV is safe to push down.3. I think registerAlias stuff should happen really at the time of creating paths and should be stored in fpinfo. Without that it will be computed every time we deparse the query. This means every time we try to EXPLAIN the query at various levels in the join tree and once for the query to be sent to the foreign server.4. The changes related to retrieved_attrs look unrelated to the patch. Either your patch should use the current method of handling retrieved_attrs or there should be a separate patch for retrieved_attrs changes. May be you want to take a look at the discussion in join pushdown thread as to why we assume retrieved_attrs to be non-NIL always.5. The blocks related to inner and outer relations in deparseFromExprForRel() look same. We should probably separate that code out into a function and call it at two places.6. ! if (is_placeholder)! errcontext(\"placeholder expression at position %d in select list\",! errpos->cur_attno);A user wouldn't know what a placeholder expression is, there is no such term in the documentation. We have to device a better way to provide an error context for an expression in general.On Fri, Aug 19, 2016 at 2:50 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp> wrote:Hi,\n\nThe postgres_fdw join pushdown in 9.6 is great, but it can't handle full joins on relations with restrictions. The reason for that is, postgres_fdw can't support deparsing subqueries when creating a remote join query. So, by adding the deparsing logic to it, I removed that limitation. Attached is a patch for that, which is based on the patch I posted before [1].\n\nAlso, by the deparsing logic, I improved the handling of PHVs so that PHVs are evaluated remotely if it's safe, as discussed in [2]. Here is an example from the regression test, which pushes down multiple levels of PHVs to the remote:\n\nEXPLAIN (VERBOSE, COSTS OFF)\nSELECT ss.*, ft2.c1 FROM ft2 LEFT JOIN (SELECT 13, q.a, ft2.c1 FROM (SELECT 13 FROM ft1 WHERE c1 = 13) q(a) RIGHT JOIN ft2 ON (q.a = ft2.c1) WHERE ft2.\\\nc1 BETWEEN 10 AND 15) ss(f1, f2, f3) ON (ft2.c1 = ss.f1) WHERE ft2.c1 BETWEEN 10 AND 15;\n \\\n QUERY PLAN \\\n\n-------------------------------------------------------------------------------------------------------------------------------------------------------\\\n-------------------------------------------------------------------------------------------------------------------------------------------------------\\\n---------------------------------------------------------------\n Foreign Scan\n Output: (13), (13), ft2_1.c1, ft2.c1\n Relations: (public.ft2) LEFT JOIN ((public.ft2) LEFT JOIN (public.ft1))\n Remote SQL: SELECT r1.\"C 1\", ss2.c1, ss2.c2, ss2.c3 FROM (\"S 1\".\"T 1\" r1 LEFT JOIN (SELECT r5.\"C 1\", 13, ss1.c1 FROM (\"S 1\".\"T 1\" r5 LEFT JOIN (SELE\\\nCT 13 FROM \"S 1\".\"T 1\" WHERE ((\"C 1\" = 13))) ss1(c1) ON (((13 = r5.\"C 1\")))) WHERE ((r5.\"C 1\" >= 10)) AND ((r5.\"C 1\" <= 15))) ss2(c1, c2, c3) ON (((r1.\\\n\"C 1\" = 13)))) WHERE ((r1.\"C 1\" >= 10)) AND ((r1.\"C 1\" <= 15))\n(4 rows)\n\nAlso, in the same way as the PHV handling, I improved the handling of whole-row reference (and system columns other than ctid), as proposed in [3]. We don't need the \"\"CASE WHEN ... IS NOT NULL THEN ROW(...) END\" conditions any more, as shown in the below example:\n\nEXPLAIN (VERBOSE, COSTS OFF)\nSELECT t1.ctid, t1, t2, t1.c1 FROM ft1 t1 JOIN ft2 t2 ON (t1.c1 = t2.c1) ORDER BY t1.c3, t1.c1 OFFSET 100 LIMIT 10;\n \\\n QUERY PLAN \\\n\n-------------------------------------------------------------------------------------------------------------------------------------------------------\\\n-------------------------------------------------------------------------------------------------------------------------------------------------------\\\n-----------------------------------------------------------\n Limit\n Output: t1.ctid, t1.*, t2.*, t1.c1, t1.c3\n -> Foreign Scan\n Output: t1.ctid, t1.*, t2.*, t1.c1, t1.c3\n Relations: (public.ft1 t1) INNER JOIN (public.ft2 t2)\n Remote SQL: SELECT ss1.c1, ss1.c2, ss1.c3, ss1.c4, ss2.c1 FROM ((SELECT ctid, ROW(\"C 1\", c2, c3, c4, c5, c6, c7, c8), \"C 1\", c3 FROM \"S 1\".\"T \\\n1\") ss1(c1, c2, c3, c4) INNER JOIN (SELECT ROW(\"C 1\", c2, c3, c4, c5, c6, c7, c8), \"C 1\" FROM \"S 1\".\"T 1\") ss2(c1, c2) ON (TRUE)) WHERE ((ss1.c3 = ss2.\\\nc2)) ORDER BY ss1.c4 ASC NULLS LAST, ss1.c3 ASC NULLS LAST\n(6 rows)\n\nI'll add this to the next CF. Comments are welcome!\n\nBest regards,\nEtsuro Fujita\n\n[1] https://www.postgresql.org/message-id/5710D7E2.7010302%40lab.ntt.co.jp\n[2] https://www.postgresql.org/message-id/b4549406-909f-7d15-dc34-499835a8f0b3%40lab.ntt.co.jp\n[3] https://www.postgresql.org/message-id/a1fa1c4c-bf96-8ea5-cff5-85b927298e73%40lab.ntt.co.jp\n\n--\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n-- Best Wishes,Ashutosh BapatEnterpriseDB CorporationThe Postgres Database Company","msg_date":"Mon, 22 Aug 2016 12:19:59 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"Hi Ashutosh,\n\nOn 2016/08/22 15:49, Ashutosh Bapat wrote:\n> 1. deparsePlaceHolderVar looks odd - each of the deparse* function is\n> named as deparse + <name of the parser node the string would parse\n> into>. PlaceHolderVar is not a parser node, so no string is going to be\n> parsed as a PlaceHolderVar. May be you want to name it as\n> deparseExerFromPlaceholderVar or something like that.\n\nThe name \"deparseExerFromPlaceholderVar\" seems long to me. How about \n\"deparsePlaceHolderExpr\"?\n\n> 2. You will need to check phlevelsup member while assessing whether a\n> PHV is safe to push down.\n\nGood catch! In addition to that, I added another check that the eval_at \nset for the PHV should be included in the relids set for the foreign \nrelation. I think that would make the shippability check more robust.\n\n> 3. I think registerAlias stuff should happen really at the time of\n> creating paths and should be stored in fpinfo. Without that it will be\n> computed every time we deparse the query. This means every time we try\n> to EXPLAIN the query at various levels in the join tree and once for the\n> query to be sent to the foreign server.\n\nHmm. I think the overhead in calculating aliases would be negligible \ncompared to the overhead in explaining each remote query for costing or \nsending the remote query for execution. So, I created aliases in the \nsame way as remote params created and stored into params_list in \ndeparse_expr_cxt. I'm not sure it's worth complicating the code.\n\n> 4. The changes related to retrieved_attrs look unrelated to the patch.\n> Either your patch should use the current method of handling\n> retrieved_attrs or there should be a separate patch for retrieved_attrs\n> changes. May be you want to take a look at the discussion in join\n> pushdown thread as to why we assume retrieved_attrs to be non-NIL always.\n\nOK, I removed those changes from the patch.\n\n> 5. The blocks related to inner and outer relations in\n> deparseFromExprForRel() look same. We should probably separate that code\n> out into a function and call it at two places.\n\nDone.\n\n> 6.\n> ! if (is_placeholder)\n> ! errcontext(\"placeholder expression at position %d in select list\",\n> ! errpos->cur_attno);\n> A user wouldn't know what a placeholder expression is, there is no such\n> term in the documentation. We have to device a better way to provide an\n> error context for an expression in general.\n\nThough I proposed that, I don't think that it's that important to let \nusers know that the expression is from a PHV. How about just saying \n\"expression\", not \"placeholder expression\", so that we have the message \n\"expression at position %d in select list\" in the context?\n\nAttached is an updated version of the patch.\n\nOther changes:\n\n* Add a bit more regression test\n* Revise code/comments\n* Cleanups\n\nThanks for the comments!\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":"Fri, 2 Sep 2016 19:25:30 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Fri, Sep 2, 2016 at 3:55 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\nwrote:\n\n> Hi Ashutosh,\n>\n> On 2016/08/22 15:49, Ashutosh Bapat wrote:\n>\n>> 1. deparsePlaceHolderVar looks odd - each of the deparse* function is\n>> named as deparse + <name of the parser node the string would parse\n>> into>. PlaceHolderVar is not a parser node, so no string is going to be\n>> parsed as a PlaceHolderVar. May be you want to name it as\n>> deparseExerFromPlaceholderVar or something like that.\n>>\n>\n> The name \"deparseExerFromPlaceholderVar\" seems long to me. How about\n> \"deparsePlaceHolderExpr\"?\n>\n\nThere isn't any node with name PlaceHolderExpr.\n\n\n> 2. You will need to check phlevelsup member while assessing whether a\n>> PHV is safe to push down.\n>>\n>\n> Good catch! In addition to that, I added another check that the eval_at\n> set for the PHV should be included in the relids set for the foreign\n> relation. I think that would make the shippability check more robust.\n>\n\nThanks.\n\n\n>\n> 3. I think registerAlias stuff should happen really at the time of\n>> creating paths and should be stored in fpinfo. Without that it will be\n>> computed every time we deparse the query. This means every time we try\n>> to EXPLAIN the query at various levels in the join tree and once for the\n>> query to be sent to the foreign server.\n>>\n>\n> Hmm. I think the overhead in calculating aliases would be negligible\n> compared to the overhead in explaining each remote query for costing or\n> sending the remote query for execution. So, I created aliases in the same\n> way as remote params created and stored into params_list in\n> deparse_expr_cxt. I'm not sure it's worth complicating the code.\n>\n\nWe defer remote parameter creation till deparse time since the the\nparameter numbers are dependent upon the sequence in which we deparse the\nquery. Creating them at the time of path creation and storing them in\nfpinfo is not possible because a. those present in the joining relations\nwill conflict with each other and need some kind of serialization at the\ntime of deparsing b. those defer for differently parameterized paths or\npaths with different pathkeys. We don't defer it because it's very light on\nperformance.\n\nThat's not true with the alias information. As long as we detect which\nrelations need subqueries, their RTIs are enough to create unique aliases\ne.g. if a relation involves RTIs 1, 2, 3 corresponding subquery can have\nalias r123 without conflicting with any other alias.\n\nHowever minimum overhead it might have, we will deparse the query every\ntime we create a path involving that kind of relation i.e. for different\npathkeys, different parameterization and different joins in which that\nrelation participates in. This number can be significant. We want to avoid\nthis overhead if we can.\n\n\n>\n> 5. The blocks related to inner and outer relations in\n>> deparseFromExprForRel() look same. We should probably separate that code\n>> out into a function and call it at two places.\n>>\n>\n> Done.\n>\n\nThanks. I see you have created function deparseOperandRelation() for the\nsame. I guess, this should be renamed as deparseRangeTblRef() since it will\ncreate RangeTblRef node when parsed back.\n\n\n>\n> 6.\n>> ! if (is_placeholder)\n>> ! errcontext(\"placeholder expression at position %d in select\n>> list\",\n>> ! errpos->cur_attno);\n>> A user wouldn't know what a placeholder expression is, there is no such\n>> term in the documentation. We have to device a better way to provide an\n>> error context for an expression in general.\n>>\n>\n> Though I proposed that, I don't think that it's that important to let\n> users know that the expression is from a PHV. How about just saying\n> \"expression\", not \"placeholder expression\", so that we have the message\n> \"expression at position %d in select list\" in the context?\n>\n> Hmm, that's better than placeholder expression, but not as explanatory as\nit should be since we won't be printing the \"select list\" in the error\ncontext and user won't have a clue as to what error context is about. But I\ndon't have any good suggestions right now. May be we should print the whole\nexpression? But that can be very long, I don't know.\n\nThis patch tries to do two things at a time 1. support join pushdown for\nFULL join when the joining relations have remote conditions 2. better\nsupport for fetching placeholder vars, whole row references and some system\ncolumns. To make reviews easy, I think we should split the patch into two\n1. supporting subqueries to be deparsed with support for one of the above\n(I will suggest FULL join support) 2. support for the other.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database Company\n\nOn Fri, Sep 2, 2016 at 3:55 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp> wrote:Hi Ashutosh,\n\nOn 2016/08/22 15:49, Ashutosh Bapat wrote:\n\n1. deparsePlaceHolderVar looks odd - each of the deparse* function is\nnamed as deparse + <name of the parser node the string would parse\ninto>. PlaceHolderVar is not a parser node, so no string is going to be\nparsed as a PlaceHolderVar. May be you want to name it as\ndeparseExerFromPlaceholderVar or something like that.\n\n\nThe name \"deparseExerFromPlaceholderVar\" seems long to me. How about \"deparsePlaceHolderExpr\"?There isn't any node with name PlaceHolderExpr.\n\n\n2. You will need to check phlevelsup member while assessing whether a\nPHV is safe to push down.\n\n\nGood catch! In addition to that, I added another check that the eval_at set for the PHV should be included in the relids set for the foreign relation. I think that would make the shippability check more robust.Thanks. \n\n\n3. I think registerAlias stuff should happen really at the time of\ncreating paths and should be stored in fpinfo. Without that it will be\ncomputed every time we deparse the query. This means every time we try\nto EXPLAIN the query at various levels in the join tree and once for the\nquery to be sent to the foreign server.\n\n\nHmm. I think the overhead in calculating aliases would be negligible compared to the overhead in explaining each remote query for costing or sending the remote query for execution. So, I created aliases in the same way as remote params created and stored into params_list in deparse_expr_cxt. I'm not sure it's worth complicating the code.We defer remote parameter creation till deparse time since the the parameter numbers are dependent upon the sequence in which we deparse the query. Creating them at the time of path creation and storing them in fpinfo is not possible because a. those present in the joining relations will conflict with each other and need some kind of serialization at the time of deparsing b. those defer for differently parameterized paths or paths with different pathkeys. We don't defer it because it's very light on performance.That's not true with the alias information. As long as we detect which relations need subqueries, their RTIs are enough to create unique aliases e.g. if a relation involves RTIs 1, 2, 3 corresponding subquery can have alias r123 without conflicting with any other alias.However minimum overhead it might have, we will deparse the query every time we create a path involving that kind of relation i.e. for different pathkeys, different parameterization and different joins in which that relation participates in. This number can be significant. We want to avoid this overhead if we can. \n\n5. The blocks related to inner and outer relations in\ndeparseFromExprForRel() look same. We should probably separate that code\nout into a function and call it at two places.\n\n\nDone.Thanks. I see you have created function deparseOperandRelation() for the same. I guess, this should be renamed as deparseRangeTblRef() since it will create RangeTblRef node when parsed back. \n\n\n6.\n! if (is_placeholder)\n! errcontext(\"placeholder expression at position %d in select list\",\n! errpos->cur_attno);\nA user wouldn't know what a placeholder expression is, there is no such\nterm in the documentation. We have to device a better way to provide an\nerror context for an expression in general.\n\n\nThough I proposed that, I don't think that it's that important to let users know that the expression is from a PHV. How about just saying \"expression\", not \"placeholder expression\", so that we have the message \"expression at position %d in select list\" in the context?\nHmm, that's better than placeholder expression, but not as explanatory as it should be since we won't be printing the \"select list\" in the error context and user won't have a clue as to what error context is about. But I don't have any good suggestions right now. May be we should print the whole expression? But that can be very long, I don't know.This patch tries to do two things at a time 1. support join pushdown for FULL join when the joining relations have remote conditions 2. better support for fetching placeholder vars, whole row references and some system columns. To make reviews easy, I think we should split the patch into two 1. supporting subqueries to be deparsed with support for one of the above (I will suggest FULL join support) 2. support for the other. -- Best Wishes,Ashutosh BapatEnterpriseDB CorporationThe Postgres Database Company","msg_date":"Tue, 6 Sep 2016 18:37:12 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/09/06 22:07, Ashutosh Bapat wrote:\n> On Fri, Sep 2, 2016 at 3:55 PM, Etsuro Fujita\n> <fujita.etsuro@lab.ntt.co.jp <mailto:fujita.etsuro@lab.ntt.co.jp>> wrote:\n\n> On 2016/08/22 15:49, Ashutosh Bapat wrote:\n> 1. deparsePlaceHolderVar looks odd - each of the deparse*\n> function is\n> named as deparse + <name of the parser node the string would parse\n> into>. PlaceHolderVar is not a parser node, so no string is\n> going to be\n> parsed as a PlaceHolderVar. May be you want to name it as\n> deparseExerFromPlaceholderVar or something like that.\n\n> The name \"deparseExerFromPlaceholderVar\" seems long to me. How\n> about \"deparsePlaceHolderExpr\"?\n\n> There isn't any node with name PlaceHolderExpr.\n\nI'll rename it to \"deparseExerInPlaceholderVar\", then.\n\n> 3. I think registerAlias stuff should happen really at the time of\n> creating paths and should be stored in fpinfo. Without that it\n> will be\n> computed every time we deparse the query. This means every time\n> we try\n> to EXPLAIN the query at various levels in the join tree and once\n> for the\n> query to be sent to the foreign server.\n\n> Hmm. I think the overhead in calculating aliases would be\n> negligible compared to the overhead in explaining each remote query\n> for costing or sending the remote query for execution. So, I\n> created aliases in the same way as remote params created and stored\n> into params_list in deparse_expr_cxt. I'm not sure it's worth\n> complicating the code.\n\n> We defer remote parameter creation till deparse time since the the\n> parameter numbers are dependent upon the sequence in which we deparse\n> the query. Creating them at the time of path creation and storing them\n> in fpinfo is not possible because a. those present in the joining\n> relations will conflict with each other and need some kind of\n> serialization at the time of deparsing b. those defer for differently\n> parameterized paths or paths with different pathkeys. We don't defer it\n> because it's very light on performance.\n>\n> That's not true with the alias information. As long as we detect which\n> relations need subqueries, their RTIs are enough to create unique\n> aliases e.g. if a relation involves RTIs 1, 2, 3 corresponding subquery\n> can have alias r123 without conflicting with any other alias.\n\nHmm. But another concern about the approach of generating an subselect \nalias for a path, if needed, at the path creation time would be that the \npath might be rejected by add_path, which would result in totally \nwasting cycles for generating the subselect alias for the path.\n\n> However minimum overhead it might have, we will deparse the query every\n> time we create a path involving that kind of relation i.e. for different\n> pathkeys, different parameterization and different joins in which that\n> relation participates in. This number can be significant. We want to\n> avoid this overhead if we can.\n\nExactly. As I said above, I think the overhead would be negligible \ncompared to the overhead in explaining each remote query for costing or \nthe overhead in sending the final remote query for execution, though.\n\n> 5. The blocks related to inner and outer relations in\n> deparseFromExprForRel() look same. We should probably separate\n> that code\n> out into a function and call it at two places.\n\n> Done.\n\n> I see you have created function deparseOperandRelation() for the\n> same. I guess, this should be renamed as deparseRangeTblRef() since it\n> will create RangeTblRef node when parsed back.\n\nOK, if there no opinions of others, I'll rename it to the name proposed \nby you, \"deparseRangeTblRef\".\n\n> 6.\n> ! if (is_placeholder)\n> ! errcontext(\"placeholder expression at position %d in\n> select list\",\n> ! errpos->cur_attno);\n> A user wouldn't know what a placeholder expression is, there is\n> no such\n> term in the documentation. We have to device a better way to\n> provide an\n> error context for an expression in general.\n\n> Though I proposed that, I don't think that it's that important to\n> let users know that the expression is from a PHV. How about just\n> saying \"expression\", not \"placeholder expression\", so that we have\n> the message \"expression at position %d in select list\" in the context?\n\n> Hmm, that's better than placeholder expression, but not as explanatory\n> as it should be since we won't be printing the \"select list\" in the\n> error context and user won't have a clue as to what error context is\n> about.\n\nI don't think so. Consider an example of the conversion error message, \nwhich is from the regression test:\n\nSELECT ft1.c1, ft2.c2, ft1 FROM ft1, ft2 WHERE ft1.c1 = ft2.c1 AND \nft1.c1 = 1;\nERROR: invalid input syntax for integer: \"foo\"\nCONTEXT: whole-row reference to foreign table \"ft1\"\n\nAs shown in the example, the error message is displayed under a remote \nquery for execution. So, ISTM it's reasonable to print something like \n\"expression at position %d in select list\" in the context if an \nexpression in a PHV.\n\n> But I don't have any good suggestions right now. May be we should\n> print the whole expression? But that can be very long, I don't know.\n\nISTM that it's a bit too expensive to print the whole expression in the \nerror context.\n\n> This patch tries to do two things at a time 1. support join pushdown for\n> FULL join when the joining relations have remote conditions 2. better\n> support for fetching placeholder vars, whole row references and some\n> system columns. To make reviews easy, I think we should split the patch\n> into two 1. supporting subqueries to be deparsed with support for one of\n> the above (I will suggest FULL join support) 2. support for the other.\n\nOK, will try.\n\nBest regards,\nEtsuro Fujita\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, 8 Sep 2016 19:51:00 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/09/08 19:51, Etsuro Fujita wrote:\n> On 2016/09/06 22:07, Ashutosh Bapat wrote:\n>> This patch tries to do two things at a time 1. support join pushdown for\n>> FULL join when the joining relations have remote conditions 2. better\n>> support for fetching placeholder vars, whole row references and some\n>> system columns. To make reviews easy, I think we should split the patch\n>> into two 1. supporting subqueries to be deparsed with support for one of\n>> the above (I will suggest FULL join support) 2. support for the other.\n\n> OK, will try.\n\nI extracted #1 from the patch. Attached is a patch for that. If that \npatch is reasonable, I'll create a patch for #2 on top of it.\n\nComments welcome!\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":"Fri, 9 Sep 2016 21:35:45 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/09/09 21:35, Etsuro Fujita wrote:\n> On 2016/09/08 19:51, Etsuro Fujita wrote:\n>> On 2016/09/06 22:07, Ashutosh Bapat wrote:\n>>> This patch tries to do two things at a time 1. support join pushdown for\n>>> FULL join when the joining relations have remote conditions 2. better\n>>> support for fetching placeholder vars, whole row references and some\n>>> system columns. To make reviews easy, I think we should split the patch\n>>> into two 1. supporting subqueries to be deparsed with support for one of\n>>> the above (I will suggest FULL join support) 2. support for the other.\n\n>> OK, will try.\n\n> I extracted #1 from the patch. Attached is a patch for that. If that\n> patch is reasonable, I'll create a patch for #2 on top of it.\n\nAttached is a patch for #2. In that patch I fixed some bugs and added a \nbit more comments. For testing, please apply the patch for #1 first.\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, 12 Sep 2016 21:35:35 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":">\n> 3. I think registerAlias stuff should happen really at the time of\n>> creating paths and should be stored in fpinfo. Without that it\n>> will be\n>> computed every time we deparse the query. This means every time\n>> we try\n>> to EXPLAIN the query at various levels in the join tree and once\n>> for the\n>> query to be sent to the foreign server.\n>>\n>\n> Hmm. I think the overhead in calculating aliases would be\n>> negligible compared to the overhead in explaining each remote query\n>> for costing or sending the remote query for execution. So, I\n>> created aliases in the same way as remote params created and stored\n>> into params_list in deparse_expr_cxt. I'm not sure it's worth\n>> complicating the code.\n>>\n>\n> We defer remote parameter creation till deparse time since the the\n>> parameter numbers are dependent upon the sequence in which we deparse\n>> the query. Creating them at the time of path creation and storing them\n>> in fpinfo is not possible because a. those present in the joining\n>> relations will conflict with each other and need some kind of\n>> serialization at the time of deparsing b. those defer for differently\n>> parameterized paths or paths with different pathkeys. We don't defer it\n>> because it's very light on performance.\n>>\n>> That's not true with the alias information. As long as we detect which\n>> relations need subqueries, their RTIs are enough to create unique\n>> aliases e.g. if a relation involves RTIs 1, 2, 3 corresponding subquery\n>> can have alias r123 without conflicting with any other alias.\n>>\n>\n> Hmm. But another concern about the approach of generating an subselect\n> alias for a path, if needed, at the path creation time would be that the\n> path might be rejected by add_path, which would result in totally wasting\n> cycles for generating the subselect alias for the path.\n>\n\nA path may get rejected but the relation is not rejected. The alias applies\nto a relation and its targetlist, which will remain same for all paths\ncreated for that relation, esp when it's a base relation or join. So, AFAIU\nthat work never gets wasted. Also, for costing paths with\nuse_remote_estimate, we deparse the query, which builds the alias\ninformation again and again for very path. That's much worse than building\nit once for a given relation.\n\n\n>\n> However minimum overhead it might have, we will deparse the query every\n>> time we create a path involving that kind of relation i.e. for different\n>> pathkeys, different parameterization and different joins in which that\n>> relation participates in. This number can be significant. We want to\n>> avoid this overhead if we can.\n>>\n>\n> Exactly. As I said above, I think the overhead would be negligible\n> compared to the overhead in explaining each remote query for costing or the\n> overhead in sending the final remote query for execution, though.\n>\n\nIt won't remain minimal as the number of paths created increases,\nincreasing the number of times a query is deparsed. We deparse query every\ntime time we cost a path for a relation with use_remote_estimates true. As\nwe try to push down more and more stuff, we will create more paths and\ndeparse the query more time.\n\nAlso, that makes the interface better. Right now, in your patch, you have\nchanged the order of deparsing in the existing code, so that the aliases\nare registered while deparsing FROM clause and before any Var nodes are\ndeparsed. If we create aliases at the time of path creation, only once in\nGetForeignJoinPaths or GetForeignPaths as appropriate, that would require\nless code churn and would save some CPU cycles as well.\n\n\n>\n> 6.\n>> ! if (is_placeholder)\n>> ! errcontext(\"placeholder expression at position %d in\n>> select list\",\n>> ! errpos->cur_attno);\n>> A user wouldn't know what a placeholder expression is, there is\n>> no such\n>> term in the documentation. We have to device a better way to\n>> provide an\n>> error context for an expression in general.\n>>\n>\n> Though I proposed that, I don't think that it's that important to\n>> let users know that the expression is from a PHV. How about just\n>> saying \"expression\", not \"placeholder expression\", so that we have\n>> the message \"expression at position %d in select list\" in the context?\n>>\n>\n> Hmm, that's better than placeholder expression, but not as explanatory\n>> as it should be since we won't be printing the \"select list\" in the\n>> error context and user won't have a clue as to what error context is\n>> about.\n>>\n>\n> I don't think so. Consider an example of the conversion error message,\n> which is from the regression test:\n>\n> SELECT ft1.c1, ft2.c2, ft1 FROM ft1, ft2 WHERE ft1.c1 = ft2.c1 AND\n> ft1.c1 = 1;\n> ERROR: invalid input syntax for integer: \"foo\"\n> CONTEXT: whole-row reference to foreign table \"ft1\"\n>\n> As shown in the example, the error message is displayed under a remote\n> query for execution. So, ISTM it's reasonable to print something like\n> \"expression at position %d in select list\" in the context if an expression\n> in a PHV.\n>\n\nI missed it. Sorry. Looks ok.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database Company\n\n\n\n\n 3. I think registerAlias stuff should happen really at the time of\n creating paths and should be stored in fpinfo. Without that it\n will be\n computed every time we deparse the query. This means every time\n we try\n to EXPLAIN the query at various levels in the join tree and once\n for the\n query to be sent to the foreign server.\n\n\n\n Hmm. I think the overhead in calculating aliases would be\n negligible compared to the overhead in explaining each remote query\n for costing or sending the remote query for execution. So, I\n created aliases in the same way as remote params created and stored\n into params_list in deparse_expr_cxt. I'm not sure it's worth\n complicating the code.\n\n\n\nWe defer remote parameter creation till deparse time since the the\nparameter numbers are dependent upon the sequence in which we deparse\nthe query. Creating them at the time of path creation and storing them\nin fpinfo is not possible because a. those present in the joining\nrelations will conflict with each other and need some kind of\nserialization at the time of deparsing b. those defer for differently\nparameterized paths or paths with different pathkeys. We don't defer it\nbecause it's very light on performance.\n\nThat's not true with the alias information. As long as we detect which\nrelations need subqueries, their RTIs are enough to create unique\naliases e.g. if a relation involves RTIs 1, 2, 3 corresponding subquery\ncan have alias r123 without conflicting with any other alias.\n\n\nHmm. But another concern about the approach of generating an subselect alias for a path, if needed, at the path creation time would be that the path might be rejected by add_path, which would result in totally wasting cycles for generating the subselect alias for the path.A path may get rejected but the relation is not rejected. The alias applies to a relation and its targetlist, which will remain same for all paths created for that relation, esp when it's a base relation or join. So, AFAIU that work never gets wasted. Also, for costing paths with use_remote_estimate, we deparse the query, which builds the alias information again and again for very path. That's much worse than building it once for a given relation. \n\n\nHowever minimum overhead it might have, we will deparse the query every\ntime we create a path involving that kind of relation i.e. for different\npathkeys, different parameterization and different joins in which that\nrelation participates in. This number can be significant. We want to\navoid this overhead if we can.\n\n\nExactly. As I said above, I think the overhead would be negligible compared to the overhead in explaining each remote query for costing or the overhead in sending the final remote query for execution, though.It won't remain minimal as the number of paths created increases, increasing the number of times a query is deparsed. We deparse query every time time we cost a path for a relation with use_remote_estimates true. As we try to push down more and more stuff, we will create more paths and deparse the query more time. Also, that makes the interface better. Right now, in your patch, you have changed the order of deparsing in the existing code, so that the aliases are registered while deparsing FROM clause and before any Var nodes are deparsed. If we create aliases at the time of path creation, only once in GetForeignJoinPaths or GetForeignPaths as appropriate, that would require less code churn and would save some CPU cycles as well. \n\n\n 6.\n ! if (is_placeholder)\n ! errcontext(\"placeholder expression at position %d in\n select list\",\n ! errpos->cur_attno);\n A user wouldn't know what a placeholder expression is, there is\n no such\n term in the documentation. We have to device a better way to\n provide an\n error context for an expression in general.\n\n\n\n Though I proposed that, I don't think that it's that important to\n let users know that the expression is from a PHV. How about just\n saying \"expression\", not \"placeholder expression\", so that we have\n the message \"expression at position %d in select list\" in the context?\n\n\n\nHmm, that's better than placeholder expression, but not as explanatory\nas it should be since we won't be printing the \"select list\" in the\nerror context and user won't have a clue as to what error context is\nabout.\n\n\nI don't think so. Consider an example of the conversion error message, which is from the regression test:\n\nSELECT ft1.c1, ft2.c2, ft1 FROM ft1, ft2 WHERE ft1.c1 = ft2.c1 AND ft1.c1 = 1;\nERROR: invalid input syntax for integer: \"foo\"\nCONTEXT: whole-row reference to foreign table \"ft1\"\n\nAs shown in the example, the error message is displayed under a remote query for execution. So, ISTM it's reasonable to print something like \"expression at position %d in select list\" in the context if an expression in a PHV.I missed it. Sorry. Looks ok.-- Best Wishes,Ashutosh BapatEnterpriseDB CorporationThe Postgres Database Company","msg_date":"Tue, 13 Sep 2016 10:47:01 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Tue, Sep 6, 2016 at 9:07 AM, Ashutosh Bapat\n<ashutosh.bapat@enterprisedb.com> wrote:\n> That's not true with the alias information. As long as we detect which\n> relations need subqueries, their RTIs are enough to create unique aliases\n> e.g. if a relation involves RTIs 1, 2, 3 corresponding subquery can have\n> alias r123 without conflicting with any other alias.\n\nWhat if RTI 123 is also used in the query?\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, 13 Sep 2016 12:58:14 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Tue, Sep 13, 2016 at 10:28 PM, Robert Haas <robertmhaas@gmail.com> wrote:\n> On Tue, Sep 6, 2016 at 9:07 AM, Ashutosh Bapat\n> <ashutosh.bapat@enterprisedb.com> wrote:\n>> That's not true with the alias information. As long as we detect which\n>> relations need subqueries, their RTIs are enough to create unique aliases\n>> e.g. if a relation involves RTIs 1, 2, 3 corresponding subquery can have\n>> alias r123 without conflicting with any other alias.\n>\n> What if RTI 123 is also used in the query?\n\nGood catch. I actually meant some combination of 1, 2 and 3, which is\nunique for a join between r1, r2 and r3. How about r1_2_3 or\nr1_r2_r3?\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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 Sep 2016 09:08:10 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Tue, Sep 13, 2016 at 11:38 PM, Ashutosh Bapat\n<ashutosh.bapat@enterprisedb.com> wrote:\n> On Tue, Sep 13, 2016 at 10:28 PM, Robert Haas <robertmhaas@gmail.com> wrote:\n>> On Tue, Sep 6, 2016 at 9:07 AM, Ashutosh Bapat\n>> <ashutosh.bapat@enterprisedb.com> wrote:\n>>> That's not true with the alias information. As long as we detect which\n>>> relations need subqueries, their RTIs are enough to create unique aliases\n>>> e.g. if a relation involves RTIs 1, 2, 3 corresponding subquery can have\n>>> alias r123 without conflicting with any other alias.\n>>\n>> What if RTI 123 is also used in the query?\n>\n> Good catch. I actually meant some combination of 1, 2 and 3, which is\n> unique for a join between r1, r2 and r3. How about r1_2_3 or\n> r1_r2_r3?\n\nSure, something like that can work, but if you have a big enough join\nthe identifier might get too long. I'm not sure why it wouldn't work\nto just use the lowest RTI involved in the join, though; the others\nwon't appear anywhere else at that query level.\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 Sep 2016 11:22:38 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Wed, Sep 14, 2016 at 8:52 PM, Robert Haas <robertmhaas@gmail.com> wrote:\n> On Tue, Sep 13, 2016 at 11:38 PM, Ashutosh Bapat\n> <ashutosh.bapat@enterprisedb.com> wrote:\n>> On Tue, Sep 13, 2016 at 10:28 PM, Robert Haas <robertmhaas@gmail.com> wrote:\n>>> On Tue, Sep 6, 2016 at 9:07 AM, Ashutosh Bapat\n>>> <ashutosh.bapat@enterprisedb.com> wrote:\n>>>> That's not true with the alias information. As long as we detect which\n>>>> relations need subqueries, their RTIs are enough to create unique aliases\n>>>> e.g. if a relation involves RTIs 1, 2, 3 corresponding subquery can have\n>>>> alias r123 without conflicting with any other alias.\n>>>\n>>> What if RTI 123 is also used in the query?\n>>\n>> Good catch. I actually meant some combination of 1, 2 and 3, which is\n>> unique for a join between r1, r2 and r3. How about r1_2_3 or\n>> r1_r2_r3?\n>\n> Sure, something like that can work, but if you have a big enough join\n> the identifier might get too long. I'm not sure why it wouldn't work\n> to just use the lowest RTI involved in the join, though; the others\n> won't appear anywhere else at that query level.\n>\n\nYes, that will work too and is much more preferred than long r1_2_3.\nThe idea is to come with a unique alias name from RTIs involved in\nthat relation. Thinking loudly, r1_2_3 is more descriptive to debug\nissues. It tells that the subquery it refers to covers RTIs 1, 2 and\n3. That information may be quite helpful to understand the remote SQL.\nr1 on the other hand can refer to relation with RTI 1 or a join\nrelation where least RTI is 1. That can be solved by using s<RTI> for\nsubquery and r<RTI> for a bare relation. But it still doesn't tell us\nwhich all relations are involved.\n\nBut since the subquery is part of remote SQL and we have to take a\nlook at it to find out meaning of individual columns, that benefit may\nbe smaller compared to the convenience of smaller alias. So +1 for\nusing the smallest RTI to indicate a subquery.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 15 Sep 2016 11:59:48 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/09/13 14:17, Ashutosh Bapat wrote:\n\n> But another concern about the approach of generating an\n> subselect alias for a path, if needed, at the path creation time\n> would be that the path might be rejected by add_path, which would\n> result in totally wasting cycles for generating the subselect alias\n> for the path.\n\n> A path may get rejected but the relation is not rejected. The alias\n> applies to a relation and its targetlist, which will remain same for all\n> paths created for that relation, esp when it's a base relation or join.\n> So, AFAIU that work never gets wasted.\n\nI think there is no guarantee that add_path won't reject foreign join \npaths. The possibility would be very low, though.\n\n> However minimum overhead it might have, we will deparse the\n> query every\n> time we create a path involving that kind of relation i.e. for\n> different\n> pathkeys, different parameterization and different joins in\n> which that\n> relation participates in. This number can be significant. We want to\n> avoid this overhead if we can.\n\n> Exactly. As I said above, I think the overhead would be negligible\n> compared to the overhead in explaining each remote query for costing\n> or the overhead in sending the final remote query for execution, though.\n\n> It won't remain minimal as the number of paths created increases,\n> increasing the number of times a query is deparsed. We deparse query\n> every time time we cost a path for a relation with use_remote_estimates\n> true. As we try to push down more and more stuff, we will create more\n> paths and deparse the query more time.\n\n> Also, that makes the interface better. Right now, in your patch, you\n> have changed the order of deparsing in the existing code, so that the\n> aliases are registered while deparsing FROM clause and before any Var\n> nodes are deparsed. If we create aliases at the time of path creation,\n> only once in GetForeignJoinPaths or GetForeignPaths as appropriate, that\n> would require less code churn and would save some CPU cycles as well.\n\nAgreed. Will fix.\n\nBest regards,\nEtsuro Fujita\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, 26 Sep 2016 16:30:22 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/09/15 15:29, Ashutosh Bapat wrote:\n> On Wed, Sep 14, 2016 at 8:52 PM, Robert Haas <robertmhaas@gmail.com> wrote:\n\n>> I'm not sure why it wouldn't work\n>> to just use the lowest RTI involved in the join, though; the others\n>> won't appear anywhere else at that query level.\n\n> So +1 for\n> using the smallest RTI to indicate a subquery.\n\n+1 for the general idea.\n\nISTM that the use of the same RTI for subqueries in multi-levels in a \nremote SQL makes the SQL a bit difficult to read. How about using the \nposition of the join rel in join_rel_list, (more precisely, the position \nplus list_length(root->parse->rtable)), instead?\n\nBest regards,\nEtsuro Fujita\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, 26 Sep 2016 16:35:17 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Mon, Sep 26, 2016 at 1:05 PM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> On 2016/09/15 15:29, Ashutosh Bapat wrote:\n>>\n>> On Wed, Sep 14, 2016 at 8:52 PM, Robert Haas <robertmhaas@gmail.com>\n>> wrote:\n>\n>\n>>> I'm not sure why it wouldn't work\n>>> to just use the lowest RTI involved in the join, though; the others\n>>> won't appear anywhere else at that query level.\n>\n>\n>> So +1 for\n>> using the smallest RTI to indicate a subquery.\n>\n>\n> +1 for the general idea.\n>\n> ISTM that the use of the same RTI for subqueries in multi-levels in a remote\n> SQL makes the SQL a bit difficult to read. How about using the position of\n> the join rel in join_rel_list, (more precisely, the position plus\n> list_length(root->parse->rtable)), instead?\n>\nWe switch to hash table to maintain the join RelOptInfos when the\nnumber of joins grows larger, where the position won't make much\nsense. We might differentiate between a base relation alias and\nsubquery alias by using different prefixes like \"r\" and \"s\" resp.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 26 Sep 2016 14:36:33 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/09/26 18:06, Ashutosh Bapat wrote:\n> On Mon, Sep 26, 2016 at 1:05 PM, Etsuro Fujita\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\n\n>> ISTM that the use of the same RTI for subqueries in multi-levels in a remote\n>> SQL makes the SQL a bit difficult to read. How about using the position of\n>> the join rel in join_rel_list, (more precisely, the position plus\n>> list_length(root->parse->rtable)), instead?\n\n> We switch to hash table to maintain the join RelOptInfos when the\n> number of joins grows larger, where the position won't make much\n> sense.\n\nThat's right, but we still store the joinrel into join_rel_list after \ncreating that hash table. That hash table is just for fast lookup. See \nbuild_join_rel.\n\n> We might differentiate between a base relation alias and\n> subquery alias by using different prefixes like \"r\" and \"s\" resp.\n\nHmm. My concern about that would the recursive use of \"s\" with the same \nRTI as subquery aliases for different level subqueries in a single \nremote SQL. For example, if those subqueries include a base rel with \nRTI=1, then \"s1\" would be used again and again within that SQL. That \nwould be logically correct, but seems confusing to me.\n\nBest regards,\nEtsuro Fujita\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, 26 Sep 2016 19:36:02 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Mon, Sep 26, 2016 at 4:06 PM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> On 2016/09/26 18:06, Ashutosh Bapat wrote:\n>>\n>> On Mon, Sep 26, 2016 at 1:05 PM, Etsuro Fujita\n>> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>\n>\n>>> ISTM that the use of the same RTI for subqueries in multi-levels in a\n>>> remote\n>>> SQL makes the SQL a bit difficult to read. How about using the position\n>>> of\n>>> the join rel in join_rel_list, (more precisely, the position plus\n>>> list_length(root->parse->rtable)), instead?\n>\n>\n>> We switch to hash table to maintain the join RelOptInfos when the\n>> number of joins grows larger, where the position won't make much\n>> sense.\n>\n>\n> That's right, but we still store the joinrel into join_rel_list after\n> creating that hash table. That hash table is just for fast lookup. See\n> build_join_rel.\n\nSorry, I wasn't clear in my reply. As the list grows, it will take\nlonger to locate the RelOptInfo of the given join. Doing that for\ncreating an alias seems an overkill.\n\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 26 Sep 2016 16:50:27 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/09/26 20:20, Ashutosh Bapat wrote:\n> On Mon, Sep 26, 2016 at 4:06 PM, Etsuro Fujita\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>> On 2016/09/26 18:06, Ashutosh Bapat wrote:\n>>> On Mon, Sep 26, 2016 at 1:05 PM, Etsuro Fujita\n>>> <fujita.etsuro@lab.ntt.co.jp> wrote:\n\n>>>> ISTM that the use of the same RTI for subqueries in multi-levels in a\n>>>> remote\n>>>> SQL makes the SQL a bit difficult to read. How about using the position\n>>>> of\n>>>> the join rel in join_rel_list, (more precisely, the position plus\n>>>> list_length(root->parse->rtable)), instead?\n\n>>> We switch to hash table to maintain the join RelOptInfos when the\n>>> number of joins grows larger, where the position won't make much\n>>> sense.\n\n>> That's right, but we still store the joinrel into join_rel_list after\n>> creating that hash table.\n\n> As the list grows, it will take\n> longer to locate the RelOptInfo of the given join. Doing that for\n> creating an alias seems an overkill.\n\nThe join rel is appended to the end of the list, so I was thinking to \nget the position info by list_length during postgresGetForeignJoinPaths.\n\nBest regards,\nEtsuro Fujita\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, 27 Sep 2016 12:18:10 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Tue, Sep 27, 2016 at 8:48 AM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> On 2016/09/26 20:20, Ashutosh Bapat wrote:\n>>\n>> On Mon, Sep 26, 2016 at 4:06 PM, Etsuro Fujita\n>> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>>>\n>>> On 2016/09/26 18:06, Ashutosh Bapat wrote:\n>>>>\n>>>> On Mon, Sep 26, 2016 at 1:05 PM, Etsuro Fujita\n>>>> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>\n>\n>>>>> ISTM that the use of the same RTI for subqueries in multi-levels in a\n>>>>> remote\n>>>>> SQL makes the SQL a bit difficult to read. How about using the\n>>>>> position\n>>>>> of\n>>>>> the join rel in join_rel_list, (more precisely, the position plus\n>>>>> list_length(root->parse->rtable)), instead?\n>\n>\n>>>> We switch to hash table to maintain the join RelOptInfos when the\n>>>> number of joins grows larger, where the position won't make much\n>>>> sense.\n>\n>\n>>> That's right, but we still store the joinrel into join_rel_list after\n>>> creating that hash table.\n>\n>\n>> As the list grows, it will take\n>> longer to locate the RelOptInfo of the given join. Doing that for\n>> creating an alias seems an overkill.\n>\n>\n> The join rel is appended to the end of the list, so I was thinking to get\n> the position info by list_length during postgresGetForeignJoinPaths.\n\nThat's true only when the paths are being added to a newly created\njoinrel. But that's not true always. We may add paths with different\njoining order to an existing joinrel, in which case list_length would\nnot give its position. Am I missing something?\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 27 Sep 2016 10:03:25 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/09/27 13:33, Ashutosh Bapat wrote:\n\nI wrote:\n>>>>>> ISTM that the use of the same RTI for subqueries in multi-levels in a\n>>>>>> remote\n>>>>>> SQL makes the SQL a bit difficult to read. How about using the\n>>>>>> position\n>>>>>> of\n>>>>>> the join rel in join_rel_list, (more precisely, the position plus\n>>>>>> list_length(root->parse->rtable)), instead?\n\nI wrote:\n>> The join rel is appended to the end of the list, so I was thinking to get\n>> the position info by list_length during postgresGetForeignJoinPaths.\n\n> That's true only when the paths are being added to a newly created\n> joinrel. But that's not true always. We may add paths with different\n> joining order to an existing joinrel, in which case list_length would\n> not give its position. Am I missing something?\n\nI think you are right, but postgresGetForeignJoinPaths only allows us to \nadd a foreign join path to a newly created joinrel. The reason is \nbecause that routine skips all its work after the first call for that \njoinrel, by checking to see if joinrel->fdw_private is not NULL. So, I \nthink it's reasonable to get the position by list_length in that routine.\n\nBest regards,\nEtsuro Fujita\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, 28 Sep 2016 18:20:05 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/09/26 16:30, Etsuro Fujita wrote:\n> On 2016/09/13 14:17, Ashutosh Bapat wrote:\n\n>> It won't remain minimal as the number of paths created increases,\n>> increasing the number of times a query is deparsed. We deparse query\n>> every time time we cost a path for a relation with use_remote_estimates\n>> true. As we try to push down more and more stuff, we will create more\n>> paths and deparse the query more time.\n\n>> Also, that makes the interface better. Right now, in your patch, you\n>> have changed the order of deparsing in the existing code, so that the\n>> aliases are registered while deparsing FROM clause and before any Var\n>> nodes are deparsed. If we create aliases at the time of path creation,\n>> only once in GetForeignJoinPaths or GetForeignPaths as appropriate, that\n>> would require less code churn and would save some CPU cycles as well.\n\n> Agreed. Will fix.\n\nDone. Attached is an updated version of the patch.\n\nI didn't create aliases at anytime. Instead, I added a logic to get \ninfo about the alias to a given expression from reltarget->exprs for \nrelations in a given join tree. See isSubqueryExpr and \ngetSubselectAliasInfo.\n\nAs proposed by you, the patch differentiates between a base relation \nalias and a subquery alias by using different prefixes \"r\" and \"s\", \nrespectively. Also, subquery aliases are indexed by RTI for baserels \nand the position in join_rel_list + the length of rtable for joinrels, \nas proposed upthread.\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":"Wed, 28 Sep 2016 18:35:45 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/09/28 18:35, Etsuro Fujita wrote:\n> Attached is an updated version of the patch.\n\nI found a minor bug in that patch; the relation_index added to \nPgFdwRelationInfo was defined as Index, but I used the modifier %d to \nprint that. So, I changed the data type to int. Also, I added a bit \nmore comments. Please find attached an updated version of the patch.\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":"Thu, 29 Sep 2016 21:12:30 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/09/29 21:12, Etsuro Fujita wrote:\n> Please find attached an updated version of the patch.\n\nHere is a new version (V6) of the patch, which is basically the same as \nthe previous patch, but I slightly modified that patch. Another patch I \nam attaching is an updated patch for pushing down PHVs to the remote, \nwhich is created on top of the V6 patch.\n\nBest regards,\nEtsuro Fujita\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":"Thu, 13 Oct 2016 19:35:40 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"Review for postgres-fdw-more-full-join-pushdown-v6 patch.\n\nThe patch applies cleanly and regression is clean (make check in\nregress directory and that in postgres_fdw).\n\nHere are some comments.\n1. Because of the following code change, for a joinrel we might end up using\nattrs_used, which will be garbage for a join rel. The assumption that tlist can\nnot be NIL for a join relation is wrong. Even for a join relation, tlist can be\nNULL. Consider query 'explain verbose select count(*) from ft1, ft2' Since\ncount(*) doesn't need any columns, the tlist is NIL. With the patch the server\ncrashes for this query.\n- if (foreignrel->reloptkind == RELOPT_JOINREL)\n+ /*\n+ * Note: tlist for a base relation might be non-NIL. For example, if the\n+ * base relation is an operand of a foreign join performing a full outer\n+ * join and has non-NIL remote_conds, the base relation will be deparsed\n+ * as a subquery, so the tlist for the base relation could be non-NIL.\n+ */\n+ if (tlist != NIL)\n {\n- /* For a join relation use the input tlist */\n\nWe can not decide whether to use deparseExplicitTargetList or\ndeparse it from attrs_used based on the tlist. SELECT lists, whether it's\ntopmost SELECT or a subquery (even on a base relation), for deparsing a\nJOIN query need to use deparseExplicitTargetList.\n\n2. The code in deparseRangeTblRef() dealing with subqueries, is very similar to\ndeparseSelectStmtForRel(). The only thing deparseRangeTblRef() does not require\nis the appending ORDER BY, which is determined by existence of pathkeys\nargument. Probably we should reuse deparseSelectStmtForRel(), instead of\nduplicating the code. This will also make the current changes to\ndeparseSelectSql unnecessary.\n\n3. Why do we need following change? The elog() just few lines above says that\nwe expect only Var nodes. Why then we use deparseExpr() instead of\ndeparseVar()?\n if (i > 0)\n appendStringInfoString(buf, \", \");\n- deparseVar(var, context);\n+ deparseExpr((Expr *) var, context);\n\n *retrieved_attrs = lappend_int(*retrieved_attrs, i + 1);\n\nAnd I get the answer for that question, somewhere down in the patch\n+ /*\n+ * If the given expression is an output column of a subquery-in-FROM,\n+ * deparse the alias to the expression instead.\n+ */\n+ if (IsA(node, Var))\n+ {\n+ int tabno;\n+ int colno;\n+\n+ if (isSubqueryExpr(node, context->foreignrel, &tabno, &colno))\n+ {\n+ appendStringInfo(context->buf, \"%s%d.%s%d\",\n+ SS_TAB_ALIAS_PREFIX, tabno,\n+ SS_COL_ALIAS_PREFIX, colno);\n+ return;\n+ }\n+ }\n+\n\nFunctionally, this code belongs to deparseVar() and not deparseExpr(). Like all\nother functions which handle Expr nodes, deparseExpr() is expected to be a\ndispatcher to the node specific function.\n\n4. This comment is good to have there, but is unrelated to this patch. May be a\nseparate patch?\n+ /* Don't generate bad syntax if no columns */\n\n5. Changed comment doesn't say anything different from the original comment. It\nmay have been good to have it this way to start with, but changing it now\ndoesn't bring anything new. You seem to have just merged prologue of the\nfunction with a comment in that function. I think, this change is unnecessary.\n- * The function constructs ... JOIN ... ON ... for join relation. For a base\n- * relation it just returns schema-qualified tablename, with the appropriate\n- * alias if so requested.\n+ * For a join relation the clause of the following form is appended to buf:\n+ * ((outer relation) <join type> (inner relation) ON (joinclauses))\n+ * For a base relation the function just adds the schema-qualified tablename,\n+ * with the appropriate alias if so requested.\n+ /* Don't generate bad syntax if no columns */\n\n6. Change may be good but unrelated to this patch. May be material for a\nseparate patch. There are few such changes in this function. While these\nchanges may be good by themselves, they distract reviewer from the goal of the\npatch.\n- appendStringInfo(buf, \"(\");\n+ appendStringInfoChar(buf, '(');\n\n7. I don't understand why you need to change this function so much. Take for\nexample the following change.\n- RelOptInfo *rel_o = fpinfo->outerrel;\n- RelOptInfo *rel_i = fpinfo->innerrel;\n- StringInfoData join_sql_o;\n- StringInfoData join_sql_i;\n+ /* Begin the FROM clause entry */\n+ appendStringInfoChar(buf, '(');\n\n /* Deparse outer relation */\n- initStringInfo(&join_sql_o);\n- deparseFromExprForRel(&join_sql_o, root, rel_o, true, params_list);\n+ deparseRangeTblRef(buf, root,\n+ fpinfo->outerrel,\n+ fpinfo->make_outerrel_subquery,\n+ params_list);\n /* Deparse outer relation */\n- initStringInfo(&join_sql_o);\n- deparseFromExprForRel(&join_sql_o, root, rel_o, true, params_list);\n+ deparseRangeTblRef(buf, root,\n+ fpinfo->outerrel,\n+ fpinfo->make_outerrel_subquery,\n+ params_list);\n\nIt removes few variables, instead directly accesses the members from fpinfo.\nAlso, deparses the individual relations in the provided buffer directly, rather\nthan in separate buffers in the original code. Instead of this, all you had to\ndo was replace a call\n- deparseFromExprForRel(&join_sql_o, root, rel_o, true, params_list);\nwith\n+ deparseRangeTblRef(&join_sql_o, root, rel_o,\nfpinfo->make_outerrel_subquery, params_list)\nSimilarly for the inner relation. Again, the changes you have done might have\nbeen good, if done in the original code, but doing those in this patch just\ncreates distractions and increases the size of the patch.\n\nSimilarly changes\n- /* Append join clause; (TRUE) if no join clause */\n+ /* Append join conditions */\n\n+ {\n+ /* No join conditions; add \"(TRUE)\" */\n appendStringInfoString(buf, \"(TRUE)\");\n+ }\n+ appendStringInfoString(buf, \" ON \");\n\n- /* End the FROM clause entry. */\n- appendStringInfo(buf, \")\");\n\n\n+ /* End the FROM clause entry */\n+ appendStringInfoChar(buf, ')');\n\n8. Why have you changed the name of variable here?\n- if (use_alias)\n+ if (add_rel_alias)\n\n9. In deparseRangeTblRef(), the targetlist for the subquery is obtained by\ncalling build_tlist_to_deparse(), but appendSubselectAlias() constructs the\ncolumn aliases based on foreignrel->reltarget->exprs. Those two are usually\nsame, but there is no code which guarantees that. Instead, pass tlist to\nappendSubselectAlias() or simply pass the number of entries in that list\n(list_length(tlist), which is all it needs. OR use foreignrel->reltarget->exprs\ndirectly instead of calling build_tlist_to_deparse(). Similar problem exists\nwith getSubselectAliasInfo().\n\n10. Deparsing a query in the form of a subquery makes it hard-to-read. The\noriginal code aimed at avoiding a subquery. Also, this change has created many\nexpected output changes, which again seem unnecessary. In fact, having the\npushable join clauses of an inner join in ON clause, which is closer to JOIN\nclause, is better than having them farther in the WHERE clause.\n- /*\n- * For an inner join, all restrictions can be treated alike. Treating the\n- * pushed down conditions as join conditions allows a top level full outer\n- * join to be deparsed without requiring subqueries.\n- */\n- if (jointype == JOIN_INNER)\n- {\n- Assert(!fpinfo->joinclauses);\n- fpinfo->joinclauses = fpinfo->remote_conds;\n- fpinfo->remote_conds = NIL;\n- }\n-\n\n11. I have reworded following comment and restructured the code that follows it\nin the attached patch.\n+ /*\n+ * Set the subquery information. If the relation performs a full outer\n+ * join and if the input relations have non-NIL remote_conds, the input\n+ * relations need to be deparsed as a subquery.\n+ */\nThe code esp. the if .. else .. block followed by another one, made it a bit\nunreadable. Hence I restructured it so that it's readable and doesn't require\nvariable \"relids\" from earlier code, which was being reused. Let me know if\nthis change looks good.\n\n12. The code below deletes the condition, which is understandable.\n- if (fpinfo_i->remote_conds || fpinfo_o->remote_conds)\nBut why does it add a new unrelated condition here? What the comment claims,\nlooks like an existing bug, unrelated to the patch. Can you please give an\nexample? If that it's an existing bug, it should be fixed as a separate patch.\nI don't understand the relation of this code with what the patch is doing.\n+ /*\n+ * We can't do anything here, and if there are any non-Vars in the\n+ * outerrel/innerrel's reltarget, give up pushing down this join,\n+ * because the deparsing logic can't support such a case currently.\n+ */\n+ if (reltarget_has_non_vars(outerrel))\n+ return false;\n+ if (reltarget_has_non_vars(innerrel))\n return false;\n\n13. The comment below is missing the main purpose i.e. creating a a unique\nalias, in case the relation gets converted into a subquery. Lowest or highest\nrelid will create a unique alias at given level of join and that would be more\nfuture proof. If we decide to consider paths for every join order, following\ncomment will no more be true.\n+ /*\n+ * Set the relation index. This is defined as the position of this\n+ * joinrel in the join_rel_list list plus the length of the rtable list.\n+ * Note that since this joinrel is at the end of the list when we are\n+ * called, we can get the position by list_length.\n+ */\n+ fpinfo->relation_index =\n+ list_length(root->parse->rtable) + list_length(root->join_rel_list);\n\n14. The function name talks about non-vars but the Assert few lines below\nasserts that every node there is a Var. Need better naming.\n+reltarget_has_non_vars(RelOptInfo *foreignrel)\n+{\n+ ListCell *lc;\n+\n+ foreach(lc, foreignrel->reltarget->exprs)\n+ {\n+ Var *var = (Var *) lfirst(lc);\n+\n+ Assert(IsA(var, Var));\nAnd also an explanation for the claim\n+ * Note: currently deparseExplicitTargetList can't properly handle such Vars.\n\n15. While deparsing every Var, we are descending the RelOptInfo hierarchy.\nInstead of that, we should probably gather all the alias information once and\nstore it in context as an array indexed by relid. While deparsing a Var we can\nget the targetlist and alias for a given relation by using var->varno as index.\nAnd then search for given Var node in the targetlist to deparse the column name\nby its position in the targetlist. For the relations that are not converted\ninto subqueries, this array will not hold any information and the Var will be\ndeparsed as it's being done today.\n\nTesting\n-------\n1. The code is changing deparser but doesn't have testcases\ntesting impact of the code. For example, we should have testcases with remote\nquery deparsed as nested subqueries or nested subqueries within subqueries and\nso on May be testcases where a relation deeper in the RelOptInfo hierarchy has\nconditions but it's immediate upper relation does not have those.\n\n2. The only testcase added by this patch, copies an existing case and adds a\nwhole row reference to one of the relations being joined. Instead we could add\nthat whole-row reference to the existing testcase itself.\n\nOn Thu, Oct 13, 2016 at 4:05 PM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> On 2016/09/29 21:12, Etsuro Fujita wrote:\n>>\n>> Please find attached an updated version of the patch.\n>\n>\n> Here is a new version (V6) of the patch, which is basically the same as the\n> previous patch, but I slightly modified that patch. Another patch I am\n> attaching is an updated patch for pushing down PHVs to the remote, which is\n> created on top of the V6 patch.\n>\n> Best regards,\n> Etsuro Fujita\n\n\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 22 Oct 2016 13:49:35 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"> 11. I have reworded following comment and restructured the code that follows it\n> in the attached patch.\n> + /*\n> + * Set the subquery information. If the relation performs a full outer\n> + * join and if the input relations have non-NIL remote_conds, the input\n> + * relations need to be deparsed as a subquery.\n> + */\n> The code esp. the if .. else .. block followed by another one, made it a bit\n> unreadable. Hence I restructured it so that it's readable and doesn't require\n> variable \"relids\" from earlier code, which was being reused. Let me know if\n> this change looks good.\n>\nSorry, forgot to attach the patch. Here it is.\n\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database Company\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, 22 Oct 2016 22:16:10 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/10/22 17:19, Ashutosh Bapat wrote:\n> Review for postgres-fdw-more-full-join-pushdown-v6 patch.\n>\n> The patch applies cleanly and regression is clean (make check in\n> regress directory and that in postgres_fdw).\n\nThanks for the review!\n\n> Here are some comments.\n> 1. Because of the following code change, for a joinrel we might end up using\n> attrs_used, which will be garbage for a join rel. The assumption that tlist can\n> not be NIL for a join relation is wrong. Even for a join relation, tlist can be\n> NULL. Consider query 'explain verbose select count(*) from ft1, ft2' Since\n> count(*) doesn't need any columns, the tlist is NIL. With the patch the server\n> crashes for this query.\n> - if (foreignrel->reloptkind == RELOPT_JOINREL)\n> + /*\n> + * Note: tlist for a base relation might be non-NIL. For example, if the\n> + * base relation is an operand of a foreign join performing a full outer\n> + * join and has non-NIL remote_conds, the base relation will be deparsed\n> + * as a subquery, so the tlist for the base relation could be non-NIL.\n> + */\n> + if (tlist != NIL)\n> {\n> - /* For a join relation use the input tlist */\n>\n> We can not decide whether to use deparseExplicitTargetList or\n> deparse it from attrs_used based on the tlist. SELECT lists, whether it's\n> topmost SELECT or a subquery (even on a base relation), for deparsing a\n> JOIN query need to use deparseExplicitTargetList.\n\nWill fix.\n\n> 2. The code in deparseRangeTblRef() dealing with subqueries, is very similar to\n> deparseSelectStmtForRel(). The only thing deparseRangeTblRef() does not require\n> is the appending ORDER BY, which is determined by existence of pathkeys\n> argument. Probably we should reuse deparseSelectStmtForRel(), instead of\n> duplicating the code. This will also make the current changes to\n> deparseSelectSql unnecessary.\n\nWill do.\n\n> 3. Why do we need following change? The elog() just few lines above says that\n> we expect only Var nodes. Why then we use deparseExpr() instead of\n> deparseVar()?\n> if (i > 0)\n> appendStringInfoString(buf, \", \");\n> - deparseVar(var, context);\n> + deparseExpr((Expr *) var, context);\n>\n> *retrieved_attrs = lappend_int(*retrieved_attrs, i + 1);\n>\n> And I get the answer for that question, somewhere down in the patch\n> + /*\n> + * If the given expression is an output column of a subquery-in-FROM,\n> + * deparse the alias to the expression instead.\n> + */\n> + if (IsA(node, Var))\n> + {\n> + int tabno;\n> + int colno;\n> +\n> + if (isSubqueryExpr(node, context->foreignrel, &tabno, &colno))\n> + {\n> + appendStringInfo(context->buf, \"%s%d.%s%d\",\n> + SS_TAB_ALIAS_PREFIX, tabno,\n> + SS_COL_ALIAS_PREFIX, colno);\n> + return;\n> + }\n> + }\n> +\n>\n> Functionally, this code belongs to deparseVar() and not deparseExpr(). Like all\n> other functions which handle Expr nodes, deparseExpr() is expected to be a\n> dispatcher to the node specific function.\n\nOK, will change that way.\n\n> 4. This comment is good to have there, but is unrelated to this patch. May be a\n> separate patch?\n> + /* Don't generate bad syntax if no columns */\n>\n> 5. Changed comment doesn't say anything different from the original comment. It\n> may have been good to have it this way to start with, but changing it now\n> doesn't bring anything new. You seem to have just merged prologue of the\n> function with a comment in that function. I think, this change is unnecessary.\n> - * The function constructs ... JOIN ... ON ... for join relation. For a base\n> - * relation it just returns schema-qualified tablename, with the appropriate\n> - * alias if so requested.\n> + * For a join relation the clause of the following form is appended to buf:\n> + * ((outer relation) <join type> (inner relation) ON (joinclauses))\n> + * For a base relation the function just adds the schema-qualified tablename,\n> + * with the appropriate alias if so requested.\n> + /* Don't generate bad syntax if no columns */\n>\n> 6. Change may be good but unrelated to this patch. May be material for a\n> separate patch. There are few such changes in this function. While these\n> changes may be good by themselves, they distract reviewer from the goal of the\n> patch.\n> - appendStringInfo(buf, \"(\");\n> + appendStringInfoChar(buf, '(');\n\nOK on #4, #5, and #6, Will remove all the changes. I will leave those \nfor separate patches.\n\n> 7. I don't understand why you need to change this function so much. Take for\n> example the following change.\n> - RelOptInfo *rel_o = fpinfo->outerrel;\n> - RelOptInfo *rel_i = fpinfo->innerrel;\n> - StringInfoData join_sql_o;\n> - StringInfoData join_sql_i;\n> + /* Begin the FROM clause entry */\n> + appendStringInfoChar(buf, '(');\n>\n> /* Deparse outer relation */\n> - initStringInfo(&join_sql_o);\n> - deparseFromExprForRel(&join_sql_o, root, rel_o, true, params_list);\n> + deparseRangeTblRef(buf, root,\n> + fpinfo->outerrel,\n> + fpinfo->make_outerrel_subquery,\n> + params_list);\n> /* Deparse outer relation */\n> - initStringInfo(&join_sql_o);\n> - deparseFromExprForRel(&join_sql_o, root, rel_o, true, params_list);\n> + deparseRangeTblRef(buf, root,\n> + fpinfo->outerrel,\n> + fpinfo->make_outerrel_subquery,\n> + params_list);\n>\n> It removes few variables, instead directly accesses the members from fpinfo.\n> Also, deparses the individual relations in the provided buffer directly, rather\n> than in separate buffers in the original code. Instead of this, all you had to\n> do was replace a call\n> - deparseFromExprForRel(&join_sql_o, root, rel_o, true, params_list);\n> with\n> + deparseRangeTblRef(&join_sql_o, root, rel_o,\n> fpinfo->make_outerrel_subquery, params_list)\n> Similarly for the inner relation. Again, the changes you have done might have\n> been good, if done in the original code, but doing those in this patch just\n> creates distractions and increases the size of the patch.\n\nI did that for efficiency because deparsing the relation directly into \nthe given buffer would save cycles, but I agree that that is another \nissue. Will remove that change, and leave that for another patch.\n\n> 8. Why have you changed the name of variable here?\n> - if (use_alias)\n> + if (add_rel_alias)\n\nSorry, I forgot to remove this change from the patch. Will fix.\n\n> 9. In deparseRangeTblRef(), the targetlist for the subquery is obtained by\n> calling build_tlist_to_deparse(), but appendSubselectAlias() constructs the\n> column aliases based on foreignrel->reltarget->exprs. Those two are usually\n> same, but there is no code which guarantees that. Instead, pass tlist to\n> appendSubselectAlias() or simply pass the number of entries in that list\n> (list_length(tlist), which is all it needs. OR use foreignrel->reltarget->exprs\n> directly instead of calling build_tlist_to_deparse(). Similar problem exists\n> with getSubselectAliasInfo().\n\nActually, the patch guarantees that since in that case (1) \nforeignrel->reltarget->exprs doesn't contain any PHVs (note that passing \nthe following check in foreign_join_ok implies that \nforeignrel->reltarget->exprs of the input rels for a given foreign join \ndoesn't contain any PHVs and (2) we have fpinfo->local_conds == NIL, so \nbuild_tlist_to_deparse() would produce a tlist equivalent to the \nforeignrel->reltarget->exprs. But as you proposed, to make the code \neasier to understand, I will change to use foreignrel->reltarget->exprs \ndirectly.\n\n /*\n * deparseExplicitTargetList() isn't smart enough to handle \nanything other\n * than a Var. In particular, if there's some PlaceHolderVar that \nwould\n * need to be evaluated within this join tree (because there's an upper\n * reference to a quantity that may go to NULL as a result of an outer\n * join), then we can't try to push the join down because we'll \nfail when\n * we get to deparseExplicitTargetList(). However, a \nPlaceHolderVar that\n * needs to be evaluated *at the top* of this join tree is OK, \nbecause we\n * can do that locally after fetching the results from the remote side.\n */\n relids = joinrel->relids;\n foreach(lc, root->placeholder_list)\n {\n PlaceHolderInfo *phinfo = lfirst(lc);\n\n if (bms_is_subset(phinfo->ph_eval_at, relids) &&\n bms_nonempty_difference(relids, phinfo->ph_eval_at))\n return false;\n }\n\n> 10. Deparsing a query in the form of a subquery makes it hard-to-read. The\n> original code aimed at avoiding a subquery. Also, this change has created many\n> expected output changes, which again seem unnecessary. In fact, having the\n> pushable join clauses of an inner join in ON clause, which is closer to JOIN\n> clause, is better than having them farther in the WHERE clause.\n> - /*\n> - * For an inner join, all restrictions can be treated alike. Treating the\n> - * pushed down conditions as join conditions allows a top level full outer\n> - * join to be deparsed without requiring subqueries.\n> - */\n> - if (jointype == JOIN_INNER)\n> - {\n> - Assert(!fpinfo->joinclauses);\n> - fpinfo->joinclauses = fpinfo->remote_conds;\n> - fpinfo->remote_conds = NIL;\n> - }\n\nI added this change for another patch that I proposed for extending the \nDML pushdown in 9.6 so that it can perform an update/delete on a join \nremotely. To create a remote query for that, I think we need to pull up \ninner join conditions mentioning the target relation in the JOIN/ON \nclauses into the WHERE clause. But I have to admit that's unrelated to \nthis patch, so I will leave that as-is.\n\n> 11. I have reworded following comment and restructured the code that follows it\n> in the attached patch.\n> + /*\n> + * Set the subquery information. If the relation performs a full outer\n> + * join and if the input relations have non-NIL remote_conds, the input\n> + * relations need to be deparsed as a subquery.\n> + */\n> The code esp. the if .. else .. block followed by another one, made it a bit\n> unreadable. Hence I restructured it so that it's readable and doesn't require\n> variable \"relids\" from earlier code, which was being reused. Let me know if\n> this change looks good.\n\nThat's a good idea. Thanks!\n\n> 12. The code below deletes the condition, which is understandable.\n> - if (fpinfo_i->remote_conds || fpinfo_o->remote_conds)\n> But why does it add a new unrelated condition here? What the comment claims,\n> looks like an existing bug, unrelated to the patch. Can you please give an\n> example? If that it's an existing bug, it should be fixed as a separate patch.\n> I don't understand the relation of this code with what the patch is doing.\n> + /*\n> + * We can't do anything here, and if there are any non-Vars in the\n> + * outerrel/innerrel's reltarget, give up pushing down this join,\n> + * because the deparsing logic can't support such a case currently.\n> + */\n> + if (reltarget_has_non_vars(outerrel))\n> + return false;\n> + if (reltarget_has_non_vars(innerrel))\n> return false;\n\nI thought the current deparsing logic wouldn't work well if the target \nlist of a given subquery contained eg, system columns other than ctid \nand oid, but I noticed that I was wrong; it can, so we don't need this \nrestriction. Will remove. (I don't think there is any bug.)\n\n> 13. The comment below is missing the main purpose i.e. creating a a unique\n> alias, in case the relation gets converted into a subquery. Lowest or highest\n> relid will create a unique alias at given level of join and that would be more\n> future proof. If we decide to consider paths for every join order, following\n> comment will no more be true.\n> + /*\n> + * Set the relation index. This is defined as the position of this\n> + * joinrel in the join_rel_list list plus the length of the rtable list.\n> + * Note that since this joinrel is at the end of the list when we are\n> + * called, we can get the position by list_length.\n> + */\n> + fpinfo->relation_index =\n> + list_length(root->parse->rtable) + list_length(root->join_rel_list);\n\nI don't agree on that point. As I said before, the approach using the \nlowest/highest relid would make a remote query having nested subqueries \ndifficult to read. And even if we decided to consider paths for every \njoin order, we could define the relation_index safely, by modifying \npostgresGetForeignJoinPaths so that it (1) sets the relation_index the \nproposed way at the first time it is called and (2) avoids setting the \nrelation_index after the first call, by checking to see \nfpinfo->relation_index > 0.\n\n> 14. The function name talks about non-vars but the Assert few lines below\n> asserts that every node there is a Var. Need better naming.\n> +reltarget_has_non_vars(RelOptInfo *foreignrel)\n> +{\n> + ListCell *lc;\n> +\n> + foreach(lc, foreignrel->reltarget->exprs)\n> + {\n> + Var *var = (Var *) lfirst(lc);\n> +\n> + Assert(IsA(var, Var));\n> And also an explanation for the claim\n> + * Note: currently deparseExplicitTargetList can't properly handle such Vars.\n\nWill remove this function for the reason described in #12.\n\n> 15. While deparsing every Var, we are descending the RelOptInfo hierarchy.\n\nRight. I think that not only avoids creating redundant data to find the \nalias of a given Var node but also would be able to find it in optimal cost.\n\n> Instead of that, we should probably gather all the alias information once and\n> store it in context as an array indexed by relid. While deparsing a Var we can\n> get the targetlist and alias for a given relation by using var->varno as index.\n> And then search for given Var node in the targetlist to deparse the column name\n> by its position in the targetlist. For the relations that are not converted\n> into subqueries, this array will not hold any information and the Var will be\n> deparsed as it's being done today.\n\nSorry, I don't understand this part. How does that work when creating a \nremote query having nested subqueries? Is that able to search for a \ngiven Var node efficiently?\n\n> Testing\n> -------\n> 1. The code is changing deparser but doesn't have testcases\n> testing impact of the code. For example, we should have testcases with remote\n> query deparsed as nested subqueries or nested subqueries within subqueries and\n> so on May be testcases where a relation deeper in the RelOptInfo hierarchy has\n> conditions but it's immediate upper relation does not have those.\n\nWill do.\n\n> 2. The only testcase added by this patch, copies an existing case and adds a\n> whole row reference to one of the relations being joined. Instead we could add\n> that whole-row reference to the existing testcase itself.\n\nWill do.\n\nBest regards,\nEtsuro Fujita\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, 24 Oct 2016 17:59:03 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":">\n>> 13. The comment below is missing the main purpose i.e. creating a a unique\n>> alias, in case the relation gets converted into a subquery. Lowest or\n>> highest\n>> relid will create a unique alias at given level of join and that would be\n>> more\n>> future proof. If we decide to consider paths for every join order,\n>> following\n>> comment will no more be true.\n>> + /*\n>> + * Set the relation index. This is defined as the position of this\n>> + * joinrel in the join_rel_list list plus the length of the rtable\n>> list.\n>> + * Note that since this joinrel is at the end of the list when we are\n>> + * called, we can get the position by list_length.\n>> + */\n>> + fpinfo->relation_index =\n>> + list_length(root->parse->rtable) +\n>> list_length(root->join_rel_list);\n>\n>\n> I don't agree on that point. As I said before, the approach using the\n> lowest/highest relid would make a remote query having nested subqueries\n> difficult to read. And even if we decided to consider paths for every join\n> order, we could define the relation_index safely, by modifying\n> postgresGetForeignJoinPaths so that it (1) sets the relation_index the\n> proposed way at the first time it is called and (2) avoids setting the\n> relation_index after the first call, by checking to see\n> fpinfo->relation_index > 0.\n\nOK. Although, the alias which contains a number, which apparently\ndoesn't show any relation with the relation (no pun intended :)) being\ndeparsed, won't make much sense in the deparsed query. But I am fine\nleaving this to the committer's judgement. May be you want to add an\nassert here making sure that relation_index is set only once.\nSomething like Assert(fpinfo->relation_index == 0) just before setting\nit.\n\n>\n>> 14. The function name talks about non-vars but the Assert few lines below\n>> asserts that every node there is a Var. Need better naming.\n>> +reltarget_has_non_vars(RelOptInfo *foreignrel)\n>> +{\n>> + ListCell *lc;\n>> +\n>> + foreach(lc, foreignrel->reltarget->exprs)\n>> + {\n>> + Var *var = (Var *) lfirst(lc);\n>> +\n>> + Assert(IsA(var, Var));\n>> And also an explanation for the claim\n>> + * Note: currently deparseExplicitTargetList can't properly handle such\n>> Vars.\n>\n>\n> Will remove this function for the reason described in #12.\n>\n>> 15. While deparsing every Var, we are descending the RelOptInfo hierarchy.\n>\n>\n> Right. I think that not only avoids creating redundant data to find the\n> alias of a given Var node but also would be able to find it in optimal cost.\n>\n>> Instead of that, we should probably gather all the alias information once\n>> and\n>> store it in context as an array indexed by relid. While deparsing a Var we\n>> can\n>> get the targetlist and alias for a given relation by using var->varno as\n>> index.\n>> And then search for given Var node in the targetlist to deparse the column\n>> name\n>> by its position in the targetlist. For the relations that are not\n>> converted\n>> into subqueries, this array will not hold any information and the Var will\n>> be\n>> deparsed as it's being done today.\n>\n>\n> Sorry, I don't understand this part. How does that work when creating a\n> remote query having nested subqueries? Is that able to search for a given\n> Var node efficiently?\n\nFor every relation that is deparsed as a subquery, we will create a\nseparate context. Each such context will have an array described\nabove. This array will contain the targetlist and aliases for all the\nrelations, covered by that relation, that are required to be deparsed\nas subqueries. These arrays bubble up to topmost relation that is not\nrequired to be deparsed as subquery. When a relation is required to be\ndeparsed as a subquery, its immediate upper relation sets the alias\nand targetlist chosen for that relation at all the indexes\ncorresponding to all the base relations covered by that relation. For\nexample, let's assume that a relation (1, 2, 3) is required to be\ndeparsed as subquery for an immediate upper relation (1, 2, 3, 4, 5)\n(thus the other joining relation being (4,5)). While deparsing for\nrelation (1,2,3,4,5), the context will contain a 5 element array, with\npositions 1, 2, 3 filled by same targetlist and alias whereas\npositions 4 and 5 will not be filled as those relations are not\ndeparsed as subqueries. Let's assume in relation (1, 2, 3), (1, 3) in\nturn requires subquery but (2) does not. Thus the context created\nwhile deparsing (1, 2, 3) will have a 3 element array with positions 1\nand 3 containing the same targetlist and alias, where as position 2\nwill be empty. When deparsing a Var node with varno = N and varattno =\nm, if the nth position in the array in the context is empty, that Var\nnode will be deparsed as rN.<column name>. But if that position is has\nalias sZ, then we search for that Var node in the targetlist and if\nit's found at kth position in the targetlist, we will deparse it as\nsZ.ck. The search in the targetlist can be performed using\ntlist_member, and then fetching the position by TargetEntry::resno.\nThis does not require any recursion and thus saves stack space and\nsome CPU cycles required for recursion. I guess, the arrays need to be\ncomputed only once for any relation when the query for that relation\nis deparsed the first time.\n\nThanks for considering other suggestions.\n\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 25 Oct 2016 15:28:07 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/10/25 18:58, Ashutosh Bapat wrote:\n\nYou wrote:\n>>> 13. The comment below is missing the main purpose i.e. creating a a unique\n>>> alias, in case the relation gets converted into a subquery. Lowest or\n>>> highest\n>>> relid will create a unique alias at given level of join and that would be\n>>> more\n>>> future proof. If we decide to consider paths for every join order,\n>>> following\n>>> comment will no more be true.\n>>> + /*\n>>> + * Set the relation index. This is defined as the position of this\n>>> + * joinrel in the join_rel_list list plus the length of the rtable\n>>> list.\n>>> + * Note that since this joinrel is at the end of the list when we are\n>>> + * called, we can get the position by list_length.\n>>> + */\n>>> + fpinfo->relation_index =\n>>> + list_length(root->parse->rtable) +\n>>> list_length(root->join_rel_list);\n\nI wrote:\n>> I don't agree on that point. As I said before, the approach using the\n>> lowest/highest relid would make a remote query having nested subqueries\n>> difficult to read. And even if we decided to consider paths for every join\n>> order, we could define the relation_index safely, by modifying\n>> postgresGetForeignJoinPaths so that it (1) sets the relation_index the\n>> proposed way at the first time it is called and (2) avoids setting the\n>> relation_index after the first call, by checking to see\n>> fpinfo->relation_index > 0.\n\n> OK. Although, the alias which contains a number, which apparently\n> doesn't show any relation with the relation (no pun intended :)) being\n> deparsed, won't make much sense in the deparsed query. But I am fine\n> leaving this to the committer's judgement.\n\nFine with me.\n\n> May be you want to add an\n> assert here making sure that relation_index is set only once.\n> Something like Assert(fpinfo->relation_index == 0) just before setting\n> it.\n\nWill do.\n\n>>> 15. While deparsing every Var, we are descending the RelOptInfo hierarchy.\n\n>> Right. I think that not only avoids creating redundant data to find the\n>> alias of a given Var node but also would be able to find it in optimal cost.\n\n>>> Instead of that, we should probably gather all the alias information once\n>>> and\n>>> store it in context as an array indexed by relid. While deparsing a Var we\n>>> can\n>>> get the targetlist and alias for a given relation by using var->varno as\n>>> index.\n>>> And then search for given Var node in the targetlist to deparse the column\n>>> name\n>>> by its position in the targetlist. For the relations that are not\n>>> converted\n>>> into subqueries, this array will not hold any information and the Var will\n>>> be\n>>> deparsed as it's being done today.\n\n>> Sorry, I don't understand this part. How does that work when creating a\n>> remote query having nested subqueries? Is that able to search for a given\n>> Var node efficiently?\n\n> For every relation that is deparsed as a subquery, we will create a\n> separate context. Each such context will have an array described\n> above. This array will contain the targetlist and aliases for all the\n> relations, covered by that relation, that are required to be deparsed\n> as subqueries. These arrays bubble up to topmost relation that is not\n> required to be deparsed as subquery. When a relation is required to be\n> deparsed as a subquery, its immediate upper relation sets the alias\n> and targetlist chosen for that relation at all the indexes\n> corresponding to all the base relations covered by that relation.\n\nThanks for the explanation!\n\n> For\n> example, let's assume that a relation (1, 2, 3) is required to be\n> deparsed as subquery for an immediate upper relation (1, 2, 3, 4, 5)\n> (thus the other joining relation being (4,5)). While deparsing for\n> relation (1,2,3,4,5), the context will contain a 5 element array, with\n> positions 1, 2, 3 filled by same targetlist and alias whereas\n> positions 4 and 5 will not be filled as those relations are not\n> deparsed as subqueries.\n\nSorry, I don't understand this. In that case, the immediate upper \nrelation (1, 2, 3, 4, 5) would need to fill the targetlist and aliases \nfor *the join relation (1, 2, 3) somewhere*, not the targetlist and \naliases for each of the component relations 1, 2, and 3, because the \njoin relation is deparsed as a subquery. Maybe I'm missing something, \nthough.\n\n> Let's assume in relation (1, 2, 3), (1, 3) in\n> turn requires subquery but (2) does not. Thus the context created\n> while deparsing (1, 2, 3) will have a 3 element array with positions 1\n> and 3 containing the same targetlist and alias, where as position 2\n> will be empty.\n\n> When deparsing a Var node with varno = N and varattno =\n> m, if the nth position in the array in the context is empty, that Var\n> node will be deparsed as rN.<column name>.\n\nWhat happens when deparsing eg, a Var with varno = 2 at the topmost \nrelation (1, 2, 3, 4, 5)? The second position of the array is empty, \nbut the join relation (1, 2, 3) is deparsed as a subquery, so the Var \nshould be deparsed as an alias of an output column of the subquery at \nthe topmost relation, I think.\n\n> But if that position is has\n> alias sZ, then we search for that Var node in the targetlist and if\n> it's found at kth position in the targetlist, we will deparse it as\n> sZ.ck. The search in the targetlist can be performed using\n> tlist_member, and then fetching the position by TargetEntry::resno.\n\n> This does not require any recursion and thus saves stack space and\n> some CPU cycles required for recursion.\n\nIs that true?\n\n> some CPU cycles required for recursion. I guess, the arrays need to be\n> computed only once for any relation when the query for that relation\n> is deparsed the first time.\n\nDoes this algorithm extend to the case where we consider paths for every \njoin order?\n\n> Thanks for considering other suggestions.\n\nYour suggestions are really helpful to improve the patch. Thanks!\n\nBest regards,\nEtsuro Fujita\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, 25 Oct 2016 21:21:42 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":">\n>\n>> For every relation that is deparsed as a subquery, we will create a\n>> separate context. Each such context will have an array described\n>> above. This array will contain the targetlist and aliases for all the\n>> relations, covered by that relation, that are required to be deparsed\n>> as subqueries. These arrays bubble up to topmost relation that is not\n>> required to be deparsed as subquery. When a relation is required to be\n>> deparsed as a subquery, its immediate upper relation sets the alias\n>> and targetlist chosen for that relation at all the indexes\n>> corresponding to all the base relations covered by that relation.\n>\n>\n> Thanks for the explanation!\n>\n>> For\n>> example, let's assume that a relation (1, 2, 3) is required to be\n>> deparsed as subquery for an immediate upper relation (1, 2, 3, 4, 5)\n>> (thus the other joining relation being (4,5)). While deparsing for\n>> relation (1,2,3,4,5), the context will contain a 5 element array, with\n>> positions 1, 2, 3 filled by same targetlist and alias whereas\n>> positions 4 and 5 will not be filled as those relations are not\n>> deparsed as subqueries.\n>\n>\n> Sorry, I don't understand this. In that case, the immediate upper relation\n> (1, 2, 3, 4, 5) would need to fill the targetlist and aliases for *the join\n> relation (1, 2, 3) somewhere*, not the targetlist and aliases for each of\n> the component relations 1, 2, and 3, because the join relation is deparsed\n> as a subquery. Maybe I'm missing something, though.\n\nThe description above does not specify \"targetlist and alias\" for each\nof (1, 2, 3). The array in the context will have positions 1, 2, 3\nfilled with *same* alias and targetlist which is derived from relation\n(1, 2, 3).\n\n>\n>> Let's assume in relation (1, 2, 3), (1, 3) in\n>> turn requires subquery but (2) does not. Thus the context created\n>> while deparsing (1, 2, 3) will have a 3 element array with positions 1\n>> and 3 containing the same targetlist and alias, where as position 2\n>> will be empty.\n>\n>\n>> When deparsing a Var node with varno = N and varattno =\n>> m, if the nth position in the array in the context is empty, that Var\n>> node will be deparsed as rN.<column name>.\n>\n>\n> What happens when deparsing eg, a Var with varno = 2 at the topmost relation\n> (1, 2, 3, 4, 5)? The second position of the array is empty, but the join\n> relation (1, 2, 3) is deparsed as a subquery, so the Var should be deparsed\n> as an alias of an output column of the subquery at the topmost relation, I\n> think.\n\nposition 2 will not be empty, it will be filled by the alias and\ntargetlist derived from relation (1, 2, 3).\n\n>\n>> But if that position is has\n>> alias sZ, then we search for that Var node in the targetlist and if\n>> it's found at kth position in the targetlist, we will deparse it as\n>> sZ.ck. The search in the targetlist can be performed using\n>> tlist_member, and then fetching the position by TargetEntry::resno.\n>\n>\n>> This does not require any recursion and thus saves stack space and\n>> some CPU cycles required for recursion.\n>\n>\n> Is that true?\n\nYes, unless you explain why is that false.\n\n>\n>> some CPU cycles required for recursion. I guess, the arrays need to be\n>> computed only once for any relation when the query for that relation\n>> is deparsed the first time.\n>\n>\n> Does this algorithm extend to the case where we consider paths for every\n> join order?\n\nYes, if we store the information about which of relations need\nsubquery and which don't for every join order.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 26 Oct 2016 12:41:44 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/10/26 16:11, Ashutosh Bapat wrote:\n\nYou wrote:\n>>> For\n>>> example, let's assume that a relation (1, 2, 3) is required to be\n>>> deparsed as subquery for an immediate upper relation (1, 2, 3, 4, 5)\n>>> (thus the other joining relation being (4,5)). While deparsing for\n>>> relation (1,2,3,4,5), the context will contain a 5 element array, with\n>>> positions 1, 2, 3 filled by same targetlist and alias whereas\n>>> positions 4 and 5 will not be filled as those relations are not\n>>> deparsed as subqueries.\n\nI wrote:\n>> Sorry, I don't understand this. In that case, the immediate upper relation\n>> (1, 2, 3, 4, 5) would need to fill the targetlist and aliases for *the join\n>> relation (1, 2, 3) somewhere*, not the targetlist and aliases for each of\n>> the component relations 1, 2, and 3, because the join relation is deparsed\n>> as a subquery. Maybe I'm missing something, though.\n\n> The description above does not specify \"targetlist and alias\" for each\n> of (1, 2, 3). The array in the context will have positions 1, 2, 3\n> filled with *same* alias and targetlist which is derived from relation\n> (1, 2, 3).\n\nOK\n\n>>> Let's assume in relation (1, 2, 3), (1, 3) in\n>>> turn requires subquery but (2) does not. Thus the context created\n>>> while deparsing (1, 2, 3) will have a 3 element array with positions 1\n>>> and 3 containing the same targetlist and alias, where as position 2\n>>> will be empty.\n\n>>> When deparsing a Var node with varno = N and varattno =\n>>> m, if the nth position in the array in the context is empty, that Var\n>>> node will be deparsed as rN.<column name>.\n\n>> What happens when deparsing eg, a Var with varno = 2 at the topmost relation\n>> (1, 2, 3, 4, 5)? The second position of the array is empty, but the join\n>> relation (1, 2, 3) is deparsed as a subquery, so the Var should be deparsed\n>> as an alias of an output column of the subquery at the topmost relation, I\n>> think.\n\n> position 2 will not be empty, it will be filled by the alias and\n> targetlist derived from relation (1, 2, 3).\n\nOK\n\n>>> But if that position is has\n>>> alias sZ, then we search for that Var node in the targetlist and if\n>>> it's found at kth position in the targetlist, we will deparse it as\n>>> sZ.ck. The search in the targetlist can be performed using\n>>> tlist_member, and then fetching the position by TargetEntry::resno.\n\n>>> This does not require any recursion and thus saves stack space and\n>>> some CPU cycles required for recursion.\n\n>> Is that true?\n\n> Yes, unless you explain why is that false.\n\nOK, that would be true, I think.\n\n>>> I guess, the arrays need to be\n>>> computed only once for any relation when the query for that relation\n>>> is deparsed the first time.\n\n>> Does this algorithm extend to the case where we consider paths for every\n>> join order?\n\n> Yes, if we store the information about which of relations need\n> subquery and which don't for every join order.\n\nHmm. Sorry, I'm not so excited about this proposal. I think (1) that \nis solving a problem that hasn't been proven to be a problem, (2) that \nwould complicate the deparser logic, and (3) the cost of creating this \narray for each relation by the bottom-up method while deparsing a remote \nquery would be not small (especially when the query is large), so that \nmight need more cycles for deparsing the query than what I proposed when \nuse_remote_estimate=false. So, I'd like to go with what I proposed, at \nleast as the first cut.\n\nBest regards,\nEtsuro Fujita\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 Oct 2016 18:00:42 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":">\n>>>> I guess, the arrays need to be\n>>>> computed only once for any relation when the query for that relation\n>>>> is deparsed the first time.\n>\n>\n>>> Does this algorithm extend to the case where we consider paths for every\n>>> join order?\n>\n>\n>> Yes, if we store the information about which of relations need\n>> subquery and which don't for every join order.\n>\n>\n> Hmm. Sorry, I'm not so excited about this proposal. I think (1) that is\n> solving a problem that hasn't been proven to be a problem, (2) that would\n> complicate the deparser logic, and (3) the cost of creating this array for\n> each relation by the bottom-up method while deparsing a remote query would\n> be not small (especially when the query is large), so that might need more\n> cycles for deparsing the query than what I proposed when\n> use_remote_estimate=false. So, I'd like to go with what I proposed, at\n> least as the first cut.\n\nThe arrays will need to computed only when there is at least one\nrelation required to be deparsed as subquery. Every relation above the\nrelation which is converted into subquery requires the array. Each\narray will be N * size of pointer long where N is the largest relid\ncovered by that relation. N will vary across the RelOptInfo hierarchy.\nAll that array holds is targetlist pointer which are required to\ncomputed anyway. So, the amount of memory is bounded by N * size of\npointer * (2N - 1), which is way lesser than what we use in other\nplaces.\n\nYour patch calls isSubqueryExpr() recursively for every Var in the\ntargetlist, which can be many for foreign tables with many columns.\nFor every such Var it may need to reach upto the relation which is\nconverted into subquery, which can as bad as reaching every base\nrelation. So, it looks like the number of recursive calls to\nisSubqueryExpr() is bounded by V * N (i.e. worst case depth of the\nRelOptInfo tree), which can be quite costly. If use_remote_estimates\nis true, we do this for every intermediate relation and for every path\ncreated. That isn't very performance efficient.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 26 Oct 2016 14:55:03 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/10/26 18:25, Ashutosh Bapat wrote:\n\n> Your patch calls isSubqueryExpr() recursively for every Var in the\n> targetlist, which can be many for foreign tables with many columns.\n> For every such Var it may need to reach upto the relation which is\n> converted into subquery, which can as bad as reaching every base\n> relation. So, it looks like the number of recursive calls to\n> isSubqueryExpr() is bounded by V * N (i.e. worst case depth of the\n> RelOptInfo tree), which can be quite costly. If use_remote_estimates\n> is true, we do this for every intermediate relation and for every path\n> created. That isn't very performance efficient.\n\nIn practice, the search cost would be negligible compared to the cost of \nexplaining/executing the query.\n\nMy concern about your proposal is: it might not be worth complicating \nthe code to solve a problem that is actually not a problem in practice.\n\nBest regards,\nEtsuro Fujita\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 Oct 2016 19:05:37 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Wed, Oct 26, 2016 at 3:35 PM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> On 2016/10/26 18:25, Ashutosh Bapat wrote:\n>\n>> Your patch calls isSubqueryExpr() recursively for every Var in the\n>> targetlist, which can be many for foreign tables with many columns.\n>> For every such Var it may need to reach upto the relation which is\n>> converted into subquery, which can as bad as reaching every base\n>> relation. So, it looks like the number of recursive calls to\n>> isSubqueryExpr() is bounded by V * N (i.e. worst case depth of the\n>> RelOptInfo tree), which can be quite costly. If use_remote_estimates\n>> is true, we do this for every intermediate relation and for every path\n>> created. That isn't very performance efficient.\n>\n>\n> In practice, the search cost would be negligible compared to the cost of\n> explaining/executing the query.\n>\n> My concern about your proposal is: it might not be worth complicating the\n> code to solve a problem that is actually not a problem in practice.\n>\n\nTo me the current code looks complicated esp. because of the recursion\ninvolved and usage of out parameters to isSubqueryExpr(). My\nsuggestion goes inline with the current method of deparsing a Var.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 26 Oct 2016 16:23:29 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/10/26 19:53, Ashutosh Bapat wrote:\n> On Wed, Oct 26, 2016 at 3:35 PM, Etsuro Fujita\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\n\n>> In practice, the search cost would be negligible compared to the cost of\n>> explaining/executing the query.\n>>\n>> My concern about your proposal is: it might not be worth complicating the\n>> code to solve a problem that is actually not a problem in practice.\n\n> To me the current code looks complicated esp. because of the recursion\n> involved and usage of out parameters to isSubqueryExpr().\n\nI don't think so. isSubqueryExpr is prety small, written in less than \n50 lines, and the code looks rather simple to me.\n\n> My\n> suggestion goes inline with the current method of deparsing a Var.\n\nYeah, I think your approach makes it easy to search for the alias to a \ngiven Var from the array you proposed. I think the complexity of your \napproach would be in extra work for building and maintaining the array \nwhile deparsing the query. I think that would probably need more \ninvasive and much larger changes to the existing code than what I proposed.\n\nI think this issue is optional, so I'd like to propose to leave this for \nthe committer's judge.\n\nBest regards,\nEtsuro Fujita\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, 27 Oct 2016 16:16:41 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Thu, Oct 27, 2016 at 12:46 PM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> On 2016/10/26 19:53, Ashutosh Bapat wrote:\n>>\n>> On Wed, Oct 26, 2016 at 3:35 PM, Etsuro Fujita\n>> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>\n>\n>>> In practice, the search cost would be negligible compared to the cost of\n>>> explaining/executing the query.\n>>>\n>>> My concern about your proposal is: it might not be worth complicating the\n>>> code to solve a problem that is actually not a problem in practice.\n>\n>\n>> To me the current code looks complicated esp. because of the recursion\n>> involved and usage of out parameters to isSubqueryExpr().\n>\n>\n> I don't think so. isSubqueryExpr is prety small, written in less than 50\n> lines, and the code looks rather simple to me.\n>\n>> My\n>> suggestion goes inline with the current method of deparsing a Var.\n>\n>\n> Yeah, I think your approach makes it easy to search for the alias to a given\n> Var from the array you proposed. I think the complexity of your approach\n> would be in extra work for building and maintaining the array while\n> deparsing the query. I think that would probably need more invasive and\n> much larger changes to the existing code than what I proposed.\n>\n> I think this issue is optional, so I'd like to propose to leave this for the\n> committer's judge.\n>\n\nFine with me.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 27 Oct 2016 14:36:04 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/10/27 18:06, Ashutosh Bapat wrote:\n> On Thu, Oct 27, 2016 at 12:46 PM, Etsuro Fujita\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>> On 2016/10/26 19:53, Ashutosh Bapat wrote:\n\n>>> On Wed, Oct 26, 2016 at 3:35 PM, Etsuro Fujita\n>>>> My concern about your proposal is: it might not be worth complicating the\n>>>> code to solve a problem that is actually not a problem in practice.\n\n>>> To me the current code looks complicated esp. because of the recursion\n>>> involved and usage of out parameters to isSubqueryExpr().\n\n>> I don't think so. isSubqueryExpr is prety small, written in less than 50\n>> lines, and the code looks rather simple to me.\n\n>>> My\n>>> suggestion goes inline with the current method of deparsing a Var.\n\n>> Yeah, I think your approach makes it easy to search for the alias to a given\n>> Var from the array you proposed. I think the complexity of your approach\n>> would be in extra work for building and maintaining the array while\n>> deparsing the query. I think that would probably need more invasive and\n>> much larger changes to the existing code than what I proposed.\n>>\n>> I think this issue is optional, so I'd like to propose to leave this for the\n>> committer's judge.\n\n> Fine with me.\n\nOK\n\nHere is the updated version, which includes the restructuring you \nproposed. Other than the above issue and the alias issue we discussed, \nI addressed all your comments except one on testing; I tried to add test \ncases where the remote query is deparsed as nested subqueries, but I \ncouldn't because IIUC, reduce_outer_joins reduced full joins to inner \njoins or left joins. So, I added two test cases: (1) the joining \nrelations are both base relations (actually, we already have that) and \n(2) one of the joining relations is a base relation and the other is a \njoin relation. I rebased the patch to HEAD, so I added a test case with \naggregate pushdown also.\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":"Thu, 27 Oct 2016 20:41:21 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/10/27 20:41, Etsuro Fujita wrote:\n> Here is the updated version, which includes the restructuring you\n> proposed. Other than the above issue and the alias issue we discussed,\n> I addressed all your comments except one on testing; I tried to add test\n> cases where the remote query is deparsed as nested subqueries, but I\n> couldn't because IIUC, reduce_outer_joins reduced full joins to inner\n> joins or left joins. So, I added two test cases: (1) the joining\n> relations are both base relations (actually, we already have that) and\n> (2) one of the joining relations is a base relation and the other is a\n> join relation. I rebased the patch to HEAD, so I added a test case with\n> aggregate pushdown also.\n\nI've updated another patch for evaluating PHVs remotely. Attached is an \nupdated version of the patch, which is created on top of the patch for \nsupporting deparsing subqueries. You can find test cases where a remote \njoin query is deparsed as nested subqueries.\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":"Tue, 1 Nov 2016 21:01:57 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"> Here is the updated version, which includes the restructuring you proposed.\n> Other than the above issue and the alias issue we discussed, I addressed all\n> your comments except one on testing; I tried to add test cases where the\n> remote query is deparsed as nested subqueries, but I couldn't because IIUC,\n> reduce_outer_joins reduced full joins to inner joins or left joins.\n\nNo always. It will do so in some cases as explained in the prologue of\nreduce_outer_joins()\n\n * More generally, an outer join can be reduced in strength if there is a\n * strict qual above it in the qual tree that constrains a Var from the\n * nullable side of the join to be non-null. (For FULL joins this applies\n * to each side separately.)\n\n> So, I\n> added two test cases: (1) the joining relations are both base relations\n> (actually, we already have that) and (2) one of the joining relations is a\n> base relation and the other is a join relation. I rebased the patch to\n> HEAD, so I added a test case with aggregate pushdown also.\n>\n\nThis patch again has a lot of unrelated changes, esp. in\ndeparseSelectStmtForRel(). What was earlier part of deparseSelectSql()\nand deparseFromExpr() is now flattened in deparseSelectStmtForRel(),\nwhich seems unnecessary. Or there are changes like\n\n-deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root, RelOptInfo *rel,\n- List *tlist, List *remote_conds, List *pathkeys,\n+deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root,\n+ RelOptInfo *foreignrel, List *tlist,\n+ List *remote_conds, List *pathkeys,\n\nwhich arise because rel has been renamed as foreignrel. The patch will\nwork even without such renaming.\n\nI think such refactoring, should be done in a separate patch.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 3 Nov 2016 15:22:33 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/11/03 18:52, Ashutosh Bapat wrote:\n>> Here is the updated version, which includes the restructuring you proposed.\n>> Other than the above issue and the alias issue we discussed, I addressed all\n>> your comments except one on testing; I tried to add test cases where the\n>> remote query is deparsed as nested subqueries, but I couldn't because IIUC,\n>> reduce_outer_joins reduced full joins to inner joins or left joins.\n\n> No always. It will do so in some cases as explained in the prologue of\n> reduce_outer_joins()\n>\n> * More generally, an outer join can be reduced in strength if there is a\n> * strict qual above it in the qual tree that constrains a Var from the\n> * nullable side of the join to be non-null. (For FULL joins this applies\n> * to each side separately.)\n\nHmm. Would it be necessary for this patch to include test cases for \nnested subqueries? As mentioned in a previous email, those test cases \ncan be added by the split patch that allow PHVs to be evaluated on the \nremote side.\n\n> This patch again has a lot of unrelated changes, esp. in\n> deparseSelectStmtForRel(). What was earlier part of deparseSelectSql()\n> and deparseFromExpr() is now flattened in deparseSelectStmtForRel(),\n> which seems unnecessary.\n\nIIUC, I think this change was done to address your comment (see the \ncomment #2 in [1]). Am I missing something?\n\n> Or there are changes like\n>\n> -deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root, RelOptInfo *rel,\n> - List *tlist, List *remote_conds, List *pathkeys,\n> +deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root,\n> + RelOptInfo *foreignrel, List *tlist,\n> + List *remote_conds, List *pathkeys,\n>\n> which arise because rel has been renamed as foreignrel. The patch will\n> work even without such renaming.\n\nI did that because we have a Relation named rel (the same name!) within \nthat function, to execute deparseTargetList in a base-relation case. \nAnother reason for that is because (1) that would match with other \nfunction definitions in deparse.c and (2) that would be consistent with \nthe existing function definition for that function in postgres_fdw.h.\n\nBest regards,\nEtsuro Fujita\n\n[1] \nhttps://www.postgresql.org/message-id/CAFjFpRfwoSsJr9418b2jA7g0nwagjZSWhPQPUmK9M6z5XSOAqQ%40mail.gmail.com\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, 4 Nov 2016 12:49:02 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Fri, Nov 4, 2016 at 9:19 AM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> On 2016/11/03 18:52, Ashutosh Bapat wrote:\n>>>\n>>> Here is the updated version, which includes the restructuring you\n>>> proposed.\n>>> Other than the above issue and the alias issue we discussed, I addressed\n>>> all\n>>> your comments except one on testing; I tried to add test cases where the\n>>> remote query is deparsed as nested subqueries, but I couldn't because\n>>> IIUC,\n>>> reduce_outer_joins reduced full joins to inner joins or left joins.\n>\n>\n>> No always. It will do so in some cases as explained in the prologue of\n>> reduce_outer_joins()\n>>\n>> * More generally, an outer join can be reduced in strength if there is a\n>> * strict qual above it in the qual tree that constrains a Var from the\n>> * nullable side of the join to be non-null. (For FULL joins this applies\n>> * to each side separately.)\n>\n>\n> Hmm. Would it be necessary for this patch to include test cases for nested\n> subqueries? As mentioned in a previous email, those test cases can be added\n> by the split patch that allow PHVs to be evaluated on the remote side.\n\nA patch should have testcases testing the functionality added. This\npatch adds functionality to deparse nested subqueries, so there should\nbe testcase to test it. If we can not come up with a testcase, then\nit's very much possible that the code related to that functionality is\nnot needed. PHV is a separate patch and we do not know whether it will\nbe committed or when it will be committed after this patch is\ncommitted. It's better to have self-sufficient patch.\n\n>\n>> This patch again has a lot of unrelated changes, esp. in\n>> deparseSelectStmtForRel(). What was earlier part of deparseSelectSql()\n>> and deparseFromExpr() is now flattened in deparseSelectStmtForRel(),\n>> which seems unnecessary.\n>\n>\n> IIUC, I think this change was done to address your comment (see the comment\n> #2 in [1]). Am I missing something?\n\nThere is some misunderstanding here. That comment basically says,\ndeparseRangeTblRef() duplicates code in deparseSelectStmtForRel(), so\nwe should either remove deparseRangeTblRef() altogether or should keep\nit to minimal avoiding duplication of code. What might have confused\nyou is the last sentence in that comment \"This will also make the\ncurrent changes to deparseSelectSql unnecessary.\" By current changes I\nmeant changes to deparseSelectSql() in your patch, not the one that's\nin the repository. I don't think we should flatten\ndeparseSelectStmtForRel() in this patch.\n\n>\n>> Or there are changes like\n>>\n>> -deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root, RelOptInfo\n>> *rel,\n>> - List *tlist, List *remote_conds, List *pathkeys,\n>> +deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root,\n>> + RelOptInfo *foreignrel, List *tlist,\n>> + List *remote_conds, List *pathkeys,\n>>\n>> which arise because rel has been renamed as foreignrel. The patch will\n>> work even without such renaming.\n>\n>\n> I did that because we have a Relation named rel (the same name!) within that\n> function, to execute deparseTargetList in a base-relation case.\n\nThat's because you have flattened deparseSelectStmtForRel(). If we\ndon't flatten it out, the variable name conflict doesn't arise.\n\n> Another\n> reason for that is because (1) that would match with other function\n> definitions in deparse.c and (2) that would be consistent with the existing\n> function definition for that function in postgres_fdw.h.\n>\n\nThat would be a separate patch, I guess.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 4 Nov 2016 09:38:17 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/11/04 13:08, Ashutosh Bapat wrote:\n> On Fri, Nov 4, 2016 at 9:19 AM, Etsuro Fujita\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>> On 2016/11/03 18:52, Ashutosh Bapat wrote:\n\nI wrote:\n>>>> Here is the updated version,\n\n>>>> Other than the above issue and the alias issue we discussed, I addressed\n>>>> all\n>>>> your comments except one on testing; I tried to add test cases where the\n>>>> remote query is deparsed as nested subqueries, but I couldn't because\n>>>> IIUC,\n>>>> reduce_outer_joins reduced full joins to inner joins or left joins.\n\n>>> No always. It will do so in some cases as explained in the prologue of\n>>> reduce_outer_joins()\n>>>\n>>> * More generally, an outer join can be reduced in strength if there is a\n>>> * strict qual above it in the qual tree that constrains a Var from the\n>>> * nullable side of the join to be non-null. (For FULL joins this applies\n>>> * to each side separately.)\n\nRight.\n\n>> Would it be necessary for this patch to include test cases for nested\n>> subqueries?\n\n> A patch should have testcases testing the functionality added. This\n> patch adds functionality to deparse nested subqueries, so there should\n> be testcase to test it.\n\nOK, I added such a test case.\n\n>>> This patch again has a lot of unrelated changes, esp. in\n>>> deparseSelectStmtForRel(). What was earlier part of deparseSelectSql()\n>>> and deparseFromExpr() is now flattened in deparseSelectStmtForRel(),\n>>> which seems unnecessary.\n\n>> IIUC, I think this change was done to address your comment (see the comment\n>> #2 in [1]). Am I missing something?\n\n> There is some misunderstanding here. That comment basically says,\n> deparseRangeTblRef() duplicates code in deparseSelectStmtForRel(), so\n> we should either remove deparseRangeTblRef() altogether or should keep\n> it to minimal avoiding duplication of code. What might have confused\n> you is the last sentence in that comment \"This will also make the\n> current changes to deparseSelectSql unnecessary.\" By current changes I\n> meant changes to deparseSelectSql() in your patch, not the one that's\n> in the repository. I don't think we should flatten\n> deparseSelectStmtForRel() in this patch.\n\nI noticed that I misunderstood your words. Sorry about that. I agree \nwith you, so I removed that change from the patch.\n\nAttached is an updated version of the patch.\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":"Fri, 4 Nov 2016 19:55:18 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/11/04 19:55, Etsuro Fujita wrote:\n> Attached is an updated version of the patch.\n\nI noticed that I have included an unrelated regression test in the \npatch. Attached is a patch with the test removed.\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, 7 Nov 2016 11:24:51 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/11/07 11:24, Etsuro Fujita wrote:\n> On 2016/11/04 19:55, Etsuro Fujita wrote:\n>> Attached is an updated version of the patch.\n\n> I noticed that I have included an unrelated regression test in the\n> patch. Attached is a patch with the test removed.\n\nI noticed that I inadvertently removed some changes from that patch, so \nI fixed that. Please find attached an updated version of the patch. \nI'm also attaching an updated version of another patch for evaluating \nPHVs remotely, which has been created on top of that patch. Changes to \nthe latter: I revised comments and added a bit more regression tests.\n\nBest regards,\nEtsuro Fujita\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":"Mon, 7 Nov 2016 21:20:49 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Mon, Nov 7, 2016 at 5:50 PM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> On 2016/11/07 11:24, Etsuro Fujita wrote:\n>>\n>> On 2016/11/04 19:55, Etsuro Fujita wrote:\n>>>\n>>> Attached is an updated version of the patch.\n>\n>\n>> I noticed that I have included an unrelated regression test in the\n>> patch. Attached is a patch with the test removed.\n>\n>\n> I noticed that I inadvertently removed some changes from that patch, so I\n> fixed that. Please find attached an updated version of the patch. I'm also\n> attaching an updated version of another patch for evaluating PHVs remotely,\n> which has been created on top of that patch. Changes to the latter: I\n> revised comments and added a bit more regression tests.\n\nThe patch looks in good shape now. Here are some comments. I have also\nmade several changes to comments correcting grammar, typos, style and\nat few places logic. Let me know if the patch looks good.\n\nI guess, below code\n+ if (!fpinfo->subquery_rels)\n+ return false;\ncan be changed to\n if (!bms_is_member(node->varno, fpinfo->subquery_rels))\n return false;\nAlso the return values from the recursive calls to isSubqueryExpr() can be\nreturned as is. I have included this change in the patch.\n\ndeparse.c seems to be using capitalized names for function which\nactually deparse something and an non-capitalized form for helper\nfunctions. From that perspective attached patch renames isSubqueryExpr\nas is_subquery_var() and getSubselectAliasInfo() as\nget_alias_id_for_var(). Actually both these functions accept a Var\nnode but somehow their names refer to expr.\n\nThis patch is using make_tlist_from_pathtarget() to create tlist to be\ndeparsed but search in RelOptInfo::reltarget::exprs for a given Var.\nAs long as the relations deparsed do not carry expressions, this might\nwork, but it will certainly break once we start deparsing relations\nwith expressions since the upper most query's tlist contains only\nVars. Instead, we should probably, create tlist and save it in fpinfo\nand use it later for searching (tlist_member()?). Possibly use using\nbuild_tlist_to_deparse(), to create the tlist similar so that\ntargetlist list creation logic is same for all the relations being\ndeparsed. I haven't included this change in the patch.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database Company\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, 11 Nov 2016 15:52:39 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/11/11 19:22, Ashutosh Bapat wrote:\n> The patch looks in good shape now.\n\nThanks for the review!\n\n> The patch looks in good shape now. Here are some comments. I have also\n> made several changes to comments correcting grammar, typos, style and\n> at few places logic. Let me know if the patch looks good.\n\nOK, will look into that.\n\n> I guess, below code\n> + if (!fpinfo->subquery_rels)\n> + return false;\n> can be changed to\n> if (!bms_is_member(node->varno, fpinfo->subquery_rels))\n> return false;\n> Also the return values from the recursive calls to isSubqueryExpr() can be\n> returned as is. I have included this change in the patch.\n\nWill look into that too.\n\n> deparse.c seems to be using capitalized names for function which\n> actually deparse something and an non-capitalized form for helper\n> functions.\n\nThat's not true. There is a function named classifyConditions(). The \nfunction naming in deparse.c is a bit arbitrary.\n\n> From that perspective attached patch renames isSubqueryExpr\n> as is_subquery_var() and getSubselectAliasInfo() as\n> get_alias_id_for_var(). Actually both these functions accept a Var\n> node but somehow their names refer to expr.\n\nThe reason why I named that function isSubqueryExpr is that I think that \nfunction would be soon extended so as to handle PHVs. See another patch \nfor evaluating PHVs remotely.\n\n> This patch is using make_tlist_from_pathtarget() to create tlist to be\n> deparsed but search in RelOptInfo::reltarget::exprs for a given Var.\n> As long as the relations deparsed do not carry expressions, this might\n> work, but it will certainly break once we start deparsing relations\n> with expressions since the upper most query's tlist contains only\n> Vars. Instead, we should probably, create tlist and save it in fpinfo\n> and use it later for searching (tlist_member()?). Possibly use using\n> build_tlist_to_deparse(), to create the tlist similar so that\n> targetlist list creation logic is same for all the relations being\n> deparsed. I haven't included this change in the patch.\n\nSeems like a good idea. Will revise.\n\nBest regards,\nEtsuro Fujita\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, 11 Nov 2016 20:50:43 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/11/11 20:50, Etsuro Fujita wrote:\n> On 2016/11/11 19:22, Ashutosh Bapat wrote:\n>> The patch looks in good shape now. Here are some comments. I have also\n>> made several changes to comments correcting grammar, typos, style and\n>> at few places logic. Let me know if the patch looks good.\n\n> OK, will look into that.\n\nDone. +1 for the changes you made, except for few things; (1) You added \nthe following comments to deparseSelectSql:\n\n+ /*\n+ * For a non-base relation, use the input tlist. If a base \nrelation is\n+ * being deparsed as a subquery, use input tlist, if the caller \nhas passed\n+ * one. The column aliases of the subquery are crafted based on \nthe input\n+ * tlist. If tlist is NIL, there will not be any expression \nreferring to\n+ * any of the columns of the base relation being deparsed. Hence \nit doesn't\n+ * matter whether the base relation is being deparsed as \nsubquery or not.\n+ */\n\nThe last sentence seems confusing to me. My understanding is: if a base \nrelation has tlist=NIL, then the base relation *must* be deparsed as \nordinary, not as a subquery. Maybe I'm missing something, though. (I \nleft that as-is, but I think we need to reword that to be more clear, at \nleast.)\n\n(2) You added the following comments to deparseRangeTblRef:\n\n > + * If make_subquery is true, deparse the relation as a subquery. \nOtherwise,\n > + * deparse it as relation name with alias.\n\nThe second sentence seems confusing to me, too, because it looks like \nthe relation being deparsed is assumed to be a base relation, but the \nrelation can be a join relation, which might join base relations, lower \njoin relations, and/or lower subqueries. So, I modified the sentence a bit.\n\n(3) I don't think we need this in isSubqueryExpr, so I removed it from \nthe patch:\n\n+\t/* Keep compiler happy. */\n+\treturn false;\n\nAlso, I revised comments you added a little bit.\n\n>> I guess, below code\n>> + if (!fpinfo->subquery_rels)\n>> + return false;\n>> can be changed to\n>> if (!bms_is_member(node->varno, fpinfo->subquery_rels))\n>> return false;\n>> Also the return values from the recursive calls to isSubqueryExpr()\n>> can be\n>> returned as is. I have included this change in the patch.\n\n> Will look into that too.\n\nDone. That's a good idea!\n\n>> deparse.c seems to be using capitalized names for function which\n>> actually deparse something and an non-capitalized form for helper\n>> functions.\n\n>> From that perspective attached patch renames isSubqueryExpr\n>> as is_subquery_var() and getSubselectAliasInfo() as\n>> get_alias_id_for_var(). Actually both these functions accept a Var\n>> node but somehow their names refer to expr.\n\nOK, I changed isSubqueryExpr to is_subquery_expr; I kept to refer to \nexpr because I think we would soon extend that function so that it can \nhandle PHVs, as I said upthread. For getSubselectAliasInfo, I changed \nthe name to get_subselect_alias_id, because (1) the word \"alias\" seems \ngeneral and (2) the \"for_var\" part would make the name a bit long.\n\n>> This patch is using make_tlist_from_pathtarget() to create tlist to be\n>> deparsed but search in RelOptInfo::reltarget::exprs for a given Var.\n>> As long as the relations deparsed do not carry expressions, this might\n>> work, but it will certainly break once we start deparsing relations\n>> with expressions since the upper most query's tlist contains only\n>> Vars. Instead, we should probably, create tlist and save it in fpinfo\n>> and use it later for searching (tlist_member()?). Possibly use using\n>> build_tlist_to_deparse(), to create the tlist similar so that\n>> targetlist list creation logic is same for all the relations being\n>> deparsed. I haven't included this change in the patch.\n\n> Seems like a good idea. Will revise.\n\nDone. I modified the patch as proposed; create the tlist by \nbuild_tlist_to_deparse in foreign_join_ok, if needed, and search the \ntlist by tlist_member. I also added a new member \"tlist\" to \nPgFdwRelationInfo to save the tlist created in foreign_join_ok.\n\nAnother idea on the \"tlist\" member would be to save a tlist created for \nEXPLAIN into that member in estimate_patch_cost_size, so that we don't \nneed to generate the tlist again in postgresGetForeignPlan, when \nuse_remote_estimate=true. But I'd like to leave that for another patch.\n\nPlease find attached an updated version of the patch.\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":"Tue, 15 Nov 2016 21:23:00 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"Thanks.\n\n> except for few things; (1) You added the\n> following comments to deparseSelectSql:\n>\n> + /*\n> + * For a non-base relation, use the input tlist. If a base relation\n> is\n> + * being deparsed as a subquery, use input tlist, if the caller has\n> passed\n> + * one. The column aliases of the subquery are crafted based on the\n> input\n> + * tlist. If tlist is NIL, there will not be any expression\n> referring to\n> + * any of the columns of the base relation being deparsed. Hence it\n> doesn't\n> + * matter whether the base relation is being deparsed as subquery or\n> not.\n> + */\n>\n> The last sentence seems confusing to me. My understanding is: if a base\n> relation has tlist=NIL, then the base relation *must* be deparsed as\n> ordinary, not as a subquery. Maybe I'm missing something, though. (I left\n> that as-is, but I think we need to reword that to be more clear, at least.)\n>\n\nHmm, I agree. I think the comment should just say, \"Use tlist to\ncreate the SELECT clause if one has been provided. For a base relation\nwith tlist = NIL, check attrs_needed information.\". Does that sound\ngood?\n\n> (2) You added the following comments to deparseRangeTblRef:\n>\n>> + * If make_subquery is true, deparse the relation as a subquery.\n>> Otherwise,\n>> + * deparse it as relation name with alias.\n>\n> The second sentence seems confusing to me, too, because it looks like the\n> relation being deparsed is assumed to be a base relation, but the relation\n> can be a join relation, which might join base relations, lower join\n> relations, and/or lower subqueries. So, I modified the sentence a bit.\n>\n\nOK.\n\n> (3) I don't think we need this in isSubqueryExpr, so I removed it from the\n> patch:\n>\n> + /* Keep compiler happy. */\n> + return false;\n>\n\nDoesn't that cause compiler warning, saying \"non-void function\nreturning nothing\" or something like that. Actually, the \"if\n(bms_is_member(node->varno, outerrel->relids))\" ends with a \"return\"\nalways. Hence we don't need to encapsulate the code in \"else\" block in\nelse { }. It could be taken outside.\n\n\n>\n>\n> OK, I changed isSubqueryExpr to is_subquery_expr; I kept to refer to expr\n> because I think we would soon extend that function so that it can handle\n> PHVs, as I said upthread. For getSubselectAliasInfo, I changed the name to\n> get_subselect_alias_id, because (1) the word \"alias\" seems general and (2)\n> the \"for_var\" part would make the name a bit long.\n\nis_subquery_expr(Var *node -- that looks odd. Either it should\nis_subquery_var(Var * ... OR it should be is_subquery_expr(Expr * ...\n. I would prefer the first one, since that's what current patch is\ndoing. When we introduce PHVs, we may change it, if required.\n\n\n>\n>\n> Done. I modified the patch as proposed; create the tlist by\n> build_tlist_to_deparse in foreign_join_ok, if needed, and search the tlist\n> by tlist_member. I also added a new member \"tlist\" to PgFdwRelationInfo to\n> save the tlist created in foreign_join_ok.\n>\n\nInstead of adding a new member, you might want to reuse grouped_tlist\nby renaming it.\n\n> Another idea on the \"tlist\" member would be to save a tlist created for\n> EXPLAIN into that member in estimate_patch_cost_size, so that we don't need\n> to generate the tlist again in postgresGetForeignPlan, when\n> use_remote_estimate=true. But I'd like to leave that for another patch.\n\nI think that will happen automatically, while deparsing, whether for\nEXPLAIN or for actual execution.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 16 Nov 2016 14:44:46 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/11/16 18:14, Ashutosh Bapat wrote:\n>> (1) You added the\n>> following comments to deparseSelectSql:\n>>\n>> + /*\n>> + * For a non-base relation, use the input tlist. If a base relation\n>> is\n>> + * being deparsed as a subquery, use input tlist, if the caller has\n>> passed\n>> + * one. The column aliases of the subquery are crafted based on the\n>> input\n>> + * tlist. If tlist is NIL, there will not be any expression\n>> referring to\n>> + * any of the columns of the base relation being deparsed. Hence it\n>> doesn't\n>> + * matter whether the base relation is being deparsed as subquery or\n>> not.\n>> + */\n>>\n>> The last sentence seems confusing to me. My understanding is: if a base\n>> relation has tlist=NIL, then the base relation *must* be deparsed as\n>> ordinary, not as a subquery. Maybe I'm missing something, though. (I left\n>> that as-is, but I think we need to reword that to be more clear, at least.)\n\n> Hmm, I agree. I think the comment should just say, \"Use tlist to\n> create the SELECT clause if one has been provided. For a base relation\n> with tlist = NIL, check attrs_needed information.\". Does that sound\n> good?\n\nI don't think that is 100% correct, because (1) tlist can be NIL for a \njoin relation, you pointed out upthread, but we need to use \ndeparseExplicitTargetList, so the first sentence is not completely \ncorrect, and (2) we need to check attrs_needed information not only for \na baserel but for an otherrel, so the second sentence is not completely \ncorrect, either. How about this, instead?:\n\n /*\n * For a join relation or an upper relation, use \ndeparseExplicitTargetList.\n * Likewise, for a base relation that is being deparsed as a \nsubquery, in\n * which case the caller would have passed tlist that is non-NIL, \nuse that\n * function. Otherwise, use deparseTargetList.\n */\n\n>> (3) I don't think we need this in isSubqueryExpr, so I removed it from the\n>> patch:\n>>\n>> + /* Keep compiler happy. */\n>> + return false;\n\n> Doesn't that cause compiler warning, saying \"non-void function\n> returning nothing\" or something like that. Actually, the \"if\n> (bms_is_member(node->varno, outerrel->relids))\" ends with a \"return\"\n> always. Hence we don't need to encapsulate the code in \"else\" block in\n> else { }. It could be taken outside.\n\nYeah, I think so too, but I like the \"if () { } else { }\" coding. That \ncoding can be found in other places in core, eg, \noperator_same_subexprs_lookup.\n\n>> OK, I changed isSubqueryExpr to is_subquery_expr; I kept to refer to expr\n>> because I think we would soon extend that function so that it can handle\n>> PHVs, as I said upthread. For getSubselectAliasInfo, I changed the name to\n>> get_subselect_alias_id, because (1) the word \"alias\" seems general and (2)\n>> the \"for_var\" part would make the name a bit long.\n\n> is_subquery_expr(Var *node -- that looks odd. Either it should\n> is_subquery_var(Var * ... OR it should be is_subquery_expr(Expr * ...\n> . I would prefer the first one, since that's what current patch is\n> doing. When we introduce PHVs, we may change it, if required.\n\nOK, I used is_subquery_var().\n\n>> Done. I modified the patch as proposed; create the tlist by\n>> build_tlist_to_deparse in foreign_join_ok, if needed, and search the tlist\n>> by tlist_member. I also added a new member \"tlist\" to PgFdwRelationInfo to\n>> save the tlist created in foreign_join_ok.\n\n> Instead of adding a new member, you might want to reuse grouped_tlist\n> by renaming it.\n\nDone.\n\n>> Another idea on the \"tlist\" member would be to save a tlist created for\n>> EXPLAIN into that member in estimate_patch_cost_size, so that we don't need\n>> to generate the tlist again in postgresGetForeignPlan, when\n>> use_remote_estimate=true. But I'd like to leave that for another patch.\n\n> I think that will happen automatically, while deparsing, whether for\n> EXPLAIN or for actual execution.\n\nReally? Anyway, I'd like to leave that as-is.\n\nPlease find attached a new version of the patch.\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":"Thu, 17 Nov 2016 14:55:21 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":">\n> /*\n> * For a join relation or an upper relation, use\n> deparseExplicitTargetList.\n> * Likewise, for a base relation that is being deparsed as a subquery,\n> in\n> * which case the caller would have passed tlist that is non-NIL, use\n> that\n> * function. Otherwise, use deparseTargetList.\n> */\n\nThis looks correct. I have modified it to make it simple in the given\npatch. But, I think we shouldn't refer to a function e.g.\ndeparseExplicitTargetlist() in the comment. Instead we should describe\nthe intent e.g. \"deparse SELECT clause from the given targetlist\" or\n\"deparse SELECT clause from attr_needed\".\n\n>\n>>> (3) I don't think we need this in isSubqueryExpr, so I removed it from\n>>> the\n>>> patch:\n>>>\n>>> + /* Keep compiler happy. */\n>>> + return false;\n>\n>\n>> Doesn't that cause compiler warning, saying \"non-void function\n>> returning nothing\" or something like that. Actually, the \"if\n>> (bms_is_member(node->varno, outerrel->relids))\" ends with a \"return\"\n>> always. Hence we don't need to encapsulate the code in \"else\" block in\n>> else { }. It could be taken outside.\n>\n>\n> Yeah, I think so too, but I like the \"if () { } else { }\" coding. That\n> coding can be found in other places in core, eg,\n> operator_same_subexprs_lookup.\n\nOK.\n\n\n>\n>>> Done. I modified the patch as proposed; create the tlist by\n>>> build_tlist_to_deparse in foreign_join_ok, if needed, and search the\n>>> tlist\n>>> by tlist_member. I also added a new member \"tlist\" to PgFdwRelationInfo\n>>> to\n>>> save the tlist created in foreign_join_ok.\n>\n>\n>> Instead of adding a new member, you might want to reuse grouped_tlist\n>> by renaming it.\n>\n>\n> Done.\n\nRight now, we are calculating tlist whether or not the ForeignPath\nemerges as the cheapest path. Instead we should calculate tlist, the\nfirst time we need it and then add it to the fpinfo.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database Company\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, 18 Nov 2016 17:10:18 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"Also another point\n\nI guess, this note doesn't add much value in the given context.\nProbably we should remove it.\n+ * Note: the tlist would have one-to-one correspondence with the\n+ * joining relation's reltarget->exprs because (1) the above test\n+ * on PHVs guarantees that the reltarget->exprs doesn't contain\n+ * any PHVs and (2) the joining relation's local_conds is NIL.\n+ * This allows us to search the targetlist entry matching a given\n+ * Var node from the tlist in get_subselect_alias_id.\n\nOn Fri, Nov 18, 2016 at 5:10 PM, Ashutosh Bapat\n<ashutosh.bapat@enterprisedb.com> wrote:\n>>\n>> /*\n>> * For a join relation or an upper relation, use\n>> deparseExplicitTargetList.\n>> * Likewise, for a base relation that is being deparsed as a subquery,\n>> in\n>> * which case the caller would have passed tlist that is non-NIL, use\n>> that\n>> * function. Otherwise, use deparseTargetList.\n>> */\n>\n> This looks correct. I have modified it to make it simple in the given\n> patch. But, I think we shouldn't refer to a function e.g.\n> deparseExplicitTargetlist() in the comment. Instead we should describe\n> the intent e.g. \"deparse SELECT clause from the given targetlist\" or\n> \"deparse SELECT clause from attr_needed\".\n>\n>>\n>>>> (3) I don't think we need this in isSubqueryExpr, so I removed it from\n>>>> the\n>>>> patch:\n>>>>\n>>>> + /* Keep compiler happy. */\n>>>> + return false;\n>>\n>>\n>>> Doesn't that cause compiler warning, saying \"non-void function\n>>> returning nothing\" or something like that. Actually, the \"if\n>>> (bms_is_member(node->varno, outerrel->relids))\" ends with a \"return\"\n>>> always. Hence we don't need to encapsulate the code in \"else\" block in\n>>> else { }. It could be taken outside.\n>>\n>>\n>> Yeah, I think so too, but I like the \"if () { } else { }\" coding. That\n>> coding can be found in other places in core, eg,\n>> operator_same_subexprs_lookup.\n>\n> OK.\n>\n>\n>>\n>>>> Done. I modified the patch as proposed; create the tlist by\n>>>> build_tlist_to_deparse in foreign_join_ok, if needed, and search the\n>>>> tlist\n>>>> by tlist_member. I also added a new member \"tlist\" to PgFdwRelationInfo\n>>>> to\n>>>> save the tlist created in foreign_join_ok.\n>>\n>>\n>>> Instead of adding a new member, you might want to reuse grouped_tlist\n>>> by renaming it.\n>>\n>>\n>> Done.\n>\n> Right now, we are calculating tlist whether or not the ForeignPath\n> emerges as the cheapest path. Instead we should calculate tlist, the\n> first time we need it and then add it to the fpinfo.\n>\n> --\n> Best Wishes,\n> Ashutosh Bapat\n> EnterpriseDB Corporation\n> The Postgres Database Company\n\n\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 18 Nov 2016 20:46:00 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/11/19 0:16, Ashutosh Bapat wrote:\n> Also another point\n>\n> I guess, this note doesn't add much value in the given context.\n> Probably we should remove it.\n> + * Note: the tlist would have one-to-one correspondence with the\n> + * joining relation's reltarget->exprs because (1) the above test\n> + * on PHVs guarantees that the reltarget->exprs doesn't contain\n> + * any PHVs and (2) the joining relation's local_conds is NIL.\n> + * This allows us to search the targetlist entry matching a given\n> + * Var node from the tlist in get_subselect_alias_id.\n\nOK, I removed.\n\n> On Fri, Nov 18, 2016 at 5:10 PM, Ashutosh Bapat\n> <ashutosh.bapat@enterprisedb.com> wrote:\n\nI wrote:\n>>> /*\n>>> * For a join relation or an upper relation, use\n>>> deparseExplicitTargetList.\n>>> * Likewise, for a base relation that is being deparsed as a subquery,\n>>> in\n>>> * which case the caller would have passed tlist that is non-NIL, use\n>>> that\n>>> * function. Otherwise, use deparseTargetList.\n>>> */\n\n>> This looks correct. I have modified it to make it simple in the given\n>> patch. But, I think we shouldn't refer to a function e.g.\n>> deparseExplicitTargetlist() in the comment. Instead we should describe\n>> the intent e.g. \"deparse SELECT clause from the given targetlist\" or\n>> \"deparse SELECT clause from attr_needed\".\n\nMy taste would be to refer to those functions, because ISTM that makes \nthe explanation more simple and direct. So, I'd like to leave that for \nthe committer's judge.\n\nI wrote:\n>>>>> Done. I modified the patch as proposed; create the tlist by\n>>>>> build_tlist_to_deparse in foreign_join_ok, if needed, and search the\n>>>>> tlist\n>>>>> by tlist_member. I also added a new member \"tlist\" to PgFdwRelationInfo\n>>>>> to\n>>>>> save the tlist created in foreign_join_ok.\n\nYou wrote:\n>>>> Instead of adding a new member, you might want to reuse grouped_tlist\n>>>> by renaming it.\n\n>>> Done.\n\n>> Right now, we are calculating tlist whether or not the ForeignPath\n>> emerges as the cheapest path.\n\nYeah, I modified the patch so, as I thought that would be consistent \nwith the aggregate pushdown patch.\n\n>> Instead we should calculate tlist, the\n>> first time we need it and then add it to the fpinfo.\n\nHaving said that, I agree on that point. I'd like to propose (1) adding \na new member to fpinfo, to store a list of output Vars from the \nsubquery, and (2) creating a tlist from it in deparseRangeTblRef, then, \nwhich would allow us to calculate the tlist only when we need it. The \nmember added to fpinfo would be also useful to address the comments on \nthe DML/UPDATE pushdown patch. See the attached patch in [1]. I named \nthe member a bit differently in that patch, though.\n\nYou modified the comments I added to deparseLockingClause into this:\n\n /*\n+ * Ignore relation if it appears in a lower subquery. Locking clause\n+ * for such a relation is included in the subquery.\n+ */\n\nI don't think the second sentence is 100% correct because a locking \nclause isn't always required for such a relation, so I modified the \nsentence a bit.\n\nBest regards,\nEtsuro Fujita\n\n[1] \nhttps://www.postgresql.org/message-id/38245b84-fabf-0899-1b24-8f94cdc5900c%40lab.ntt.co.jp\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, 21 Nov 2016 21:15:49 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":">\n> Yeah, I modified the patch so, as I thought that would be consistent with\n> the aggregate pushdown patch.\n\naggregate pushdown needs the tlist to judge the shippability of\ntargetlist. For joins that's not required, so we should defer, if we\ncan.\n\n>\n>>> Instead we should calculate tlist, the\n>>> first time we need it and then add it to the fpinfo.\n>\n>\n> Having said that, I agree on that point. I'd like to propose (1) adding a\n> new member to fpinfo, to store a list of output Vars from the subquery, and\n> (2) creating a tlist from it in deparseRangeTblRef, then, which would allow\n> us to calculate the tlist only when we need it. The member added to fpinfo\n> would be also useful to address the comments on the DML/UPDATE pushdown\n> patch. See the attached patch in [1]. I named the member a bit differently\n> in that patch, though.\n\nAgain the list of Vars will be wasted if we don't choose that path for\nfinal planning. So, I don't see the point of adding list of Vars\nthere. It looks like we are replacing one list with the other when\nnone of those are useful, if the path doesn't get chosen for the final\nplan. If you think that the member is useful for DML/UDPATE pushdown,\nyou may want to add it in the other patch.\n\n>\n> You modified the comments I added to deparseLockingClause into this:\n>\n> /*\n> + * Ignore relation if it appears in a lower subquery. Locking clause\n> + * for such a relation is included in the subquery.\n> + */\n>\n> I don't think the second sentence is 100% correct because a locking clause\n> isn't always required for such a relation, so I modified the sentence a bit.\n>\n\nI guess, \"if required\" part was implicit in construct \"such a\nrelation\". Your version seems to make it explicit. I am fine with it.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 21 Nov 2016 18:32:12 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/11/21 22:02, Ashutosh Bapat wrote:\n\nYou wrote:\n>>>> Instead we should calculate tlist, the\n>>>> first time we need it and then add it to the fpinfo.\n\nI wrote:\n>> Having said that, I agree on that point. I'd like to propose (1) adding a\n>> new member to fpinfo, to store a list of output Vars from the subquery, and\n>> (2) creating a tlist from it in deparseRangeTblRef, then, which would allow\n>> us to calculate the tlist only when we need it. The member added to fpinfo\n>> would be also useful to address the comments on the DML/UPDATE pushdown\n>> patch. See the attached patch in [1]. I named the member a bit differently\n>> in that patch, though.\n\n> Again the list of Vars will be wasted if we don't choose that path for\n> final planning. So, I don't see the point of adding list of Vars\n> there.\n\n> If you think that the member is useful for DML/UDPATE pushdown,\n> you may want to add it in the other patch.\n\nOK, I'd like to propose referencing to foreignrel->reltarget->exprs \ndirectly in deparseRangeTblRef and get_subselect_alias_id, then, which \nis the same as what I proposed in the first version of the patch, but \nI'd also like to change deparseRangeTblRef to use add_to_flat_tlist, not \nmake_tlist_from_pathtarget, to create a tlist of the subquery, as you \nproposed.\n\n>> You modified the comments I added to deparseLockingClause into this:\n>>\n>> /*\n>> + * Ignore relation if it appears in a lower subquery. Locking clause\n>> + * for such a relation is included in the subquery.\n>> + */\n>>\n>> I don't think the second sentence is 100% correct because a locking clause\n>> isn't always required for such a relation, so I modified the sentence a bit.\n\n> I guess, \"if required\" part was implicit in construct \"such a\n> relation\". Your version seems to make it explicit. I am fine with it.\n\nOK, let's leave that for the committer's judge.\n\nPlease find attached an updated version of the patch.\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":"Tue, 22 Nov 2016 15:55:14 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"The comments should explain why is the assertion true.\n+ /* Shouldn't be NIL */\n+ Assert(tlist != NIL);\n+ /* Should be same length */\n+ Assert(list_length(tlist) ==\nlist_length(foreignrel->reltarget->exprs));\n\n\n>\n> OK, I'd like to propose referencing to foreignrel->reltarget->exprs directly\n> in deparseRangeTblRef and get_subselect_alias_id, then, which is the same as\n> what I proposed in the first version of the patch, but I'd also like to\n> change deparseRangeTblRef to use add_to_flat_tlist, not\n> make_tlist_from_pathtarget, to create a tlist of the subquery, as you\n> proposed.\n\nThere is a already a function to build targetlist for a given relation\nbuild_tlist_to_deparse(), which does the same thing as this code for a join or\nbase relation and when there are no local conditions. Why don't we use that\nfunction instead of duplicating that logic? If tomorrow, the logic to build the\ntargetlist changes, we will need to duplicate those changes. We should avoid\nthat.\n+ /* Build a tlist from the subquery. */\n+ tlist = add_to_flat_tlist(tlist, foreignrel->reltarget->exprs);\n\nThe comment below says \"get the aliases\", but what the function really returns\nis the identifiers used for creating aliases. Please correct the comment.\n+/*\n+ * Get the relation and column alias for a given Var node, which belongs to\n+ * input foreignrel. They are returned in *tabno and *colno respectively.\n+ */\n\nWe discussed that we have to deparse and search from the same targetlist. And\nthat the targetlist should be saved in fpinfo, the first time it gets created.\nBut the patch seems to be searching in foreignrel->reltarget->exprs and\ndeparsing from the tlist returned by add_to_flat_tlist(tlist,\nforeignrel->reltarget->exprs).\n+ foreach(lc, foreignrel->reltarget->exprs)\n+ {\n+ if (equal(lfirst(lc), (Node *) node))\n+ {\n+ *colno = i;\n+ return;\n+ }\n+ i++;\n+ }\nI guess, the reason why you are doing it this way, is SELECT clause for the\noutermost query gets deparsed before FROM clause. For later we call\ndeparseRangeTblRef(), which builds the tlist. So, while deparsing SELECT\nclause, we do not have tlist to build from. In that case, I guess, we have to\nbuild the tlist in get_subselect_alias_id() if it's not available and stick it\nin fpinfo. Subsequent calls to get_subselect_alias_id() should find tlist\nthere. Then in deparseRangeTblRef() assert that there's a tlist in fpinfo\nand use it to build the SELECT clause of subquery. That way, we don't build\ntlist unless it's needed and also use the same tlist for all searches. Please\nuse tlist_member() to search into the tlist.\n\nThe name get_subselect_alias_id() seems to suggest that the function returns\nidentifier for subselect alias, which isn't correct. It returns the alias\nidentifiers for deparsing a Var node. But I guess, we have left this to the\ncommitter's judgement.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 22 Nov 2016 14:58:50 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/11/22 18:28, Ashutosh Bapat wrote:\n> The comments should explain why is the assertion true.\n> + /* Shouldn't be NIL */\n> + Assert(tlist != NIL);\n> + /* Should be same length */\n> + Assert(list_length(tlist) ==\n> list_length(foreignrel->reltarget->exprs));\n\nWill revise.\n\n>> OK, I'd like to propose referencing to foreignrel->reltarget->exprs directly\n>> in deparseRangeTblRef and get_subselect_alias_id, then, which is the same as\n>> what I proposed in the first version of the patch, but I'd also like to\n>> change deparseRangeTblRef to use add_to_flat_tlist, not\n>> make_tlist_from_pathtarget, to create a tlist of the subquery, as you\n>> proposed.\n\n> There is a already a function to build targetlist for a given relation\n> build_tlist_to_deparse(), which does the same thing as this code for a join or\n> base relation and when there are no local conditions. Why don't we use that\n> function instead of duplicating that logic? If tomorrow, the logic to build the\n> targetlist changes, we will need to duplicate those changes. We should avoid\n> that.\n> + /* Build a tlist from the subquery. */\n> + tlist = add_to_flat_tlist(tlist, foreignrel->reltarget->exprs);\n\nWill modify the patch to use build_tlist_to_deparse. Actually, the \nearly versions of the patch used that function, but it looks like I \nchanged not to use that function, as I misunderstood about your comments \non this part at some point. Sorry for that.\n\n> The comment below says \"get the aliases\", but what the function really returns\n> is the identifiers used for creating aliases. Please correct the comment.\n> +/*\n> + * Get the relation and column alias for a given Var node, which belongs to\n> + * input foreignrel. They are returned in *tabno and *colno respectively.\n> + */\n\nActually, this was rewritten into the above by *you*. The original \ncomment I added was:\n\n+ /*\n+ * Get info about the subselect alias to given expression.\n+ *\n+ * The subselect table and column numbers are returned to *tabno and \n*colno,\n+ * respectively.\n+ */\n\nI'd like to change the comment into something like the original one.\n\n> We discussed that we have to deparse and search from the same targetlist. And\n> that the targetlist should be saved in fpinfo, the first time it gets created.\n> But the patch seems to be searching in foreignrel->reltarget->exprs and\n> deparsing from the tlist returned by add_to_flat_tlist(tlist,\n> foreignrel->reltarget->exprs).\n> + foreach(lc, foreignrel->reltarget->exprs)\n> + {\n> + if (equal(lfirst(lc), (Node *) node))\n> + {\n> + *colno = i;\n> + return;\n> + }\n> + i++;\n> + }\n> I guess, the reason why you are doing it this way, is SELECT clause for the\n> outermost query gets deparsed before FROM clause. For later we call\n> deparseRangeTblRef(), which builds the tlist. So, while deparsing SELECT\n> clause, we do not have tlist to build from.\n\nThat's right.\n\n> In that case, I guess, we have to\n> build the tlist in get_subselect_alias_id() if it's not available and stick it\n> in fpinfo. Subsequent calls to get_subselect_alias_id() should find tlist\n> there. Then in deparseRangeTblRef() assert that there's a tlist in fpinfo\n> and use it to build the SELECT clause of subquery. That way, we don't build\n> tlist unless it's needed and also use the same tlist for all searches. Please\n> use tlist_member() to search into the tlist.\n\nI don't think that's a good idea because that would make the code \nunnecessarily complicated and inefficient. I think the direct search \ninto the foreignrel->reltarget->exprs shown above would be better \nbecause that's more simple and efficient than what you proposed. Note \nthat since (1) the foreignrel->reltarget->exprs doesn't contain any PHVs \nand (2) the local_conds is empty, the tlist created for the subquery by \nbuild_tlist_to_deparse is guaranteed to have one-to-one correspondence \nwith the foreignrel->reltarget->exprs, so the direct search works well.\n\n> The name get_subselect_alias_id() seems to suggest that the function returns\n> identifier for subselect alias, which isn't correct. It returns the alias\n> identifiers for deparsing a Var node. But I guess, we have left this to the\n> committer's judgement.\n\nFine with me.\n\nBest regards,\nEtsuro Fujita\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, 22 Nov 2016 20:19:04 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":">\n>\n>> There is a already a function to build targetlist for a given relation\n>> build_tlist_to_deparse(), which does the same thing as this code for a\n>> join or\n>> base relation and when there are no local conditions. Why don't we use\n>> that\n>> function instead of duplicating that logic? If tomorrow, the logic to\n>> build the\n>> targetlist changes, we will need to duplicate those changes. We should\n>> avoid\n>> that.\n>> + /* Build a tlist from the subquery. */\n>> + tlist = add_to_flat_tlist(tlist, foreignrel->reltarget->exprs);\n>\n>\n> Will modify the patch to use build_tlist_to_deparse. Actually, the early\n> versions of the patch used that function, but it looks like I changed not to\n> use that function, as I misunderstood about your comments on this part at\n> some point. Sorry for that.\n>\n>> The comment below says \"get the aliases\", but what the function really\n>> returns\n>> is the identifiers used for creating aliases. Please correct the comment.\n>> +/*\n>> + * Get the relation and column alias for a given Var node, which belongs\n>> to\n>> + * input foreignrel. They are returned in *tabno and *colno respectively.\n>> + */\n>\n>\n> Actually, this was rewritten into the above by *you*. The original comment\n> I added was:\n>\n> + /*\n> + * Get info about the subselect alias to given expression.\n> + *\n> + * The subselect table and column numbers are returned to *tabno and\n> *colno,\n> + * respectively.\n> + */\n>\n> I'd like to change the comment into something like the original one.\n\nSorry. I think the current version is better than previous one. The\nterm \"subselect alias\" is confusing in the previous version. In the\ncurrent version, \"Get the relation and column alias for a given Var\nnode,\" we need to add word \"identifiers\" like \"Get the relation and\ncolumn identifiers for a given Var node\".\n\n>\n>> We discussed that we have to deparse and search from the same targetlist.\n>> And\n>> that the targetlist should be saved in fpinfo, the first time it gets\n>> created.\n>> But the patch seems to be searching in foreignrel->reltarget->exprs and\n>> deparsing from the tlist returned by add_to_flat_tlist(tlist,\n>> foreignrel->reltarget->exprs).\n>> + foreach(lc, foreignrel->reltarget->exprs)\n>> + {\n>> + if (equal(lfirst(lc), (Node *) node))\n>> + {\n>> + *colno = i;\n>> + return;\n>> + }\n>> + i++;\n>> + }\n>> I guess, the reason why you are doing it this way, is SELECT clause for\n>> the\n>> outermost query gets deparsed before FROM clause. For later we call\n>> deparseRangeTblRef(), which builds the tlist. So, while deparsing SELECT\n>> clause, we do not have tlist to build from.\n>\n>\n> That's right.\n>\n>> In that case, I guess, we have to\n>> build the tlist in get_subselect_alias_id() if it's not available and\n>> stick it\n>> in fpinfo. Subsequent calls to get_subselect_alias_id() should find tlist\n>> there. Then in deparseRangeTblRef() assert that there's a tlist in fpinfo\n>> and use it to build the SELECT clause of subquery. That way, we don't\n>> build\n>> tlist unless it's needed and also use the same tlist for all searches.\n>> Please\n>> use tlist_member() to search into the tlist.\n>\n>\n> I don't think that's a good idea because that would make the code\n> unnecessarily complicated and inefficient. I think the direct search into\n> the foreignrel->reltarget->exprs shown above would be better because that's\n> more simple and efficient than what you proposed. Note that since (1) the\n> foreignrel->reltarget->exprs doesn't contain any PHVs and (2) the\n> local_conds is empty, the tlist created for the subquery by\n> build_tlist_to_deparse is guaranteed to have one-to-one correspondence with\n> the foreignrel->reltarget->exprs, so the direct search works well.\n\nIf we deparse from and search into different objects, that's not very\nmaintainable code. Changes to any one of them require changes to the\nother, thus creating avenues for bugs. Even if slighly expensive, we\nshould search into and deparse from the same targetlist. I think I\nhave explained this before.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 22 Nov 2016 21:00:30 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/11/23 0:30, Ashutosh Bapat wrote:\n\nYou wrote:\n>>> The comment below says \"get the aliases\", but what the function really\n>>> returns\n>>> is the identifiers used for creating aliases. Please correct the comment.\n>>> +/*\n>>> + * Get the relation and column alias for a given Var node, which belongs\n>>> to\n>>> + * input foreignrel. They are returned in *tabno and *colno respectively.\n>>> + */\n\nI wrote:\n>> Actually, this was rewritten into the above by *you*. The original comment\n>> I added was:\n>>\n>> + /*\n>> + * Get info about the subselect alias to given expression.\n>> + *\n>> + * The subselect table and column numbers are returned to *tabno and\n>> *colno,\n>> + * respectively.\n>> + */\n>>\n>> I'd like to change the comment into something like the original one.\n\n> Sorry. I think the current version is better than previous one. The\n> term \"subselect alias\" is confusing in the previous version. In the\n> current version, \"Get the relation and column alias for a given Var\n> node,\" we need to add word \"identifiers\" like \"Get the relation and\n> column identifiers for a given Var node\".\n\nOK, but one thing I'm concerned about is the term \"relation alias\" seems \na bit confusing because we already used the term for the alias of the \nform \"rN\". To avoid that, how about saying \"table alias\", not \"relation \nalias\"? (in which case, the comment would be something like \"Get the \ntable and column identifiers for a given Var node\".)\n\n>>> We discussed that we have to deparse and search from the same targetlist.\n>>> And\n>>> that the targetlist should be saved in fpinfo, the first time it gets\n>>> created.\n>>> But the patch seems to be searching in foreignrel->reltarget->exprs and\n>>> deparsing from the tlist returned by add_to_flat_tlist(tlist,\n>>> foreignrel->reltarget->exprs).\n>>> + foreach(lc, foreignrel->reltarget->exprs)\n>>> + {\n>>> + if (equal(lfirst(lc), (Node *) node))\n>>> + {\n>>> + *colno = i;\n>>> + return;\n>>> + }\n>>> + i++;\n>>> + }\n>>> I guess, the reason why you are doing it this way, is SELECT clause for\n>>> the\n>>> outermost query gets deparsed before FROM clause. For later we call\n>>> deparseRangeTblRef(), which builds the tlist. So, while deparsing SELECT\n>>> clause, we do not have tlist to build from.\n\n>> That's right.\n\n>>> In that case, I guess, we have to\n>>> build the tlist in get_subselect_alias_id() if it's not available and\n>>> stick it\n>>> in fpinfo. Subsequent calls to get_subselect_alias_id() should find tlist\n>>> there. Then in deparseRangeTblRef() assert that there's a tlist in fpinfo\n>>> and use it to build the SELECT clause of subquery. That way, we don't\n>>> build\n>>> tlist unless it's needed and also use the same tlist for all searches.\n>>> Please\n>>> use tlist_member() to search into the tlist.\n\n>> I don't think that's a good idea because that would make the code\n>> unnecessarily complicated and inefficient. I think the direct search into\n>> the foreignrel->reltarget->exprs shown above would be better because that's\n>> more simple and efficient than what you proposed. Note that since (1) the\n>> foreignrel->reltarget->exprs doesn't contain any PHVs and (2) the\n>> local_conds is empty, the tlist created for the subquery by\n>> build_tlist_to_deparse is guaranteed to have one-to-one correspondence with\n>> the foreignrel->reltarget->exprs, so the direct search works well.\n\n> If we deparse from and search into different objects, that's not very\n> maintainable code. Changes to any one of them require changes to the\n> other, thus creating avenues for bugs. Even if slighly expensive, we\n> should search into and deparse from the same targetlist.\n\nI don't think so; in the current version, we essentially deparse from \nand search into the same object, ie, foreignrel->reltarget->exprs, since \nthe tlist created by build_tlist_to_deparse is build from the \nforeignrel->reltarget->exprs. Also, since it is just created for \ndeparsing the relation as a subquery in deparseRangeTblRef and isn't \nstored into fpinfo or anywhere alse, we would need no concern about \ncreating such avenues. IMO I think adding comments would be enough for \nthis. Anyway, I think this is an optional issue, so I'd like to leave \nthis for the committer's judge.\n\n> I think I\n> have explained this before.\n\nMy apologies for having misunderstood your words.\n\nBest regards,\nEtsuro Fujita\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, 24 Nov 2016 16:30:24 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":">\n>> Sorry. I think the current version is better than previous one. The\n>> term \"subselect alias\" is confusing in the previous version. In the\n>> current version, \"Get the relation and column alias for a given Var\n>> node,\" we need to add word \"identifiers\" like \"Get the relation and\n>> column identifiers for a given Var node\".\n>\n>\n> OK, but one thing I'm concerned about is the term \"relation alias\" seems a\n> bit confusing because we already used the term for the alias of the form\n> \"rN\". To avoid that, how about saying \"table alias\", not \"relation alias\"?\n> (in which case, the comment would be something like \"Get the table and\n> column identifiers for a given Var node\".)\n\ntable will be misleading as subquery can represent a join and\ncorresponding alias would represent the join. Relation is better term.\n\n\n>\n>\n>> If we deparse from and search into different objects, that's not very\n>> maintainable code. Changes to any one of them require changes to the\n>> other, thus creating avenues for bugs. Even if slighly expensive, we\n>> should search into and deparse from the same targetlist.\n>\n>\n> I don't think so; in the current version, we essentially deparse from and\n> search into the same object, ie, foreignrel->reltarget->exprs, since the\n> tlist created by build_tlist_to_deparse is build from the\n> foreignrel->reltarget->exprs. Also, since it is just created for deparsing\n> the relation as a subquery in deparseRangeTblRef and isn't stored into\n> fpinfo or anywhere alse, we would need no concern about creating such\n> avenues. IMO I think adding comments would be enough for this.\n\nbuild_tlist_to_depase() calls pull_var_nodes() before creating the\ntlist, whereas the code that searches does not do that. Code-wise\nthose are not the same things.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 24 Nov 2016 13:16:14 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/11/24 16:46, Ashutosh Bapat wrote:\n>>> Sorry. I think the current version is better than previous one. The\n>>> term \"subselect alias\" is confusing in the previous version. In the\n>>> current version, \"Get the relation and column alias for a given Var\n>>> node,\" we need to add word \"identifiers\" like \"Get the relation and\n>>> column identifiers for a given Var node\".\n\nI wrote:\n>> OK, but one thing I'm concerned about is the term \"relation alias\" seems a\n>> bit confusing because we already used the term for the alias of the form\n>> \"rN\". To avoid that, how about saying \"table alias\", not \"relation alias\"?\n>> (in which case, the comment would be something like \"Get the table and\n>> column identifiers for a given Var node\".)\n\n> table will be misleading as subquery can represent a join and\n> corresponding alias would represent the join. Relation is better term.\n\nBut the documentation uses the word \"table\" for a join relation. See \n7.2.1.2. Table and Column Aliases:\nhttps://www.postgresql.org/docs/9.6/static/queries-table-expressions.html#QUERIES-TABLE-ALIASES\n\n>> I don't think so; in the current version, we essentially deparse from and\n>> search into the same object, ie, foreignrel->reltarget->exprs, since the\n>> tlist created by build_tlist_to_deparse is build from the\n>> foreignrel->reltarget->exprs. Also, since it is just created for deparsing\n>> the relation as a subquery in deparseRangeTblRef and isn't stored into\n>> fpinfo or anywhere alse, we would need no concern about creating such\n>> avenues. IMO I think adding comments would be enough for this.\n\n> build_tlist_to_depase() calls pull_var_nodes() before creating the\n> tlist, whereas the code that searches does not do that. Code-wise\n> those are not the same things.\n\nYou missed the point; the foreignrel->reltarget->exprs doesn't contain \nany PHVs, so the tlist created by build_tlist_to_depase will be \nguaranteed to be one-to-one with the foreignrel->reltarget->exprs.\n\nBest regards,\nEtsuro Fujita\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, 24 Nov 2016 16:57:20 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Thu, Nov 24, 2016 at 1:27 PM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> On 2016/11/24 16:46, Ashutosh Bapat wrote:\n>>>>\n>>>> Sorry. I think the current version is better than previous one. The\n>>>> term \"subselect alias\" is confusing in the previous version. In the\n>>>> current version, \"Get the relation and column alias for a given Var\n>>>> node,\" we need to add word \"identifiers\" like \"Get the relation and\n>>>> column identifiers for a given Var node\".\n>\n>\n> I wrote:\n>>>\n>>> OK, but one thing I'm concerned about is the term \"relation alias\" seems\n>>> a\n>>> bit confusing because we already used the term for the alias of the form\n>>> \"rN\". To avoid that, how about saying \"table alias\", not \"relation\n>>> alias\"?\n>>> (in which case, the comment would be something like \"Get the table and\n>>> column identifiers for a given Var node\".)\n>\n>\n>> table will be misleading as subquery can represent a join and\n>> corresponding alias would represent the join. Relation is better term.\n>\n>\n> But the documentation uses the word \"table\" for a join relation. See\n> 7.2.1.2. Table and Column Aliases:\n> https://www.postgresql.org/docs/9.6/static/queries-table-expressions.html#QUERIES-TABLE-ALIASES\n\nThe definition there says \"A temporary name can be given to tables and\ncomplex table references to be used for references to the derived\ntable in the rest of the query. This is called a table alias.\", please\nnote the usage \"complex table references\". Within the code, we do not\nrefer to a relation as table. So, the comment in this code should not\nrefer to a relation as table either.\n\n>\n>>> I don't think so; in the current version, we essentially deparse from and\n>>> search into the same object, ie, foreignrel->reltarget->exprs, since the\n>>> tlist created by build_tlist_to_deparse is build from the\n>>> foreignrel->reltarget->exprs. Also, since it is just created for\n>>> deparsing\n>>> the relation as a subquery in deparseRangeTblRef and isn't stored into\n>>> fpinfo or anywhere alse, we would need no concern about creating such\n>>> avenues. IMO I think adding comments would be enough for this.\n>\n>\n>> build_tlist_to_depase() calls pull_var_nodes() before creating the\n>> tlist, whereas the code that searches does not do that. Code-wise\n>> those are not the same things.\n>\n>\n> You missed the point; the foreignrel->reltarget->exprs doesn't contain any\n> PHVs, so the tlist created by build_tlist_to_depase will be guaranteed to be\n> one-to-one with the foreignrel->reltarget->exprs.\n\nIt's guaranteed now, but can not be forever. This means anybody who\nsupports PHV tomorrow, will have to \"remember\" to update this code as\nwell. If s/he misses it, a bug will be introduced. That's the avenue\nfor bug, I am talking about.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 24 Nov 2016 14:09:27 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/11/24 17:39, Ashutosh Bapat wrote:\n> On Thu, Nov 24, 2016 at 1:27 PM, Etsuro Fujita\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>> On 2016/11/24 16:46, Ashutosh Bapat wrote:\n>>> table will be misleading as subquery can represent a join and\n>>> corresponding alias would represent the join. Relation is better term.\n\n>> But the documentation uses the word \"table\" for a join relation. See\n>> 7.2.1.2. Table and Column Aliases:\n>> https://www.postgresql.org/docs/9.6/static/queries-table-expressions.html#QUERIES-TABLE-ALIASES\n\n> The definition there says \"A temporary name can be given to tables and\n> complex table references to be used for references to the derived\n> table in the rest of the query. This is called a table alias.\", please\n> note the usage \"complex table references\". Within the code, we do not\n> refer to a relation as table. So, the comment in this code should not\n> refer to a relation as table either.\n\nOK, will keep using the term \"relation\".\n\nI wrote:\n>>>> I don't think so; in the current version, we essentially deparse from and\n>>>> search into the same object, ie, foreignrel->reltarget->exprs, since the\n>>>> tlist created by build_tlist_to_deparse is build from the\n>>>> foreignrel->reltarget->exprs. Also, since it is just created for\n>>>> deparsing\n>>>> the relation as a subquery in deparseRangeTblRef and isn't stored into\n>>>> fpinfo or anywhere alse, we would need no concern about creating such\n>>>> avenues. IMO I think adding comments would be enough for this.\n\n>>> build_tlist_to_depase() calls pull_var_nodes() before creating the\n>>> tlist, whereas the code that searches does not do that. Code-wise\n>>> those are not the same things.\n\n>> You missed the point; the foreignrel->reltarget->exprs doesn't contain any\n>> PHVs, so the tlist created by build_tlist_to_depase will be guaranteed to be\n>> one-to-one with the foreignrel->reltarget->exprs.\n\n> It's guaranteed now, but can not be forever. This means anybody who\n> supports PHV tomorrow, will have to \"remember\" to update this code as\n> well. If s/he misses it, a bug will be introduced. That's the avenue\n> for bug, I am talking about.\n\nIt *can* be guaranteed. See another patch for supporting evaluating \nPHVs remotely.\n\nBest regards,\nEtsuro Fujita\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, 24 Nov 2016 17:51:50 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":">\n>\n>>>> build_tlist_to_depase() calls pull_var_nodes() before creating the\n>>>> tlist, whereas the code that searches does not do that. Code-wise\n>>>> those are not the same things.\n>\n>\n>>> You missed the point; the foreignrel->reltarget->exprs doesn't contain\n>>> any\n>>> PHVs, so the tlist created by build_tlist_to_depase will be guaranteed to\n>>> be\n>>> one-to-one with the foreignrel->reltarget->exprs.\n>\n>\n>> It's guaranteed now, but can not be forever. This means anybody who\n>> supports PHV tomorrow, will have to \"remember\" to update this code as\n>> well. If s/he misses it, a bug will be introduced. That's the avenue\n>> for bug, I am talking about.\n>\n>\n> It *can* be guaranteed. See another patch for supporting evaluating PHVs\n> remotely.\n\nWe can't base assumptions in this patch on something in the other\npatch, esp. when that's not even reviewed once. PHV is just one case,\nsubqueries involving aggregates is other and there are others. But\nthat's not really the point. The point is\nadd_to_flat_tlist(pull_var_clause(rel->reltarget->exprs) can not be\nproved to be same as rel->reltarget->exprs in general. So, we should\nbase our code on an assumption that can not be proved.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 24 Nov 2016 14:50:31 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/11/24 18:20, Ashutosh Bapat wrote:\n\nI wrote:\n>>>> You missed the point; the foreignrel->reltarget->exprs doesn't contain\n>>>> any\n>>>> PHVs, so the tlist created by build_tlist_to_depase will be guaranteed to\n>>>> be\n>>>> one-to-one with the foreignrel->reltarget->exprs.\n\nYou wrote:\n>>> It's guaranteed now, but can not be forever. This means anybody who\n>>> supports PHV tomorrow, will have to \"remember\" to update this code as\n>>> well. If s/he misses it, a bug will be introduced. That's the avenue\n>>> for bug, I am talking about.\n\nI wrote:\n>> It *can* be guaranteed. See another patch for supporting evaluating PHVs\n>> remotely.\n\n> We can't base assumptions in this patch on something in the other\n> patch, esp. when that's not even reviewed once. PHV is just one case,\n> subqueries involving aggregates is other and there are others.\n\nLet's discuss this together with the patch for PHVs. That was one of \nthe reasons why I had merged the two patches into one. I'd like to \nleave enhancements such as subqueries involving aggregates for future \nwork, though.\n\n> But\n> that's not really the point. The point is\n> add_to_flat_tlist(pull_var_clause(rel->reltarget->exprs) can not be\n> proved to be same as rel->reltarget->exprs in general.\n\nReally? As mentioned in comments for RelOptInfo in relation.h shown \nbelow, the rel->reltarget->exprs for a base/join relation only contains \nVars and PHVs, so the two things would be proved to be one-to-one. \n(Note: we don't need to care about the appendrel-child-relation case \nbecause appendrel child relations wouldn't appear in the main join tree.)\n\n * reltarget - Default Path output tlist for this rel; normally \ncontains\n * Var and PlaceHolderVar nodes for the values we need to\n * output from this relation.\n * List is in no particular order, but all rels of an\n * appendrel set must use corresponding orders.\n * NOTE: in an appendrel child relation, may contain\n * arbitrary expressions pulled up from a subquery!\n\nBest regards,\nEtsuro Fujita\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, 24 Nov 2016 19:04:46 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/11/22 18:28, Ashutosh Bapat wrote:\n> The comments should explain why is the assertion true.\n> + /* Shouldn't be NIL */\n> + Assert(tlist != NIL);\n\nI noticed that I was wrong; in the Assertion the tlist can be empty. An \nexample for such a case is:\n\nSELECT 1 FROM (SELECT c1 FROM ft4 WHERE c1 between 50 and 60) t1 FULL \nJOIN (SELECT c1 FROM ft5 WHERE c1 between 50 and 60) t2 ON (TRUE);\n\nIn this example any columns of the relations ft4 and ft5 wouldn't be \nneeded for the join or the final output, so both the tlists for the \nrelations created in deparseRangeTblRef would be empty. Attached is an \nupdated version fixing this bug. Changes are:\n\n* I removed make_outerrel_subquery/make_innerrel_subquery from fpinfo \nand added a new member \"is_subquery_rel\" to fpinfo, as a flag on whether \nto deparse the relation as a subquery.\n* I modified deparseRangeTblRef, deparseSelectSql, and is_subquery_var \nto see the is_subquery_rel, not \nmake_outerrel_subquery/make_innerrel_subquery.\n* I modified appendSubselectAlias to handle the case where ncols = 0 \nproperly.\n* I renamed subquery_rels in fpinfo to lower_subquery_rels, to make it \neasy to distinguish this from is_subquery_rel clearly.\n* I added regression tests for that.\n\nThe rest of the patch is the same as the previous version, but:\n\n> + /* Should be same length */\n> + Assert(list_length(tlist) ==\n> list_length(foreignrel->reltarget->exprs));\n\nI added comments explaining the reason.\n\n> The name get_subselect_alias_id() seems to suggest that the function returns\n> identifier for subselect alias, which isn't correct. It returns the alias\n> identifiers for deparsing a Var node. But I guess, we have left this to the\n> committer's judgement.\n\nI agree that the name isn't good, so I changed it to \nget_relation_column_alias_ids(). Let me know if it may be better.\n\nI also revised code and comments a bit, just for consistency.\n\nI will update another patch for PHVs on top of the attached one.\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":"Thu, 24 Nov 2016 21:45:40 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/11/24 21:45, Etsuro Fujita wrote:\n> On 2016/11/22 18:28, Ashutosh Bapat wrote:\n>> The comments should explain why is the assertion true.\n>> + /* Shouldn't be NIL */\n>> + Assert(tlist != NIL);\n\n> I noticed that I was wrong; in the Assertion the tlist can be empty. An\n> example for such a case is:\n>\n> SELECT 1 FROM (SELECT c1 FROM ft4 WHERE c1 between 50 and 60) t1 FULL\n> JOIN (SELECT c1 FROM ft5 WHERE c1 between 50 and 60) t2 ON (TRUE);\n>\n> In this example any columns of the relations ft4 and ft5 wouldn't be\n> needed for the join or the final output, so both the tlists for the\n> relations created in deparseRangeTblRef would be empty. Attached is an\n> updated version fixing this bug. Changes are:\n>\n> * I removed make_outerrel_subquery/make_innerrel_subquery from fpinfo\n> and added a new member \"is_subquery_rel\" to fpinfo, as a flag on whether\n> to deparse the relation as a subquery.\n> * I modified deparseRangeTblRef, deparseSelectSql, and is_subquery_var\n> to see the is_subquery_rel, not\n> make_outerrel_subquery/make_innerrel_subquery.\n> * I modified appendSubselectAlias to handle the case where ncols = 0\n> properly.\n> * I renamed subquery_rels in fpinfo to lower_subquery_rels, to make it\n> easy to distinguish this from is_subquery_rel clearly.\n> * I added regression tests for that.\n\nI noticed that the above changes are not adequate; the updated patch \nbreaks EXPLAIN outputs for EvalPlanQual subplans of foreign joins in \nwhich foreign tables are full joined and in the remote join query the \nforeign tables are deparsed as subqueries. Consider:\n\npostgres=# explain verbose select * from test, ((select * from ft1 where \nb between 1 and 3) t1 full join (select * from ft2 where b between 1 and \n3) t2\n on (t1.a = t2.a)) ss for update of test;\n \n \nQUERY PLA\nN\n-------------------------------------------------------------------------------------------------------------------------------------------------------\n------------------------------------------------------------------------------------------------------------------------------------------------\n LockRows (cost=100.00..287.33 rows=6780 width=94)\n Output: test.a, test.b, ft1.a, ft1.b, ft2.a, ft2.b, test.ctid, \nft1.*, ft2.*\n -> Nested Loop (cost=100.00..219.53 rows=6780 width=94)\n Output: test.a, test.b, ft1.a, ft1.b, ft2.a, ft2.b, test.ctid, \nft1.*, ft2.*\n -> Seq Scan on public.test (cost=0.00..32.60 rows=2260 width=14)\n Output: test.a, test.b, test.ctid\n -> Materialize (cost=100.00..102.19 rows=3 width=80)\n Output: ft1.a, ft1.b, ft1.*, ft2.a, ft2.b, ft2.*\n -> Foreign Scan (cost=100.00..102.17 rows=3 width=80)\n Output: ft1.a, ft1.b, ft1.*, ft2.a, ft2.b, ft2.*\n Relations: (public.ft1) FULL JOIN (public.ft2)\n Remote SQL: SELECT s5.c1, s5.c2, s5.c3, s6.c1, \ns6.c2, s6.c3 FROM ((SELECT a, b, ROW(a, b) FROM public.t1 WHERE ((b >= \n1)) AND ((b\n<= 3))) s5(c1, c2, c3) FULL JOIN (SELECT a, b, ROW(a, b) FROM public.t2 \nWHERE ((b >= 1)) AND ((b <= 3))) s6(c1, c2, c3) ON (((s5.c1 = s6.c1))))\n -> Hash Full Join (cost=201.14..202.29 rows=3 \nwidth=80)\n Output: ft1.a, ft1.b, ft1.*, ft2.a, ft2.b, ft2.*\n Hash Cond: (ft1.a = ft2.a)\n -> Foreign Scan on public.ft1 \n(cost=100.00..101.11 rows=3 width=40)\n Output: ft1.a, ft1.b, ft1.*\n Remote SQL: SELECT NULL FROM public.t1 \nWHERE ((b >= 1)) AND ((b <= 3))\n -> Hash (cost=101.11..101.11 rows=3 width=40)\n Output: ft2.a, ft2.b, ft2.*\n -> Foreign Scan on public.ft2 \n(cost=100.00..101.11 rows=3 width=40)\n Output: ft2.a, ft2.b, ft2.*\n Remote SQL: SELECT NULL FROM \npublic.t2 WHERE ((b >= 1)) AND ((b <= 3))\n(23 rows)\n\nThe SELECT clauses of the remote queries for the Foreign Scans in the \nEPQ subplan are deparsed incorrectly into \"SELECT NULL\". The reason for \nthat is that since the foreign tables' fpinfo->is_subquery_rel has been \nset true at the time those clauses are deparsed (due to that the foreign \ntables have been deparsed as subqueries in the main remote join query), \ndeparseSelectSql calls deparseExplicitTargetList with an empty tlist, to \nconstruct the clauses, so it deparses the tlist into \"NULL\". This is \nharmless because we don't use the remote queries for the Foreign Scans \nduring an EPQ recheck, but that would be confusing for users, so I \nmodified the patch a bit further to make the EXPLAIN output as-is. \nAttached is a new version of the patch.\n\n> I also revised code and comments a bit, just for consistency.\n\nAlso, I renamed appendSubselectAlias to appendSubqueryAlias, because we \nuse the term \"subquery\" for other places.\n\n> I will update another patch for PHVs on top of the attached one.\n\nDone. I also revised some code and comments. I'm also attaching the \nupdated version of the patch for PHVs.\n\nBest regards,\nEtsuro Fujita\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":"Mon, 28 Nov 2016 19:22:03 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Mon, Nov 28, 2016 at 3:52 PM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> On 2016/11/24 21:45, Etsuro Fujita wrote:\n>>\n>> On 2016/11/22 18:28, Ashutosh Bapat wrote:\n>>>\n>>> The comments should explain why is the assertion true.\n>>> + /* Shouldn't be NIL */\n>>> + Assert(tlist != NIL);\n>\n>\n>> I noticed that I was wrong; in the Assertion the tlist can be empty. An\n>> example for such a case is:\n>>\n>> SELECT 1 FROM (SELECT c1 FROM ft4 WHERE c1 between 50 and 60) t1 FULL\n>> JOIN (SELECT c1 FROM ft5 WHERE c1 between 50 and 60) t2 ON (TRUE);\n>>\n>> In this example any columns of the relations ft4 and ft5 wouldn't be\n>> needed for the join or the final output, so both the tlists for the\n>> relations created in deparseRangeTblRef would be empty. Attached is an\n>> updated version fixing this bug. Changes are:\n>>\n>> * I removed make_outerrel_subquery/make_innerrel_subquery from fpinfo\n>> and added a new member \"is_subquery_rel\" to fpinfo, as a flag on whether\n>> to deparse the relation as a subquery.\n\nReplacing make_outerrel_subquery/make_innerrel_subquery with is_subquery_rel\nseems to be a bad idea. Whether a relation needs to be deparsed as a subquery\nor not depends upon the relation which joins it with another relation. Take for\nexample a case when a join ABC can pull up the conditions on AB, but ABD can\nnot. In this case we will mark that AB in ABD needs to be deparsed as a\nsubquery but will not mark so in ABC. So, if we choose to join ABCD as (ABC)D\nwe don't need AB to be deparsed as a subquery. If we choose to join ABCD as\n(ABD)C, we will need AB to deparsed as a subquery. Regression doesn't have a\ntestcase like that hence the code seems to be working.\n\nSome more comments\n1. The output of the following query\n+SELECT 1 FROM (SELECT c1 FROM ft4 WHERE c1 between 50 and 60) t1 FULL\nJOIN (SELECT c1 FROM ft5 WHERE c1 between 50 and 60) t2 ON (TRUE);\nproduces 21 rows all containing a \"1\". That's not quite good use of expected\noutput space, given that all that the test tests is the empty\ntargetlists are deparsed\ncorrectly. You might want to either just test EXPLAIN output or make \"between 50\nand 60\" condition more stringent to reduce the number of rows output.\n\n2. Got lost here. I guess, all you need to say is \"test deparsing FOR UPDATE\nclause with subqueries.\". Anyway, the sentence is very hard to read and needs\nsimplification.\n+-- d. test deparsing the remote queries for simple foreign table scans in\n+-- an EPQ subplan of a foreign join in which the foreign tables are full\n+-- joined and in the remote join query the foreign tables are deparsed as\n+-- subqueries\n\n3. Why is foreignrel variable changed to rel?\n-extern void deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root,\n- RelOptInfo *foreignrel, List *tlist,\n- List *remote_conds, List *pathkeys,\n- List **retrieved_attrs, List **params_list);\n+extern void deparseSelectStmtForRel(StringInfo buf, PlannerInfo\n*root, RelOptInfo *rel,\n+ List *tlist, List *remote_conds, List *pathkeys,\n+ bool is_subquery, List **retrieved_attrs,\n+ List **params_list);\n\n4. I am still not happy with this change\n+ /*\n+ * Since (1) the expressions in foreignrel's reltarget doesn't contain\n+ * any PHVs and (2) foreignrel's local_conds is empty, the tlist\n+ * created by build_tlist_to_deparse must be one-to-one with the\n+ * expressions.\n+ */\n+ Assert(list_length(tlist) ==\nlist_length(foreignrel->reltarget->exprs));\nthe assertion only checks that the number of elements in both the lists are\nsame but does not check whether those lists are same i.e. they contain the same\nelements in the same order. This equality is crucial to deparsing logic. If\nsomehow build_tlist_to_deparse() breaks that assumption in future, we have no\nway to detect it, unless a regression test fails.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 5 Dec 2016 16:50:45 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Mon, Dec 5, 2016 at 6:20 AM, Ashutosh Bapat\n<ashutosh.bapat@enterprisedb.com> wrote:\n> 4. I am still not happy with this change\n> + /*\n> + * Since (1) the expressions in foreignrel's reltarget doesn't contain\n> + * any PHVs and (2) foreignrel's local_conds is empty, the tlist\n> + * created by build_tlist_to_deparse must be one-to-one with the\n> + * expressions.\n> + */\n> + Assert(list_length(tlist) ==\n> list_length(foreignrel->reltarget->exprs));\n> the assertion only checks that the number of elements in both the lists are\n> same but does not check whether those lists are same i.e. they contain the same\n> elements in the same order. This equality is crucial to deparsing logic. If\n> somehow build_tlist_to_deparse() breaks that assumption in future, we have no\n> way to detect it, unless a regression test fails.\n\nIf there's an easy way to do a more exact comparison, great. But if\nwe can't get an awesome Assert(), a helpful Assert() is still better\nthan a kick in the head.\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, 6 Dec 2016 15:27:06 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/12/07 5:27, Robert Haas wrote:\n> On Mon, Dec 5, 2016 at 6:20 AM, Ashutosh Bapat\n> <ashutosh.bapat@enterprisedb.com> wrote:\n>> 4. I am still not happy with this change\n>> + /*\n>> + * Since (1) the expressions in foreignrel's reltarget doesn't contain\n>> + * any PHVs and (2) foreignrel's local_conds is empty, the tlist\n>> + * created by build_tlist_to_deparse must be one-to-one with the\n>> + * expressions.\n>> + */\n>> + Assert(list_length(tlist) ==\n>> list_length(foreignrel->reltarget->exprs));\n>> the assertion only checks that the number of elements in both the lists are\n>> same but does not check whether those lists are same i.e. they contain the same\n>> elements in the same order. This equality is crucial to deparsing logic. If\n>> somehow build_tlist_to_deparse() breaks that assumption in future, we have no\n>> way to detect it, unless a regression test fails.\n\n> If there's an easy way to do a more exact comparison, great.\n\nAshutosh proposed this to do the comparison:\n\nOn 2016/11/22 18:28, Ashutosh Bapat wrote:\n> I guess, the reason why you are doing it this way, is SELECT clause for the\n> outermost query gets deparsed before FROM clause. For later we call\n> deparseRangeTblRef(), which builds the tlist. So, while deparsing SELECT\n> clause, we do not have tlist to build from. In that case, I guess, we have to\n> build the tlist in get_subselect_alias_id() if it's not available and stick it\n> in fpinfo. Subsequent calls to get_subselect_alias_id() should find tlist\n> there. Then in deparseRangeTblRef() assert that there's a tlist in fpinfo\n> and use it to build the SELECT clause of subquery. That way, we don't build\n> tlist unless it's needed and also use the same tlist for all searches. Please\n> use tlist_member() to search into the tlist.\n\nThis would probably work, but seems to me a bit complicated. Instead, \nI'd like to propose that we build the tlist for each relation being \ndeparsed as a subquery in a given join tree, right before deparsing the \nSELECT clause in deparseSelectStmtForRel, if is_subquery is false and \nlower_subquery_rels isn't NULL, and store the tlist into the relation's \nfpinfo. That would allow us to build the tlist only when we need it, \nand to use tlist_member for the exact comparison. I think it would be \nmuch easier to implement that.\n\nBest regards,\nEtsuro Fujita\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, 7 Dec 2016 12:15:41 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Wed, Dec 7, 2016 at 1:57 AM, Robert Haas <robertmhaas@gmail.com> wrote:\n> On Mon, Dec 5, 2016 at 6:20 AM, Ashutosh Bapat\n> <ashutosh.bapat@enterprisedb.com> wrote:\n>> 4. I am still not happy with this change\n>> + /*\n>> + * Since (1) the expressions in foreignrel's reltarget doesn't contain\n>> + * any PHVs and (2) foreignrel's local_conds is empty, the tlist\n>> + * created by build_tlist_to_deparse must be one-to-one with the\n>> + * expressions.\n>> + */\n>> + Assert(list_length(tlist) ==\n>> list_length(foreignrel->reltarget->exprs));\n>> the assertion only checks that the number of elements in both the lists are\n>> same but does not check whether those lists are same i.e. they contain the same\n>> elements in the same order. This equality is crucial to deparsing logic. If\n>> somehow build_tlist_to_deparse() breaks that assumption in future, we have no\n>> way to detect it, unless a regression test fails.\n>\n> If there's an easy way to do a more exact comparison, great. But if\n> we can't get an awesome Assert(), a helpful Assert() is still better\n> than a kick in the head.\n\nThe assert is not a problem in itself, but the reason we have to add\nthe assert. The problem is explained in [1], point #9.\n\n[1] https://www.postgresql.org/message-id/CAFjFpRfwoSsJr9418b2jA7g0nwagjZSWhPQPUmK9M6z5XSOAqQ%40mail.gmail.com\n\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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 Dec 2016 12:02:35 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":">\n>\n> Ashutosh proposed this to do the comparison:\n>\n> On 2016/11/22 18:28, Ashutosh Bapat wrote:\n>>\n>> I guess, the reason why you are doing it this way, is SELECT clause for\n>> the\n>> outermost query gets deparsed before FROM clause. For later we call\n>> deparseRangeTblRef(), which builds the tlist. So, while deparsing SELECT\n>> clause, we do not have tlist to build from. In that case, I guess, we have\n>> to\n>> build the tlist in get_subselect_alias_id() if it's not available and\n>> stick it\n>> in fpinfo. Subsequent calls to get_subselect_alias_id() should find tlist\n>> there. Then in deparseRangeTblRef() assert that there's a tlist in fpinfo\n>> and use it to build the SELECT clause of subquery. That way, we don't\n>> build\n>> tlist unless it's needed and also use the same tlist for all searches.\n>> Please\n>> use tlist_member() to search into the tlist.\n>\n>\n> This would probably work, but seems to me a bit complicated. Instead, I'd\n> like to propose that we build the tlist for each relation being deparsed as\n> a subquery in a given join tree, right before deparsing the SELECT clause in\n> deparseSelectStmtForRel, if is_subquery is false and lower_subquery_rels\n> isn't NULL, and store the tlist into the relation's fpinfo. That would\n> allow us to build the tlist only when we need it, and to use tlist_member\n> for the exact comparison. I think it would be much easier to implement\n> that.\n>\n\nIIRC, this is inline with my original proposal of creating tlists\nbefore deparsing anything. Along-with that I also proposed to bubble\nthis array of tlist up the join hierarchy to avoid recursion [1] point\n#15, further elaborated in [2]\n\n[1] https://www.postgresql.org/message-id/ad449b25-66ee-4c06-568c-0eb2e1bef9f9%40lab.ntt.co.jp\n[2] https://www.postgresql.org/message-id/CAFjFpRcn7%3DDNOXm-PJ_jVDqAmghKVf6tApQC%2B_RgMZ8tOPExcA%40mail.gmail.com\n\nWe seem to be moving towards that solution, but as discussed we have\nleft it to committer's judgement.\n--\nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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 Dec 2016 12:09:01 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/12/05 20:20, Ashutosh Bapat wrote:\n> On Mon, Nov 28, 2016 at 3:52 PM, Etsuro Fujita\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>> On 2016/11/24 21:45, Etsuro Fujita wrote:\n>>> * I removed make_outerrel_subquery/make_innerrel_subquery from fpinfo\n>>> and added a new member \"is_subquery_rel\" to fpinfo, as a flag on whether\n>>> to deparse the relation as a subquery.\n\n> Replacing make_outerrel_subquery/make_innerrel_subquery with is_subquery_rel\n> seems to be a bad idea. Whether a relation needs to be deparsed as a subquery\n> or not depends upon the relation which joins it with another relation. Take for\n> example a case when a join ABC can pull up the conditions on AB, but ABD can\n> not. In this case we will mark that AB in ABD needs to be deparsed as a\n> subquery but will not mark so in ABC. So, if we choose to join ABCD as (ABC)D\n> we don't need AB to be deparsed as a subquery. If we choose to join ABCD as\n> (ABD)C, we will need AB to deparsed as a subquery.\n\nThis is not true; since (1) currently, a relation needs to be deparsed \nas a subquery only when the relation is full-joined with any other \nrelation, and (2) the join ordering for full joins is preserved, we \nwouldn't have such an example. (You might think we would have that when \nextending the patch to handle PHVs, but the answer would be no, because \nthe patch for PHVs would determine whether a relation emitting PHVs \nneeds to be deparsed as a subquery, depending on the relation itself. \nSee the patch for PHVs.) I like is_subquery_rel more than \nmake_outerrel_subquery/make_innerrel_subquery, because it reduces the \nnumber of members in fpinfo. But the choice would be arbitrary, so I \nwouldn't object to going back to \nmake_outerrel_subquery/make_innerrel_subquery.\n\n> Some more comments\n> 1. The output of the following query\n> +SELECT 1 FROM (SELECT c1 FROM ft4 WHERE c1 between 50 and 60) t1 FULL\n> JOIN (SELECT c1 FROM ft5 WHERE c1 between 50 and 60) t2 ON (TRUE);\n> produces 21 rows all containing a \"1\". That's not quite good use of expected\n> output space, given that all that the test tests is the empty\n> targetlists are deparsed\n> correctly. You might want to either just test EXPLAIN output or make \"between 50\n> and 60\" condition more stringent to reduce the number of rows output.\n\nWill do.\n\n> 2. Got lost here. I guess, all you need to say is \"test deparsing FOR UPDATE\n> clause with subqueries.\". Anyway, the sentence is very hard to read and needs\n> simplification.\n> +-- d. test deparsing the remote queries for simple foreign table scans in\n> +-- an EPQ subplan of a foreign join in which the foreign tables are full\n> +-- joined and in the remote join query the foreign tables are deparsed as\n> +-- subqueries\n\nWill do.\n\n> 3. Why is foreignrel variable changed to rel?\n> -extern void deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root,\n> - RelOptInfo *foreignrel, List *tlist,\n> - List *remote_conds, List *pathkeys,\n> - List **retrieved_attrs, List **params_list);\n> +extern void deparseSelectStmtForRel(StringInfo buf, PlannerInfo\n> *root, RelOptInfo *rel,\n> + List *tlist, List *remote_conds, List *pathkeys,\n> + bool is_subquery, List **retrieved_attrs,\n> + List **params_list);\n\nI changed the name that way to match with the function definition in \ndeparse.c.\n\nThanks for the comments!\n\nBest regards,\nEtsuro Fujita\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, 7 Dec 2016 20:21:58 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/12/07 15:39, Ashutosh Bapat wrote:\n\n>> On 2016/11/22 18:28, Ashutosh Bapat wrote:\n>>>\n>>> I guess, the reason why you are doing it this way, is SELECT clause for\n>>> the\n>>> outermost query gets deparsed before FROM clause. For later we call\n>>> deparseRangeTblRef(), which builds the tlist. So, while deparsing SELECT\n>>> clause, we do not have tlist to build from. In that case, I guess, we have\n>>> to\n>>> build the tlist in get_subselect_alias_id() if it's not available and\n>>> stick it\n>>> in fpinfo. Subsequent calls to get_subselect_alias_id() should find tlist\n>>> there. Then in deparseRangeTblRef() assert that there's a tlist in fpinfo\n>>> and use it to build the SELECT clause of subquery. That way, we don't\n>>> build\n>>> tlist unless it's needed and also use the same tlist for all searches.\n>>> Please\n>>> use tlist_member() to search into the tlist.\n\nI wrote:\n>> This would probably work, but seems to me a bit complicated. Instead, I'd\n>> like to propose that we build the tlist for each relation being deparsed as\n>> a subquery in a given join tree, right before deparsing the SELECT clause in\n>> deparseSelectStmtForRel, if is_subquery is false and lower_subquery_rels\n>> isn't NULL, and store the tlist into the relation's fpinfo. That would\n>> allow us to build the tlist only when we need it, and to use tlist_member\n>> for the exact comparison. I think it would be much easier to implement\n>> that.\n\n> IIRC, this is inline with my original proposal of creating tlists\n> before deparsing anything. Along-with that I also proposed to bubble\n> this array of tlist up the join hierarchy to avoid recursion [1] point\n> #15, further elaborated in [2]\n>\n> [1] https://www.postgresql.org/message-id/ad449b25-66ee-4c06-568c-0eb2e1bef9f9%40lab.ntt.co.jp\n> [2] https://www.postgresql.org/message-id/CAFjFpRcn7%3DDNOXm-PJ_jVDqAmghKVf6tApQC%2B_RgMZ8tOPExcA%40mail.gmail.com\n>\n> We seem to be moving towards that solution, but as discussed we have\n> left it to committer's judgement.\n\nMy proposal here would be a bit different from what you proposed; right \nbefore deparseSelectSql in deparseSelectStmtForRel, build the tlists for \nrelations present in the given jointree that will be deparsed as \nsubqueries, by (1) traversing the given jointree and (2) applying \nbuild_tlist_to_deparse to each relation to be deparsed as a subquery and \nsaving the result in that relation's fpinfo. I think that would be \nimplemented easily, and the overhead would be small.\n\nBest regards,\nEtsuro Fujita\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, 7 Dec 2016 20:23:41 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Wed, Dec 7, 2016 at 4:51 PM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> On 2016/12/05 20:20, Ashutosh Bapat wrote:\n>>\n>> On Mon, Nov 28, 2016 at 3:52 PM, Etsuro Fujita\n>> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>>>\n>>> On 2016/11/24 21:45, Etsuro Fujita wrote:\n>>>>\n>>>> * I removed make_outerrel_subquery/make_innerrel_subquery from fpinfo\n>>>> and added a new member \"is_subquery_rel\" to fpinfo, as a flag on whether\n>>>> to deparse the relation as a subquery.\n>\n>\n>> Replacing make_outerrel_subquery/make_innerrel_subquery with\n>> is_subquery_rel\n>> seems to be a bad idea. Whether a relation needs to be deparsed as a\n>> subquery\n>> or not depends upon the relation which joins it with another relation.\n>> Take for\n>> example a case when a join ABC can pull up the conditions on AB, but ABD\n>> can\n>> not. In this case we will mark that AB in ABD needs to be deparsed as a\n>> subquery but will not mark so in ABC. So, if we choose to join ABCD as\n>> (ABC)D\n>> we don't need AB to be deparsed as a subquery. If we choose to join ABCD\n>> as\n>> (ABD)C, we will need AB to deparsed as a subquery.\n>\n>\n> This is not true; since (1) currently, a relation needs to be deparsed as a\n> subquery only when the relation is full-joined with any other relation, and\n> (2) the join ordering for full joins is preserved, we wouldn't have such an\n> example. (You might think we would have that when extending the patch to\n> handle PHVs, but the answer would be no, because the patch for PHVs would\n> determine whether a relation emitting PHVs needs to be deparsed as a\n> subquery, depending on the relation itself. See the patch for PHVs.) I like\n> is_subquery_rel more than make_outerrel_subquery/make_innerrel_subquery,\n> because it reduces the number of members in fpinfo. But the choice would be\n> arbitrary, so I wouldn't object to going back to\n> make_outerrel_subquery/make_innerrel_subquery.\n\nI think you are right. And even if we were to deparse a relation as\nsubquery, when it shouldn't be, we will only loose a syntactic\noptimization. So, it shouldn't result in a bug.\n\n\n>> 3. Why is foreignrel variable changed to rel?\n>> -extern void deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root,\n>> - RelOptInfo *foreignrel, List *tlist,\n>> - List *remote_conds, List *pathkeys,\n>> - List **retrieved_attrs, List **params_list);\n>> +extern void deparseSelectStmtForRel(StringInfo buf, PlannerInfo\n>> *root, RelOptInfo *rel,\n>> + List *tlist, List *remote_conds, List *pathkeys,\n>> + bool is_subquery, List **retrieved_attrs,\n>> + List **params_list);\n>\n>\n> I changed the name that way to match with the function definition in\n> deparse.c.\n>\n\nThat's something good to do but it's unrelated to this patch. I have\nrepeated this line several times in this thread :)\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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 Dec 2016 17:41:02 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/12/07 21:11, Ashutosh Bapat wrote:\n> On Wed, Dec 7, 2016 at 4:51 PM, Etsuro Fujita\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>> On 2016/12/05 20:20, Ashutosh Bapat wrote:\n\n>>> 3. Why is foreignrel variable changed to rel?\n>>> -extern void deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root,\n>>> - RelOptInfo *foreignrel, List *tlist,\n>>> - List *remote_conds, List *pathkeys,\n>>> - List **retrieved_attrs, List **params_list);\n>>> +extern void deparseSelectStmtForRel(StringInfo buf, PlannerInfo\n>>> *root, RelOptInfo *rel,\n>>> + List *tlist, List *remote_conds, List *pathkeys,\n>>> + bool is_subquery, List **retrieved_attrs,\n>>> + List **params_list);\n\n>> I changed the name that way to match with the function definition in\n>> deparse.c.\n\n> That's something good to do but it's unrelated to this patch. I have\n> repeated this line several times in this thread :)\n\nOK.\n\nBest regards,\nEtsuro Fujita\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, 8 Dec 2016 21:04:57 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/12/07 20:23, Etsuro Fujita wrote:\n> On 2016/12/07 15:39, Ashutosh Bapat wrote:\n\n>>> On 2016/11/22 18:28, Ashutosh Bapat wrote:\n>>>>\n>>>> I guess, the reason why you are doing it this way, is SELECT clause for\n>>>> the\n>>>> outermost query gets deparsed before FROM clause. For later we call\n>>>> deparseRangeTblRef(), which builds the tlist. So, while deparsing\n>>>> SELECT\n>>>> clause, we do not have tlist to build from. In that case, I guess,\n>>>> we have\n>>>> to\n>>>> build the tlist in get_subselect_alias_id() if it's not available and\n>>>> stick it\n>>>> in fpinfo. Subsequent calls to get_subselect_alias_id() should find\n>>>> tlist\n>>>> there. Then in deparseRangeTblRef() assert that there's a tlist in\n>>>> fpinfo\n>>>> and use it to build the SELECT clause of subquery. That way, we don't\n>>>> build\n>>>> tlist unless it's needed and also use the same tlist for all searches.\n>>>> Please\n>>>> use tlist_member() to search into the tlist.\n\n> I wrote:\n>>> This would probably work, but seems to me a bit complicated.\n>>> Instead, I'd\n>>> like to propose that we build the tlist for each relation being\n>>> deparsed as\n>>> a subquery in a given join tree, right before deparsing the SELECT\n>>> clause in\n>>> deparseSelectStmtForRel, if is_subquery is false and lower_subquery_rels\n>>> isn't NULL, and store the tlist into the relation's fpinfo. That would\n>>> allow us to build the tlist only when we need it, and to use\n>>> tlist_member\n>>> for the exact comparison. I think it would be much easier to implement\n>>> that.\n\n>> IIRC, this is inline with my original proposal of creating tlists\n>> before deparsing anything. Along-with that I also proposed to bubble\n>> this array of tlist up the join hierarchy to avoid recursion [1] point\n>> #15, further elaborated in [2]\n>>\n>> [1]\n>> https://www.postgresql.org/message-id/ad449b25-66ee-4c06-568c-0eb2e1bef9f9%40lab.ntt.co.jp\n>>\n>> [2]\n>> https://www.postgresql.org/message-id/CAFjFpRcn7%3DDNOXm-PJ_jVDqAmghKVf6tApQC%2B_RgMZ8tOPExcA%40mail.gmail.com\n\n> My proposal here would be a bit different from what you proposed; right\n> before deparseSelectSql in deparseSelectStmtForRel, build the tlists for\n> relations present in the given jointree that will be deparsed as\n> subqueries, by (1) traversing the given jointree and (2) applying\n> build_tlist_to_deparse to each relation to be deparsed as a subquery and\n> saving the result in that relation's fpinfo. I think that would be\n> implemented easily, and the overhead would be small.\n\nI implemented that to address your concern:\n* deparseRangeTblRef uses the tlist created as above, to build the \nSELECT clause of the subquery. (You said \"Then in deparseRangeTblRef() \nassert that there's a tlist in fpinfo\", but the tlist may be empty, so I \ndidn't add any assertion to that function.)\n* get_relation_column_alias_ids uses tlist_member with the tlist.\n\nI also addressed the comments #1, #2 and #3 in [1]. For #1, I added \n\"LIMIT 10\" to the query. Attached is the new version of the patch.\n\nOther changes:\n* As discussed before, renamed grouped_tlist in fpinfo to \"tlist\" and \nused it to store the tlist created as above.\n* Also, renamed SS_REL_ALIAS_PREFIX to SUBQUERY_REL_ALIAS_PREFIX \n(Likewise for SS_COL_ALIAS_PREFIX).\n* Relocated some functions.\n* Revised comments a little bit.\n\nBest regards,\nEtsuro Fujita\n\n[1] \nhttps://www.postgresql.org/message-id/CAFjFpRfU4-gxqZ8ahoKM1ZtDJEThe3K8Lb_6beRKa8mmP%3Dv%3DfA%40mail.gmail.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, 8 Dec 2016 21:08:57 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2016/12/08 21:08, Etsuro Fujita wrote:\n> On 2016/12/07 20:23, Etsuro Fujita wrote:\n>> My proposal here would be a bit different from what you proposed; right\n>> before deparseSelectSql in deparseSelectStmtForRel, build the tlists for\n>> relations present in the given jointree that will be deparsed as\n>> subqueries, by (1) traversing the given jointree and (2) applying\n>> build_tlist_to_deparse to each relation to be deparsed as a subquery and\n>> saving the result in that relation's fpinfo. I think that would be\n>> implemented easily, and the overhead would be small.\n\n> I implemented that to address your concern:\n> * deparseRangeTblRef uses the tlist created as above, to build the\n> SELECT clause of the subquery. (You said \"Then in deparseRangeTblRef()\n> assert that there's a tlist in fpinfo\", but the tlist may be empty, so I\n> didn't add any assertion to that function.)\n> * get_relation_column_alias_ids uses tlist_member with the tlist.\n>\n> I also addressed the comments #1, #2 and #3 in [1]. For #1, I added\n> \"LIMIT 10\" to the query. Attached is the new version of the patch.\n>\n> Other changes:\n> * As discussed before, renamed grouped_tlist in fpinfo to \"tlist\" and\n> used it to store the tlist created as above.\n> * Also, renamed SS_REL_ALIAS_PREFIX to SUBQUERY_REL_ALIAS_PREFIX\n> (Likewise for SS_COL_ALIAS_PREFIX).\n> * Relocated some functions.\n> * Revised comments a little bit.\n\nI updated the patch a bit further: simplified the function name \n(s/build_subquery_rel_tlists/build_subquery_tlists/), and revised \ncomments a little bit. Attached is an updated version \n(postgres-fdw-subquery-support-v14.patch). And I rebased another patch \nfor PHVs against that patch, which is also attached \n(postgres-fdw-phv-pushdown-v14.patch).\n\nBest regards,\nEtsuro Fujita\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, 27 Dec 2016 21:51:06 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":">\n> I updated the patch a bit further: simplified the function name\n> (s/build_subquery_rel_tlists/build_subquery_tlists/), and revised comments a\n> little bit. Attached is an updated version\n> (postgres-fdw-subquery-support-v14.patch). And I rebased another patch for\n> PHVs against that patch, which is also attached\n> (postgres-fdw-phv-pushdown-v14.patch).\n>\nFew comments\n\nIn build_subquery_tlists(), why don't we handle base relations?\n+ if (foreignrel->reloptkind != RELOPT_JOINREL)\n+ return;\n\nAlso, in this function, if fpinfo->tlist is already set, why do we want to\nbuild it again?\n\nIn build_tlist_to_deparse(), if fpinfo->tlist for the given relation is set, we\nshould just return it rather than constructing it again.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 3 Jan 2017 13:58:23 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2017/01/03 17:28, Ashutosh Bapat wrote:\n\nI wrote:\n>> I updated the patch a bit further: simplified the function name\n>> (s/build_subquery_rel_tlists/build_subquery_tlists/), and revised comments a\n>> little bit. Attached is an updated version\n>> (postgres-fdw-subquery-support-v14.patch).\n\n> Few comments\n\nThanks for the comments!\n\n> In build_subquery_tlists(), why don't we handle base relations?\n> + if (foreignrel->reloptkind != RELOPT_JOINREL)\n> + return;\n\nThe reason for that is we don't need to handle the baserel cases; the \ntlist for a base relation, if needed, would be created while recursing \ninto a join relation that joins the base relation to other base/join \nrelation.\n\n> Also, in this function, if fpinfo->tlist is already set, why do we want to\n> build it again?\n\nWhen this function gets called, fpinfo->tlist isn't set for any base or \njoin relation that needs to build the tlist, so we always need to build \nit for each such relation.\n\n> In build_tlist_to_deparse(), if fpinfo->tlist for the given relation is set, we\n> should just return it rather than constructing it again.\n\nIn that function we wouldn't have such cases for base or join relations \nneeding the tlist.\n\nBest regards,\nEtsuro Fujita\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, 5 Jan 2017 20:44:51 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Thu, Jan 5, 2017 at 5:14 PM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> On 2017/01/03 17:28, Ashutosh Bapat wrote:\n>\n> I wrote:\n>>>\n>>> I updated the patch a bit further: simplified the function name\n>>> (s/build_subquery_rel_tlists/build_subquery_tlists/), and revised\n>>> comments a\n>>> little bit. Attached is an updated version\n>>> (postgres-fdw-subquery-support-v14.patch).\n>\n>\n>> Few comments\n>\n>\n> Thanks for the comments!\n>\n>> In build_subquery_tlists(), why don't we handle base relations?\n>> + if (foreignrel->reloptkind != RELOPT_JOINREL)\n>> + return;\n>\n>\n> The reason for that is we don't need to handle the baserel cases; the tlist\n> for a base relation, if needed, would be created while recursing into a join\n> relation that joins the base relation to other base/join relation.\n\nRight. Sorry, I misunderstood the code. May be a comment would help.\n\n>\n>> Also, in this function, if fpinfo->tlist is already set, why do we want to\n>> build it again?\n>\n>\n> When this function gets called, fpinfo->tlist isn't set for any base or join\n> relation that needs to build the tlist, so we always need to build it for\n> each such relation.\n\nIIUC, for a relation with use_remote_estimates we will deparse the\nquery twice and will build the targetlist twice.\n\n>\n>> In build_tlist_to_deparse(), if fpinfo->tlist for the given relation is\n>> set, we\n>> should just return it rather than constructing it again.\n>\n>\n> In that function we wouldn't have such cases for base or join relations\n> needing the tlist.\n\nSame explanation as above.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 5 Jan 2017 17:41:26 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2017/01/05 21:11, Ashutosh Bapat wrote:\n> On Thu, Jan 5, 2017 at 5:14 PM, Etsuro Fujita\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>> On 2017/01/03 17:28, Ashutosh Bapat wrote:\n>>> In build_subquery_tlists(), why don't we handle base relations?\n>>> + if (foreignrel->reloptkind != RELOPT_JOINREL)\n>>> + return;\n\n>> The reason for that is we don't need to handle the baserel cases; the tlist\n>> for a base relation, if needed, would be created while recursing into a join\n>> relation that joins the base relation to other base/join relation.\n\n> Right. Sorry, I misunderstood the code. May be a comment would help.\n\nWill add the comment.\n\n>>> Also, in this function, if fpinfo->tlist is already set, why do we want to\n>>> build it again?\n\n>> When this function gets called, fpinfo->tlist isn't set for any base or join\n>> relation that needs to build the tlist, so we always need to build it for\n>> each such relation.\n\n> IIUC, for a relation with use_remote_estimates we will deparse the\n> query twice and will build the targetlist twice.\n\nThat's right. We could avoid the duplicate work the way you proposed, \nbut I was thinking to leave that for another patch. Should we do that \nin this patch?\n\n>>> In build_tlist_to_deparse(), if fpinfo->tlist for the given relation is\n>>> set, we\n>>> should just return it rather than constructing it again.\n\n>> In that function we wouldn't have such cases for base or join relations\n>> needing the tlist.\n\n> Same explanation as above.\n\nWill revise if it's better to do that in this patch.\n\nBest regards,\nEtsuro Fujita\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, 5 Jan 2017 21:21:56 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Thu, Jan 5, 2017 at 5:51 PM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> On 2017/01/05 21:11, Ashutosh Bapat wrote:\n>>\n>> On Thu, Jan 5, 2017 at 5:14 PM, Etsuro Fujita\n>> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>>>\n>>> On 2017/01/03 17:28, Ashutosh Bapat wrote:\n>>>>\n>>>> In build_subquery_tlists(), why don't we handle base relations?\n>>>> + if (foreignrel->reloptkind != RELOPT_JOINREL)\n>>>> + return;\n>\n>\n>>> The reason for that is we don't need to handle the baserel cases; the\n>>> tlist\n>>> for a base relation, if needed, would be created while recursing into a\n>>> join\n>>> relation that joins the base relation to other base/join relation.\n>\n>\n>> Right. Sorry, I misunderstood the code. May be a comment would help.\n>\n>\n> Will add the comment.\n>\n>>>> Also, in this function, if fpinfo->tlist is already set, why do we want\n>>>> to\n>>>> build it again?\n>\n>\n>>> When this function gets called, fpinfo->tlist isn't set for any base or\n>>> join\n>>> relation that needs to build the tlist, so we always need to build it for\n>>> each such relation.\n>\n>\n>> IIUC, for a relation with use_remote_estimates we will deparse the\n>> query twice and will build the targetlist twice.\n>\n>\n> That's right. We could avoid the duplicate work the way you proposed, but I\n> was thinking to leave that for another patch. Should we do that in this\n> patch?\n\nIf you are agree that the change is needed, it's better to do it in\nthis patch itself if we can, instead of a one liner patch.\n>\n>>>> In build_tlist_to_deparse(), if fpinfo->tlist for the given relation is\n>>>> set, we\n>>>> should just return it rather than constructing it again.\n>\n>\n>>> In that function we wouldn't have such cases for base or join relations\n>>> needing the tlist.\n>\n>\n>> Same explanation as above.\n>\n>\n> Will revise if it's better to do that in this patch.\n\nThanks.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 5 Jan 2017 18:08:17 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2017/01/05 21:38, Ashutosh Bapat wrote:\n> On Thu, Jan 5, 2017 at 5:51 PM, Etsuro Fujita\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>> On 2017/01/05 21:11, Ashutosh Bapat wrote:\n\n>>>> On 2017/01/03 17:28, Ashutosh Bapat wrote:\n>>>>> Also, in this function, if fpinfo->tlist is already set, why do we want\n>>>>> to\n>>>>> build it again?\n\n>>> IIUC, for a relation with use_remote_estimates we will deparse the\n>>> query twice and will build the targetlist twice.\n\n>> That's right. We could avoid the duplicate work the way you proposed, but I\n>> was thinking to leave that for another patch. Should we do that in this\n>> patch?\n\n> If you are agree that the change is needed, it's better to do it in\n> this patch itself if we can, instead of a one liner patch.\n\nWhile working on this, I noticed a weird case. Consider:\n\npostgres=# explain verbose select 1 from ft1 left join ft2 on (ft1.a = \nft2.a) inner join test on (true);\n QUERY PLAN\n-------------------------------------------------------------------------------------------------\n Nested Loop (cost=100.00..103.06 rows=1 width=4)\n Output: 1\n -> Foreign Scan (cost=100.00..102.04 rows=1 width=0)\n Relations: (public.ft1) LEFT JOIN (public.ft2)\n Remote SQL: SELECT NULL FROM (public.t1 r1 LEFT JOIN public.t2 \nr2 ON (((r1.a = r2.a))))\n -> Seq Scan on public.test (cost=0.00..1.01 rows=1 width=0)\n Output: test.a, test.b\n(7 rows)\n\nIn this case the fpinfo->tlist of the foreign join is NIL, so whether or \nnot the tlist is already built cannot be discriminated by the \nfpinfo->tlist. We might need another flag to show that the tlist has \nbeen built already. Since this is an existing issue and we would need \nto give careful thought to this, so I'd like to leave this for another \npatch.\n\nBest regards,\nEtsuro Fujita\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, 11 Jan 2017 21:15:34 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Wed, Jan 11, 2017 at 5:45 PM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> On 2017/01/05 21:38, Ashutosh Bapat wrote:\n>>\n>> On Thu, Jan 5, 2017 at 5:51 PM, Etsuro Fujita\n>> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>>>\n>>> On 2017/01/05 21:11, Ashutosh Bapat wrote:\n>\n>\n>>>>> On 2017/01/03 17:28, Ashutosh Bapat wrote:\n>>>>>>\n>>>>>> Also, in this function, if fpinfo->tlist is already set, why do we\n>>>>>> want\n>>>>>> to\n>>>>>> build it again?\n>\n>\n>>>> IIUC, for a relation with use_remote_estimates we will deparse the\n>>>> query twice and will build the targetlist twice.\n>\n>\n>>> That's right. We could avoid the duplicate work the way you proposed,\n>>> but I\n>>> was thinking to leave that for another patch. Should we do that in this\n>>> patch?\n>\n>\n>> If you are agree that the change is needed, it's better to do it in\n>> this patch itself if we can, instead of a one liner patch.\n>\n>\n> While working on this, I noticed a weird case. Consider:\n>\n> postgres=# explain verbose select 1 from ft1 left join ft2 on (ft1.a =\n> ft2.a) inner join test on (true);\n> QUERY PLAN\n> -------------------------------------------------------------------------------------------------\n> Nested Loop (cost=100.00..103.06 rows=1 width=4)\n> Output: 1\n> -> Foreign Scan (cost=100.00..102.04 rows=1 width=0)\n> Relations: (public.ft1) LEFT JOIN (public.ft2)\n> Remote SQL: SELECT NULL FROM (public.t1 r1 LEFT JOIN public.t2 r2\n> ON (((r1.a = r2.a))))\n> -> Seq Scan on public.test (cost=0.00..1.01 rows=1 width=0)\n> Output: test.a, test.b\n> (7 rows)\n>\n> In this case the fpinfo->tlist of the foreign join is NIL, so whether or not\n> the tlist is already built cannot be discriminated by the fpinfo->tlist. We\n> might need another flag to show that the tlist has been built already.\n> Since this is an existing issue and we would need to give careful thought to\n> this, so I'd like to leave this for another patch.\n\nI think in that case, relation's targetlist will also be NIL or\ncontain no Var node. It wouldn't be expensive to build it again and\nagain. That's better than maintaining a new flag. This is just a\nsuggestion, for an additional check, you might want to check\nrel->reltarget->exprs for NIL. But I think we don't need it.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 12 Jan 2017 14:55:01 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2017/01/12 18:25, Ashutosh Bapat wrote:\n> On Wed, Jan 11, 2017 at 5:45 PM, Etsuro Fujita\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\n\n>>>> On 2017/01/05 21:11, Ashutosh Bapat wrote:\n>>>>> IIUC, for a relation with use_remote_estimates we will deparse the\n>>>>> query twice and will build the targetlist twice.\n\n>> While working on this, I noticed a weird case. Consider:\n>>\n>> postgres=# explain verbose select 1 from ft1 left join ft2 on (ft1.a =\n>> ft2.a) inner join test on (true);\n>> QUERY PLAN\n>> -------------------------------------------------------------------------------------------------\n>> Nested Loop (cost=100.00..103.06 rows=1 width=4)\n>> Output: 1\n>> -> Foreign Scan (cost=100.00..102.04 rows=1 width=0)\n>> Relations: (public.ft1) LEFT JOIN (public.ft2)\n>> Remote SQL: SELECT NULL FROM (public.t1 r1 LEFT JOIN public.t2 r2\n>> ON (((r1.a = r2.a))))\n>> -> Seq Scan on public.test (cost=0.00..1.01 rows=1 width=0)\n>> Output: test.a, test.b\n>> (7 rows)\n>>\n>> In this case the fpinfo->tlist of the foreign join is NIL, so whether or not\n>> the tlist is already built cannot be discriminated by the fpinfo->tlist. We\n>> might need another flag to show that the tlist has been built already.\n>> Since this is an existing issue and we would need to give careful thought to\n>> this, so I'd like to leave this for another patch.\n\n> I think in that case, relation's targetlist will also be NIL or\n> contain no Var node. It wouldn't be expensive to build it again and\n> again. That's better than maintaining a new flag.\n\nI think that's ugly. A more clean way I'm thinking is: (1) in \npostgresGetForeignJoinPaths(), create a tlist by \nbuild_tlist_to_deparse() and save it in fpinfo->tlist before \nestimate_path_cost_size() if use_remote_estimates=true, (2) in \nestimate_path_cost_size(), use the fpinfo->tlist if \nuse_remote_estimates=true, and (3) in postgresGetForeignPlan(), use the \nfpinfo->tlist as the fdw_scan_tlist if use_remote_estimates=true, \notherwise create a tlist as the fdw_scan_tlist by \nbuild_tlist_to_deparse(), like the attached.\n\nWhat do you think about that?\n\nAnother change is: I simplified build_tlist_to_deparse() a bit and put \nthat in postgres_fdw.c because that is used only in postgres_fdw.c.\n\nI still think we should discuss this separately because this is an \nexisting issue and that makes it easy to review the patch. If the \nattached is the right way to go, I'll update the join-pushdown patch on \ntop of the patch.\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":"Fri, 13 Jan 2017 16:00:19 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Fri, Jan 13, 2017 at 12:30 PM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> On 2017/01/12 18:25, Ashutosh Bapat wrote:\n>>\n>> On Wed, Jan 11, 2017 at 5:45 PM, Etsuro Fujita\n>> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>\n>\n>>>>> On 2017/01/05 21:11, Ashutosh Bapat wrote:\n>>>>>>\n>>>>>> IIUC, for a relation with use_remote_estimates we will deparse the\n>>>>>> query twice and will build the targetlist twice.\n>\n>\n>>> While working on this, I noticed a weird case. Consider:\n>>>\n>>> postgres=# explain verbose select 1 from ft1 left join ft2 on (ft1.a =\n>>> ft2.a) inner join test on (true);\n>>> QUERY PLAN\n>>>\n>>> -------------------------------------------------------------------------------------------------\n>>> Nested Loop (cost=100.00..103.06 rows=1 width=4)\n>>> Output: 1\n>>> -> Foreign Scan (cost=100.00..102.04 rows=1 width=0)\n>>> Relations: (public.ft1) LEFT JOIN (public.ft2)\n>>> Remote SQL: SELECT NULL FROM (public.t1 r1 LEFT JOIN public.t2\n>>> r2\n>>> ON (((r1.a = r2.a))))\n>>> -> Seq Scan on public.test (cost=0.00..1.01 rows=1 width=0)\n>>> Output: test.a, test.b\n>>> (7 rows)\n>>>\n>>> In this case the fpinfo->tlist of the foreign join is NIL, so whether or\n>>> not\n>>> the tlist is already built cannot be discriminated by the fpinfo->tlist.\n>>> We\n>>> might need another flag to show that the tlist has been built already.\n>>> Since this is an existing issue and we would need to give careful thought\n>>> to\n>>> this, so I'd like to leave this for another patch.\n>\n>\n>> I think in that case, relation's targetlist will also be NIL or\n>> contain no Var node. It wouldn't be expensive to build it again and\n>> again. That's better than maintaining a new flag.\n>\n>\n> I think that's ugly. A more clean way I'm thinking is: (1) in\n> postgresGetForeignJoinPaths(), create a tlist by build_tlist_to_deparse()\n> and save it in fpinfo->tlist before estimate_path_cost_size() if\n> use_remote_estimates=true, (2) in estimate_path_cost_size(), use the\n> fpinfo->tlist if use_remote_estimates=true, and (3) in\n> postgresGetForeignPlan(), use the fpinfo->tlist as the fdw_scan_tlist if\n> use_remote_estimates=true, otherwise create a tlist as the fdw_scan_tlist by\n> build_tlist_to_deparse(), like the attached.\n>\n> What do you think about that?\n>\n> Another change is: I simplified build_tlist_to_deparse() a bit and put that\n> in postgres_fdw.c because that is used only in postgres_fdw.c.\n>\n> I still think we should discuss this separately because this is an existing\n> issue and that makes it easy to review the patch. If the attached is the\n> right way to go, I'll update the join-pushdown patch on top of the patch.\n\nI don't think it's right to assume that the targetlist will be\navailable only when use_remote_estimate is true; for grouping it's\ncertainly not.\n\nBut I don't see this discussion getting anywhere. I will leave it to\nthe committer's judgement. I think we should pick up your patch on\n27th December, update the comment per your mail on 5th Jan. I will\nreview it once and list down the things left to committer's judgement.\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 27 Jan 2017 16:34:06 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2017/01/27 20:04, Ashutosh Bapat wrote:\n> On Fri, Jan 13, 2017 at 12:30 PM, Etsuro Fujita\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>> A more clean way I'm thinking is: (1) in\n>> postgresGetForeignJoinPaths(), create a tlist by build_tlist_to_deparse()\n>> and save it in fpinfo->tlist before estimate_path_cost_size() if\n>> use_remote_estimates=true, (2) in estimate_path_cost_size(), use the\n>> fpinfo->tlist if use_remote_estimates=true, and (3) in\n>> postgresGetForeignPlan(), use the fpinfo->tlist as the fdw_scan_tlist if\n>> use_remote_estimates=true, otherwise create a tlist as the fdw_scan_tlist by\n>> build_tlist_to_deparse(), like the attached.\n>>\n>> What do you think about that?\n>>\n>> Another change is: I simplified build_tlist_to_deparse() a bit and put that\n>> in postgres_fdw.c because that is used only in postgres_fdw.c.\n>>\n>> I still think we should discuss this separately because this is an existing\n>> issue and that makes it easy to review the patch. If the attached is the\n>> right way to go, I'll update the join-pushdown patch on top of the patch.\n\n> I don't think it's right to assume that the targetlist will be\n> available only when use_remote_estimate is true; for grouping it's\n> certainly not.\n\nMy explanation was not enough. Sorry about that. My proposal described \nabove was for join relations, not upper relations. We build the \ntargetlist of an upper relation during postgresGetForeignUpperPaths, so \nfor grouping I think we should use that targetlist in \nestimate_path_cost_size and postgresGetForeignPlan. The patch was \ncreated that way.\n\n> But I don't see this discussion getting anywhere. I will leave it to\n> the committer's judgement.\n\nI'm fine with that.\n\n> I think we should pick up your patch on\n> 27th December, update the comment per your mail on 5th Jan. I will\n> review it once and list down the things left to committer's judgement.\n\nSorry, I started thinking we went in the wrong direction. I added to \ndeparseSelectStmtForRel build_subquery_tlists, which creates a tlist for \neach subquery present in a given join tree prior to deparsing a whole \nremote query. But that's nothing but an overhead; we need to create a \ntlist for the top-level query because we use it as fdw_scan_tlist, but \nnot for subqueries, and we need to create retrieved_attrs for the \ntop-level query while deparsing the targetlist in \ndeparseExplicitTargetList, but not for subqueries. So, we should need \nsome work to avoid such a useless overhead.\n\nBest regards,\nEtsuro Fujita\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, 27 Jan 2017 21:25:49 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2017/01/27 21:25, Etsuro Fujita wrote:\n> On 2017/01/27 20:04, Ashutosh Bapat wrote:\n>> I think we should pick up your patch on\n>> 27th December, update the comment per your mail on 5th Jan. I will\n>> review it once and list down the things left to committer's judgement.\n\n> Sorry, I started thinking we went in the wrong direction. I added to\n> deparseSelectStmtForRel build_subquery_tlists, which creates a tlist for\n> each subquery present in a given join tree prior to deparsing a whole\n> remote query. But that's nothing but an overhead; we need to create a\n> tlist for the top-level query because we use it as fdw_scan_tlist, but\n> not for subqueries, and we need to create retrieved_attrs for the\n> top-level query while deparsing the targetlist in\n> deparseExplicitTargetList, but not for subqueries. So, we should need\n> some work to avoid such a useless overhead.\n\nI think we can avoid the useless overhead by adding a new function, \ndeparseSubqueryTargetList, that deparses expressions in the given \nrelation's reltarget, not the tlist, as a SELECT clause of the subquery \nrepresenting the given relation. That would also allow us to make the \n1-to-1 relationship between the subquery output columns and their \naliases more explicit, which was your original comment. Please find \nattached the new version. (The patch doesn't need the patch to avoid \nduplicate construction of the tlist, discussed upthread.)\n\nOther changes:\n* I went back to make_outerrel_subquery and make_innerrel_subquery, \nwhich are flags to indicate whether to deparse the input relations as \nsubqueries. is_subquery_rel would work well for handling the cases of \nfull joins with restrictions on the input relations, but I noticed that \nthat wouldn't work well when extending to handle the cases where we \ndeparse the input relations as subqueries to evaluate PHVs remotely.\n* Since appendSubqueryAlias in the previous patch is pretty small, I \nincluded the code into deparseRangeTblRef.\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, 30 Jan 2017 20:30:51 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Mon, Jan 30, 2017 at 5:00 PM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> On 2017/01/27 21:25, Etsuro Fujita wrote:\n>>\n>> On 2017/01/27 20:04, Ashutosh Bapat wrote:\n>>>\n>>> I think we should pick up your patch on\n>>> 27th December, update the comment per your mail on 5th Jan. I will\n>>> review it once and list down the things left to committer's judgement.\n>\n>\n>> Sorry, I started thinking we went in the wrong direction. I added to\n>> deparseSelectStmtForRel build_subquery_tlists, which creates a tlist for\n>> each subquery present in a given join tree prior to deparsing a whole\n>> remote query. But that's nothing but an overhead; we need to create a\n>> tlist for the top-level query because we use it as fdw_scan_tlist, but\n>> not for subqueries, and we need to create retrieved_attrs for the\n>> top-level query while deparsing the targetlist in\n>> deparseExplicitTargetList, but not for subqueries. So, we should need\n>> some work to avoid such a useless overhead.\n>\n>\n> I think we can avoid the useless overhead by adding a new function,\n> deparseSubqueryTargetList, that deparses expressions in the given relation's\n> reltarget, not the tlist, as a SELECT clause of the subquery representing\n> the given relation. That would also allow us to make the 1-to-1\n> relationship between the subquery output columns and their aliases more\n> explicit, which was your original comment. Please find attached the new\n> version. (The patch doesn't need the patch to avoid duplicate construction\n> of the tlist, discussed upthread.)\n\nI have not looked at the patch, but the reason we use a tlist instead\nof list of expressions is because fdw_scan_tlist is expected in that\nform and we don't want two different representations one to deparse\nSELECT and one to interpret results from the foreign server. What you\ndescribe above seems to introduce exactly that hazard.\n\n>\n> Other changes:\n> * I went back to make_outerrel_subquery and make_innerrel_subquery, which\n> are flags to indicate whether to deparse the input relations as subqueries.\n> is_subquery_rel would work well for handling the cases of full joins with\n> restrictions on the input relations, but I noticed that that wouldn't work\n> well when extending to handle the cases where we deparse the input relations\n> as subqueries to evaluate PHVs remotely.\n\nI had objected to using a single variable instead of two previously\nand you had argued against it in [1]. There you had mentioned that PHV\ndoesn't need two variables, but now you are taking the other stand,\nwithout any apparent reason. Can you please clarify it?\n\n[1] https://www.postgresql.org/message-id/1eb58ee4-8ffa-7c40-1229-c8973e6498ea@lab.ntt.co.jp\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database 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, 30 Jan 2017 17:35:18 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2017/01/30 21:05, Ashutosh Bapat wrote:\n> On Mon, Jan 30, 2017 at 5:00 PM, Etsuro Fujita\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>> On 2017/01/27 21:25, Etsuro Fujita wrote:\n>>> Sorry, I started thinking we went in the wrong direction. I added to\n>>> deparseSelectStmtForRel build_subquery_tlists, which creates a tlist for\n>>> each subquery present in a given join tree prior to deparsing a whole\n>>> remote query. But that's nothing but an overhead; we need to create a\n>>> tlist for the top-level query because we use it as fdw_scan_tlist, but\n>>> not for subqueries, and we need to create retrieved_attrs for the\n>>> top-level query while deparsing the targetlist in\n>>> deparseExplicitTargetList, but not for subqueries. So, we should need\n>>> some work to avoid such a useless overhead.\n\n>> I think we can avoid the useless overhead by adding a new function,\n>> deparseSubqueryTargetList, that deparses expressions in the given relation's\n>> reltarget, not the tlist, as a SELECT clause of the subquery representing\n>> the given relation. That would also allow us to make the 1-to-1\n>> relationship between the subquery output columns and their aliases more\n>> explicit, which was your original comment. Please find attached the new\n>> version. (The patch doesn't need the patch to avoid duplicate construction\n>> of the tlist, discussed upthread.)\n\n> I have not looked at the patch, but the reason we use a tlist instead\n> of list of expressions is because fdw_scan_tlist is expected in that\n> form\n\nActually, we wouldn't need that to deparse a remote join query; a list \nof expressions would be enough.\n\n> and we don't want two different representations one to deparse\n> SELECT and one to interpret results from the foreign server. What you\n> describe above seems to introduce exactly that hazard.\n\nI agree with you to some extent, BUT:\n* I don't think it's a good idea to create a tlist for each base/join \nrelation that is deparsed as a subquery, to just avoid that hazard. As \nI said above, that's nothing but an overhead.\n* I think we would need to have two different representations for at \nleast base relations; we use fpinfo->attrs_used to deparse a simple \nforeign table scan query for a base relation, but when deparsing the \nbase relation as a subquery, we would need to use the list of \nexpressions in the base relation's reltarget, to deparse a SELECT clause \nof the subquery, because we need to deparse a whole-row reference to the \nbase relation as ROW, not all the user columns expanded, as shown in \nthis extracted from the regression tests in the patch:\n\n+ EXPLAIN (VERBOSE, COSTS OFF)\n+ SELECT t1.c1, ss.a, ss.b FROM (SELECT c1 FROM \"S 1\".\"T 3\" WHERE c1 = \n50) t1 INNER JOIN (SELECT t2.c1, t3.c1 FROM (SELECT c1 FROM ft4 WHERE c1 \nbetween 50 and 60) t2 FULL JOIN (SELECT c1 FROM ft5 WHERE c1 between 50 \nand 60) t3 ON (t2.c1 = t3.c1) WHERE t2.c1 IS NULL OR t2.c1 IS NOT NULL) \nss(a, b) ON (TRUE) ORDER BY t1.c1, ss.a, ss.b FOR UPDATE OF t1;\n+ \n \n QUERY PLAN \n \n \n\n+ \n----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------\n+ LockRows\n+ Output: \"T 3\".c1, ft4.c1, ft5.c1, \"T 3\".ctid, ft4.*, ft5.*\n+ -> Nested Loop\n+ Output: \"T 3\".c1, ft4.c1, ft5.c1, \"T 3\".ctid, ft4.*, ft5.*\n+ -> Foreign Scan\n+ Output: ft4.c1, ft4.*, ft5.c1, ft5.*\n+ Relations: (public.ft4) FULL JOIN (public.ft5)\n+ Remote SQL: SELECT s8.c1, s8.c2, s9.c1, s9.c2 FROM \n((SELECT c1, ROW(c1, c2, c3) FROM \"S 1\".\"T 3\" WHERE ((c1 >= 50)) AND \n((c1 <= 60))) s8(c1, c2) FULL JOIN (SELECT c1, ROW(c1, c2, c3) FROM \"S \n1\".\"T 4\" WHERE ((c1 >= 50)) AND ((c1 <= 60))) s9(c1, c2) ON (((s8.c1 = \ns9.c1)))) WHERE (((s8.c1 IS NULL) OR (s8.c1 IS NOT NULL))) ORDER BY \ns8.c1 ASC NULLS LAST, s9.c1 ASC NULLS LAST\n+ -> Hash Full Join\n+ Output: ft4.c1, ft4.*, ft5.c1, ft5.*\n+ Hash Cond: (ft4.c1 = ft5.c1)\n+ Filter: ((ft4.c1 IS NULL) OR (ft4.c1 IS NOT NULL))\n+ -> Foreign Scan on public.ft4\n+ Output: ft4.c1, ft4.*\n+ Remote SQL: SELECT c1, c2, c3 FROM \"S 1\".\"T \n3\" WHERE ((c1 >= 50)) AND ((c1 <= 60))\n+ -> Hash\n+ Output: ft5.c1, ft5.*\n+ -> Foreign Scan on public.ft5\n+ Output: ft5.c1, ft5.*\n+ Remote SQL: SELECT c1, c2, c3 FROM \"S \n1\".\"T 4\" WHERE ((c1 >= 50)) AND ((c1 <= 60))\n+ -> Materialize\n+ Output: \"T 3\".c1, \"T 3\".ctid\n+ -> Seq Scan on \"S 1\".\"T 3\"\n+ Output: \"T 3\".c1, \"T 3\".ctid\n+ Filter: (\"T 3\".c1 = 50)\n+ (25 rows)\n\n>> Other changes:\n>> * I went back to make_outerrel_subquery and make_innerrel_subquery, which\n>> are flags to indicate whether to deparse the input relations as subqueries.\n>> is_subquery_rel would work well for handling the cases of full joins with\n>> restrictions on the input relations, but I noticed that that wouldn't work\n>> well when extending to handle the cases where we deparse the input relations\n>> as subqueries to evaluate PHVs remotely.\n\n> I had objected to using a single variable instead of two previously\n> and you had argued against it in [1]. There you had mentioned that PHV\n> doesn't need two variables, but now you are taking the other stand,\n> without any apparent reason. Can you please clarify it?\n\nSorry, I missed some cases. Consider the join {A, B, C} satisfying the \nidentity 3 in optimizer/README, ie,\n\n (A leftjoin B on (Pab)) leftjoin C on (Pbc)\n = A leftjoin (B leftjoin C on (Pbc)) on (Pab)\n\nwhere predicate Pbc fails for all-null B rows. Assume that B has a PHV \nin the relation's reltarget. While considering 2-way joins, \nforeign_join_ok would determine to deparse B as a subquery and hence set \nthe B's is_subquery_rel to true for the join relation {A, B}. However, \nif the planner selects the join order {A leftjoin (B leftjoin C on \n(Pbc)) on (Pab)} as the cheapest path for the join {A, B, C}, we would \ndeparse B as a subquery according to the is_subquery_rel flag while \ncreating the FROM clause entry for the join relation {B, C}. That \nwouldn't look good. That would be logically correct, though. To avoid \nthis, I'd like to go back to the two variables previously proposed.\n\nBest regards,\nEtsuro Fujita\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, 31 Jan 2017 17:50:51 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Mon, Jan 30, 2017 at 8:30 PM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> Other changes:\n> * I went back to make_outerrel_subquery and make_innerrel_subquery, which\n> are flags to indicate whether to deparse the input relations as subqueries.\n> is_subquery_rel would work well for handling the cases of full joins with\n> restrictions on the input relations, but I noticed that that wouldn't work\n> well when extending to handle the cases where we deparse the input relations\n> as subqueries to evaluate PHVs remotely.\n> * Since appendSubqueryAlias in the previous patch is pretty small, I\n> included the code into deparseRangeTblRef.\n\nThe patch is very fresh, so moved to CF 2017-03.\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, 1 Feb 2017 13:51:03 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 1/30/17 6:30 AM, Etsuro Fujita wrote:\n> On 2017/01/27 21:25, Etsuro Fujita wrote:\n>> On 2017/01/27 20:04, Ashutosh Bapat wrote:\n>>> I think we should pick up your patch on\n>>> 27th December, update the comment per your mail on 5th Jan. I will\n>>> review it once and list down the things left to committer's judgement.\n> \n>> Sorry, I started thinking we went in the wrong direction. I added to\n>> deparseSelectStmtForRel build_subquery_tlists, which creates a tlist for\n>> each subquery present in a given join tree prior to deparsing a whole\n>> remote query. But that's nothing but an overhead; we need to create a\n>> tlist for the top-level query because we use it as fdw_scan_tlist, but\n>> not for subqueries, and we need to create retrieved_attrs for the\n>> top-level query while deparsing the targetlist in\n>> deparseExplicitTargetList, but not for subqueries. So, we should need\n>> some work to avoid such a useless overhead.\n> \n> I think we can avoid the useless overhead by adding a new function,\n> deparseSubqueryTargetList, that deparses expressions in the given\n> relation's reltarget, not the tlist, as a SELECT clause of the subquery\n> representing the given relation. That would also allow us to make the\n> 1-to-1 relationship between the subquery output columns and their\n> aliases more explicit, which was your original comment. Please find\n> attached the new version. (The patch doesn't need the patch to avoid\n> duplicate construction of the tlist, discussed upthread.)\n> \n> Other changes:\n> * I went back to make_outerrel_subquery and make_innerrel_subquery,\n> which are flags to indicate whether to deparse the input relations as\n> subqueries. is_subquery_rel would work well for handling the cases of\n> full joins with restrictions on the input relations, but I noticed that\n> that wouldn't work well when extending to handle the cases where we\n> deparse the input relations as subqueries to evaluate PHVs remotely.\n> * Since appendSubqueryAlias in the previous patch is pretty small, I\n> included the code into deparseRangeTblRef.\n\nThis patch does not apply cleanly at cccbdde:\n\n$ patch -p1 < ../other/postgres-fdw-subquery-support-v15.patch\npatching file contrib/postgres_fdw/deparse.c\nHunk #11 succeeded at 1371 (offset -3 lines).\nHunk #12 succeeded at 1419 (offset -3 lines).\nHunk #13 succeeded at 1486 (offset -3 lines).\nHunk #14 succeeded at 2186 (offset -3 lines).\nHunk #15 succeeded at 3082 (offset -3 lines).\npatching file contrib/postgres_fdw/expected/postgres_fdw.out\npatching file contrib/postgres_fdw/postgres_fdw.c\nHunk #1 succeeded at 669 (offset 1 line).\nHunk #2 succeeded at 1245 (offset -1 lines).\nHunk #3 succeeded at 2557 (offset -1 lines).\nHunk #4 succeeded at 4157 (offset 3 lines).\nHunk #5 succeeded at 4183 (offset 3 lines).\nHunk #6 succeeded at 4212 (offset 3 lines).\nHunk #7 succeeded at 4315 (offset 3 lines).\npatching file contrib/postgres_fdw/postgres_fdw.h\npatching file contrib/postgres_fdw/sql/postgres_fdw.sql\n\nSince these are just offsets I'll leave the patch as \"Needs review\" but\nan updated patch would be appreciated.\n\nThanks,\n-- \n-David\ndavid@pgmasters.net\n\n\n-- \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, 16 Mar 2017 11:46:22 -0400","msg_from":"David Steele <david@pgmasters.net>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Thu, Mar 16, 2017 at 11:46 AM, David Steele <david@pgmasters.net> wrote:\n> $ patch -p1 < ../other/postgres-fdw-subquery-support-v15.patch\n> patching file contrib/postgres_fdw/deparse.c\n> Hunk #11 succeeded at 1371 (offset -3 lines).\n> Hunk #12 succeeded at 1419 (offset -3 lines).\n> Hunk #13 succeeded at 1486 (offset -3 lines).\n> Hunk #14 succeeded at 2186 (offset -3 lines).\n> Hunk #15 succeeded at 3082 (offset -3 lines).\n> patching file contrib/postgres_fdw/expected/postgres_fdw.out\n> patching file contrib/postgres_fdw/postgres_fdw.c\n> Hunk #1 succeeded at 669 (offset 1 line).\n> Hunk #2 succeeded at 1245 (offset -1 lines).\n> Hunk #3 succeeded at 2557 (offset -1 lines).\n> Hunk #4 succeeded at 4157 (offset 3 lines).\n> Hunk #5 succeeded at 4183 (offset 3 lines).\n> Hunk #6 succeeded at 4212 (offset 3 lines).\n> Hunk #7 succeeded at 4315 (offset 3 lines).\n> patching file contrib/postgres_fdw/postgres_fdw.h\n> patching file contrib/postgres_fdw/sql/postgres_fdw.sql\n>\n> Since these are just offsets I'll leave the patch as \"Needs review\" but\n> an updated patch would be appreciated.\n\nI don't think that's really needed. Offsets don't hurt anything.\nEven fuzz is OK. As long as the hunks are applying, I think it's\nfine.\n\nIncidentally, I'm reading through this one now.\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, 16 Mar 2017 12:48:43 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On Thu, Mar 16, 2017 at 12:48 PM, Robert Haas <robertmhaas@gmail.com> wrote:\n> On Thu, Mar 16, 2017 at 11:46 AM, David Steele <david@pgmasters.net> wrote:\n>> $ patch -p1 < ../other/postgres-fdw-subquery-support-v15.patch\n>> patching file contrib/postgres_fdw/deparse.c\n>> Hunk #11 succeeded at 1371 (offset -3 lines).\n>> Hunk #12 succeeded at 1419 (offset -3 lines).\n>> Hunk #13 succeeded at 1486 (offset -3 lines).\n>> Hunk #14 succeeded at 2186 (offset -3 lines).\n>> Hunk #15 succeeded at 3082 (offset -3 lines).\n>> patching file contrib/postgres_fdw/expected/postgres_fdw.out\n>> patching file contrib/postgres_fdw/postgres_fdw.c\n>> Hunk #1 succeeded at 669 (offset 1 line).\n>> Hunk #2 succeeded at 1245 (offset -1 lines).\n>> Hunk #3 succeeded at 2557 (offset -1 lines).\n>> Hunk #4 succeeded at 4157 (offset 3 lines).\n>> Hunk #5 succeeded at 4183 (offset 3 lines).\n>> Hunk #6 succeeded at 4212 (offset 3 lines).\n>> Hunk #7 succeeded at 4315 (offset 3 lines).\n>> patching file contrib/postgres_fdw/postgres_fdw.h\n>> patching file contrib/postgres_fdw/sql/postgres_fdw.sql\n>>\n>> Since these are just offsets I'll leave the patch as \"Needs review\" but\n>> an updated patch would be appreciated.\n>\n> I don't think that's really needed. Offsets don't hurt anything.\n> Even fuzz is OK. As long as the hunks are applying, I think it's\n> fine.\n>\n> Incidentally, I'm reading through this one now.\n\nAnd ... I don't see anything to complain about, so, committed.\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, 16 Mar 2017 13:35:23 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Push down more full joins in postgres_fdw"},{"msg_contents":"On 2017/03/17 2:35, Robert Haas wrote:\n> And ... I don't see anything to complain about, so, committed.\n\nThanks for committing, Robert! Thanks for reviewing, Ashutosh and David!\n\nBest regards,\nEtsuro Fujita\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 Mar 2017 18:42:49 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Push down more full joins in postgres_fdw"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nThe postgres_fdw join pushdown in 9.6 is great, but it can't handle full \\njoins on relations with restrictions. The reason for that is, \\npostgres_fdw can't support deparsing subqueries when creating a remote \\njoin query. So, by adding the deparsing logic to it, I removed that \\nlimitation. Attached is a patch for that, which is based on the patch I \\nposted before [1].\\n\\nAlso, by the deparsing logic, I improved the handling of PHVs so that \\nPHVs are evaluated remotely if it's safe, as discussed in [2]. Here is \\nan example from the regression test, which pushes down multiple levels \\nof PHVs to the remote:\\n\\nEXPLAIN (VERBOSE, COSTS OFF)\\nSELECT ss.*, ft2.c1 FROM ft2 LEFT JOIN (SELECT 13, q.a, ft2.c1 FROM \\n(SELECT 13 FROM ft1 WHERE c1 = 13) q(a) RIGHT JOIN ft2 ON (q.a = ft2.c1) \\nWHERE ft2.\\\\\\nc1 BETWEEN 10 AND 15) ss(f1, f2, f3) ON (ft2.c1 = ss.f1) WHERE ft2.c1 \\nBETWEEN 10 AND 15;\\n \\\\\\n QUERY PLAN \\\\\\n\\n-------------------------------------------------------------------------------------------------------------------------------------------------------\\\\\\n-------------------------------------------------------------------------------------------------------------------------------------------------------\\\\\\n---------------------------------------------------------------\\n Foreign Scan\\n Output: (13), (13), ft2_1.c1, ft2.c1\\n Relations: (public.ft2) LEFT JOIN ((public.ft2) LEFT JOIN (public.ft1))\\n Remote SQL: SELECT r1.\\\"C 1\\\", ss2.c1, ss2.c2, ss2.c3 FROM (\\\"S 1\\\".\\\"T \\n1\\\" r1 LEFT JOIN (SELECT r5.\\\"C 1\\\", 13, ss1.c1 FROM (\\\"S 1\\\".\\\"T 1\\\" r5 LEFT \\nJOIN (SELE\\\\\\nCT 13 FROM \\\"S 1\\\".\\\"T 1\\\" WHERE ((\\\"C 1\\\" = 13))) ss1(c1) ON (((13 = r5.\\\"C \\n1\\\")))) WHERE ((r5.\\\"C 1\\\" >= 10)) AND ((r5.\\\"C 1\\\" <= 15))) ss2(c1, c2, c3) \\nON (((r1.\\\\\\n\\\"C 1\\\" = 13)))) WHERE ((r1.\\\"C 1\\\" >= 10)) AND ((r1.\\\"C 1\\\" <= 15))\\n(4 rows)\\n\\nAlso, in the same way as the PHV handling, I improved the handling of \\nwhole-row reference (and system columns other than ctid), as proposed in \\n[3]. We don't need the \\\"\\\"CASE WHEN ... IS NOT NULL THEN ROW(...) END\\\" \\nconditions any more, as shown in the below example:\\n\\nEXPLAIN (VERBOSE, COSTS OFF)\\nSELECT t1.ctid, t1, t2, t1.c1 FROM ft1 t1 JOIN ft2 t2 ON (t1.c1 = t2.c1) \\nORDER BY t1.c3, t1.c1 OFFSET 100 LIMIT 10;\\n \\\\\\n QUERY PLAN \\\\\\n\\n-------------------------------------------------------------------------------------------------------------------------------------------------------\\\\\\n-------------------------------------------------------------------------------------------------------------------------------------------------------\\\\\\n-----------------------------------------------------------\\n Limit\\n Output: t1.ctid, t1.*, t2.*, t1.c1, t1.c3\\n -> Foreign Scan\\n Output: t1.ctid, t1.*, t2.*, t1.c1, t1.c3\\n Relations: (public.ft1 t1) INNER JOIN (public.ft2 t2)\\n Remote SQL: SELECT ss1.c1, ss1.c2, ss1.c3, ss1.c4, ss2.c1 \\nFROM ((SELECT ctid, ROW(\\\"C 1\\\", c2, c3, c4, c5, c6, c7, c8), \\\"C 1\\\", c3 \\nFROM \\\"S 1\\\".\\\"T \\\\\\n1\\\") ss1(c1, c2, c3, c4) INNER JOIN (SELECT ROW(\\\"C 1\\\", c2, c3, c4, c5, \\nc6, c7, c8), \\\"C 1\\\" FROM \\\"S 1\\\".\\\"T 1\\\") ss2(c1, c2) ON (TRUE)) WHERE \\n((ss1.c3 = ss2.\\\\\\nc2)) ORDER BY ss1.c4 ASC NULLS LAST, ss1.c3 ASC NULLS LAST\\n(6 rows)\\n\\nI'll add this to the next CF. Comments are welcome!\\n\\nBest regards,\\nEtsuro Fujita\\n\\n[1] https://www.postgresql.org/message-id/5710D7E2.7010302%40lab.ntt.co.jp\\n[2] \\nhttps://www.postgresql.org/message-id/b4549406-909f-7d15-dc34-499835a8f0b3%40lab.ntt.co.jp\\n[3] \\nhttps://www.postgresql.org/message-id/a1fa1c4c-bf96-8ea5-cff5-85b927298e73%40lab.ntt.co.jp\\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, 19 Aug 2016 18:20:32 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"Thanks Fujita-san for working on this. I took a quick look at the patch.\\nHere are some quick comments.\\n\\n1. deparsePlaceHolderVar looks odd - each of the deparse* function is named\\nas deparse + <name of the parser node the string would parse into>.\\nPlaceHolderVar is not a parser node, so no string is going to be parsed as\\na PlaceHolderVar. May be you want to name it as\\ndeparseExerFromPlaceholderVar or something like that.\\n\\n2. You will need to check phlevelsup member while assessing whether a PHV\\nis safe to push down.\\n\\n3. I think registerAlias stuff should happen really at the time of creating\\npaths and should be stored in fpinfo. Without that it will be computed\\nevery time we deparse the query. This means every time we try to EXPLAIN\\nthe query at various levels in the join tree and once for the query to be\\nsent to the foreign server.\\n\\n4. The changes related to retrieved_attrs look unrelated to the patch.\\nEither your patch should use the current method of handling retrieved_attrs\\nor there should be a separate patch for retrieved_attrs changes. May be you\\nwant to take a look at the discussion in join pushdown thread as to why we\\nassume retrieved_attrs to be non-NIL always.\\n\\n5. The blocks related to inner and outer relations in\\ndeparseFromExprForRel() look same. We should probably separate that code\\nout into a function and call it at two places.\\n\\n6.\\n! if (is_placeholder)\\n! errcontext(\\\"placeholder expression at position %d in select list\\\",\\n! errpos->cur_attno);\\nA user wouldn't know what a placeholder expression is, there is no such\\nterm in the documentation. We have to device a better way to provide an\\nerror context for an expression in general.\\n\\n\\n\\n\\nOn Fri, Aug 19, 2016 at 2:50 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\\nwrote:\\n\\n> Hi,\\n>\\n> The postgres_fdw join pushdown in 9.6 is great, but it can't handle full\\n> joins on relations with restrictions. The reason for that is, postgres_fdw\\n> can't support deparsing subqueries when creating a remote join query. So,\\n> by adding the deparsing logic to it, I removed that limitation. Attached\\n> is a patch for that, which is based on the patch I posted before [1].\\n>\\n> Also, by the deparsing logic, I improved the handling of PHVs so that PHVs\\n> are evaluated remotely if it's safe, as discussed in [2]. Here is an\\n> example from the regression test, which pushes down multiple levels of PHVs\\n> to the remote:\\n>\\n> EXPLAIN (VERBOSE, COSTS OFF)\\n> SELECT ss.*, ft2.c1 FROM ft2 LEFT JOIN (SELECT 13, q.a, ft2.c1 FROM\\n> (SELECT 13 FROM ft1 WHERE c1 = 13) q(a) RIGHT JOIN ft2 ON (q.a = ft2.c1)\\n> WHERE ft2.\\\\\\n> c1 BETWEEN 10 AND 15) ss(f1, f2, f3) ON (ft2.c1 = ss.f1) WHERE ft2.c1\\n> BETWEEN 10 AND 15;\\n> \\\\\\n> QUERY PLAN \\\\\\n>\\n> ------------------------------------------------------------\\n> ------------------------------------------------------------\\n> -------------------------------\\\\\\n> ------------------------------------------------------------\\n> ------------------------------------------------------------\\n> -------------------------------\\\\\\n> ---------------------------------------------------------------\\n> Foreign Scan\\n> Output: (13), (13), ft2_1.c1, ft2.c1\\n> Relations: (public.ft2) LEFT JOIN ((public.ft2) LEFT JOIN (public.ft1))\\n> Remote SQL: SELECT r1.\\\"C 1\\\", ss2.c1, ss2.c2, ss2.c3 FROM (\\\"S 1\\\".\\\"T 1\\\"\\n> r1 LEFT JOIN (SELECT r5.\\\"C 1\\\", 13, ss1.c1 FROM (\\\"S 1\\\".\\\"T 1\\\" r5 LEFT JOIN\\n> (SELE\\\\\\n> CT 13 FROM \\\"S 1\\\".\\\"T 1\\\" WHERE ((\\\"C 1\\\" = 13))) ss1(c1) ON (((13 = r5.\\\"C\\n> 1\\\")))) WHERE ((r5.\\\"C 1\\\" >= 10)) AND ((r5.\\\"C 1\\\" <= 15))) ss2(c1, c2, c3) ON\\n> (((r1.\\\\\\n> \\\"C 1\\\" = 13)))) WHERE ((r1.\\\"C 1\\\" >= 10)) AND ((r1.\\\"C 1\\\" <= 15))\\n> (4 rows)\\n>\\n> Also, in the same way as the PHV handling, I improved the handling of\\n> whole-row reference (and system columns other than ctid), as proposed in\\n> [3]. We don't need the \\\"\\\"CASE WHEN ... IS NOT NULL THEN ROW(...) END\\\"\\n> conditions any more, as shown in the below example:\\n>\\n> EXPLAIN (VERBOSE, COSTS OFF)\\n> SELECT t1.ctid, t1, t2, t1.c1 FROM ft1 t1 JOIN ft2 t2 ON (t1.c1 = t2.c1)\\n> ORDER BY t1.c3, t1.c1 OFFSET 100 LIMIT 10;\\n> \\\\\\n> QUERY PLAN \\\\\\n>\\n> ------------------------------------------------------------\\n> ------------------------------------------------------------\\n> -------------------------------\\\\\\n> ------------------------------------------------------------\\n> ------------------------------------------------------------\\n> -------------------------------\\\\\\n> -----------------------------------------------------------\\n> Limit\\n> Output: t1.ctid, t1.*, t2.*, t1.c1, t1.c3\\n> -> Foreign Scan\\n> Output: t1.ctid, t1.*, t2.*, t1.c1, t1.c3\\n> Relations: (public.ft1 t1) INNER JOIN (public.ft2 t2)\\n> Remote SQL: SELECT ss1.c1, ss1.c2, ss1.c3, ss1.c4, ss2.c1 FROM\\n> ((SELECT ctid, ROW(\\\"C 1\\\", c2, c3, c4, c5, c6, c7, c8), \\\"C 1\\\", c3 FROM \\\"S\\n> 1\\\".\\\"T \\\\\\n> 1\\\") ss1(c1, c2, c3, c4) INNER JOIN (SELECT ROW(\\\"C 1\\\", c2, c3, c4, c5, c6,\\n> c7, c8), \\\"C 1\\\" FROM \\\"S 1\\\".\\\"T 1\\\") ss2(c1, c2) ON (TRUE)) WHERE ((ss1.c3 =\\n> ss2.\\\\\\n> c2)) ORDER BY ss1.c4 ASC NULLS LAST, ss1.c3 ASC NULLS LAST\\n> (6 rows)\\n>\\n> I'll add this to the next CF. Comments are welcome!\\n>\\n> Best regards,\\n> Etsuro Fujita\\n>\\n> [1] https://www.postgresql.org/message-id/5710D7E2.7010302%40lab.ntt.co.jp\\n> [2] https://www.postgresql.org/message-id/b4549406-909f-7d15-dc3\\n> 4-499835a8f0b3%40lab.ntt.co.jp\\n> [3] https://www.postgresql.org/message-id/a1fa1c4c-bf96-8ea5-cff\\n> 5-85b927298e73%40lab.ntt.co.jp\\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>\\n\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database Company\\n\\nThanks Fujita-san for working on this. I took a quick look at the patch. Here are some quick comments.1. deparsePlaceHolderVar looks odd - each of the deparse* function is named as deparse + <name of the parser node the string would parse into>. PlaceHolderVar is not a parser node, so no string is going to be parsed as a PlaceHolderVar. May be you want to name it as deparseExerFromPlaceholderVar or something like that.2. You will need to check phlevelsup member while assessing whether a PHV is safe to push down.3. I think registerAlias stuff should happen really at the time of creating paths and should be stored in fpinfo. Without that it will be computed every time we deparse the query. This means every time we try to EXPLAIN the query at various levels in the join tree and once for the query to be sent to the foreign server.4. The changes related to retrieved_attrs look unrelated to the patch. Either your patch should use the current method of handling retrieved_attrs or there should be a separate patch for retrieved_attrs changes. May be you want to take a look at the discussion in join pushdown thread as to why we assume retrieved_attrs to be non-NIL always.5. The blocks related to inner and outer relations in deparseFromExprForRel() look same. We should probably separate that code out into a function and call it at two places.6. ! if (is_placeholder)! errcontext(\\\"placeholder expression at position %d in select list\\\",! errpos->cur_attno);A user wouldn't know what a placeholder expression is, there is no such term in the documentation. We have to device a better way to provide an error context for an expression in general.On Fri, Aug 19, 2016 at 2:50 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp> wrote:Hi,\\n\\nThe postgres_fdw join pushdown in 9.6 is great, but it can't handle full joins on relations with restrictions. The reason for that is, postgres_fdw can't support deparsing subqueries when creating a remote join query. So, by adding the deparsing logic to it, I removed that limitation. Attached is a patch for that, which is based on the patch I posted before [1].\\n\\nAlso, by the deparsing logic, I improved the handling of PHVs so that PHVs are evaluated remotely if it's safe, as discussed in [2]. Here is an example from the regression test, which pushes down multiple levels of PHVs to the remote:\\n\\nEXPLAIN (VERBOSE, COSTS OFF)\\nSELECT ss.*, ft2.c1 FROM ft2 LEFT JOIN (SELECT 13, q.a, ft2.c1 FROM (SELECT 13 FROM ft1 WHERE c1 = 13) q(a) RIGHT JOIN ft2 ON (q.a = ft2.c1) WHERE ft2.\\\\\\nc1 BETWEEN 10 AND 15) ss(f1, f2, f3) ON (ft2.c1 = ss.f1) WHERE ft2.c1 BETWEEN 10 AND 15;\\n \\\\\\n QUERY PLAN \\\\\\n\\n-------------------------------------------------------------------------------------------------------------------------------------------------------\\\\\\n-------------------------------------------------------------------------------------------------------------------------------------------------------\\\\\\n---------------------------------------------------------------\\n Foreign Scan\\n Output: (13), (13), ft2_1.c1, ft2.c1\\n Relations: (public.ft2) LEFT JOIN ((public.ft2) LEFT JOIN (public.ft1))\\n Remote SQL: SELECT r1.\\\"C 1\\\", ss2.c1, ss2.c2, ss2.c3 FROM (\\\"S 1\\\".\\\"T 1\\\" r1 LEFT JOIN (SELECT r5.\\\"C 1\\\", 13, ss1.c1 FROM (\\\"S 1\\\".\\\"T 1\\\" r5 LEFT JOIN (SELE\\\\\\nCT 13 FROM \\\"S 1\\\".\\\"T 1\\\" WHERE ((\\\"C 1\\\" = 13))) ss1(c1) ON (((13 = r5.\\\"C 1\\\")))) WHERE ((r5.\\\"C 1\\\" >= 10)) AND ((r5.\\\"C 1\\\" <= 15))) ss2(c1, c2, c3) ON (((r1.\\\\\\n\\\"C 1\\\" = 13)))) WHERE ((r1.\\\"C 1\\\" >= 10)) AND ((r1.\\\"C 1\\\" <= 15))\\n(4 rows)\\n\\nAlso, in the same way as the PHV handling, I improved the handling of whole-row reference (and system columns other than ctid), as proposed in [3]. We don't need the \\\"\\\"CASE WHEN ... IS NOT NULL THEN ROW(...) END\\\" conditions any more, as shown in the below example:\\n\\nEXPLAIN (VERBOSE, COSTS OFF)\\nSELECT t1.ctid, t1, t2, t1.c1 FROM ft1 t1 JOIN ft2 t2 ON (t1.c1 = t2.c1) ORDER BY t1.c3, t1.c1 OFFSET 100 LIMIT 10;\\n \\\\\\n QUERY PLAN \\\\\\n\\n-------------------------------------------------------------------------------------------------------------------------------------------------------\\\\\\n-------------------------------------------------------------------------------------------------------------------------------------------------------\\\\\\n-----------------------------------------------------------\\n Limit\\n Output: t1.ctid, t1.*, t2.*, t1.c1, t1.c3\\n -> Foreign Scan\\n Output: t1.ctid, t1.*, t2.*, t1.c1, t1.c3\\n Relations: (public.ft1 t1) INNER JOIN (public.ft2 t2)\\n Remote SQL: SELECT ss1.c1, ss1.c2, ss1.c3, ss1.c4, ss2.c1 FROM ((SELECT ctid, ROW(\\\"C 1\\\", c2, c3, c4, c5, c6, c7, c8), \\\"C 1\\\", c3 FROM \\\"S 1\\\".\\\"T \\\\\\n1\\\") ss1(c1, c2, c3, c4) INNER JOIN (SELECT ROW(\\\"C 1\\\", c2, c3, c4, c5, c6, c7, c8), \\\"C 1\\\" FROM \\\"S 1\\\".\\\"T 1\\\") ss2(c1, c2) ON (TRUE)) WHERE ((ss1.c3 = ss2.\\\\\\nc2)) ORDER BY ss1.c4 ASC NULLS LAST, ss1.c3 ASC NULLS LAST\\n(6 rows)\\n\\nI'll add this to the next CF. Comments are welcome!\\n\\nBest regards,\\nEtsuro Fujita\\n\\n[1] https://www.postgresql.org/message-id/5710D7E2.7010302%40lab.ntt.co.jp\\n[2] https://www.postgresql.org/message-id/b4549406-909f-7d15-dc34-499835a8f0b3%40lab.ntt.co.jp\\n[3] https://www.postgresql.org/message-id/a1fa1c4c-bf96-8ea5-cff5-85b927298e73%40lab.ntt.co.jp\\n\\n--\\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n-- Best Wishes,Ashutosh BapatEnterpriseDB CorporationThe Postgres Database Company\",\n \"msg_date\": \"Mon, 22 Aug 2016 12:19:59 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"Hi Ashutosh,\\n\\nOn 2016/08/22 15:49, Ashutosh Bapat wrote:\\n> 1. deparsePlaceHolderVar looks odd - each of the deparse* function is\\n> named as deparse + <name of the parser node the string would parse\\n> into>. PlaceHolderVar is not a parser node, so no string is going to be\\n> parsed as a PlaceHolderVar. May be you want to name it as\\n> deparseExerFromPlaceholderVar or something like that.\\n\\nThe name \\\"deparseExerFromPlaceholderVar\\\" seems long to me. How about \\n\\\"deparsePlaceHolderExpr\\\"?\\n\\n> 2. You will need to check phlevelsup member while assessing whether a\\n> PHV is safe to push down.\\n\\nGood catch! In addition to that, I added another check that the eval_at \\nset for the PHV should be included in the relids set for the foreign \\nrelation. I think that would make the shippability check more robust.\\n\\n> 3. I think registerAlias stuff should happen really at the time of\\n> creating paths and should be stored in fpinfo. Without that it will be\\n> computed every time we deparse the query. This means every time we try\\n> to EXPLAIN the query at various levels in the join tree and once for the\\n> query to be sent to the foreign server.\\n\\nHmm. I think the overhead in calculating aliases would be negligible \\ncompared to the overhead in explaining each remote query for costing or \\nsending the remote query for execution. So, I created aliases in the \\nsame way as remote params created and stored into params_list in \\ndeparse_expr_cxt. I'm not sure it's worth complicating the code.\\n\\n> 4. The changes related to retrieved_attrs look unrelated to the patch.\\n> Either your patch should use the current method of handling\\n> retrieved_attrs or there should be a separate patch for retrieved_attrs\\n> changes. May be you want to take a look at the discussion in join\\n> pushdown thread as to why we assume retrieved_attrs to be non-NIL always.\\n\\nOK, I removed those changes from the patch.\\n\\n> 5. The blocks related to inner and outer relations in\\n> deparseFromExprForRel() look same. We should probably separate that code\\n> out into a function and call it at two places.\\n\\nDone.\\n\\n> 6.\\n> ! if (is_placeholder)\\n> ! errcontext(\\\"placeholder expression at position %d in select list\\\",\\n> ! errpos->cur_attno);\\n> A user wouldn't know what a placeholder expression is, there is no such\\n> term in the documentation. We have to device a better way to provide an\\n> error context for an expression in general.\\n\\nThough I proposed that, I don't think that it's that important to let \\nusers know that the expression is from a PHV. How about just saying \\n\\\"expression\\\", not \\\"placeholder expression\\\", so that we have the message \\n\\\"expression at position %d in select list\\\" in the context?\\n\\nAttached is an updated version of the patch.\\n\\nOther changes:\\n\\n* Add a bit more regression test\\n* Revise code/comments\\n* Cleanups\\n\\nThanks for the comments!\\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\": \"Fri, 2 Sep 2016 19:25:30 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Fri, Sep 2, 2016 at 3:55 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\\nwrote:\\n\\n> Hi Ashutosh,\\n>\\n> On 2016/08/22 15:49, Ashutosh Bapat wrote:\\n>\\n>> 1. deparsePlaceHolderVar looks odd - each of the deparse* function is\\n>> named as deparse + <name of the parser node the string would parse\\n>> into>. PlaceHolderVar is not a parser node, so no string is going to be\\n>> parsed as a PlaceHolderVar. May be you want to name it as\\n>> deparseExerFromPlaceholderVar or something like that.\\n>>\\n>\\n> The name \\\"deparseExerFromPlaceholderVar\\\" seems long to me. How about\\n> \\\"deparsePlaceHolderExpr\\\"?\\n>\\n\\nThere isn't any node with name PlaceHolderExpr.\\n\\n\\n> 2. You will need to check phlevelsup member while assessing whether a\\n>> PHV is safe to push down.\\n>>\\n>\\n> Good catch! In addition to that, I added another check that the eval_at\\n> set for the PHV should be included in the relids set for the foreign\\n> relation. I think that would make the shippability check more robust.\\n>\\n\\nThanks.\\n\\n\\n>\\n> 3. I think registerAlias stuff should happen really at the time of\\n>> creating paths and should be stored in fpinfo. Without that it will be\\n>> computed every time we deparse the query. This means every time we try\\n>> to EXPLAIN the query at various levels in the join tree and once for the\\n>> query to be sent to the foreign server.\\n>>\\n>\\n> Hmm. I think the overhead in calculating aliases would be negligible\\n> compared to the overhead in explaining each remote query for costing or\\n> sending the remote query for execution. So, I created aliases in the same\\n> way as remote params created and stored into params_list in\\n> deparse_expr_cxt. I'm not sure it's worth complicating the code.\\n>\\n\\nWe defer remote parameter creation till deparse time since the the\\nparameter numbers are dependent upon the sequence in which we deparse the\\nquery. Creating them at the time of path creation and storing them in\\nfpinfo is not possible because a. those present in the joining relations\\nwill conflict with each other and need some kind of serialization at the\\ntime of deparsing b. those defer for differently parameterized paths or\\npaths with different pathkeys. We don't defer it because it's very light on\\nperformance.\\n\\nThat's not true with the alias information. As long as we detect which\\nrelations need subqueries, their RTIs are enough to create unique aliases\\ne.g. if a relation involves RTIs 1, 2, 3 corresponding subquery can have\\nalias r123 without conflicting with any other alias.\\n\\nHowever minimum overhead it might have, we will deparse the query every\\ntime we create a path involving that kind of relation i.e. for different\\npathkeys, different parameterization and different joins in which that\\nrelation participates in. This number can be significant. We want to avoid\\nthis overhead if we can.\\n\\n\\n>\\n> 5. The blocks related to inner and outer relations in\\n>> deparseFromExprForRel() look same. We should probably separate that code\\n>> out into a function and call it at two places.\\n>>\\n>\\n> Done.\\n>\\n\\nThanks. I see you have created function deparseOperandRelation() for the\\nsame. I guess, this should be renamed as deparseRangeTblRef() since it will\\ncreate RangeTblRef node when parsed back.\\n\\n\\n>\\n> 6.\\n>> ! if (is_placeholder)\\n>> ! errcontext(\\\"placeholder expression at position %d in select\\n>> list\\\",\\n>> ! errpos->cur_attno);\\n>> A user wouldn't know what a placeholder expression is, there is no such\\n>> term in the documentation. We have to device a better way to provide an\\n>> error context for an expression in general.\\n>>\\n>\\n> Though I proposed that, I don't think that it's that important to let\\n> users know that the expression is from a PHV. How about just saying\\n> \\\"expression\\\", not \\\"placeholder expression\\\", so that we have the message\\n> \\\"expression at position %d in select list\\\" in the context?\\n>\\n> Hmm, that's better than placeholder expression, but not as explanatory as\\nit should be since we won't be printing the \\\"select list\\\" in the error\\ncontext and user won't have a clue as to what error context is about. But I\\ndon't have any good suggestions right now. May be we should print the whole\\nexpression? But that can be very long, I don't know.\\n\\nThis patch tries to do two things at a time 1. support join pushdown for\\nFULL join when the joining relations have remote conditions 2. better\\nsupport for fetching placeholder vars, whole row references and some system\\ncolumns. To make reviews easy, I think we should split the patch into two\\n1. supporting subqueries to be deparsed with support for one of the above\\n(I will suggest FULL join support) 2. support for the other.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database Company\\n\\nOn Fri, Sep 2, 2016 at 3:55 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp> wrote:Hi Ashutosh,\\n\\nOn 2016/08/22 15:49, Ashutosh Bapat wrote:\\n\\n1. deparsePlaceHolderVar looks odd - each of the deparse* function is\\nnamed as deparse + <name of the parser node the string would parse\\ninto>. PlaceHolderVar is not a parser node, so no string is going to be\\nparsed as a PlaceHolderVar. May be you want to name it as\\ndeparseExerFromPlaceholderVar or something like that.\\n\\n\\nThe name \\\"deparseExerFromPlaceholderVar\\\" seems long to me. How about \\\"deparsePlaceHolderExpr\\\"?There isn't any node with name PlaceHolderExpr.\\n\\n\\n2. You will need to check phlevelsup member while assessing whether a\\nPHV is safe to push down.\\n\\n\\nGood catch! In addition to that, I added another check that the eval_at set for the PHV should be included in the relids set for the foreign relation. I think that would make the shippability check more robust.Thanks. \\n\\n\\n3. I think registerAlias stuff should happen really at the time of\\ncreating paths and should be stored in fpinfo. Without that it will be\\ncomputed every time we deparse the query. This means every time we try\\nto EXPLAIN the query at various levels in the join tree and once for the\\nquery to be sent to the foreign server.\\n\\n\\nHmm. I think the overhead in calculating aliases would be negligible compared to the overhead in explaining each remote query for costing or sending the remote query for execution. So, I created aliases in the same way as remote params created and stored into params_list in deparse_expr_cxt. I'm not sure it's worth complicating the code.We defer remote parameter creation till deparse time since the the parameter numbers are dependent upon the sequence in which we deparse the query. Creating them at the time of path creation and storing them in fpinfo is not possible because a. those present in the joining relations will conflict with each other and need some kind of serialization at the time of deparsing b. those defer for differently parameterized paths or paths with different pathkeys. We don't defer it because it's very light on performance.That's not true with the alias information. As long as we detect which relations need subqueries, their RTIs are enough to create unique aliases e.g. if a relation involves RTIs 1, 2, 3 corresponding subquery can have alias r123 without conflicting with any other alias.However minimum overhead it might have, we will deparse the query every time we create a path involving that kind of relation i.e. for different pathkeys, different parameterization and different joins in which that relation participates in. This number can be significant. We want to avoid this overhead if we can. \\n\\n5. The blocks related to inner and outer relations in\\ndeparseFromExprForRel() look same. We should probably separate that code\\nout into a function and call it at two places.\\n\\n\\nDone.Thanks. I see you have created function deparseOperandRelation() for the same. I guess, this should be renamed as deparseRangeTblRef() since it will create RangeTblRef node when parsed back. \\n\\n\\n6.\\n! if (is_placeholder)\\n! errcontext(\\\"placeholder expression at position %d in select list\\\",\\n! errpos->cur_attno);\\nA user wouldn't know what a placeholder expression is, there is no such\\nterm in the documentation. We have to device a better way to provide an\\nerror context for an expression in general.\\n\\n\\nThough I proposed that, I don't think that it's that important to let users know that the expression is from a PHV. How about just saying \\\"expression\\\", not \\\"placeholder expression\\\", so that we have the message \\\"expression at position %d in select list\\\" in the context?\\nHmm, that's better than placeholder expression, but not as explanatory as it should be since we won't be printing the \\\"select list\\\" in the error context and user won't have a clue as to what error context is about. But I don't have any good suggestions right now. May be we should print the whole expression? But that can be very long, I don't know.This patch tries to do two things at a time 1. support join pushdown for FULL join when the joining relations have remote conditions 2. better support for fetching placeholder vars, whole row references and some system columns. To make reviews easy, I think we should split the patch into two 1. supporting subqueries to be deparsed with support for one of the above (I will suggest FULL join support) 2. support for the other. -- Best Wishes,Ashutosh BapatEnterpriseDB CorporationThe Postgres Database Company\",\n \"msg_date\": \"Tue, 6 Sep 2016 18:37:12 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/09/06 22:07, Ashutosh Bapat wrote:\\n> On Fri, Sep 2, 2016 at 3:55 PM, Etsuro Fujita\\n> <fujita.etsuro@lab.ntt.co.jp <mailto:fujita.etsuro@lab.ntt.co.jp>> wrote:\\n\\n> On 2016/08/22 15:49, Ashutosh Bapat wrote:\\n> 1. deparsePlaceHolderVar looks odd - each of the deparse*\\n> function is\\n> named as deparse + <name of the parser node the string would parse\\n> into>. PlaceHolderVar is not a parser node, so no string is\\n> going to be\\n> parsed as a PlaceHolderVar. May be you want to name it as\\n> deparseExerFromPlaceholderVar or something like that.\\n\\n> The name \\\"deparseExerFromPlaceholderVar\\\" seems long to me. How\\n> about \\\"deparsePlaceHolderExpr\\\"?\\n\\n> There isn't any node with name PlaceHolderExpr.\\n\\nI'll rename it to \\\"deparseExerInPlaceholderVar\\\", then.\\n\\n> 3. I think registerAlias stuff should happen really at the time of\\n> creating paths and should be stored in fpinfo. Without that it\\n> will be\\n> computed every time we deparse the query. This means every time\\n> we try\\n> to EXPLAIN the query at various levels in the join tree and once\\n> for the\\n> query to be sent to the foreign server.\\n\\n> Hmm. I think the overhead in calculating aliases would be\\n> negligible compared to the overhead in explaining each remote query\\n> for costing or sending the remote query for execution. So, I\\n> created aliases in the same way as remote params created and stored\\n> into params_list in deparse_expr_cxt. I'm not sure it's worth\\n> complicating the code.\\n\\n> We defer remote parameter creation till deparse time since the the\\n> parameter numbers are dependent upon the sequence in which we deparse\\n> the query. Creating them at the time of path creation and storing them\\n> in fpinfo is not possible because a. those present in the joining\\n> relations will conflict with each other and need some kind of\\n> serialization at the time of deparsing b. those defer for differently\\n> parameterized paths or paths with different pathkeys. We don't defer it\\n> because it's very light on performance.\\n>\\n> That's not true with the alias information. As long as we detect which\\n> relations need subqueries, their RTIs are enough to create unique\\n> aliases e.g. if a relation involves RTIs 1, 2, 3 corresponding subquery\\n> can have alias r123 without conflicting with any other alias.\\n\\nHmm. But another concern about the approach of generating an subselect \\nalias for a path, if needed, at the path creation time would be that the \\npath might be rejected by add_path, which would result in totally \\nwasting cycles for generating the subselect alias for the path.\\n\\n> However minimum overhead it might have, we will deparse the query every\\n> time we create a path involving that kind of relation i.e. for different\\n> pathkeys, different parameterization and different joins in which that\\n> relation participates in. This number can be significant. We want to\\n> avoid this overhead if we can.\\n\\nExactly. As I said above, I think the overhead would be negligible \\ncompared to the overhead in explaining each remote query for costing or \\nthe overhead in sending the final remote query for execution, though.\\n\\n> 5. The blocks related to inner and outer relations in\\n> deparseFromExprForRel() look same. We should probably separate\\n> that code\\n> out into a function and call it at two places.\\n\\n> Done.\\n\\n> I see you have created function deparseOperandRelation() for the\\n> same. I guess, this should be renamed as deparseRangeTblRef() since it\\n> will create RangeTblRef node when parsed back.\\n\\nOK, if there no opinions of others, I'll rename it to the name proposed \\nby you, \\\"deparseRangeTblRef\\\".\\n\\n> 6.\\n> ! if (is_placeholder)\\n> ! errcontext(\\\"placeholder expression at position %d in\\n> select list\\\",\\n> ! errpos->cur_attno);\\n> A user wouldn't know what a placeholder expression is, there is\\n> no such\\n> term in the documentation. We have to device a better way to\\n> provide an\\n> error context for an expression in general.\\n\\n> Though I proposed that, I don't think that it's that important to\\n> let users know that the expression is from a PHV. How about just\\n> saying \\\"expression\\\", not \\\"placeholder expression\\\", so that we have\\n> the message \\\"expression at position %d in select list\\\" in the context?\\n\\n> Hmm, that's better than placeholder expression, but not as explanatory\\n> as it should be since we won't be printing the \\\"select list\\\" in the\\n> error context and user won't have a clue as to what error context is\\n> about.\\n\\nI don't think so. Consider an example of the conversion error message, \\nwhich is from the regression test:\\n\\nSELECT ft1.c1, ft2.c2, ft1 FROM ft1, ft2 WHERE ft1.c1 = ft2.c1 AND \\nft1.c1 = 1;\\nERROR: invalid input syntax for integer: \\\"foo\\\"\\nCONTEXT: whole-row reference to foreign table \\\"ft1\\\"\\n\\nAs shown in the example, the error message is displayed under a remote \\nquery for execution. So, ISTM it's reasonable to print something like \\n\\\"expression at position %d in select list\\\" in the context if an \\nexpression in a PHV.\\n\\n> But I don't have any good suggestions right now. May be we should\\n> print the whole expression? But that can be very long, I don't know.\\n\\nISTM that it's a bit too expensive to print the whole expression in the \\nerror context.\\n\\n> This patch tries to do two things at a time 1. support join pushdown for\\n> FULL join when the joining relations have remote conditions 2. better\\n> support for fetching placeholder vars, whole row references and some\\n> system columns. To make reviews easy, I think we should split the patch\\n> into two 1. supporting subqueries to be deparsed with support for one of\\n> the above (I will suggest FULL join support) 2. support for the other.\\n\\nOK, will try.\\n\\nBest regards,\\nEtsuro Fujita\\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, 8 Sep 2016 19:51:00 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/09/08 19:51, Etsuro Fujita wrote:\\n> On 2016/09/06 22:07, Ashutosh Bapat wrote:\\n>> This patch tries to do two things at a time 1. support join pushdown for\\n>> FULL join when the joining relations have remote conditions 2. better\\n>> support for fetching placeholder vars, whole row references and some\\n>> system columns. To make reviews easy, I think we should split the patch\\n>> into two 1. supporting subqueries to be deparsed with support for one of\\n>> the above (I will suggest FULL join support) 2. support for the other.\\n\\n> OK, will try.\\n\\nI extracted #1 from the patch. Attached is a patch for that. If that \\npatch is reasonable, I'll create a patch for #2 on top of it.\\n\\nComments welcome!\\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\": \"Fri, 9 Sep 2016 21:35:45 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/09/09 21:35, Etsuro Fujita wrote:\\n> On 2016/09/08 19:51, Etsuro Fujita wrote:\\n>> On 2016/09/06 22:07, Ashutosh Bapat wrote:\\n>>> This patch tries to do two things at a time 1. support join pushdown for\\n>>> FULL join when the joining relations have remote conditions 2. better\\n>>> support for fetching placeholder vars, whole row references and some\\n>>> system columns. To make reviews easy, I think we should split the patch\\n>>> into two 1. supporting subqueries to be deparsed with support for one of\\n>>> the above (I will suggest FULL join support) 2. support for the other.\\n\\n>> OK, will try.\\n\\n> I extracted #1 from the patch. Attached is a patch for that. If that\\n> patch is reasonable, I'll create a patch for #2 on top of it.\\n\\nAttached is a patch for #2. In that patch I fixed some bugs and added a \\nbit more comments. For testing, please apply the patch for #1 first.\\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, 12 Sep 2016 21:35:35 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \">\\n> 3. I think registerAlias stuff should happen really at the time of\\n>> creating paths and should be stored in fpinfo. Without that it\\n>> will be\\n>> computed every time we deparse the query. This means every time\\n>> we try\\n>> to EXPLAIN the query at various levels in the join tree and once\\n>> for the\\n>> query to be sent to the foreign server.\\n>>\\n>\\n> Hmm. I think the overhead in calculating aliases would be\\n>> negligible compared to the overhead in explaining each remote query\\n>> for costing or sending the remote query for execution. So, I\\n>> created aliases in the same way as remote params created and stored\\n>> into params_list in deparse_expr_cxt. I'm not sure it's worth\\n>> complicating the code.\\n>>\\n>\\n> We defer remote parameter creation till deparse time since the the\\n>> parameter numbers are dependent upon the sequence in which we deparse\\n>> the query. Creating them at the time of path creation and storing them\\n>> in fpinfo is not possible because a. those present in the joining\\n>> relations will conflict with each other and need some kind of\\n>> serialization at the time of deparsing b. those defer for differently\\n>> parameterized paths or paths with different pathkeys. We don't defer it\\n>> because it's very light on performance.\\n>>\\n>> That's not true with the alias information. As long as we detect which\\n>> relations need subqueries, their RTIs are enough to create unique\\n>> aliases e.g. if a relation involves RTIs 1, 2, 3 corresponding subquery\\n>> can have alias r123 without conflicting with any other alias.\\n>>\\n>\\n> Hmm. But another concern about the approach of generating an subselect\\n> alias for a path, if needed, at the path creation time would be that the\\n> path might be rejected by add_path, which would result in totally wasting\\n> cycles for generating the subselect alias for the path.\\n>\\n\\nA path may get rejected but the relation is not rejected. The alias applies\\nto a relation and its targetlist, which will remain same for all paths\\ncreated for that relation, esp when it's a base relation or join. So, AFAIU\\nthat work never gets wasted. Also, for costing paths with\\nuse_remote_estimate, we deparse the query, which builds the alias\\ninformation again and again for very path. That's much worse than building\\nit once for a given relation.\\n\\n\\n>\\n> However minimum overhead it might have, we will deparse the query every\\n>> time we create a path involving that kind of relation i.e. for different\\n>> pathkeys, different parameterization and different joins in which that\\n>> relation participates in. This number can be significant. We want to\\n>> avoid this overhead if we can.\\n>>\\n>\\n> Exactly. As I said above, I think the overhead would be negligible\\n> compared to the overhead in explaining each remote query for costing or the\\n> overhead in sending the final remote query for execution, though.\\n>\\n\\nIt won't remain minimal as the number of paths created increases,\\nincreasing the number of times a query is deparsed. We deparse query every\\ntime time we cost a path for a relation with use_remote_estimates true. As\\nwe try to push down more and more stuff, we will create more paths and\\ndeparse the query more time.\\n\\nAlso, that makes the interface better. Right now, in your patch, you have\\nchanged the order of deparsing in the existing code, so that the aliases\\nare registered while deparsing FROM clause and before any Var nodes are\\ndeparsed. If we create aliases at the time of path creation, only once in\\nGetForeignJoinPaths or GetForeignPaths as appropriate, that would require\\nless code churn and would save some CPU cycles as well.\\n\\n\\n>\\n> 6.\\n>> ! if (is_placeholder)\\n>> ! errcontext(\\\"placeholder expression at position %d in\\n>> select list\\\",\\n>> ! errpos->cur_attno);\\n>> A user wouldn't know what a placeholder expression is, there is\\n>> no such\\n>> term in the documentation. We have to device a better way to\\n>> provide an\\n>> error context for an expression in general.\\n>>\\n>\\n> Though I proposed that, I don't think that it's that important to\\n>> let users know that the expression is from a PHV. How about just\\n>> saying \\\"expression\\\", not \\\"placeholder expression\\\", so that we have\\n>> the message \\\"expression at position %d in select list\\\" in the context?\\n>>\\n>\\n> Hmm, that's better than placeholder expression, but not as explanatory\\n>> as it should be since we won't be printing the \\\"select list\\\" in the\\n>> error context and user won't have a clue as to what error context is\\n>> about.\\n>>\\n>\\n> I don't think so. Consider an example of the conversion error message,\\n> which is from the regression test:\\n>\\n> SELECT ft1.c1, ft2.c2, ft1 FROM ft1, ft2 WHERE ft1.c1 = ft2.c1 AND\\n> ft1.c1 = 1;\\n> ERROR: invalid input syntax for integer: \\\"foo\\\"\\n> CONTEXT: whole-row reference to foreign table \\\"ft1\\\"\\n>\\n> As shown in the example, the error message is displayed under a remote\\n> query for execution. So, ISTM it's reasonable to print something like\\n> \\\"expression at position %d in select list\\\" in the context if an expression\\n> in a PHV.\\n>\\n\\nI missed it. Sorry. Looks ok.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database Company\\n\\n\\n\\n\\n 3. I think registerAlias stuff should happen really at the time of\\n creating paths and should be stored in fpinfo. Without that it\\n will be\\n computed every time we deparse the query. This means every time\\n we try\\n to EXPLAIN the query at various levels in the join tree and once\\n for the\\n query to be sent to the foreign server.\\n\\n\\n\\n Hmm. I think the overhead in calculating aliases would be\\n negligible compared to the overhead in explaining each remote query\\n for costing or sending the remote query for execution. So, I\\n created aliases in the same way as remote params created and stored\\n into params_list in deparse_expr_cxt. I'm not sure it's worth\\n complicating the code.\\n\\n\\n\\nWe defer remote parameter creation till deparse time since the the\\nparameter numbers are dependent upon the sequence in which we deparse\\nthe query. Creating them at the time of path creation and storing them\\nin fpinfo is not possible because a. those present in the joining\\nrelations will conflict with each other and need some kind of\\nserialization at the time of deparsing b. those defer for differently\\nparameterized paths or paths with different pathkeys. We don't defer it\\nbecause it's very light on performance.\\n\\nThat's not true with the alias information. As long as we detect which\\nrelations need subqueries, their RTIs are enough to create unique\\naliases e.g. if a relation involves RTIs 1, 2, 3 corresponding subquery\\ncan have alias r123 without conflicting with any other alias.\\n\\n\\nHmm. But another concern about the approach of generating an subselect alias for a path, if needed, at the path creation time would be that the path might be rejected by add_path, which would result in totally wasting cycles for generating the subselect alias for the path.A path may get rejected but the relation is not rejected. The alias applies to a relation and its targetlist, which will remain same for all paths created for that relation, esp when it's a base relation or join. So, AFAIU that work never gets wasted. Also, for costing paths with use_remote_estimate, we deparse the query, which builds the alias information again and again for very path. That's much worse than building it once for a given relation. \\n\\n\\nHowever minimum overhead it might have, we will deparse the query every\\ntime we create a path involving that kind of relation i.e. for different\\npathkeys, different parameterization and different joins in which that\\nrelation participates in. This number can be significant. We want to\\navoid this overhead if we can.\\n\\n\\nExactly. As I said above, I think the overhead would be negligible compared to the overhead in explaining each remote query for costing or the overhead in sending the final remote query for execution, though.It won't remain minimal as the number of paths created increases, increasing the number of times a query is deparsed. We deparse query every time time we cost a path for a relation with use_remote_estimates true. As we try to push down more and more stuff, we will create more paths and deparse the query more time. Also, that makes the interface better. Right now, in your patch, you have changed the order of deparsing in the existing code, so that the aliases are registered while deparsing FROM clause and before any Var nodes are deparsed. If we create aliases at the time of path creation, only once in GetForeignJoinPaths or GetForeignPaths as appropriate, that would require less code churn and would save some CPU cycles as well. \\n\\n\\n 6.\\n ! if (is_placeholder)\\n ! errcontext(\\\"placeholder expression at position %d in\\n select list\\\",\\n ! errpos->cur_attno);\\n A user wouldn't know what a placeholder expression is, there is\\n no such\\n term in the documentation. We have to device a better way to\\n provide an\\n error context for an expression in general.\\n\\n\\n\\n Though I proposed that, I don't think that it's that important to\\n let users know that the expression is from a PHV. How about just\\n saying \\\"expression\\\", not \\\"placeholder expression\\\", so that we have\\n the message \\\"expression at position %d in select list\\\" in the context?\\n\\n\\n\\nHmm, that's better than placeholder expression, but not as explanatory\\nas it should be since we won't be printing the \\\"select list\\\" in the\\nerror context and user won't have a clue as to what error context is\\nabout.\\n\\n\\nI don't think so. Consider an example of the conversion error message, which is from the regression test:\\n\\nSELECT ft1.c1, ft2.c2, ft1 FROM ft1, ft2 WHERE ft1.c1 = ft2.c1 AND ft1.c1 = 1;\\nERROR: invalid input syntax for integer: \\\"foo\\\"\\nCONTEXT: whole-row reference to foreign table \\\"ft1\\\"\\n\\nAs shown in the example, the error message is displayed under a remote query for execution. So, ISTM it's reasonable to print something like \\\"expression at position %d in select list\\\" in the context if an expression in a PHV.I missed it. Sorry. Looks ok.-- Best Wishes,Ashutosh BapatEnterpriseDB CorporationThe Postgres Database Company\",\n \"msg_date\": \"Tue, 13 Sep 2016 10:47:01 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Tue, Sep 6, 2016 at 9:07 AM, Ashutosh Bapat\\n<ashutosh.bapat@enterprisedb.com> wrote:\\n> That's not true with the alias information. As long as we detect which\\n> relations need subqueries, their RTIs are enough to create unique aliases\\n> e.g. if a relation involves RTIs 1, 2, 3 corresponding subquery can have\\n> alias r123 without conflicting with any other alias.\\n\\nWhat if RTI 123 is also used in the query?\\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, 13 Sep 2016 12:58:14 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Tue, Sep 13, 2016 at 10:28 PM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Tue, Sep 6, 2016 at 9:07 AM, Ashutosh Bapat\\n> <ashutosh.bapat@enterprisedb.com> wrote:\\n>> That's not true with the alias information. As long as we detect which\\n>> relations need subqueries, their RTIs are enough to create unique aliases\\n>> e.g. if a relation involves RTIs 1, 2, 3 corresponding subquery can have\\n>> alias r123 without conflicting with any other alias.\\n>\\n> What if RTI 123 is also used in the query?\\n\\nGood catch. I actually meant some combination of 1, 2 and 3, which is\\nunique for a join between r1, r2 and r3. How about r1_2_3 or\\nr1_r2_r3?\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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 Sep 2016 09:08:10 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Tue, Sep 13, 2016 at 11:38 PM, Ashutosh Bapat\\n<ashutosh.bapat@enterprisedb.com> wrote:\\n> On Tue, Sep 13, 2016 at 10:28 PM, Robert Haas <robertmhaas@gmail.com> wrote:\\n>> On Tue, Sep 6, 2016 at 9:07 AM, Ashutosh Bapat\\n>> <ashutosh.bapat@enterprisedb.com> wrote:\\n>>> That's not true with the alias information. As long as we detect which\\n>>> relations need subqueries, their RTIs are enough to create unique aliases\\n>>> e.g. if a relation involves RTIs 1, 2, 3 corresponding subquery can have\\n>>> alias r123 without conflicting with any other alias.\\n>>\\n>> What if RTI 123 is also used in the query?\\n>\\n> Good catch. I actually meant some combination of 1, 2 and 3, which is\\n> unique for a join between r1, r2 and r3. How about r1_2_3 or\\n> r1_r2_r3?\\n\\nSure, something like that can work, but if you have a big enough join\\nthe identifier might get too long. I'm not sure why it wouldn't work\\nto just use the lowest RTI involved in the join, though; the others\\nwon't appear anywhere else at that query level.\\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 Sep 2016 11:22:38 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Wed, Sep 14, 2016 at 8:52 PM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Tue, Sep 13, 2016 at 11:38 PM, Ashutosh Bapat\\n> <ashutosh.bapat@enterprisedb.com> wrote:\\n>> On Tue, Sep 13, 2016 at 10:28 PM, Robert Haas <robertmhaas@gmail.com> wrote:\\n>>> On Tue, Sep 6, 2016 at 9:07 AM, Ashutosh Bapat\\n>>> <ashutosh.bapat@enterprisedb.com> wrote:\\n>>>> That's not true with the alias information. As long as we detect which\\n>>>> relations need subqueries, their RTIs are enough to create unique aliases\\n>>>> e.g. if a relation involves RTIs 1, 2, 3 corresponding subquery can have\\n>>>> alias r123 without conflicting with any other alias.\\n>>>\\n>>> What if RTI 123 is also used in the query?\\n>>\\n>> Good catch. I actually meant some combination of 1, 2 and 3, which is\\n>> unique for a join between r1, r2 and r3. How about r1_2_3 or\\n>> r1_r2_r3?\\n>\\n> Sure, something like that can work, but if you have a big enough join\\n> the identifier might get too long. I'm not sure why it wouldn't work\\n> to just use the lowest RTI involved in the join, though; the others\\n> won't appear anywhere else at that query level.\\n>\\n\\nYes, that will work too and is much more preferred than long r1_2_3.\\nThe idea is to come with a unique alias name from RTIs involved in\\nthat relation. Thinking loudly, r1_2_3 is more descriptive to debug\\nissues. It tells that the subquery it refers to covers RTIs 1, 2 and\\n3. That information may be quite helpful to understand the remote SQL.\\nr1 on the other hand can refer to relation with RTI 1 or a join\\nrelation where least RTI is 1. That can be solved by using s<RTI> for\\nsubquery and r<RTI> for a bare relation. But it still doesn't tell us\\nwhich all relations are involved.\\n\\nBut since the subquery is part of remote SQL and we have to take a\\nlook at it to find out meaning of individual columns, that benefit may\\nbe smaller compared to the convenience of smaller alias. So +1 for\\nusing the smallest RTI to indicate a subquery.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 15 Sep 2016 11:59:48 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/09/13 14:17, Ashutosh Bapat wrote:\\n\\n> But another concern about the approach of generating an\\n> subselect alias for a path, if needed, at the path creation time\\n> would be that the path might be rejected by add_path, which would\\n> result in totally wasting cycles for generating the subselect alias\\n> for the path.\\n\\n> A path may get rejected but the relation is not rejected. The alias\\n> applies to a relation and its targetlist, which will remain same for all\\n> paths created for that relation, esp when it's a base relation or join.\\n> So, AFAIU that work never gets wasted.\\n\\nI think there is no guarantee that add_path won't reject foreign join \\npaths. The possibility would be very low, though.\\n\\n> However minimum overhead it might have, we will deparse the\\n> query every\\n> time we create a path involving that kind of relation i.e. for\\n> different\\n> pathkeys, different parameterization and different joins in\\n> which that\\n> relation participates in. This number can be significant. We want to\\n> avoid this overhead if we can.\\n\\n> Exactly. As I said above, I think the overhead would be negligible\\n> compared to the overhead in explaining each remote query for costing\\n> or the overhead in sending the final remote query for execution, though.\\n\\n> It won't remain minimal as the number of paths created increases,\\n> increasing the number of times a query is deparsed. We deparse query\\n> every time time we cost a path for a relation with use_remote_estimates\\n> true. As we try to push down more and more stuff, we will create more\\n> paths and deparse the query more time.\\n\\n> Also, that makes the interface better. Right now, in your patch, you\\n> have changed the order of deparsing in the existing code, so that the\\n> aliases are registered while deparsing FROM clause and before any Var\\n> nodes are deparsed. If we create aliases at the time of path creation,\\n> only once in GetForeignJoinPaths or GetForeignPaths as appropriate, that\\n> would require less code churn and would save some CPU cycles as well.\\n\\nAgreed. Will fix.\\n\\nBest regards,\\nEtsuro Fujita\\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, 26 Sep 2016 16:30:22 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/09/15 15:29, Ashutosh Bapat wrote:\\n> On Wed, Sep 14, 2016 at 8:52 PM, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n>> I'm not sure why it wouldn't work\\n>> to just use the lowest RTI involved in the join, though; the others\\n>> won't appear anywhere else at that query level.\\n\\n> So +1 for\\n> using the smallest RTI to indicate a subquery.\\n\\n+1 for the general idea.\\n\\nISTM that the use of the same RTI for subqueries in multi-levels in a \\nremote SQL makes the SQL a bit difficult to read. How about using the \\nposition of the join rel in join_rel_list, (more precisely, the position \\nplus list_length(root->parse->rtable)), instead?\\n\\nBest regards,\\nEtsuro Fujita\\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, 26 Sep 2016 16:35:17 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Mon, Sep 26, 2016 at 1:05 PM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> On 2016/09/15 15:29, Ashutosh Bapat wrote:\\n>>\\n>> On Wed, Sep 14, 2016 at 8:52 PM, Robert Haas <robertmhaas@gmail.com>\\n>> wrote:\\n>\\n>\\n>>> I'm not sure why it wouldn't work\\n>>> to just use the lowest RTI involved in the join, though; the others\\n>>> won't appear anywhere else at that query level.\\n>\\n>\\n>> So +1 for\\n>> using the smallest RTI to indicate a subquery.\\n>\\n>\\n> +1 for the general idea.\\n>\\n> ISTM that the use of the same RTI for subqueries in multi-levels in a remote\\n> SQL makes the SQL a bit difficult to read. How about using the position of\\n> the join rel in join_rel_list, (more precisely, the position plus\\n> list_length(root->parse->rtable)), instead?\\n>\\nWe switch to hash table to maintain the join RelOptInfos when the\\nnumber of joins grows larger, where the position won't make much\\nsense. We might differentiate between a base relation alias and\\nsubquery alias by using different prefixes like \\\"r\\\" and \\\"s\\\" resp.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 26 Sep 2016 14:36:33 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/09/26 18:06, Ashutosh Bapat wrote:\\n> On Mon, Sep 26, 2016 at 1:05 PM, Etsuro Fujita\\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n\\n>> ISTM that the use of the same RTI for subqueries in multi-levels in a remote\\n>> SQL makes the SQL a bit difficult to read. How about using the position of\\n>> the join rel in join_rel_list, (more precisely, the position plus\\n>> list_length(root->parse->rtable)), instead?\\n\\n> We switch to hash table to maintain the join RelOptInfos when the\\n> number of joins grows larger, where the position won't make much\\n> sense.\\n\\nThat's right, but we still store the joinrel into join_rel_list after \\ncreating that hash table. That hash table is just for fast lookup. See \\nbuild_join_rel.\\n\\n> We might differentiate between a base relation alias and\\n> subquery alias by using different prefixes like \\\"r\\\" and \\\"s\\\" resp.\\n\\nHmm. My concern about that would the recursive use of \\\"s\\\" with the same \\nRTI as subquery aliases for different level subqueries in a single \\nremote SQL. For example, if those subqueries include a base rel with \\nRTI=1, then \\\"s1\\\" would be used again and again within that SQL. That \\nwould be logically correct, but seems confusing to me.\\n\\nBest regards,\\nEtsuro Fujita\\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, 26 Sep 2016 19:36:02 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Mon, Sep 26, 2016 at 4:06 PM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> On 2016/09/26 18:06, Ashutosh Bapat wrote:\\n>>\\n>> On Mon, Sep 26, 2016 at 1:05 PM, Etsuro Fujita\\n>> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>\\n>\\n>>> ISTM that the use of the same RTI for subqueries in multi-levels in a\\n>>> remote\\n>>> SQL makes the SQL a bit difficult to read. How about using the position\\n>>> of\\n>>> the join rel in join_rel_list, (more precisely, the position plus\\n>>> list_length(root->parse->rtable)), instead?\\n>\\n>\\n>> We switch to hash table to maintain the join RelOptInfos when the\\n>> number of joins grows larger, where the position won't make much\\n>> sense.\\n>\\n>\\n> That's right, but we still store the joinrel into join_rel_list after\\n> creating that hash table. That hash table is just for fast lookup. See\\n> build_join_rel.\\n\\nSorry, I wasn't clear in my reply. As the list grows, it will take\\nlonger to locate the RelOptInfo of the given join. Doing that for\\ncreating an alias seems an overkill.\\n\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 26 Sep 2016 16:50:27 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/09/26 20:20, Ashutosh Bapat wrote:\\n> On Mon, Sep 26, 2016 at 4:06 PM, Etsuro Fujita\\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>> On 2016/09/26 18:06, Ashutosh Bapat wrote:\\n>>> On Mon, Sep 26, 2016 at 1:05 PM, Etsuro Fujita\\n>>> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n\\n>>>> ISTM that the use of the same RTI for subqueries in multi-levels in a\\n>>>> remote\\n>>>> SQL makes the SQL a bit difficult to read. How about using the position\\n>>>> of\\n>>>> the join rel in join_rel_list, (more precisely, the position plus\\n>>>> list_length(root->parse->rtable)), instead?\\n\\n>>> We switch to hash table to maintain the join RelOptInfos when the\\n>>> number of joins grows larger, where the position won't make much\\n>>> sense.\\n\\n>> That's right, but we still store the joinrel into join_rel_list after\\n>> creating that hash table.\\n\\n> As the list grows, it will take\\n> longer to locate the RelOptInfo of the given join. Doing that for\\n> creating an alias seems an overkill.\\n\\nThe join rel is appended to the end of the list, so I was thinking to \\nget the position info by list_length during postgresGetForeignJoinPaths.\\n\\nBest regards,\\nEtsuro Fujita\\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, 27 Sep 2016 12:18:10 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Tue, Sep 27, 2016 at 8:48 AM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> On 2016/09/26 20:20, Ashutosh Bapat wrote:\\n>>\\n>> On Mon, Sep 26, 2016 at 4:06 PM, Etsuro Fujita\\n>> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>>>\\n>>> On 2016/09/26 18:06, Ashutosh Bapat wrote:\\n>>>>\\n>>>> On Mon, Sep 26, 2016 at 1:05 PM, Etsuro Fujita\\n>>>> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>\\n>\\n>>>>> ISTM that the use of the same RTI for subqueries in multi-levels in a\\n>>>>> remote\\n>>>>> SQL makes the SQL a bit difficult to read. How about using the\\n>>>>> position\\n>>>>> of\\n>>>>> the join rel in join_rel_list, (more precisely, the position plus\\n>>>>> list_length(root->parse->rtable)), instead?\\n>\\n>\\n>>>> We switch to hash table to maintain the join RelOptInfos when the\\n>>>> number of joins grows larger, where the position won't make much\\n>>>> sense.\\n>\\n>\\n>>> That's right, but we still store the joinrel into join_rel_list after\\n>>> creating that hash table.\\n>\\n>\\n>> As the list grows, it will take\\n>> longer to locate the RelOptInfo of the given join. Doing that for\\n>> creating an alias seems an overkill.\\n>\\n>\\n> The join rel is appended to the end of the list, so I was thinking to get\\n> the position info by list_length during postgresGetForeignJoinPaths.\\n\\nThat's true only when the paths are being added to a newly created\\njoinrel. But that's not true always. We may add paths with different\\njoining order to an existing joinrel, in which case list_length would\\nnot give its position. Am I missing something?\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 27 Sep 2016 10:03:25 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/09/27 13:33, Ashutosh Bapat wrote:\\n\\nI wrote:\\n>>>>>> ISTM that the use of the same RTI for subqueries in multi-levels in a\\n>>>>>> remote\\n>>>>>> SQL makes the SQL a bit difficult to read. How about using the\\n>>>>>> position\\n>>>>>> of\\n>>>>>> the join rel in join_rel_list, (more precisely, the position plus\\n>>>>>> list_length(root->parse->rtable)), instead?\\n\\nI wrote:\\n>> The join rel is appended to the end of the list, so I was thinking to get\\n>> the position info by list_length during postgresGetForeignJoinPaths.\\n\\n> That's true only when the paths are being added to a newly created\\n> joinrel. But that's not true always. We may add paths with different\\n> joining order to an existing joinrel, in which case list_length would\\n> not give its position. Am I missing something?\\n\\nI think you are right, but postgresGetForeignJoinPaths only allows us to \\nadd a foreign join path to a newly created joinrel. The reason is \\nbecause that routine skips all its work after the first call for that \\njoinrel, by checking to see if joinrel->fdw_private is not NULL. So, I \\nthink it's reasonable to get the position by list_length in that routine.\\n\\nBest regards,\\nEtsuro Fujita\\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\": \"Wed, 28 Sep 2016 18:20:05 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/09/26 16:30, Etsuro Fujita wrote:\\n> On 2016/09/13 14:17, Ashutosh Bapat wrote:\\n\\n>> It won't remain minimal as the number of paths created increases,\\n>> increasing the number of times a query is deparsed. We deparse query\\n>> every time time we cost a path for a relation with use_remote_estimates\\n>> true. As we try to push down more and more stuff, we will create more\\n>> paths and deparse the query more time.\\n\\n>> Also, that makes the interface better. Right now, in your patch, you\\n>> have changed the order of deparsing in the existing code, so that the\\n>> aliases are registered while deparsing FROM clause and before any Var\\n>> nodes are deparsed. If we create aliases at the time of path creation,\\n>> only once in GetForeignJoinPaths or GetForeignPaths as appropriate, that\\n>> would require less code churn and would save some CPU cycles as well.\\n\\n> Agreed. Will fix.\\n\\nDone. Attached is an updated version of the patch.\\n\\nI didn't create aliases at anytime. Instead, I added a logic to get \\ninfo about the alias to a given expression from reltarget->exprs for \\nrelations in a given join tree. See isSubqueryExpr and \\ngetSubselectAliasInfo.\\n\\nAs proposed by you, the patch differentiates between a base relation \\nalias and a subquery alias by using different prefixes \\\"r\\\" and \\\"s\\\", \\nrespectively. Also, subquery aliases are indexed by RTI for baserels \\nand the position in join_rel_list + the length of rtable for joinrels, \\nas proposed upthread.\\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\": \"Wed, 28 Sep 2016 18:35:45 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/09/28 18:35, Etsuro Fujita wrote:\\n> Attached is an updated version of the patch.\\n\\nI found a minor bug in that patch; the relation_index added to \\nPgFdwRelationInfo was defined as Index, but I used the modifier %d to \\nprint that. So, I changed the data type to int. Also, I added a bit \\nmore comments. Please find attached an updated version of the patch.\\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\": \"Thu, 29 Sep 2016 21:12:30 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/09/29 21:12, Etsuro Fujita wrote:\\n> Please find attached an updated version of the patch.\\n\\nHere is a new version (V6) of the patch, which is basically the same as \\nthe previous patch, but I slightly modified that patch. Another patch I \\nam attaching is an updated patch for pushing down PHVs to the remote, \\nwhich is created on top of the V6 patch.\\n\\nBest regards,\\nEtsuro Fujita\\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, 13 Oct 2016 19:35:40 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"Review for postgres-fdw-more-full-join-pushdown-v6 patch.\\n\\nThe patch applies cleanly and regression is clean (make check in\\nregress directory and that in postgres_fdw).\\n\\nHere are some comments.\\n1. Because of the following code change, for a joinrel we might end up using\\nattrs_used, which will be garbage for a join rel. The assumption that tlist can\\nnot be NIL for a join relation is wrong. Even for a join relation, tlist can be\\nNULL. Consider query 'explain verbose select count(*) from ft1, ft2' Since\\ncount(*) doesn't need any columns, the tlist is NIL. With the patch the server\\ncrashes for this query.\\n- if (foreignrel->reloptkind == RELOPT_JOINREL)\\n+ /*\\n+ * Note: tlist for a base relation might be non-NIL. For example, if the\\n+ * base relation is an operand of a foreign join performing a full outer\\n+ * join and has non-NIL remote_conds, the base relation will be deparsed\\n+ * as a subquery, so the tlist for the base relation could be non-NIL.\\n+ */\\n+ if (tlist != NIL)\\n {\\n- /* For a join relation use the input tlist */\\n\\nWe can not decide whether to use deparseExplicitTargetList or\\ndeparse it from attrs_used based on the tlist. SELECT lists, whether it's\\ntopmost SELECT or a subquery (even on a base relation), for deparsing a\\nJOIN query need to use deparseExplicitTargetList.\\n\\n2. The code in deparseRangeTblRef() dealing with subqueries, is very similar to\\ndeparseSelectStmtForRel(). The only thing deparseRangeTblRef() does not require\\nis the appending ORDER BY, which is determined by existence of pathkeys\\nargument. Probably we should reuse deparseSelectStmtForRel(), instead of\\nduplicating the code. This will also make the current changes to\\ndeparseSelectSql unnecessary.\\n\\n3. Why do we need following change? The elog() just few lines above says that\\nwe expect only Var nodes. Why then we use deparseExpr() instead of\\ndeparseVar()?\\n if (i > 0)\\n appendStringInfoString(buf, \\\", \\\");\\n- deparseVar(var, context);\\n+ deparseExpr((Expr *) var, context);\\n\\n *retrieved_attrs = lappend_int(*retrieved_attrs, i + 1);\\n\\nAnd I get the answer for that question, somewhere down in the patch\\n+ /*\\n+ * If the given expression is an output column of a subquery-in-FROM,\\n+ * deparse the alias to the expression instead.\\n+ */\\n+ if (IsA(node, Var))\\n+ {\\n+ int tabno;\\n+ int colno;\\n+\\n+ if (isSubqueryExpr(node, context->foreignrel, &tabno, &colno))\\n+ {\\n+ appendStringInfo(context->buf, \\\"%s%d.%s%d\\\",\\n+ SS_TAB_ALIAS_PREFIX, tabno,\\n+ SS_COL_ALIAS_PREFIX, colno);\\n+ return;\\n+ }\\n+ }\\n+\\n\\nFunctionally, this code belongs to deparseVar() and not deparseExpr(). Like all\\nother functions which handle Expr nodes, deparseExpr() is expected to be a\\ndispatcher to the node specific function.\\n\\n4. This comment is good to have there, but is unrelated to this patch. May be a\\nseparate patch?\\n+ /* Don't generate bad syntax if no columns */\\n\\n5. Changed comment doesn't say anything different from the original comment. It\\nmay have been good to have it this way to start with, but changing it now\\ndoesn't bring anything new. You seem to have just merged prologue of the\\nfunction with a comment in that function. I think, this change is unnecessary.\\n- * The function constructs ... JOIN ... ON ... for join relation. For a base\\n- * relation it just returns schema-qualified tablename, with the appropriate\\n- * alias if so requested.\\n+ * For a join relation the clause of the following form is appended to buf:\\n+ * ((outer relation) <join type> (inner relation) ON (joinclauses))\\n+ * For a base relation the function just adds the schema-qualified tablename,\\n+ * with the appropriate alias if so requested.\\n+ /* Don't generate bad syntax if no columns */\\n\\n6. Change may be good but unrelated to this patch. May be material for a\\nseparate patch. There are few such changes in this function. While these\\nchanges may be good by themselves, they distract reviewer from the goal of the\\npatch.\\n- appendStringInfo(buf, \\\"(\\\");\\n+ appendStringInfoChar(buf, '(');\\n\\n7. I don't understand why you need to change this function so much. Take for\\nexample the following change.\\n- RelOptInfo *rel_o = fpinfo->outerrel;\\n- RelOptInfo *rel_i = fpinfo->innerrel;\\n- StringInfoData join_sql_o;\\n- StringInfoData join_sql_i;\\n+ /* Begin the FROM clause entry */\\n+ appendStringInfoChar(buf, '(');\\n\\n /* Deparse outer relation */\\n- initStringInfo(&join_sql_o);\\n- deparseFromExprForRel(&join_sql_o, root, rel_o, true, params_list);\\n+ deparseRangeTblRef(buf, root,\\n+ fpinfo->outerrel,\\n+ fpinfo->make_outerrel_subquery,\\n+ params_list);\\n /* Deparse outer relation */\\n- initStringInfo(&join_sql_o);\\n- deparseFromExprForRel(&join_sql_o, root, rel_o, true, params_list);\\n+ deparseRangeTblRef(buf, root,\\n+ fpinfo->outerrel,\\n+ fpinfo->make_outerrel_subquery,\\n+ params_list);\\n\\nIt removes few variables, instead directly accesses the members from fpinfo.\\nAlso, deparses the individual relations in the provided buffer directly, rather\\nthan in separate buffers in the original code. Instead of this, all you had to\\ndo was replace a call\\n- deparseFromExprForRel(&join_sql_o, root, rel_o, true, params_list);\\nwith\\n+ deparseRangeTblRef(&join_sql_o, root, rel_o,\\nfpinfo->make_outerrel_subquery, params_list)\\nSimilarly for the inner relation. Again, the changes you have done might have\\nbeen good, if done in the original code, but doing those in this patch just\\ncreates distractions and increases the size of the patch.\\n\\nSimilarly changes\\n- /* Append join clause; (TRUE) if no join clause */\\n+ /* Append join conditions */\\n\\n+ {\\n+ /* No join conditions; add \\\"(TRUE)\\\" */\\n appendStringInfoString(buf, \\\"(TRUE)\\\");\\n+ }\\n+ appendStringInfoString(buf, \\\" ON \\\");\\n\\n- /* End the FROM clause entry. */\\n- appendStringInfo(buf, \\\")\\\");\\n\\n\\n+ /* End the FROM clause entry */\\n+ appendStringInfoChar(buf, ')');\\n\\n8. Why have you changed the name of variable here?\\n- if (use_alias)\\n+ if (add_rel_alias)\\n\\n9. In deparseRangeTblRef(), the targetlist for the subquery is obtained by\\ncalling build_tlist_to_deparse(), but appendSubselectAlias() constructs the\\ncolumn aliases based on foreignrel->reltarget->exprs. Those two are usually\\nsame, but there is no code which guarantees that. Instead, pass tlist to\\nappendSubselectAlias() or simply pass the number of entries in that list\\n(list_length(tlist), which is all it needs. OR use foreignrel->reltarget->exprs\\ndirectly instead of calling build_tlist_to_deparse(). Similar problem exists\\nwith getSubselectAliasInfo().\\n\\n10. Deparsing a query in the form of a subquery makes it hard-to-read. The\\noriginal code aimed at avoiding a subquery. Also, this change has created many\\nexpected output changes, which again seem unnecessary. In fact, having the\\npushable join clauses of an inner join in ON clause, which is closer to JOIN\\nclause, is better than having them farther in the WHERE clause.\\n- /*\\n- * For an inner join, all restrictions can be treated alike. Treating the\\n- * pushed down conditions as join conditions allows a top level full outer\\n- * join to be deparsed without requiring subqueries.\\n- */\\n- if (jointype == JOIN_INNER)\\n- {\\n- Assert(!fpinfo->joinclauses);\\n- fpinfo->joinclauses = fpinfo->remote_conds;\\n- fpinfo->remote_conds = NIL;\\n- }\\n-\\n\\n11. I have reworded following comment and restructured the code that follows it\\nin the attached patch.\\n+ /*\\n+ * Set the subquery information. If the relation performs a full outer\\n+ * join and if the input relations have non-NIL remote_conds, the input\\n+ * relations need to be deparsed as a subquery.\\n+ */\\nThe code esp. the if .. else .. block followed by another one, made it a bit\\nunreadable. Hence I restructured it so that it's readable and doesn't require\\nvariable \\\"relids\\\" from earlier code, which was being reused. Let me know if\\nthis change looks good.\\n\\n12. The code below deletes the condition, which is understandable.\\n- if (fpinfo_i->remote_conds || fpinfo_o->remote_conds)\\nBut why does it add a new unrelated condition here? What the comment claims,\\nlooks like an existing bug, unrelated to the patch. Can you please give an\\nexample? If that it's an existing bug, it should be fixed as a separate patch.\\nI don't understand the relation of this code with what the patch is doing.\\n+ /*\\n+ * We can't do anything here, and if there are any non-Vars in the\\n+ * outerrel/innerrel's reltarget, give up pushing down this join,\\n+ * because the deparsing logic can't support such a case currently.\\n+ */\\n+ if (reltarget_has_non_vars(outerrel))\\n+ return false;\\n+ if (reltarget_has_non_vars(innerrel))\\n return false;\\n\\n13. The comment below is missing the main purpose i.e. creating a a unique\\nalias, in case the relation gets converted into a subquery. Lowest or highest\\nrelid will create a unique alias at given level of join and that would be more\\nfuture proof. If we decide to consider paths for every join order, following\\ncomment will no more be true.\\n+ /*\\n+ * Set the relation index. This is defined as the position of this\\n+ * joinrel in the join_rel_list list plus the length of the rtable list.\\n+ * Note that since this joinrel is at the end of the list when we are\\n+ * called, we can get the position by list_length.\\n+ */\\n+ fpinfo->relation_index =\\n+ list_length(root->parse->rtable) + list_length(root->join_rel_list);\\n\\n14. The function name talks about non-vars but the Assert few lines below\\nasserts that every node there is a Var. Need better naming.\\n+reltarget_has_non_vars(RelOptInfo *foreignrel)\\n+{\\n+ ListCell *lc;\\n+\\n+ foreach(lc, foreignrel->reltarget->exprs)\\n+ {\\n+ Var *var = (Var *) lfirst(lc);\\n+\\n+ Assert(IsA(var, Var));\\nAnd also an explanation for the claim\\n+ * Note: currently deparseExplicitTargetList can't properly handle such Vars.\\n\\n15. While deparsing every Var, we are descending the RelOptInfo hierarchy.\\nInstead of that, we should probably gather all the alias information once and\\nstore it in context as an array indexed by relid. While deparsing a Var we can\\nget the targetlist and alias for a given relation by using var->varno as index.\\nAnd then search for given Var node in the targetlist to deparse the column name\\nby its position in the targetlist. For the relations that are not converted\\ninto subqueries, this array will not hold any information and the Var will be\\ndeparsed as it's being done today.\\n\\nTesting\\n-------\\n1. The code is changing deparser but doesn't have testcases\\ntesting impact of the code. For example, we should have testcases with remote\\nquery deparsed as nested subqueries or nested subqueries within subqueries and\\nso on May be testcases where a relation deeper in the RelOptInfo hierarchy has\\nconditions but it's immediate upper relation does not have those.\\n\\n2. The only testcase added by this patch, copies an existing case and adds a\\nwhole row reference to one of the relations being joined. Instead we could add\\nthat whole-row reference to the existing testcase itself.\\n\\nOn Thu, Oct 13, 2016 at 4:05 PM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> On 2016/09/29 21:12, Etsuro Fujita wrote:\\n>>\\n>> Please find attached an updated version of the patch.\\n>\\n>\\n> Here is a new version (V6) of the patch, which is basically the same as the\\n> previous patch, but I slightly modified that patch. Another patch I am\\n> attaching is an updated patch for pushing down PHVs to the remote, which is\\n> created on top of the V6 patch.\\n>\\n> Best regards,\\n> Etsuro Fujita\\n\\n\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 22 Oct 2016 13:49:35 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"> 11. I have reworded following comment and restructured the code that follows it\\n> in the attached patch.\\n> + /*\\n> + * Set the subquery information. If the relation performs a full outer\\n> + * join and if the input relations have non-NIL remote_conds, the input\\n> + * relations need to be deparsed as a subquery.\\n> + */\\n> The code esp. the if .. else .. block followed by another one, made it a bit\\n> unreadable. Hence I restructured it so that it's readable and doesn't require\\n> variable \\\"relids\\\" from earlier code, which was being reused. Let me know if\\n> this change looks good.\\n>\\nSorry, forgot to attach the patch. Here it is.\\n\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database Company\\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, 22 Oct 2016 22:16:10 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/10/22 17:19, Ashutosh Bapat wrote:\\n> Review for postgres-fdw-more-full-join-pushdown-v6 patch.\\n>\\n> The patch applies cleanly and regression is clean (make check in\\n> regress directory and that in postgres_fdw).\\n\\nThanks for the review!\\n\\n> Here are some comments.\\n> 1. Because of the following code change, for a joinrel we might end up using\\n> attrs_used, which will be garbage for a join rel. The assumption that tlist can\\n> not be NIL for a join relation is wrong. Even for a join relation, tlist can be\\n> NULL. Consider query 'explain verbose select count(*) from ft1, ft2' Since\\n> count(*) doesn't need any columns, the tlist is NIL. With the patch the server\\n> crashes for this query.\\n> - if (foreignrel->reloptkind == RELOPT_JOINREL)\\n> + /*\\n> + * Note: tlist for a base relation might be non-NIL. For example, if the\\n> + * base relation is an operand of a foreign join performing a full outer\\n> + * join and has non-NIL remote_conds, the base relation will be deparsed\\n> + * as a subquery, so the tlist for the base relation could be non-NIL.\\n> + */\\n> + if (tlist != NIL)\\n> {\\n> - /* For a join relation use the input tlist */\\n>\\n> We can not decide whether to use deparseExplicitTargetList or\\n> deparse it from attrs_used based on the tlist. SELECT lists, whether it's\\n> topmost SELECT or a subquery (even on a base relation), for deparsing a\\n> JOIN query need to use deparseExplicitTargetList.\\n\\nWill fix.\\n\\n> 2. The code in deparseRangeTblRef() dealing with subqueries, is very similar to\\n> deparseSelectStmtForRel(). The only thing deparseRangeTblRef() does not require\\n> is the appending ORDER BY, which is determined by existence of pathkeys\\n> argument. Probably we should reuse deparseSelectStmtForRel(), instead of\\n> duplicating the code. This will also make the current changes to\\n> deparseSelectSql unnecessary.\\n\\nWill do.\\n\\n> 3. Why do we need following change? The elog() just few lines above says that\\n> we expect only Var nodes. Why then we use deparseExpr() instead of\\n> deparseVar()?\\n> if (i > 0)\\n> appendStringInfoString(buf, \\\", \\\");\\n> - deparseVar(var, context);\\n> + deparseExpr((Expr *) var, context);\\n>\\n> *retrieved_attrs = lappend_int(*retrieved_attrs, i + 1);\\n>\\n> And I get the answer for that question, somewhere down in the patch\\n> + /*\\n> + * If the given expression is an output column of a subquery-in-FROM,\\n> + * deparse the alias to the expression instead.\\n> + */\\n> + if (IsA(node, Var))\\n> + {\\n> + int tabno;\\n> + int colno;\\n> +\\n> + if (isSubqueryExpr(node, context->foreignrel, &tabno, &colno))\\n> + {\\n> + appendStringInfo(context->buf, \\\"%s%d.%s%d\\\",\\n> + SS_TAB_ALIAS_PREFIX, tabno,\\n> + SS_COL_ALIAS_PREFIX, colno);\\n> + return;\\n> + }\\n> + }\\n> +\\n>\\n> Functionally, this code belongs to deparseVar() and not deparseExpr(). Like all\\n> other functions which handle Expr nodes, deparseExpr() is expected to be a\\n> dispatcher to the node specific function.\\n\\nOK, will change that way.\\n\\n> 4. This comment is good to have there, but is unrelated to this patch. May be a\\n> separate patch?\\n> + /* Don't generate bad syntax if no columns */\\n>\\n> 5. Changed comment doesn't say anything different from the original comment. It\\n> may have been good to have it this way to start with, but changing it now\\n> doesn't bring anything new. You seem to have just merged prologue of the\\n> function with a comment in that function. I think, this change is unnecessary.\\n> - * The function constructs ... JOIN ... ON ... for join relation. For a base\\n> - * relation it just returns schema-qualified tablename, with the appropriate\\n> - * alias if so requested.\\n> + * For a join relation the clause of the following form is appended to buf:\\n> + * ((outer relation) <join type> (inner relation) ON (joinclauses))\\n> + * For a base relation the function just adds the schema-qualified tablename,\\n> + * with the appropriate alias if so requested.\\n> + /* Don't generate bad syntax if no columns */\\n>\\n> 6. Change may be good but unrelated to this patch. May be material for a\\n> separate patch. There are few such changes in this function. While these\\n> changes may be good by themselves, they distract reviewer from the goal of the\\n> patch.\\n> - appendStringInfo(buf, \\\"(\\\");\\n> + appendStringInfoChar(buf, '(');\\n\\nOK on #4, #5, and #6, Will remove all the changes. I will leave those \\nfor separate patches.\\n\\n> 7. I don't understand why you need to change this function so much. Take for\\n> example the following change.\\n> - RelOptInfo *rel_o = fpinfo->outerrel;\\n> - RelOptInfo *rel_i = fpinfo->innerrel;\\n> - StringInfoData join_sql_o;\\n> - StringInfoData join_sql_i;\\n> + /* Begin the FROM clause entry */\\n> + appendStringInfoChar(buf, '(');\\n>\\n> /* Deparse outer relation */\\n> - initStringInfo(&join_sql_o);\\n> - deparseFromExprForRel(&join_sql_o, root, rel_o, true, params_list);\\n> + deparseRangeTblRef(buf, root,\\n> + fpinfo->outerrel,\\n> + fpinfo->make_outerrel_subquery,\\n> + params_list);\\n> /* Deparse outer relation */\\n> - initStringInfo(&join_sql_o);\\n> - deparseFromExprForRel(&join_sql_o, root, rel_o, true, params_list);\\n> + deparseRangeTblRef(buf, root,\\n> + fpinfo->outerrel,\\n> + fpinfo->make_outerrel_subquery,\\n> + params_list);\\n>\\n> It removes few variables, instead directly accesses the members from fpinfo.\\n> Also, deparses the individual relations in the provided buffer directly, rather\\n> than in separate buffers in the original code. Instead of this, all you had to\\n> do was replace a call\\n> - deparseFromExprForRel(&join_sql_o, root, rel_o, true, params_list);\\n> with\\n> + deparseRangeTblRef(&join_sql_o, root, rel_o,\\n> fpinfo->make_outerrel_subquery, params_list)\\n> Similarly for the inner relation. Again, the changes you have done might have\\n> been good, if done in the original code, but doing those in this patch just\\n> creates distractions and increases the size of the patch.\\n\\nI did that for efficiency because deparsing the relation directly into \\nthe given buffer would save cycles, but I agree that that is another \\nissue. Will remove that change, and leave that for another patch.\\n\\n> 8. Why have you changed the name of variable here?\\n> - if (use_alias)\\n> + if (add_rel_alias)\\n\\nSorry, I forgot to remove this change from the patch. Will fix.\\n\\n> 9. In deparseRangeTblRef(), the targetlist for the subquery is obtained by\\n> calling build_tlist_to_deparse(), but appendSubselectAlias() constructs the\\n> column aliases based on foreignrel->reltarget->exprs. Those two are usually\\n> same, but there is no code which guarantees that. Instead, pass tlist to\\n> appendSubselectAlias() or simply pass the number of entries in that list\\n> (list_length(tlist), which is all it needs. OR use foreignrel->reltarget->exprs\\n> directly instead of calling build_tlist_to_deparse(). Similar problem exists\\n> with getSubselectAliasInfo().\\n\\nActually, the patch guarantees that since in that case (1) \\nforeignrel->reltarget->exprs doesn't contain any PHVs (note that passing \\nthe following check in foreign_join_ok implies that \\nforeignrel->reltarget->exprs of the input rels for a given foreign join \\ndoesn't contain any PHVs and (2) we have fpinfo->local_conds == NIL, so \\nbuild_tlist_to_deparse() would produce a tlist equivalent to the \\nforeignrel->reltarget->exprs. But as you proposed, to make the code \\neasier to understand, I will change to use foreignrel->reltarget->exprs \\ndirectly.\\n\\n /*\\n * deparseExplicitTargetList() isn't smart enough to handle \\nanything other\\n * than a Var. In particular, if there's some PlaceHolderVar that \\nwould\\n * need to be evaluated within this join tree (because there's an upper\\n * reference to a quantity that may go to NULL as a result of an outer\\n * join), then we can't try to push the join down because we'll \\nfail when\\n * we get to deparseExplicitTargetList(). However, a \\nPlaceHolderVar that\\n * needs to be evaluated *at the top* of this join tree is OK, \\nbecause we\\n * can do that locally after fetching the results from the remote side.\\n */\\n relids = joinrel->relids;\\n foreach(lc, root->placeholder_list)\\n {\\n PlaceHolderInfo *phinfo = lfirst(lc);\\n\\n if (bms_is_subset(phinfo->ph_eval_at, relids) &&\\n bms_nonempty_difference(relids, phinfo->ph_eval_at))\\n return false;\\n }\\n\\n> 10. Deparsing a query in the form of a subquery makes it hard-to-read. The\\n> original code aimed at avoiding a subquery. Also, this change has created many\\n> expected output changes, which again seem unnecessary. In fact, having the\\n> pushable join clauses of an inner join in ON clause, which is closer to JOIN\\n> clause, is better than having them farther in the WHERE clause.\\n> - /*\\n> - * For an inner join, all restrictions can be treated alike. Treating the\\n> - * pushed down conditions as join conditions allows a top level full outer\\n> - * join to be deparsed without requiring subqueries.\\n> - */\\n> - if (jointype == JOIN_INNER)\\n> - {\\n> - Assert(!fpinfo->joinclauses);\\n> - fpinfo->joinclauses = fpinfo->remote_conds;\\n> - fpinfo->remote_conds = NIL;\\n> - }\\n\\nI added this change for another patch that I proposed for extending the \\nDML pushdown in 9.6 so that it can perform an update/delete on a join \\nremotely. To create a remote query for that, I think we need to pull up \\ninner join conditions mentioning the target relation in the JOIN/ON \\nclauses into the WHERE clause. But I have to admit that's unrelated to \\nthis patch, so I will leave that as-is.\\n\\n> 11. I have reworded following comment and restructured the code that follows it\\n> in the attached patch.\\n> + /*\\n> + * Set the subquery information. If the relation performs a full outer\\n> + * join and if the input relations have non-NIL remote_conds, the input\\n> + * relations need to be deparsed as a subquery.\\n> + */\\n> The code esp. the if .. else .. block followed by another one, made it a bit\\n> unreadable. Hence I restructured it so that it's readable and doesn't require\\n> variable \\\"relids\\\" from earlier code, which was being reused. Let me know if\\n> this change looks good.\\n\\nThat's a good idea. Thanks!\\n\\n> 12. The code below deletes the condition, which is understandable.\\n> - if (fpinfo_i->remote_conds || fpinfo_o->remote_conds)\\n> But why does it add a new unrelated condition here? What the comment claims,\\n> looks like an existing bug, unrelated to the patch. Can you please give an\\n> example? If that it's an existing bug, it should be fixed as a separate patch.\\n> I don't understand the relation of this code with what the patch is doing.\\n> + /*\\n> + * We can't do anything here, and if there are any non-Vars in the\\n> + * outerrel/innerrel's reltarget, give up pushing down this join,\\n> + * because the deparsing logic can't support such a case currently.\\n> + */\\n> + if (reltarget_has_non_vars(outerrel))\\n> + return false;\\n> + if (reltarget_has_non_vars(innerrel))\\n> return false;\\n\\nI thought the current deparsing logic wouldn't work well if the target \\nlist of a given subquery contained eg, system columns other than ctid \\nand oid, but I noticed that I was wrong; it can, so we don't need this \\nrestriction. Will remove. (I don't think there is any bug.)\\n\\n> 13. The comment below is missing the main purpose i.e. creating a a unique\\n> alias, in case the relation gets converted into a subquery. Lowest or highest\\n> relid will create a unique alias at given level of join and that would be more\\n> future proof. If we decide to consider paths for every join order, following\\n> comment will no more be true.\\n> + /*\\n> + * Set the relation index. This is defined as the position of this\\n> + * joinrel in the join_rel_list list plus the length of the rtable list.\\n> + * Note that since this joinrel is at the end of the list when we are\\n> + * called, we can get the position by list_length.\\n> + */\\n> + fpinfo->relation_index =\\n> + list_length(root->parse->rtable) + list_length(root->join_rel_list);\\n\\nI don't agree on that point. As I said before, the approach using the \\nlowest/highest relid would make a remote query having nested subqueries \\ndifficult to read. And even if we decided to consider paths for every \\njoin order, we could define the relation_index safely, by modifying \\npostgresGetForeignJoinPaths so that it (1) sets the relation_index the \\nproposed way at the first time it is called and (2) avoids setting the \\nrelation_index after the first call, by checking to see \\nfpinfo->relation_index > 0.\\n\\n> 14. The function name talks about non-vars but the Assert few lines below\\n> asserts that every node there is a Var. Need better naming.\\n> +reltarget_has_non_vars(RelOptInfo *foreignrel)\\n> +{\\n> + ListCell *lc;\\n> +\\n> + foreach(lc, foreignrel->reltarget->exprs)\\n> + {\\n> + Var *var = (Var *) lfirst(lc);\\n> +\\n> + Assert(IsA(var, Var));\\n> And also an explanation for the claim\\n> + * Note: currently deparseExplicitTargetList can't properly handle such Vars.\\n\\nWill remove this function for the reason described in #12.\\n\\n> 15. While deparsing every Var, we are descending the RelOptInfo hierarchy.\\n\\nRight. I think that not only avoids creating redundant data to find the \\nalias of a given Var node but also would be able to find it in optimal cost.\\n\\n> Instead of that, we should probably gather all the alias information once and\\n> store it in context as an array indexed by relid. While deparsing a Var we can\\n> get the targetlist and alias for a given relation by using var->varno as index.\\n> And then search for given Var node in the targetlist to deparse the column name\\n> by its position in the targetlist. For the relations that are not converted\\n> into subqueries, this array will not hold any information and the Var will be\\n> deparsed as it's being done today.\\n\\nSorry, I don't understand this part. How does that work when creating a \\nremote query having nested subqueries? Is that able to search for a \\ngiven Var node efficiently?\\n\\n> Testing\\n> -------\\n> 1. The code is changing deparser but doesn't have testcases\\n> testing impact of the code. For example, we should have testcases with remote\\n> query deparsed as nested subqueries or nested subqueries within subqueries and\\n> so on May be testcases where a relation deeper in the RelOptInfo hierarchy has\\n> conditions but it's immediate upper relation does not have those.\\n\\nWill do.\\n\\n> 2. The only testcase added by this patch, copies an existing case and adds a\\n> whole row reference to one of the relations being joined. Instead we could add\\n> that whole-row reference to the existing testcase itself.\\n\\nWill do.\\n\\nBest regards,\\nEtsuro Fujita\\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, 24 Oct 2016 17:59:03 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \">\\n>> 13. The comment below is missing the main purpose i.e. creating a a unique\\n>> alias, in case the relation gets converted into a subquery. Lowest or\\n>> highest\\n>> relid will create a unique alias at given level of join and that would be\\n>> more\\n>> future proof. If we decide to consider paths for every join order,\\n>> following\\n>> comment will no more be true.\\n>> + /*\\n>> + * Set the relation index. This is defined as the position of this\\n>> + * joinrel in the join_rel_list list plus the length of the rtable\\n>> list.\\n>> + * Note that since this joinrel is at the end of the list when we are\\n>> + * called, we can get the position by list_length.\\n>> + */\\n>> + fpinfo->relation_index =\\n>> + list_length(root->parse->rtable) +\\n>> list_length(root->join_rel_list);\\n>\\n>\\n> I don't agree on that point. As I said before, the approach using the\\n> lowest/highest relid would make a remote query having nested subqueries\\n> difficult to read. And even if we decided to consider paths for every join\\n> order, we could define the relation_index safely, by modifying\\n> postgresGetForeignJoinPaths so that it (1) sets the relation_index the\\n> proposed way at the first time it is called and (2) avoids setting the\\n> relation_index after the first call, by checking to see\\n> fpinfo->relation_index > 0.\\n\\nOK. Although, the alias which contains a number, which apparently\\ndoesn't show any relation with the relation (no pun intended :)) being\\ndeparsed, won't make much sense in the deparsed query. But I am fine\\nleaving this to the committer's judgement. May be you want to add an\\nassert here making sure that relation_index is set only once.\\nSomething like Assert(fpinfo->relation_index == 0) just before setting\\nit.\\n\\n>\\n>> 14. The function name talks about non-vars but the Assert few lines below\\n>> asserts that every node there is a Var. Need better naming.\\n>> +reltarget_has_non_vars(RelOptInfo *foreignrel)\\n>> +{\\n>> + ListCell *lc;\\n>> +\\n>> + foreach(lc, foreignrel->reltarget->exprs)\\n>> + {\\n>> + Var *var = (Var *) lfirst(lc);\\n>> +\\n>> + Assert(IsA(var, Var));\\n>> And also an explanation for the claim\\n>> + * Note: currently deparseExplicitTargetList can't properly handle such\\n>> Vars.\\n>\\n>\\n> Will remove this function for the reason described in #12.\\n>\\n>> 15. While deparsing every Var, we are descending the RelOptInfo hierarchy.\\n>\\n>\\n> Right. I think that not only avoids creating redundant data to find the\\n> alias of a given Var node but also would be able to find it in optimal cost.\\n>\\n>> Instead of that, we should probably gather all the alias information once\\n>> and\\n>> store it in context as an array indexed by relid. While deparsing a Var we\\n>> can\\n>> get the targetlist and alias for a given relation by using var->varno as\\n>> index.\\n>> And then search for given Var node in the targetlist to deparse the column\\n>> name\\n>> by its position in the targetlist. For the relations that are not\\n>> converted\\n>> into subqueries, this array will not hold any information and the Var will\\n>> be\\n>> deparsed as it's being done today.\\n>\\n>\\n> Sorry, I don't understand this part. How does that work when creating a\\n> remote query having nested subqueries? Is that able to search for a given\\n> Var node efficiently?\\n\\nFor every relation that is deparsed as a subquery, we will create a\\nseparate context. Each such context will have an array described\\nabove. This array will contain the targetlist and aliases for all the\\nrelations, covered by that relation, that are required to be deparsed\\nas subqueries. These arrays bubble up to topmost relation that is not\\nrequired to be deparsed as subquery. When a relation is required to be\\ndeparsed as a subquery, its immediate upper relation sets the alias\\nand targetlist chosen for that relation at all the indexes\\ncorresponding to all the base relations covered by that relation. For\\nexample, let's assume that a relation (1, 2, 3) is required to be\\ndeparsed as subquery for an immediate upper relation (1, 2, 3, 4, 5)\\n(thus the other joining relation being (4,5)). While deparsing for\\nrelation (1,2,3,4,5), the context will contain a 5 element array, with\\npositions 1, 2, 3 filled by same targetlist and alias whereas\\npositions 4 and 5 will not be filled as those relations are not\\ndeparsed as subqueries. Let's assume in relation (1, 2, 3), (1, 3) in\\nturn requires subquery but (2) does not. Thus the context created\\nwhile deparsing (1, 2, 3) will have a 3 element array with positions 1\\nand 3 containing the same targetlist and alias, where as position 2\\nwill be empty. When deparsing a Var node with varno = N and varattno =\\nm, if the nth position in the array in the context is empty, that Var\\nnode will be deparsed as rN.<column name>. But if that position is has\\nalias sZ, then we search for that Var node in the targetlist and if\\nit's found at kth position in the targetlist, we will deparse it as\\nsZ.ck. The search in the targetlist can be performed using\\ntlist_member, and then fetching the position by TargetEntry::resno.\\nThis does not require any recursion and thus saves stack space and\\nsome CPU cycles required for recursion. I guess, the arrays need to be\\ncomputed only once for any relation when the query for that relation\\nis deparsed the first time.\\n\\nThanks for considering other suggestions.\\n\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 25 Oct 2016 15:28:07 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/10/25 18:58, Ashutosh Bapat wrote:\\n\\nYou wrote:\\n>>> 13. The comment below is missing the main purpose i.e. creating a a unique\\n>>> alias, in case the relation gets converted into a subquery. Lowest or\\n>>> highest\\n>>> relid will create a unique alias at given level of join and that would be\\n>>> more\\n>>> future proof. If we decide to consider paths for every join order,\\n>>> following\\n>>> comment will no more be true.\\n>>> + /*\\n>>> + * Set the relation index. This is defined as the position of this\\n>>> + * joinrel in the join_rel_list list plus the length of the rtable\\n>>> list.\\n>>> + * Note that since this joinrel is at the end of the list when we are\\n>>> + * called, we can get the position by list_length.\\n>>> + */\\n>>> + fpinfo->relation_index =\\n>>> + list_length(root->parse->rtable) +\\n>>> list_length(root->join_rel_list);\\n\\nI wrote:\\n>> I don't agree on that point. As I said before, the approach using the\\n>> lowest/highest relid would make a remote query having nested subqueries\\n>> difficult to read. And even if we decided to consider paths for every join\\n>> order, we could define the relation_index safely, by modifying\\n>> postgresGetForeignJoinPaths so that it (1) sets the relation_index the\\n>> proposed way at the first time it is called and (2) avoids setting the\\n>> relation_index after the first call, by checking to see\\n>> fpinfo->relation_index > 0.\\n\\n> OK. Although, the alias which contains a number, which apparently\\n> doesn't show any relation with the relation (no pun intended :)) being\\n> deparsed, won't make much sense in the deparsed query. But I am fine\\n> leaving this to the committer's judgement.\\n\\nFine with me.\\n\\n> May be you want to add an\\n> assert here making sure that relation_index is set only once.\\n> Something like Assert(fpinfo->relation_index == 0) just before setting\\n> it.\\n\\nWill do.\\n\\n>>> 15. While deparsing every Var, we are descending the RelOptInfo hierarchy.\\n\\n>> Right. I think that not only avoids creating redundant data to find the\\n>> alias of a given Var node but also would be able to find it in optimal cost.\\n\\n>>> Instead of that, we should probably gather all the alias information once\\n>>> and\\n>>> store it in context as an array indexed by relid. While deparsing a Var we\\n>>> can\\n>>> get the targetlist and alias for a given relation by using var->varno as\\n>>> index.\\n>>> And then search for given Var node in the targetlist to deparse the column\\n>>> name\\n>>> by its position in the targetlist. For the relations that are not\\n>>> converted\\n>>> into subqueries, this array will not hold any information and the Var will\\n>>> be\\n>>> deparsed as it's being done today.\\n\\n>> Sorry, I don't understand this part. How does that work when creating a\\n>> remote query having nested subqueries? Is that able to search for a given\\n>> Var node efficiently?\\n\\n> For every relation that is deparsed as a subquery, we will create a\\n> separate context. Each such context will have an array described\\n> above. This array will contain the targetlist and aliases for all the\\n> relations, covered by that relation, that are required to be deparsed\\n> as subqueries. These arrays bubble up to topmost relation that is not\\n> required to be deparsed as subquery. When a relation is required to be\\n> deparsed as a subquery, its immediate upper relation sets the alias\\n> and targetlist chosen for that relation at all the indexes\\n> corresponding to all the base relations covered by that relation.\\n\\nThanks for the explanation!\\n\\n> For\\n> example, let's assume that a relation (1, 2, 3) is required to be\\n> deparsed as subquery for an immediate upper relation (1, 2, 3, 4, 5)\\n> (thus the other joining relation being (4,5)). While deparsing for\\n> relation (1,2,3,4,5), the context will contain a 5 element array, with\\n> positions 1, 2, 3 filled by same targetlist and alias whereas\\n> positions 4 and 5 will not be filled as those relations are not\\n> deparsed as subqueries.\\n\\nSorry, I don't understand this. In that case, the immediate upper \\nrelation (1, 2, 3, 4, 5) would need to fill the targetlist and aliases \\nfor *the join relation (1, 2, 3) somewhere*, not the targetlist and \\naliases for each of the component relations 1, 2, and 3, because the \\njoin relation is deparsed as a subquery. Maybe I'm missing something, \\nthough.\\n\\n> Let's assume in relation (1, 2, 3), (1, 3) in\\n> turn requires subquery but (2) does not. Thus the context created\\n> while deparsing (1, 2, 3) will have a 3 element array with positions 1\\n> and 3 containing the same targetlist and alias, where as position 2\\n> will be empty.\\n\\n> When deparsing a Var node with varno = N and varattno =\\n> m, if the nth position in the array in the context is empty, that Var\\n> node will be deparsed as rN.<column name>.\\n\\nWhat happens when deparsing eg, a Var with varno = 2 at the topmost \\nrelation (1, 2, 3, 4, 5)? The second position of the array is empty, \\nbut the join relation (1, 2, 3) is deparsed as a subquery, so the Var \\nshould be deparsed as an alias of an output column of the subquery at \\nthe topmost relation, I think.\\n\\n> But if that position is has\\n> alias sZ, then we search for that Var node in the targetlist and if\\n> it's found at kth position in the targetlist, we will deparse it as\\n> sZ.ck. The search in the targetlist can be performed using\\n> tlist_member, and then fetching the position by TargetEntry::resno.\\n\\n> This does not require any recursion and thus saves stack space and\\n> some CPU cycles required for recursion.\\n\\nIs that true?\\n\\n> some CPU cycles required for recursion. I guess, the arrays need to be\\n> computed only once for any relation when the query for that relation\\n> is deparsed the first time.\\n\\nDoes this algorithm extend to the case where we consider paths for every \\njoin order?\\n\\n> Thanks for considering other suggestions.\\n\\nYour suggestions are really helpful to improve the patch. Thanks!\\n\\nBest regards,\\nEtsuro Fujita\\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, 25 Oct 2016 21:21:42 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \">\\n>\\n>> For every relation that is deparsed as a subquery, we will create a\\n>> separate context. Each such context will have an array described\\n>> above. This array will contain the targetlist and aliases for all the\\n>> relations, covered by that relation, that are required to be deparsed\\n>> as subqueries. These arrays bubble up to topmost relation that is not\\n>> required to be deparsed as subquery. When a relation is required to be\\n>> deparsed as a subquery, its immediate upper relation sets the alias\\n>> and targetlist chosen for that relation at all the indexes\\n>> corresponding to all the base relations covered by that relation.\\n>\\n>\\n> Thanks for the explanation!\\n>\\n>> For\\n>> example, let's assume that a relation (1, 2, 3) is required to be\\n>> deparsed as subquery for an immediate upper relation (1, 2, 3, 4, 5)\\n>> (thus the other joining relation being (4,5)). While deparsing for\\n>> relation (1,2,3,4,5), the context will contain a 5 element array, with\\n>> positions 1, 2, 3 filled by same targetlist and alias whereas\\n>> positions 4 and 5 will not be filled as those relations are not\\n>> deparsed as subqueries.\\n>\\n>\\n> Sorry, I don't understand this. In that case, the immediate upper relation\\n> (1, 2, 3, 4, 5) would need to fill the targetlist and aliases for *the join\\n> relation (1, 2, 3) somewhere*, not the targetlist and aliases for each of\\n> the component relations 1, 2, and 3, because the join relation is deparsed\\n> as a subquery. Maybe I'm missing something, though.\\n\\nThe description above does not specify \\\"targetlist and alias\\\" for each\\nof (1, 2, 3). The array in the context will have positions 1, 2, 3\\nfilled with *same* alias and targetlist which is derived from relation\\n(1, 2, 3).\\n\\n>\\n>> Let's assume in relation (1, 2, 3), (1, 3) in\\n>> turn requires subquery but (2) does not. Thus the context created\\n>> while deparsing (1, 2, 3) will have a 3 element array with positions 1\\n>> and 3 containing the same targetlist and alias, where as position 2\\n>> will be empty.\\n>\\n>\\n>> When deparsing a Var node with varno = N and varattno =\\n>> m, if the nth position in the array in the context is empty, that Var\\n>> node will be deparsed as rN.<column name>.\\n>\\n>\\n> What happens when deparsing eg, a Var with varno = 2 at the topmost relation\\n> (1, 2, 3, 4, 5)? The second position of the array is empty, but the join\\n> relation (1, 2, 3) is deparsed as a subquery, so the Var should be deparsed\\n> as an alias of an output column of the subquery at the topmost relation, I\\n> think.\\n\\nposition 2 will not be empty, it will be filled by the alias and\\ntargetlist derived from relation (1, 2, 3).\\n\\n>\\n>> But if that position is has\\n>> alias sZ, then we search for that Var node in the targetlist and if\\n>> it's found at kth position in the targetlist, we will deparse it as\\n>> sZ.ck. The search in the targetlist can be performed using\\n>> tlist_member, and then fetching the position by TargetEntry::resno.\\n>\\n>\\n>> This does not require any recursion and thus saves stack space and\\n>> some CPU cycles required for recursion.\\n>\\n>\\n> Is that true?\\n\\nYes, unless you explain why is that false.\\n\\n>\\n>> some CPU cycles required for recursion. I guess, the arrays need to be\\n>> computed only once for any relation when the query for that relation\\n>> is deparsed the first time.\\n>\\n>\\n> Does this algorithm extend to the case where we consider paths for every\\n> join order?\\n\\nYes, if we store the information about which of relations need\\nsubquery and which don't for every join order.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 26 Oct 2016 12:41:44 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/10/26 16:11, Ashutosh Bapat wrote:\\n\\nYou wrote:\\n>>> For\\n>>> example, let's assume that a relation (1, 2, 3) is required to be\\n>>> deparsed as subquery for an immediate upper relation (1, 2, 3, 4, 5)\\n>>> (thus the other joining relation being (4,5)). While deparsing for\\n>>> relation (1,2,3,4,5), the context will contain a 5 element array, with\\n>>> positions 1, 2, 3 filled by same targetlist and alias whereas\\n>>> positions 4 and 5 will not be filled as those relations are not\\n>>> deparsed as subqueries.\\n\\nI wrote:\\n>> Sorry, I don't understand this. In that case, the immediate upper relation\\n>> (1, 2, 3, 4, 5) would need to fill the targetlist and aliases for *the join\\n>> relation (1, 2, 3) somewhere*, not the targetlist and aliases for each of\\n>> the component relations 1, 2, and 3, because the join relation is deparsed\\n>> as a subquery. Maybe I'm missing something, though.\\n\\n> The description above does not specify \\\"targetlist and alias\\\" for each\\n> of (1, 2, 3). The array in the context will have positions 1, 2, 3\\n> filled with *same* alias and targetlist which is derived from relation\\n> (1, 2, 3).\\n\\nOK\\n\\n>>> Let's assume in relation (1, 2, 3), (1, 3) in\\n>>> turn requires subquery but (2) does not. Thus the context created\\n>>> while deparsing (1, 2, 3) will have a 3 element array with positions 1\\n>>> and 3 containing the same targetlist and alias, where as position 2\\n>>> will be empty.\\n\\n>>> When deparsing a Var node with varno = N and varattno =\\n>>> m, if the nth position in the array in the context is empty, that Var\\n>>> node will be deparsed as rN.<column name>.\\n\\n>> What happens when deparsing eg, a Var with varno = 2 at the topmost relation\\n>> (1, 2, 3, 4, 5)? The second position of the array is empty, but the join\\n>> relation (1, 2, 3) is deparsed as a subquery, so the Var should be deparsed\\n>> as an alias of an output column of the subquery at the topmost relation, I\\n>> think.\\n\\n> position 2 will not be empty, it will be filled by the alias and\\n> targetlist derived from relation (1, 2, 3).\\n\\nOK\\n\\n>>> But if that position is has\\n>>> alias sZ, then we search for that Var node in the targetlist and if\\n>>> it's found at kth position in the targetlist, we will deparse it as\\n>>> sZ.ck. The search in the targetlist can be performed using\\n>>> tlist_member, and then fetching the position by TargetEntry::resno.\\n\\n>>> This does not require any recursion and thus saves stack space and\\n>>> some CPU cycles required for recursion.\\n\\n>> Is that true?\\n\\n> Yes, unless you explain why is that false.\\n\\nOK, that would be true, I think.\\n\\n>>> I guess, the arrays need to be\\n>>> computed only once for any relation when the query for that relation\\n>>> is deparsed the first time.\\n\\n>> Does this algorithm extend to the case where we consider paths for every\\n>> join order?\\n\\n> Yes, if we store the information about which of relations need\\n> subquery and which don't for every join order.\\n\\nHmm. Sorry, I'm not so excited about this proposal. I think (1) that \\nis solving a problem that hasn't been proven to be a problem, (2) that \\nwould complicate the deparser logic, and (3) the cost of creating this \\narray for each relation by the bottom-up method while deparsing a remote \\nquery would be not small (especially when the query is large), so that \\nmight need more cycles for deparsing the query than what I proposed when \\nuse_remote_estimate=false. So, I'd like to go with what I proposed, at \\nleast as the first cut.\\n\\nBest regards,\\nEtsuro Fujita\\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\": \"Wed, 26 Oct 2016 18:00:42 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \">\\n>>>> I guess, the arrays need to be\\n>>>> computed only once for any relation when the query for that relation\\n>>>> is deparsed the first time.\\n>\\n>\\n>>> Does this algorithm extend to the case where we consider paths for every\\n>>> join order?\\n>\\n>\\n>> Yes, if we store the information about which of relations need\\n>> subquery and which don't for every join order.\\n>\\n>\\n> Hmm. Sorry, I'm not so excited about this proposal. I think (1) that is\\n> solving a problem that hasn't been proven to be a problem, (2) that would\\n> complicate the deparser logic, and (3) the cost of creating this array for\\n> each relation by the bottom-up method while deparsing a remote query would\\n> be not small (especially when the query is large), so that might need more\\n> cycles for deparsing the query than what I proposed when\\n> use_remote_estimate=false. So, I'd like to go with what I proposed, at\\n> least as the first cut.\\n\\nThe arrays will need to computed only when there is at least one\\nrelation required to be deparsed as subquery. Every relation above the\\nrelation which is converted into subquery requires the array. Each\\narray will be N * size of pointer long where N is the largest relid\\ncovered by that relation. N will vary across the RelOptInfo hierarchy.\\nAll that array holds is targetlist pointer which are required to\\ncomputed anyway. So, the amount of memory is bounded by N * size of\\npointer * (2N - 1), which is way lesser than what we use in other\\nplaces.\\n\\nYour patch calls isSubqueryExpr() recursively for every Var in the\\ntargetlist, which can be many for foreign tables with many columns.\\nFor every such Var it may need to reach upto the relation which is\\nconverted into subquery, which can as bad as reaching every base\\nrelation. So, it looks like the number of recursive calls to\\nisSubqueryExpr() is bounded by V * N (i.e. worst case depth of the\\nRelOptInfo tree), which can be quite costly. If use_remote_estimates\\nis true, we do this for every intermediate relation and for every path\\ncreated. That isn't very performance efficient.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 26 Oct 2016 14:55:03 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/10/26 18:25, Ashutosh Bapat wrote:\\n\\n> Your patch calls isSubqueryExpr() recursively for every Var in the\\n> targetlist, which can be many for foreign tables with many columns.\\n> For every such Var it may need to reach upto the relation which is\\n> converted into subquery, which can as bad as reaching every base\\n> relation. So, it looks like the number of recursive calls to\\n> isSubqueryExpr() is bounded by V * N (i.e. worst case depth of the\\n> RelOptInfo tree), which can be quite costly. If use_remote_estimates\\n> is true, we do this for every intermediate relation and for every path\\n> created. That isn't very performance efficient.\\n\\nIn practice, the search cost would be negligible compared to the cost of \\nexplaining/executing the query.\\n\\nMy concern about your proposal is: it might not be worth complicating \\nthe code to solve a problem that is actually not a problem in practice.\\n\\nBest regards,\\nEtsuro Fujita\\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\": \"Wed, 26 Oct 2016 19:05:37 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 26, 2016 at 3:35 PM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> On 2016/10/26 18:25, Ashutosh Bapat wrote:\\n>\\n>> Your patch calls isSubqueryExpr() recursively for every Var in the\\n>> targetlist, which can be many for foreign tables with many columns.\\n>> For every such Var it may need to reach upto the relation which is\\n>> converted into subquery, which can as bad as reaching every base\\n>> relation. So, it looks like the number of recursive calls to\\n>> isSubqueryExpr() is bounded by V * N (i.e. worst case depth of the\\n>> RelOptInfo tree), which can be quite costly. If use_remote_estimates\\n>> is true, we do this for every intermediate relation and for every path\\n>> created. That isn't very performance efficient.\\n>\\n>\\n> In practice, the search cost would be negligible compared to the cost of\\n> explaining/executing the query.\\n>\\n> My concern about your proposal is: it might not be worth complicating the\\n> code to solve a problem that is actually not a problem in practice.\\n>\\n\\nTo me the current code looks complicated esp. because of the recursion\\ninvolved and usage of out parameters to isSubqueryExpr(). My\\nsuggestion goes inline with the current method of deparsing a Var.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 26 Oct 2016 16:23:29 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/10/26 19:53, Ashutosh Bapat wrote:\\n> On Wed, Oct 26, 2016 at 3:35 PM, Etsuro Fujita\\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n\\n>> In practice, the search cost would be negligible compared to the cost of\\n>> explaining/executing the query.\\n>>\\n>> My concern about your proposal is: it might not be worth complicating the\\n>> code to solve a problem that is actually not a problem in practice.\\n\\n> To me the current code looks complicated esp. because of the recursion\\n> involved and usage of out parameters to isSubqueryExpr().\\n\\nI don't think so. isSubqueryExpr is prety small, written in less than \\n50 lines, and the code looks rather simple to me.\\n\\n> My\\n> suggestion goes inline with the current method of deparsing a Var.\\n\\nYeah, I think your approach makes it easy to search for the alias to a \\ngiven Var from the array you proposed. I think the complexity of your \\napproach would be in extra work for building and maintaining the array \\nwhile deparsing the query. I think that would probably need more \\ninvasive and much larger changes to the existing code than what I proposed.\\n\\nI think this issue is optional, so I'd like to propose to leave this for \\nthe committer's judge.\\n\\nBest regards,\\nEtsuro Fujita\\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, 27 Oct 2016 16:16:41 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 27, 2016 at 12:46 PM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> On 2016/10/26 19:53, Ashutosh Bapat wrote:\\n>>\\n>> On Wed, Oct 26, 2016 at 3:35 PM, Etsuro Fujita\\n>> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>\\n>\\n>>> In practice, the search cost would be negligible compared to the cost of\\n>>> explaining/executing the query.\\n>>>\\n>>> My concern about your proposal is: it might not be worth complicating the\\n>>> code to solve a problem that is actually not a problem in practice.\\n>\\n>\\n>> To me the current code looks complicated esp. because of the recursion\\n>> involved and usage of out parameters to isSubqueryExpr().\\n>\\n>\\n> I don't think so. isSubqueryExpr is prety small, written in less than 50\\n> lines, and the code looks rather simple to me.\\n>\\n>> My\\n>> suggestion goes inline with the current method of deparsing a Var.\\n>\\n>\\n> Yeah, I think your approach makes it easy to search for the alias to a given\\n> Var from the array you proposed. I think the complexity of your approach\\n> would be in extra work for building and maintaining the array while\\n> deparsing the query. I think that would probably need more invasive and\\n> much larger changes to the existing code than what I proposed.\\n>\\n> I think this issue is optional, so I'd like to propose to leave this for the\\n> committer's judge.\\n>\\n\\nFine with me.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 27 Oct 2016 14:36:04 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/10/27 18:06, Ashutosh Bapat wrote:\\n> On Thu, Oct 27, 2016 at 12:46 PM, Etsuro Fujita\\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>> On 2016/10/26 19:53, Ashutosh Bapat wrote:\\n\\n>>> On Wed, Oct 26, 2016 at 3:35 PM, Etsuro Fujita\\n>>>> My concern about your proposal is: it might not be worth complicating the\\n>>>> code to solve a problem that is actually not a problem in practice.\\n\\n>>> To me the current code looks complicated esp. because of the recursion\\n>>> involved and usage of out parameters to isSubqueryExpr().\\n\\n>> I don't think so. isSubqueryExpr is prety small, written in less than 50\\n>> lines, and the code looks rather simple to me.\\n\\n>>> My\\n>>> suggestion goes inline with the current method of deparsing a Var.\\n\\n>> Yeah, I think your approach makes it easy to search for the alias to a given\\n>> Var from the array you proposed. I think the complexity of your approach\\n>> would be in extra work for building and maintaining the array while\\n>> deparsing the query. I think that would probably need more invasive and\\n>> much larger changes to the existing code than what I proposed.\\n>>\\n>> I think this issue is optional, so I'd like to propose to leave this for the\\n>> committer's judge.\\n\\n> Fine with me.\\n\\nOK\\n\\nHere is the updated version, which includes the restructuring you \\nproposed. Other than the above issue and the alias issue we discussed, \\nI addressed all your comments except one on testing; I tried to add test \\ncases where the remote query is deparsed as nested subqueries, but I \\ncouldn't because IIUC, reduce_outer_joins reduced full joins to inner \\njoins or left joins. So, I added two test cases: (1) the joining \\nrelations are both base relations (actually, we already have that) and \\n(2) one of the joining relations is a base relation and the other is a \\njoin relation. I rebased the patch to HEAD, so I added a test case with \\naggregate pushdown also.\\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\": \"Thu, 27 Oct 2016 20:41:21 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/10/27 20:41, Etsuro Fujita wrote:\\n> Here is the updated version, which includes the restructuring you\\n> proposed. Other than the above issue and the alias issue we discussed,\\n> I addressed all your comments except one on testing; I tried to add test\\n> cases where the remote query is deparsed as nested subqueries, but I\\n> couldn't because IIUC, reduce_outer_joins reduced full joins to inner\\n> joins or left joins. So, I added two test cases: (1) the joining\\n> relations are both base relations (actually, we already have that) and\\n> (2) one of the joining relations is a base relation and the other is a\\n> join relation. I rebased the patch to HEAD, so I added a test case with\\n> aggregate pushdown also.\\n\\nI've updated another patch for evaluating PHVs remotely. Attached is an \\nupdated version of the patch, which is created on top of the patch for \\nsupporting deparsing subqueries. You can find test cases where a remote \\njoin query is deparsed as nested subqueries.\\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\": \"Tue, 1 Nov 2016 21:01:57 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"> Here is the updated version, which includes the restructuring you proposed.\\n> Other than the above issue and the alias issue we discussed, I addressed all\\n> your comments except one on testing; I tried to add test cases where the\\n> remote query is deparsed as nested subqueries, but I couldn't because IIUC,\\n> reduce_outer_joins reduced full joins to inner joins or left joins.\\n\\nNo always. It will do so in some cases as explained in the prologue of\\nreduce_outer_joins()\\n\\n * More generally, an outer join can be reduced in strength if there is a\\n * strict qual above it in the qual tree that constrains a Var from the\\n * nullable side of the join to be non-null. (For FULL joins this applies\\n * to each side separately.)\\n\\n> So, I\\n> added two test cases: (1) the joining relations are both base relations\\n> (actually, we already have that) and (2) one of the joining relations is a\\n> base relation and the other is a join relation. I rebased the patch to\\n> HEAD, so I added a test case with aggregate pushdown also.\\n>\\n\\nThis patch again has a lot of unrelated changes, esp. in\\ndeparseSelectStmtForRel(). What was earlier part of deparseSelectSql()\\nand deparseFromExpr() is now flattened in deparseSelectStmtForRel(),\\nwhich seems unnecessary. Or there are changes like\\n\\n-deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root, RelOptInfo *rel,\\n- List *tlist, List *remote_conds, List *pathkeys,\\n+deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root,\\n+ RelOptInfo *foreignrel, List *tlist,\\n+ List *remote_conds, List *pathkeys,\\n\\nwhich arise because rel has been renamed as foreignrel. The patch will\\nwork even without such renaming.\\n\\nI think such refactoring, should be done in a separate patch.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 3 Nov 2016 15:22:33 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/11/03 18:52, Ashutosh Bapat wrote:\\n>> Here is the updated version, which includes the restructuring you proposed.\\n>> Other than the above issue and the alias issue we discussed, I addressed all\\n>> your comments except one on testing; I tried to add test cases where the\\n>> remote query is deparsed as nested subqueries, but I couldn't because IIUC,\\n>> reduce_outer_joins reduced full joins to inner joins or left joins.\\n\\n> No always. It will do so in some cases as explained in the prologue of\\n> reduce_outer_joins()\\n>\\n> * More generally, an outer join can be reduced in strength if there is a\\n> * strict qual above it in the qual tree that constrains a Var from the\\n> * nullable side of the join to be non-null. (For FULL joins this applies\\n> * to each side separately.)\\n\\nHmm. Would it be necessary for this patch to include test cases for \\nnested subqueries? As mentioned in a previous email, those test cases \\ncan be added by the split patch that allow PHVs to be evaluated on the \\nremote side.\\n\\n> This patch again has a lot of unrelated changes, esp. in\\n> deparseSelectStmtForRel(). What was earlier part of deparseSelectSql()\\n> and deparseFromExpr() is now flattened in deparseSelectStmtForRel(),\\n> which seems unnecessary.\\n\\nIIUC, I think this change was done to address your comment (see the \\ncomment #2 in [1]). Am I missing something?\\n\\n> Or there are changes like\\n>\\n> -deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root, RelOptInfo *rel,\\n> - List *tlist, List *remote_conds, List *pathkeys,\\n> +deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root,\\n> + RelOptInfo *foreignrel, List *tlist,\\n> + List *remote_conds, List *pathkeys,\\n>\\n> which arise because rel has been renamed as foreignrel. The patch will\\n> work even without such renaming.\\n\\nI did that because we have a Relation named rel (the same name!) within \\nthat function, to execute deparseTargetList in a base-relation case. \\nAnother reason for that is because (1) that would match with other \\nfunction definitions in deparse.c and (2) that would be consistent with \\nthe existing function definition for that function in postgres_fdw.h.\\n\\nBest regards,\\nEtsuro Fujita\\n\\n[1] \\nhttps://www.postgresql.org/message-id/CAFjFpRfwoSsJr9418b2jA7g0nwagjZSWhPQPUmK9M6z5XSOAqQ%40mail.gmail.com\\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, 4 Nov 2016 12:49:02 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 4, 2016 at 9:19 AM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> On 2016/11/03 18:52, Ashutosh Bapat wrote:\\n>>>\\n>>> Here is the updated version, which includes the restructuring you\\n>>> proposed.\\n>>> Other than the above issue and the alias issue we discussed, I addressed\\n>>> all\\n>>> your comments except one on testing; I tried to add test cases where the\\n>>> remote query is deparsed as nested subqueries, but I couldn't because\\n>>> IIUC,\\n>>> reduce_outer_joins reduced full joins to inner joins or left joins.\\n>\\n>\\n>> No always. It will do so in some cases as explained in the prologue of\\n>> reduce_outer_joins()\\n>>\\n>> * More generally, an outer join can be reduced in strength if there is a\\n>> * strict qual above it in the qual tree that constrains a Var from the\\n>> * nullable side of the join to be non-null. (For FULL joins this applies\\n>> * to each side separately.)\\n>\\n>\\n> Hmm. Would it be necessary for this patch to include test cases for nested\\n> subqueries? As mentioned in a previous email, those test cases can be added\\n> by the split patch that allow PHVs to be evaluated on the remote side.\\n\\nA patch should have testcases testing the functionality added. This\\npatch adds functionality to deparse nested subqueries, so there should\\nbe testcase to test it. If we can not come up with a testcase, then\\nit's very much possible that the code related to that functionality is\\nnot needed. PHV is a separate patch and we do not know whether it will\\nbe committed or when it will be committed after this patch is\\ncommitted. It's better to have self-sufficient patch.\\n\\n>\\n>> This patch again has a lot of unrelated changes, esp. in\\n>> deparseSelectStmtForRel(). What was earlier part of deparseSelectSql()\\n>> and deparseFromExpr() is now flattened in deparseSelectStmtForRel(),\\n>> which seems unnecessary.\\n>\\n>\\n> IIUC, I think this change was done to address your comment (see the comment\\n> #2 in [1]). Am I missing something?\\n\\nThere is some misunderstanding here. That comment basically says,\\ndeparseRangeTblRef() duplicates code in deparseSelectStmtForRel(), so\\nwe should either remove deparseRangeTblRef() altogether or should keep\\nit to minimal avoiding duplication of code. What might have confused\\nyou is the last sentence in that comment \\\"This will also make the\\ncurrent changes to deparseSelectSql unnecessary.\\\" By current changes I\\nmeant changes to deparseSelectSql() in your patch, not the one that's\\nin the repository. I don't think we should flatten\\ndeparseSelectStmtForRel() in this patch.\\n\\n>\\n>> Or there are changes like\\n>>\\n>> -deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root, RelOptInfo\\n>> *rel,\\n>> - List *tlist, List *remote_conds, List *pathkeys,\\n>> +deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root,\\n>> + RelOptInfo *foreignrel, List *tlist,\\n>> + List *remote_conds, List *pathkeys,\\n>>\\n>> which arise because rel has been renamed as foreignrel. The patch will\\n>> work even without such renaming.\\n>\\n>\\n> I did that because we have a Relation named rel (the same name!) within that\\n> function, to execute deparseTargetList in a base-relation case.\\n\\nThat's because you have flattened deparseSelectStmtForRel(). If we\\ndon't flatten it out, the variable name conflict doesn't arise.\\n\\n> Another\\n> reason for that is because (1) that would match with other function\\n> definitions in deparse.c and (2) that would be consistent with the existing\\n> function definition for that function in postgres_fdw.h.\\n>\\n\\nThat would be a separate patch, I guess.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 4 Nov 2016 09:38:17 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/11/04 13:08, Ashutosh Bapat wrote:\\n> On Fri, Nov 4, 2016 at 9:19 AM, Etsuro Fujita\\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>> On 2016/11/03 18:52, Ashutosh Bapat wrote:\\n\\nI wrote:\\n>>>> Here is the updated version,\\n\\n>>>> Other than the above issue and the alias issue we discussed, I addressed\\n>>>> all\\n>>>> your comments except one on testing; I tried to add test cases where the\\n>>>> remote query is deparsed as nested subqueries, but I couldn't because\\n>>>> IIUC,\\n>>>> reduce_outer_joins reduced full joins to inner joins or left joins.\\n\\n>>> No always. It will do so in some cases as explained in the prologue of\\n>>> reduce_outer_joins()\\n>>>\\n>>> * More generally, an outer join can be reduced in strength if there is a\\n>>> * strict qual above it in the qual tree that constrains a Var from the\\n>>> * nullable side of the join to be non-null. (For FULL joins this applies\\n>>> * to each side separately.)\\n\\nRight.\\n\\n>> Would it be necessary for this patch to include test cases for nested\\n>> subqueries?\\n\\n> A patch should have testcases testing the functionality added. This\\n> patch adds functionality to deparse nested subqueries, so there should\\n> be testcase to test it.\\n\\nOK, I added such a test case.\\n\\n>>> This patch again has a lot of unrelated changes, esp. in\\n>>> deparseSelectStmtForRel(). What was earlier part of deparseSelectSql()\\n>>> and deparseFromExpr() is now flattened in deparseSelectStmtForRel(),\\n>>> which seems unnecessary.\\n\\n>> IIUC, I think this change was done to address your comment (see the comment\\n>> #2 in [1]). Am I missing something?\\n\\n> There is some misunderstanding here. That comment basically says,\\n> deparseRangeTblRef() duplicates code in deparseSelectStmtForRel(), so\\n> we should either remove deparseRangeTblRef() altogether or should keep\\n> it to minimal avoiding duplication of code. What might have confused\\n> you is the last sentence in that comment \\\"This will also make the\\n> current changes to deparseSelectSql unnecessary.\\\" By current changes I\\n> meant changes to deparseSelectSql() in your patch, not the one that's\\n> in the repository. I don't think we should flatten\\n> deparseSelectStmtForRel() in this patch.\\n\\nI noticed that I misunderstood your words. Sorry about that. I agree \\nwith you, so I removed that change from the patch.\\n\\nAttached is an updated version of the patch.\\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\": \"Fri, 4 Nov 2016 19:55:18 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/11/04 19:55, Etsuro Fujita wrote:\\n> Attached is an updated version of the patch.\\n\\nI noticed that I have included an unrelated regression test in the \\npatch. Attached is a patch with the test removed.\\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, 7 Nov 2016 11:24:51 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/11/07 11:24, Etsuro Fujita wrote:\\n> On 2016/11/04 19:55, Etsuro Fujita wrote:\\n>> Attached is an updated version of the patch.\\n\\n> I noticed that I have included an unrelated regression test in the\\n> patch. Attached is a patch with the test removed.\\n\\nI noticed that I inadvertently removed some changes from that patch, so \\nI fixed that. Please find attached an updated version of the patch. \\nI'm also attaching an updated version of another patch for evaluating \\nPHVs remotely, which has been created on top of that patch. Changes to \\nthe latter: I revised comments and added a bit more regression tests.\\n\\nBest regards,\\nEtsuro Fujita\\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, 7 Nov 2016 21:20:49 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 7, 2016 at 5:50 PM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> On 2016/11/07 11:24, Etsuro Fujita wrote:\\n>>\\n>> On 2016/11/04 19:55, Etsuro Fujita wrote:\\n>>>\\n>>> Attached is an updated version of the patch.\\n>\\n>\\n>> I noticed that I have included an unrelated regression test in the\\n>> patch. Attached is a patch with the test removed.\\n>\\n>\\n> I noticed that I inadvertently removed some changes from that patch, so I\\n> fixed that. Please find attached an updated version of the patch. I'm also\\n> attaching an updated version of another patch for evaluating PHVs remotely,\\n> which has been created on top of that patch. Changes to the latter: I\\n> revised comments and added a bit more regression tests.\\n\\nThe patch looks in good shape now. Here are some comments. I have also\\nmade several changes to comments correcting grammar, typos, style and\\nat few places logic. Let me know if the patch looks good.\\n\\nI guess, below code\\n+ if (!fpinfo->subquery_rels)\\n+ return false;\\ncan be changed to\\n if (!bms_is_member(node->varno, fpinfo->subquery_rels))\\n return false;\\nAlso the return values from the recursive calls to isSubqueryExpr() can be\\nreturned as is. I have included this change in the patch.\\n\\ndeparse.c seems to be using capitalized names for function which\\nactually deparse something and an non-capitalized form for helper\\nfunctions. From that perspective attached patch renames isSubqueryExpr\\nas is_subquery_var() and getSubselectAliasInfo() as\\nget_alias_id_for_var(). Actually both these functions accept a Var\\nnode but somehow their names refer to expr.\\n\\nThis patch is using make_tlist_from_pathtarget() to create tlist to be\\ndeparsed but search in RelOptInfo::reltarget::exprs for a given Var.\\nAs long as the relations deparsed do not carry expressions, this might\\nwork, but it will certainly break once we start deparsing relations\\nwith expressions since the upper most query's tlist contains only\\nVars. Instead, we should probably, create tlist and save it in fpinfo\\nand use it later for searching (tlist_member()?). Possibly use using\\nbuild_tlist_to_deparse(), to create the tlist similar so that\\ntargetlist list creation logic is same for all the relations being\\ndeparsed. I haven't included this change in the patch.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database Company\\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, 11 Nov 2016 15:52:39 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/11/11 19:22, Ashutosh Bapat wrote:\\n> The patch looks in good shape now.\\n\\nThanks for the review!\\n\\n> The patch looks in good shape now. Here are some comments. I have also\\n> made several changes to comments correcting grammar, typos, style and\\n> at few places logic. Let me know if the patch looks good.\\n\\nOK, will look into that.\\n\\n> I guess, below code\\n> + if (!fpinfo->subquery_rels)\\n> + return false;\\n> can be changed to\\n> if (!bms_is_member(node->varno, fpinfo->subquery_rels))\\n> return false;\\n> Also the return values from the recursive calls to isSubqueryExpr() can be\\n> returned as is. I have included this change in the patch.\\n\\nWill look into that too.\\n\\n> deparse.c seems to be using capitalized names for function which\\n> actually deparse something and an non-capitalized form for helper\\n> functions.\\n\\nThat's not true. There is a function named classifyConditions(). The \\nfunction naming in deparse.c is a bit arbitrary.\\n\\n> From that perspective attached patch renames isSubqueryExpr\\n> as is_subquery_var() and getSubselectAliasInfo() as\\n> get_alias_id_for_var(). Actually both these functions accept a Var\\n> node but somehow their names refer to expr.\\n\\nThe reason why I named that function isSubqueryExpr is that I think that \\nfunction would be soon extended so as to handle PHVs. See another patch \\nfor evaluating PHVs remotely.\\n\\n> This patch is using make_tlist_from_pathtarget() to create tlist to be\\n> deparsed but search in RelOptInfo::reltarget::exprs for a given Var.\\n> As long as the relations deparsed do not carry expressions, this might\\n> work, but it will certainly break once we start deparsing relations\\n> with expressions since the upper most query's tlist contains only\\n> Vars. Instead, we should probably, create tlist and save it in fpinfo\\n> and use it later for searching (tlist_member()?). Possibly use using\\n> build_tlist_to_deparse(), to create the tlist similar so that\\n> targetlist list creation logic is same for all the relations being\\n> deparsed. I haven't included this change in the patch.\\n\\nSeems like a good idea. Will revise.\\n\\nBest regards,\\nEtsuro Fujita\\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, 11 Nov 2016 20:50:43 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/11/11 20:50, Etsuro Fujita wrote:\\n> On 2016/11/11 19:22, Ashutosh Bapat wrote:\\n>> The patch looks in good shape now. Here are some comments. I have also\\n>> made several changes to comments correcting grammar, typos, style and\\n>> at few places logic. Let me know if the patch looks good.\\n\\n> OK, will look into that.\\n\\nDone. +1 for the changes you made, except for few things; (1) You added \\nthe following comments to deparseSelectSql:\\n\\n+ /*\\n+ * For a non-base relation, use the input tlist. If a base \\nrelation is\\n+ * being deparsed as a subquery, use input tlist, if the caller \\nhas passed\\n+ * one. The column aliases of the subquery are crafted based on \\nthe input\\n+ * tlist. If tlist is NIL, there will not be any expression \\nreferring to\\n+ * any of the columns of the base relation being deparsed. Hence \\nit doesn't\\n+ * matter whether the base relation is being deparsed as \\nsubquery or not.\\n+ */\\n\\nThe last sentence seems confusing to me. My understanding is: if a base \\nrelation has tlist=NIL, then the base relation *must* be deparsed as \\nordinary, not as a subquery. Maybe I'm missing something, though. (I \\nleft that as-is, but I think we need to reword that to be more clear, at \\nleast.)\\n\\n(2) You added the following comments to deparseRangeTblRef:\\n\\n > + * If make_subquery is true, deparse the relation as a subquery. \\nOtherwise,\\n > + * deparse it as relation name with alias.\\n\\nThe second sentence seems confusing to me, too, because it looks like \\nthe relation being deparsed is assumed to be a base relation, but the \\nrelation can be a join relation, which might join base relations, lower \\njoin relations, and/or lower subqueries. So, I modified the sentence a bit.\\n\\n(3) I don't think we need this in isSubqueryExpr, so I removed it from \\nthe patch:\\n\\n+\\t/* Keep compiler happy. */\\n+\\treturn false;\\n\\nAlso, I revised comments you added a little bit.\\n\\n>> I guess, below code\\n>> + if (!fpinfo->subquery_rels)\\n>> + return false;\\n>> can be changed to\\n>> if (!bms_is_member(node->varno, fpinfo->subquery_rels))\\n>> return false;\\n>> Also the return values from the recursive calls to isSubqueryExpr()\\n>> can be\\n>> returned as is. I have included this change in the patch.\\n\\n> Will look into that too.\\n\\nDone. That's a good idea!\\n\\n>> deparse.c seems to be using capitalized names for function which\\n>> actually deparse something and an non-capitalized form for helper\\n>> functions.\\n\\n>> From that perspective attached patch renames isSubqueryExpr\\n>> as is_subquery_var() and getSubselectAliasInfo() as\\n>> get_alias_id_for_var(). Actually both these functions accept a Var\\n>> node but somehow their names refer to expr.\\n\\nOK, I changed isSubqueryExpr to is_subquery_expr; I kept to refer to \\nexpr because I think we would soon extend that function so that it can \\nhandle PHVs, as I said upthread. For getSubselectAliasInfo, I changed \\nthe name to get_subselect_alias_id, because (1) the word \\\"alias\\\" seems \\ngeneral and (2) the \\\"for_var\\\" part would make the name a bit long.\\n\\n>> This patch is using make_tlist_from_pathtarget() to create tlist to be\\n>> deparsed but search in RelOptInfo::reltarget::exprs for a given Var.\\n>> As long as the relations deparsed do not carry expressions, this might\\n>> work, but it will certainly break once we start deparsing relations\\n>> with expressions since the upper most query's tlist contains only\\n>> Vars. Instead, we should probably, create tlist and save it in fpinfo\\n>> and use it later for searching (tlist_member()?). Possibly use using\\n>> build_tlist_to_deparse(), to create the tlist similar so that\\n>> targetlist list creation logic is same for all the relations being\\n>> deparsed. I haven't included this change in the patch.\\n\\n> Seems like a good idea. Will revise.\\n\\nDone. I modified the patch as proposed; create the tlist by \\nbuild_tlist_to_deparse in foreign_join_ok, if needed, and search the \\ntlist by tlist_member. I also added a new member \\\"tlist\\\" to \\nPgFdwRelationInfo to save the tlist created in foreign_join_ok.\\n\\nAnother idea on the \\\"tlist\\\" member would be to save a tlist created for \\nEXPLAIN into that member in estimate_patch_cost_size, so that we don't \\nneed to generate the tlist again in postgresGetForeignPlan, when \\nuse_remote_estimate=true. But I'd like to leave that for another patch.\\n\\nPlease find attached an updated version of the patch.\\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\": \"Tue, 15 Nov 2016 21:23:00 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"Thanks.\\n\\n> except for few things; (1) You added the\\n> following comments to deparseSelectSql:\\n>\\n> + /*\\n> + * For a non-base relation, use the input tlist. If a base relation\\n> is\\n> + * being deparsed as a subquery, use input tlist, if the caller has\\n> passed\\n> + * one. The column aliases of the subquery are crafted based on the\\n> input\\n> + * tlist. If tlist is NIL, there will not be any expression\\n> referring to\\n> + * any of the columns of the base relation being deparsed. Hence it\\n> doesn't\\n> + * matter whether the base relation is being deparsed as subquery or\\n> not.\\n> + */\\n>\\n> The last sentence seems confusing to me. My understanding is: if a base\\n> relation has tlist=NIL, then the base relation *must* be deparsed as\\n> ordinary, not as a subquery. Maybe I'm missing something, though. (I left\\n> that as-is, but I think we need to reword that to be more clear, at least.)\\n>\\n\\nHmm, I agree. I think the comment should just say, \\\"Use tlist to\\ncreate the SELECT clause if one has been provided. For a base relation\\nwith tlist = NIL, check attrs_needed information.\\\". Does that sound\\ngood?\\n\\n> (2) You added the following comments to deparseRangeTblRef:\\n>\\n>> + * If make_subquery is true, deparse the relation as a subquery.\\n>> Otherwise,\\n>> + * deparse it as relation name with alias.\\n>\\n> The second sentence seems confusing to me, too, because it looks like the\\n> relation being deparsed is assumed to be a base relation, but the relation\\n> can be a join relation, which might join base relations, lower join\\n> relations, and/or lower subqueries. So, I modified the sentence a bit.\\n>\\n\\nOK.\\n\\n> (3) I don't think we need this in isSubqueryExpr, so I removed it from the\\n> patch:\\n>\\n> + /* Keep compiler happy. */\\n> + return false;\\n>\\n\\nDoesn't that cause compiler warning, saying \\\"non-void function\\nreturning nothing\\\" or something like that. Actually, the \\\"if\\n(bms_is_member(node->varno, outerrel->relids))\\\" ends with a \\\"return\\\"\\nalways. Hence we don't need to encapsulate the code in \\\"else\\\" block in\\nelse { }. It could be taken outside.\\n\\n\\n>\\n>\\n> OK, I changed isSubqueryExpr to is_subquery_expr; I kept to refer to expr\\n> because I think we would soon extend that function so that it can handle\\n> PHVs, as I said upthread. For getSubselectAliasInfo, I changed the name to\\n> get_subselect_alias_id, because (1) the word \\\"alias\\\" seems general and (2)\\n> the \\\"for_var\\\" part would make the name a bit long.\\n\\nis_subquery_expr(Var *node -- that looks odd. Either it should\\nis_subquery_var(Var * ... OR it should be is_subquery_expr(Expr * ...\\n. I would prefer the first one, since that's what current patch is\\ndoing. When we introduce PHVs, we may change it, if required.\\n\\n\\n>\\n>\\n> Done. I modified the patch as proposed; create the tlist by\\n> build_tlist_to_deparse in foreign_join_ok, if needed, and search the tlist\\n> by tlist_member. I also added a new member \\\"tlist\\\" to PgFdwRelationInfo to\\n> save the tlist created in foreign_join_ok.\\n>\\n\\nInstead of adding a new member, you might want to reuse grouped_tlist\\nby renaming it.\\n\\n> Another idea on the \\\"tlist\\\" member would be to save a tlist created for\\n> EXPLAIN into that member in estimate_patch_cost_size, so that we don't need\\n> to generate the tlist again in postgresGetForeignPlan, when\\n> use_remote_estimate=true. But I'd like to leave that for another patch.\\n\\nI think that will happen automatically, while deparsing, whether for\\nEXPLAIN or for actual execution.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 16 Nov 2016 14:44:46 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/11/16 18:14, Ashutosh Bapat wrote:\\n>> (1) You added the\\n>> following comments to deparseSelectSql:\\n>>\\n>> + /*\\n>> + * For a non-base relation, use the input tlist. If a base relation\\n>> is\\n>> + * being deparsed as a subquery, use input tlist, if the caller has\\n>> passed\\n>> + * one. The column aliases of the subquery are crafted based on the\\n>> input\\n>> + * tlist. If tlist is NIL, there will not be any expression\\n>> referring to\\n>> + * any of the columns of the base relation being deparsed. Hence it\\n>> doesn't\\n>> + * matter whether the base relation is being deparsed as subquery or\\n>> not.\\n>> + */\\n>>\\n>> The last sentence seems confusing to me. My understanding is: if a base\\n>> relation has tlist=NIL, then the base relation *must* be deparsed as\\n>> ordinary, not as a subquery. Maybe I'm missing something, though. (I left\\n>> that as-is, but I think we need to reword that to be more clear, at least.)\\n\\n> Hmm, I agree. I think the comment should just say, \\\"Use tlist to\\n> create the SELECT clause if one has been provided. For a base relation\\n> with tlist = NIL, check attrs_needed information.\\\". Does that sound\\n> good?\\n\\nI don't think that is 100% correct, because (1) tlist can be NIL for a \\njoin relation, you pointed out upthread, but we need to use \\ndeparseExplicitTargetList, so the first sentence is not completely \\ncorrect, and (2) we need to check attrs_needed information not only for \\na baserel but for an otherrel, so the second sentence is not completely \\ncorrect, either. How about this, instead?:\\n\\n /*\\n * For a join relation or an upper relation, use \\ndeparseExplicitTargetList.\\n * Likewise, for a base relation that is being deparsed as a \\nsubquery, in\\n * which case the caller would have passed tlist that is non-NIL, \\nuse that\\n * function. Otherwise, use deparseTargetList.\\n */\\n\\n>> (3) I don't think we need this in isSubqueryExpr, so I removed it from the\\n>> patch:\\n>>\\n>> + /* Keep compiler happy. */\\n>> + return false;\\n\\n> Doesn't that cause compiler warning, saying \\\"non-void function\\n> returning nothing\\\" or something like that. Actually, the \\\"if\\n> (bms_is_member(node->varno, outerrel->relids))\\\" ends with a \\\"return\\\"\\n> always. Hence we don't need to encapsulate the code in \\\"else\\\" block in\\n> else { }. It could be taken outside.\\n\\nYeah, I think so too, but I like the \\\"if () { } else { }\\\" coding. That \\ncoding can be found in other places in core, eg, \\noperator_same_subexprs_lookup.\\n\\n>> OK, I changed isSubqueryExpr to is_subquery_expr; I kept to refer to expr\\n>> because I think we would soon extend that function so that it can handle\\n>> PHVs, as I said upthread. For getSubselectAliasInfo, I changed the name to\\n>> get_subselect_alias_id, because (1) the word \\\"alias\\\" seems general and (2)\\n>> the \\\"for_var\\\" part would make the name a bit long.\\n\\n> is_subquery_expr(Var *node -- that looks odd. Either it should\\n> is_subquery_var(Var * ... OR it should be is_subquery_expr(Expr * ...\\n> . I would prefer the first one, since that's what current patch is\\n> doing. When we introduce PHVs, we may change it, if required.\\n\\nOK, I used is_subquery_var().\\n\\n>> Done. I modified the patch as proposed; create the tlist by\\n>> build_tlist_to_deparse in foreign_join_ok, if needed, and search the tlist\\n>> by tlist_member. I also added a new member \\\"tlist\\\" to PgFdwRelationInfo to\\n>> save the tlist created in foreign_join_ok.\\n\\n> Instead of adding a new member, you might want to reuse grouped_tlist\\n> by renaming it.\\n\\nDone.\\n\\n>> Another idea on the \\\"tlist\\\" member would be to save a tlist created for\\n>> EXPLAIN into that member in estimate_patch_cost_size, so that we don't need\\n>> to generate the tlist again in postgresGetForeignPlan, when\\n>> use_remote_estimate=true. But I'd like to leave that for another patch.\\n\\n> I think that will happen automatically, while deparsing, whether for\\n> EXPLAIN or for actual execution.\\n\\nReally? Anyway, I'd like to leave that as-is.\\n\\nPlease find attached a new version of the patch.\\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\": \"Thu, 17 Nov 2016 14:55:21 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \">\\n> /*\\n> * For a join relation or an upper relation, use\\n> deparseExplicitTargetList.\\n> * Likewise, for a base relation that is being deparsed as a subquery,\\n> in\\n> * which case the caller would have passed tlist that is non-NIL, use\\n> that\\n> * function. Otherwise, use deparseTargetList.\\n> */\\n\\nThis looks correct. I have modified it to make it simple in the given\\npatch. But, I think we shouldn't refer to a function e.g.\\ndeparseExplicitTargetlist() in the comment. Instead we should describe\\nthe intent e.g. \\\"deparse SELECT clause from the given targetlist\\\" or\\n\\\"deparse SELECT clause from attr_needed\\\".\\n\\n>\\n>>> (3) I don't think we need this in isSubqueryExpr, so I removed it from\\n>>> the\\n>>> patch:\\n>>>\\n>>> + /* Keep compiler happy. */\\n>>> + return false;\\n>\\n>\\n>> Doesn't that cause compiler warning, saying \\\"non-void function\\n>> returning nothing\\\" or something like that. Actually, the \\\"if\\n>> (bms_is_member(node->varno, outerrel->relids))\\\" ends with a \\\"return\\\"\\n>> always. Hence we don't need to encapsulate the code in \\\"else\\\" block in\\n>> else { }. It could be taken outside.\\n>\\n>\\n> Yeah, I think so too, but I like the \\\"if () { } else { }\\\" coding. That\\n> coding can be found in other places in core, eg,\\n> operator_same_subexprs_lookup.\\n\\nOK.\\n\\n\\n>\\n>>> Done. I modified the patch as proposed; create the tlist by\\n>>> build_tlist_to_deparse in foreign_join_ok, if needed, and search the\\n>>> tlist\\n>>> by tlist_member. I also added a new member \\\"tlist\\\" to PgFdwRelationInfo\\n>>> to\\n>>> save the tlist created in foreign_join_ok.\\n>\\n>\\n>> Instead of adding a new member, you might want to reuse grouped_tlist\\n>> by renaming it.\\n>\\n>\\n> Done.\\n\\nRight now, we are calculating tlist whether or not the ForeignPath\\nemerges as the cheapest path. Instead we should calculate tlist, the\\nfirst time we need it and then add it to the fpinfo.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database Company\\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, 18 Nov 2016 17:10:18 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"Also another point\\n\\nI guess, this note doesn't add much value in the given context.\\nProbably we should remove it.\\n+ * Note: the tlist would have one-to-one correspondence with the\\n+ * joining relation's reltarget->exprs because (1) the above test\\n+ * on PHVs guarantees that the reltarget->exprs doesn't contain\\n+ * any PHVs and (2) the joining relation's local_conds is NIL.\\n+ * This allows us to search the targetlist entry matching a given\\n+ * Var node from the tlist in get_subselect_alias_id.\\n\\nOn Fri, Nov 18, 2016 at 5:10 PM, Ashutosh Bapat\\n<ashutosh.bapat@enterprisedb.com> wrote:\\n>>\\n>> /*\\n>> * For a join relation or an upper relation, use\\n>> deparseExplicitTargetList.\\n>> * Likewise, for a base relation that is being deparsed as a subquery,\\n>> in\\n>> * which case the caller would have passed tlist that is non-NIL, use\\n>> that\\n>> * function. Otherwise, use deparseTargetList.\\n>> */\\n>\\n> This looks correct. I have modified it to make it simple in the given\\n> patch. But, I think we shouldn't refer to a function e.g.\\n> deparseExplicitTargetlist() in the comment. Instead we should describe\\n> the intent e.g. \\\"deparse SELECT clause from the given targetlist\\\" or\\n> \\\"deparse SELECT clause from attr_needed\\\".\\n>\\n>>\\n>>>> (3) I don't think we need this in isSubqueryExpr, so I removed it from\\n>>>> the\\n>>>> patch:\\n>>>>\\n>>>> + /* Keep compiler happy. */\\n>>>> + return false;\\n>>\\n>>\\n>>> Doesn't that cause compiler warning, saying \\\"non-void function\\n>>> returning nothing\\\" or something like that. Actually, the \\\"if\\n>>> (bms_is_member(node->varno, outerrel->relids))\\\" ends with a \\\"return\\\"\\n>>> always. Hence we don't need to encapsulate the code in \\\"else\\\" block in\\n>>> else { }. It could be taken outside.\\n>>\\n>>\\n>> Yeah, I think so too, but I like the \\\"if () { } else { }\\\" coding. That\\n>> coding can be found in other places in core, eg,\\n>> operator_same_subexprs_lookup.\\n>\\n> OK.\\n>\\n>\\n>>\\n>>>> Done. I modified the patch as proposed; create the tlist by\\n>>>> build_tlist_to_deparse in foreign_join_ok, if needed, and search the\\n>>>> tlist\\n>>>> by tlist_member. I also added a new member \\\"tlist\\\" to PgFdwRelationInfo\\n>>>> to\\n>>>> save the tlist created in foreign_join_ok.\\n>>\\n>>\\n>>> Instead of adding a new member, you might want to reuse grouped_tlist\\n>>> by renaming it.\\n>>\\n>>\\n>> Done.\\n>\\n> Right now, we are calculating tlist whether or not the ForeignPath\\n> emerges as the cheapest path. Instead we should calculate tlist, the\\n> first time we need it and then add it to the fpinfo.\\n>\\n> --\\n> Best Wishes,\\n> Ashutosh Bapat\\n> EnterpriseDB Corporation\\n> The Postgres Database Company\\n\\n\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 18 Nov 2016 20:46:00 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/11/19 0:16, Ashutosh Bapat wrote:\\n> Also another point\\n>\\n> I guess, this note doesn't add much value in the given context.\\n> Probably we should remove it.\\n> + * Note: the tlist would have one-to-one correspondence with the\\n> + * joining relation's reltarget->exprs because (1) the above test\\n> + * on PHVs guarantees that the reltarget->exprs doesn't contain\\n> + * any PHVs and (2) the joining relation's local_conds is NIL.\\n> + * This allows us to search the targetlist entry matching a given\\n> + * Var node from the tlist in get_subselect_alias_id.\\n\\nOK, I removed.\\n\\n> On Fri, Nov 18, 2016 at 5:10 PM, Ashutosh Bapat\\n> <ashutosh.bapat@enterprisedb.com> wrote:\\n\\nI wrote:\\n>>> /*\\n>>> * For a join relation or an upper relation, use\\n>>> deparseExplicitTargetList.\\n>>> * Likewise, for a base relation that is being deparsed as a subquery,\\n>>> in\\n>>> * which case the caller would have passed tlist that is non-NIL, use\\n>>> that\\n>>> * function. Otherwise, use deparseTargetList.\\n>>> */\\n\\n>> This looks correct. I have modified it to make it simple in the given\\n>> patch. But, I think we shouldn't refer to a function e.g.\\n>> deparseExplicitTargetlist() in the comment. Instead we should describe\\n>> the intent e.g. \\\"deparse SELECT clause from the given targetlist\\\" or\\n>> \\\"deparse SELECT clause from attr_needed\\\".\\n\\nMy taste would be to refer to those functions, because ISTM that makes \\nthe explanation more simple and direct. So, I'd like to leave that for \\nthe committer's judge.\\n\\nI wrote:\\n>>>>> Done. I modified the patch as proposed; create the tlist by\\n>>>>> build_tlist_to_deparse in foreign_join_ok, if needed, and search the\\n>>>>> tlist\\n>>>>> by tlist_member. I also added a new member \\\"tlist\\\" to PgFdwRelationInfo\\n>>>>> to\\n>>>>> save the tlist created in foreign_join_ok.\\n\\nYou wrote:\\n>>>> Instead of adding a new member, you might want to reuse grouped_tlist\\n>>>> by renaming it.\\n\\n>>> Done.\\n\\n>> Right now, we are calculating tlist whether or not the ForeignPath\\n>> emerges as the cheapest path.\\n\\nYeah, I modified the patch so, as I thought that would be consistent \\nwith the aggregate pushdown patch.\\n\\n>> Instead we should calculate tlist, the\\n>> first time we need it and then add it to the fpinfo.\\n\\nHaving said that, I agree on that point. I'd like to propose (1) adding \\na new member to fpinfo, to store a list of output Vars from the \\nsubquery, and (2) creating a tlist from it in deparseRangeTblRef, then, \\nwhich would allow us to calculate the tlist only when we need it. The \\nmember added to fpinfo would be also useful to address the comments on \\nthe DML/UPDATE pushdown patch. See the attached patch in [1]. I named \\nthe member a bit differently in that patch, though.\\n\\nYou modified the comments I added to deparseLockingClause into this:\\n\\n /*\\n+ * Ignore relation if it appears in a lower subquery. Locking clause\\n+ * for such a relation is included in the subquery.\\n+ */\\n\\nI don't think the second sentence is 100% correct because a locking \\nclause isn't always required for such a relation, so I modified the \\nsentence a bit.\\n\\nBest regards,\\nEtsuro Fujita\\n\\n[1] \\nhttps://www.postgresql.org/message-id/38245b84-fabf-0899-1b24-8f94cdc5900c%40lab.ntt.co.jp\\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, 21 Nov 2016 21:15:49 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \">\\n> Yeah, I modified the patch so, as I thought that would be consistent with\\n> the aggregate pushdown patch.\\n\\naggregate pushdown needs the tlist to judge the shippability of\\ntargetlist. For joins that's not required, so we should defer, if we\\ncan.\\n\\n>\\n>>> Instead we should calculate tlist, the\\n>>> first time we need it and then add it to the fpinfo.\\n>\\n>\\n> Having said that, I agree on that point. I'd like to propose (1) adding a\\n> new member to fpinfo, to store a list of output Vars from the subquery, and\\n> (2) creating a tlist from it in deparseRangeTblRef, then, which would allow\\n> us to calculate the tlist only when we need it. The member added to fpinfo\\n> would be also useful to address the comments on the DML/UPDATE pushdown\\n> patch. See the attached patch in [1]. I named the member a bit differently\\n> in that patch, though.\\n\\nAgain the list of Vars will be wasted if we don't choose that path for\\nfinal planning. So, I don't see the point of adding list of Vars\\nthere. It looks like we are replacing one list with the other when\\nnone of those are useful, if the path doesn't get chosen for the final\\nplan. If you think that the member is useful for DML/UDPATE pushdown,\\nyou may want to add it in the other patch.\\n\\n>\\n> You modified the comments I added to deparseLockingClause into this:\\n>\\n> /*\\n> + * Ignore relation if it appears in a lower subquery. Locking clause\\n> + * for such a relation is included in the subquery.\\n> + */\\n>\\n> I don't think the second sentence is 100% correct because a locking clause\\n> isn't always required for such a relation, so I modified the sentence a bit.\\n>\\n\\nI guess, \\\"if required\\\" part was implicit in construct \\\"such a\\nrelation\\\". Your version seems to make it explicit. I am fine with it.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 21 Nov 2016 18:32:12 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/11/21 22:02, Ashutosh Bapat wrote:\\n\\nYou wrote:\\n>>>> Instead we should calculate tlist, the\\n>>>> first time we need it and then add it to the fpinfo.\\n\\nI wrote:\\n>> Having said that, I agree on that point. I'd like to propose (1) adding a\\n>> new member to fpinfo, to store a list of output Vars from the subquery, and\\n>> (2) creating a tlist from it in deparseRangeTblRef, then, which would allow\\n>> us to calculate the tlist only when we need it. The member added to fpinfo\\n>> would be also useful to address the comments on the DML/UPDATE pushdown\\n>> patch. See the attached patch in [1]. I named the member a bit differently\\n>> in that patch, though.\\n\\n> Again the list of Vars will be wasted if we don't choose that path for\\n> final planning. So, I don't see the point of adding list of Vars\\n> there.\\n\\n> If you think that the member is useful for DML/UDPATE pushdown,\\n> you may want to add it in the other patch.\\n\\nOK, I'd like to propose referencing to foreignrel->reltarget->exprs \\ndirectly in deparseRangeTblRef and get_subselect_alias_id, then, which \\nis the same as what I proposed in the first version of the patch, but \\nI'd also like to change deparseRangeTblRef to use add_to_flat_tlist, not \\nmake_tlist_from_pathtarget, to create a tlist of the subquery, as you \\nproposed.\\n\\n>> You modified the comments I added to deparseLockingClause into this:\\n>>\\n>> /*\\n>> + * Ignore relation if it appears in a lower subquery. Locking clause\\n>> + * for such a relation is included in the subquery.\\n>> + */\\n>>\\n>> I don't think the second sentence is 100% correct because a locking clause\\n>> isn't always required for such a relation, so I modified the sentence a bit.\\n\\n> I guess, \\\"if required\\\" part was implicit in construct \\\"such a\\n> relation\\\". Your version seems to make it explicit. I am fine with it.\\n\\nOK, let's leave that for the committer's judge.\\n\\nPlease find attached an updated version of the patch.\\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\": \"Tue, 22 Nov 2016 15:55:14 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"The comments should explain why is the assertion true.\\n+ /* Shouldn't be NIL */\\n+ Assert(tlist != NIL);\\n+ /* Should be same length */\\n+ Assert(list_length(tlist) ==\\nlist_length(foreignrel->reltarget->exprs));\\n\\n\\n>\\n> OK, I'd like to propose referencing to foreignrel->reltarget->exprs directly\\n> in deparseRangeTblRef and get_subselect_alias_id, then, which is the same as\\n> what I proposed in the first version of the patch, but I'd also like to\\n> change deparseRangeTblRef to use add_to_flat_tlist, not\\n> make_tlist_from_pathtarget, to create a tlist of the subquery, as you\\n> proposed.\\n\\nThere is a already a function to build targetlist for a given relation\\nbuild_tlist_to_deparse(), which does the same thing as this code for a join or\\nbase relation and when there are no local conditions. Why don't we use that\\nfunction instead of duplicating that logic? If tomorrow, the logic to build the\\ntargetlist changes, we will need to duplicate those changes. We should avoid\\nthat.\\n+ /* Build a tlist from the subquery. */\\n+ tlist = add_to_flat_tlist(tlist, foreignrel->reltarget->exprs);\\n\\nThe comment below says \\\"get the aliases\\\", but what the function really returns\\nis the identifiers used for creating aliases. Please correct the comment.\\n+/*\\n+ * Get the relation and column alias for a given Var node, which belongs to\\n+ * input foreignrel. They are returned in *tabno and *colno respectively.\\n+ */\\n\\nWe discussed that we have to deparse and search from the same targetlist. And\\nthat the targetlist should be saved in fpinfo, the first time it gets created.\\nBut the patch seems to be searching in foreignrel->reltarget->exprs and\\ndeparsing from the tlist returned by add_to_flat_tlist(tlist,\\nforeignrel->reltarget->exprs).\\n+ foreach(lc, foreignrel->reltarget->exprs)\\n+ {\\n+ if (equal(lfirst(lc), (Node *) node))\\n+ {\\n+ *colno = i;\\n+ return;\\n+ }\\n+ i++;\\n+ }\\nI guess, the reason why you are doing it this way, is SELECT clause for the\\noutermost query gets deparsed before FROM clause. For later we call\\ndeparseRangeTblRef(), which builds the tlist. So, while deparsing SELECT\\nclause, we do not have tlist to build from. In that case, I guess, we have to\\nbuild the tlist in get_subselect_alias_id() if it's not available and stick it\\nin fpinfo. Subsequent calls to get_subselect_alias_id() should find tlist\\nthere. Then in deparseRangeTblRef() assert that there's a tlist in fpinfo\\nand use it to build the SELECT clause of subquery. That way, we don't build\\ntlist unless it's needed and also use the same tlist for all searches. Please\\nuse tlist_member() to search into the tlist.\\n\\nThe name get_subselect_alias_id() seems to suggest that the function returns\\nidentifier for subselect alias, which isn't correct. It returns the alias\\nidentifiers for deparsing a Var node. But I guess, we have left this to the\\ncommitter's judgement.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 22 Nov 2016 14:58:50 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/11/22 18:28, Ashutosh Bapat wrote:\\n> The comments should explain why is the assertion true.\\n> + /* Shouldn't be NIL */\\n> + Assert(tlist != NIL);\\n> + /* Should be same length */\\n> + Assert(list_length(tlist) ==\\n> list_length(foreignrel->reltarget->exprs));\\n\\nWill revise.\\n\\n>> OK, I'd like to propose referencing to foreignrel->reltarget->exprs directly\\n>> in deparseRangeTblRef and get_subselect_alias_id, then, which is the same as\\n>> what I proposed in the first version of the patch, but I'd also like to\\n>> change deparseRangeTblRef to use add_to_flat_tlist, not\\n>> make_tlist_from_pathtarget, to create a tlist of the subquery, as you\\n>> proposed.\\n\\n> There is a already a function to build targetlist for a given relation\\n> build_tlist_to_deparse(), which does the same thing as this code for a join or\\n> base relation and when there are no local conditions. Why don't we use that\\n> function instead of duplicating that logic? If tomorrow, the logic to build the\\n> targetlist changes, we will need to duplicate those changes. We should avoid\\n> that.\\n> + /* Build a tlist from the subquery. */\\n> + tlist = add_to_flat_tlist(tlist, foreignrel->reltarget->exprs);\\n\\nWill modify the patch to use build_tlist_to_deparse. Actually, the \\nearly versions of the patch used that function, but it looks like I \\nchanged not to use that function, as I misunderstood about your comments \\non this part at some point. Sorry for that.\\n\\n> The comment below says \\\"get the aliases\\\", but what the function really returns\\n> is the identifiers used for creating aliases. Please correct the comment.\\n> +/*\\n> + * Get the relation and column alias for a given Var node, which belongs to\\n> + * input foreignrel. They are returned in *tabno and *colno respectively.\\n> + */\\n\\nActually, this was rewritten into the above by *you*. The original \\ncomment I added was:\\n\\n+ /*\\n+ * Get info about the subselect alias to given expression.\\n+ *\\n+ * The subselect table and column numbers are returned to *tabno and \\n*colno,\\n+ * respectively.\\n+ */\\n\\nI'd like to change the comment into something like the original one.\\n\\n> We discussed that we have to deparse and search from the same targetlist. And\\n> that the targetlist should be saved in fpinfo, the first time it gets created.\\n> But the patch seems to be searching in foreignrel->reltarget->exprs and\\n> deparsing from the tlist returned by add_to_flat_tlist(tlist,\\n> foreignrel->reltarget->exprs).\\n> + foreach(lc, foreignrel->reltarget->exprs)\\n> + {\\n> + if (equal(lfirst(lc), (Node *) node))\\n> + {\\n> + *colno = i;\\n> + return;\\n> + }\\n> + i++;\\n> + }\\n> I guess, the reason why you are doing it this way, is SELECT clause for the\\n> outermost query gets deparsed before FROM clause. For later we call\\n> deparseRangeTblRef(), which builds the tlist. So, while deparsing SELECT\\n> clause, we do not have tlist to build from.\\n\\nThat's right.\\n\\n> In that case, I guess, we have to\\n> build the tlist in get_subselect_alias_id() if it's not available and stick it\\n> in fpinfo. Subsequent calls to get_subselect_alias_id() should find tlist\\n> there. Then in deparseRangeTblRef() assert that there's a tlist in fpinfo\\n> and use it to build the SELECT clause of subquery. That way, we don't build\\n> tlist unless it's needed and also use the same tlist for all searches. Please\\n> use tlist_member() to search into the tlist.\\n\\nI don't think that's a good idea because that would make the code \\nunnecessarily complicated and inefficient. I think the direct search \\ninto the foreignrel->reltarget->exprs shown above would be better \\nbecause that's more simple and efficient than what you proposed. Note \\nthat since (1) the foreignrel->reltarget->exprs doesn't contain any PHVs \\nand (2) the local_conds is empty, the tlist created for the subquery by \\nbuild_tlist_to_deparse is guaranteed to have one-to-one correspondence \\nwith the foreignrel->reltarget->exprs, so the direct search works well.\\n\\n> The name get_subselect_alias_id() seems to suggest that the function returns\\n> identifier for subselect alias, which isn't correct. It returns the alias\\n> identifiers for deparsing a Var node. But I guess, we have left this to the\\n> committer's judgement.\\n\\nFine with me.\\n\\nBest regards,\\nEtsuro Fujita\\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, 22 Nov 2016 20:19:04 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \">\\n>\\n>> There is a already a function to build targetlist for a given relation\\n>> build_tlist_to_deparse(), which does the same thing as this code for a\\n>> join or\\n>> base relation and when there are no local conditions. Why don't we use\\n>> that\\n>> function instead of duplicating that logic? If tomorrow, the logic to\\n>> build the\\n>> targetlist changes, we will need to duplicate those changes. We should\\n>> avoid\\n>> that.\\n>> + /* Build a tlist from the subquery. */\\n>> + tlist = add_to_flat_tlist(tlist, foreignrel->reltarget->exprs);\\n>\\n>\\n> Will modify the patch to use build_tlist_to_deparse. Actually, the early\\n> versions of the patch used that function, but it looks like I changed not to\\n> use that function, as I misunderstood about your comments on this part at\\n> some point. Sorry for that.\\n>\\n>> The comment below says \\\"get the aliases\\\", but what the function really\\n>> returns\\n>> is the identifiers used for creating aliases. Please correct the comment.\\n>> +/*\\n>> + * Get the relation and column alias for a given Var node, which belongs\\n>> to\\n>> + * input foreignrel. They are returned in *tabno and *colno respectively.\\n>> + */\\n>\\n>\\n> Actually, this was rewritten into the above by *you*. The original comment\\n> I added was:\\n>\\n> + /*\\n> + * Get info about the subselect alias to given expression.\\n> + *\\n> + * The subselect table and column numbers are returned to *tabno and\\n> *colno,\\n> + * respectively.\\n> + */\\n>\\n> I'd like to change the comment into something like the original one.\\n\\nSorry. I think the current version is better than previous one. The\\nterm \\\"subselect alias\\\" is confusing in the previous version. In the\\ncurrent version, \\\"Get the relation and column alias for a given Var\\nnode,\\\" we need to add word \\\"identifiers\\\" like \\\"Get the relation and\\ncolumn identifiers for a given Var node\\\".\\n\\n>\\n>> We discussed that we have to deparse and search from the same targetlist.\\n>> And\\n>> that the targetlist should be saved in fpinfo, the first time it gets\\n>> created.\\n>> But the patch seems to be searching in foreignrel->reltarget->exprs and\\n>> deparsing from the tlist returned by add_to_flat_tlist(tlist,\\n>> foreignrel->reltarget->exprs).\\n>> + foreach(lc, foreignrel->reltarget->exprs)\\n>> + {\\n>> + if (equal(lfirst(lc), (Node *) node))\\n>> + {\\n>> + *colno = i;\\n>> + return;\\n>> + }\\n>> + i++;\\n>> + }\\n>> I guess, the reason why you are doing it this way, is SELECT clause for\\n>> the\\n>> outermost query gets deparsed before FROM clause. For later we call\\n>> deparseRangeTblRef(), which builds the tlist. So, while deparsing SELECT\\n>> clause, we do not have tlist to build from.\\n>\\n>\\n> That's right.\\n>\\n>> In that case, I guess, we have to\\n>> build the tlist in get_subselect_alias_id() if it's not available and\\n>> stick it\\n>> in fpinfo. Subsequent calls to get_subselect_alias_id() should find tlist\\n>> there. Then in deparseRangeTblRef() assert that there's a tlist in fpinfo\\n>> and use it to build the SELECT clause of subquery. That way, we don't\\n>> build\\n>> tlist unless it's needed and also use the same tlist for all searches.\\n>> Please\\n>> use tlist_member() to search into the tlist.\\n>\\n>\\n> I don't think that's a good idea because that would make the code\\n> unnecessarily complicated and inefficient. I think the direct search into\\n> the foreignrel->reltarget->exprs shown above would be better because that's\\n> more simple and efficient than what you proposed. Note that since (1) the\\n> foreignrel->reltarget->exprs doesn't contain any PHVs and (2) the\\n> local_conds is empty, the tlist created for the subquery by\\n> build_tlist_to_deparse is guaranteed to have one-to-one correspondence with\\n> the foreignrel->reltarget->exprs, so the direct search works well.\\n\\nIf we deparse from and search into different objects, that's not very\\nmaintainable code. Changes to any one of them require changes to the\\nother, thus creating avenues for bugs. Even if slighly expensive, we\\nshould search into and deparse from the same targetlist. I think I\\nhave explained this before.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 22 Nov 2016 21:00:30 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/11/23 0:30, Ashutosh Bapat wrote:\\n\\nYou wrote:\\n>>> The comment below says \\\"get the aliases\\\", but what the function really\\n>>> returns\\n>>> is the identifiers used for creating aliases. Please correct the comment.\\n>>> +/*\\n>>> + * Get the relation and column alias for a given Var node, which belongs\\n>>> to\\n>>> + * input foreignrel. They are returned in *tabno and *colno respectively.\\n>>> + */\\n\\nI wrote:\\n>> Actually, this was rewritten into the above by *you*. The original comment\\n>> I added was:\\n>>\\n>> + /*\\n>> + * Get info about the subselect alias to given expression.\\n>> + *\\n>> + * The subselect table and column numbers are returned to *tabno and\\n>> *colno,\\n>> + * respectively.\\n>> + */\\n>>\\n>> I'd like to change the comment into something like the original one.\\n\\n> Sorry. I think the current version is better than previous one. The\\n> term \\\"subselect alias\\\" is confusing in the previous version. In the\\n> current version, \\\"Get the relation and column alias for a given Var\\n> node,\\\" we need to add word \\\"identifiers\\\" like \\\"Get the relation and\\n> column identifiers for a given Var node\\\".\\n\\nOK, but one thing I'm concerned about is the term \\\"relation alias\\\" seems \\na bit confusing because we already used the term for the alias of the \\nform \\\"rN\\\". To avoid that, how about saying \\\"table alias\\\", not \\\"relation \\nalias\\\"? (in which case, the comment would be something like \\\"Get the \\ntable and column identifiers for a given Var node\\\".)\\n\\n>>> We discussed that we have to deparse and search from the same targetlist.\\n>>> And\\n>>> that the targetlist should be saved in fpinfo, the first time it gets\\n>>> created.\\n>>> But the patch seems to be searching in foreignrel->reltarget->exprs and\\n>>> deparsing from the tlist returned by add_to_flat_tlist(tlist,\\n>>> foreignrel->reltarget->exprs).\\n>>> + foreach(lc, foreignrel->reltarget->exprs)\\n>>> + {\\n>>> + if (equal(lfirst(lc), (Node *) node))\\n>>> + {\\n>>> + *colno = i;\\n>>> + return;\\n>>> + }\\n>>> + i++;\\n>>> + }\\n>>> I guess, the reason why you are doing it this way, is SELECT clause for\\n>>> the\\n>>> outermost query gets deparsed before FROM clause. For later we call\\n>>> deparseRangeTblRef(), which builds the tlist. So, while deparsing SELECT\\n>>> clause, we do not have tlist to build from.\\n\\n>> That's right.\\n\\n>>> In that case, I guess, we have to\\n>>> build the tlist in get_subselect_alias_id() if it's not available and\\n>>> stick it\\n>>> in fpinfo. Subsequent calls to get_subselect_alias_id() should find tlist\\n>>> there. Then in deparseRangeTblRef() assert that there's a tlist in fpinfo\\n>>> and use it to build the SELECT clause of subquery. That way, we don't\\n>>> build\\n>>> tlist unless it's needed and also use the same tlist for all searches.\\n>>> Please\\n>>> use tlist_member() to search into the tlist.\\n\\n>> I don't think that's a good idea because that would make the code\\n>> unnecessarily complicated and inefficient. I think the direct search into\\n>> the foreignrel->reltarget->exprs shown above would be better because that's\\n>> more simple and efficient than what you proposed. Note that since (1) the\\n>> foreignrel->reltarget->exprs doesn't contain any PHVs and (2) the\\n>> local_conds is empty, the tlist created for the subquery by\\n>> build_tlist_to_deparse is guaranteed to have one-to-one correspondence with\\n>> the foreignrel->reltarget->exprs, so the direct search works well.\\n\\n> If we deparse from and search into different objects, that's not very\\n> maintainable code. Changes to any one of them require changes to the\\n> other, thus creating avenues for bugs. Even if slighly expensive, we\\n> should search into and deparse from the same targetlist.\\n\\nI don't think so; in the current version, we essentially deparse from \\nand search into the same object, ie, foreignrel->reltarget->exprs, since \\nthe tlist created by build_tlist_to_deparse is build from the \\nforeignrel->reltarget->exprs. Also, since it is just created for \\ndeparsing the relation as a subquery in deparseRangeTblRef and isn't \\nstored into fpinfo or anywhere alse, we would need no concern about \\ncreating such avenues. IMO I think adding comments would be enough for \\nthis. Anyway, I think this is an optional issue, so I'd like to leave \\nthis for the committer's judge.\\n\\n> I think I\\n> have explained this before.\\n\\nMy apologies for having misunderstood your words.\\n\\nBest regards,\\nEtsuro Fujita\\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, 24 Nov 2016 16:30:24 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \">\\n>> Sorry. I think the current version is better than previous one. The\\n>> term \\\"subselect alias\\\" is confusing in the previous version. In the\\n>> current version, \\\"Get the relation and column alias for a given Var\\n>> node,\\\" we need to add word \\\"identifiers\\\" like \\\"Get the relation and\\n>> column identifiers for a given Var node\\\".\\n>\\n>\\n> OK, but one thing I'm concerned about is the term \\\"relation alias\\\" seems a\\n> bit confusing because we already used the term for the alias of the form\\n> \\\"rN\\\". To avoid that, how about saying \\\"table alias\\\", not \\\"relation alias\\\"?\\n> (in which case, the comment would be something like \\\"Get the table and\\n> column identifiers for a given Var node\\\".)\\n\\ntable will be misleading as subquery can represent a join and\\ncorresponding alias would represent the join. Relation is better term.\\n\\n\\n>\\n>\\n>> If we deparse from and search into different objects, that's not very\\n>> maintainable code. Changes to any one of them require changes to the\\n>> other, thus creating avenues for bugs. Even if slighly expensive, we\\n>> should search into and deparse from the same targetlist.\\n>\\n>\\n> I don't think so; in the current version, we essentially deparse from and\\n> search into the same object, ie, foreignrel->reltarget->exprs, since the\\n> tlist created by build_tlist_to_deparse is build from the\\n> foreignrel->reltarget->exprs. Also, since it is just created for deparsing\\n> the relation as a subquery in deparseRangeTblRef and isn't stored into\\n> fpinfo or anywhere alse, we would need no concern about creating such\\n> avenues. IMO I think adding comments would be enough for this.\\n\\nbuild_tlist_to_depase() calls pull_var_nodes() before creating the\\ntlist, whereas the code that searches does not do that. Code-wise\\nthose are not the same things.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 24 Nov 2016 13:16:14 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/11/24 16:46, Ashutosh Bapat wrote:\\n>>> Sorry. I think the current version is better than previous one. The\\n>>> term \\\"subselect alias\\\" is confusing in the previous version. In the\\n>>> current version, \\\"Get the relation and column alias for a given Var\\n>>> node,\\\" we need to add word \\\"identifiers\\\" like \\\"Get the relation and\\n>>> column identifiers for a given Var node\\\".\\n\\nI wrote:\\n>> OK, but one thing I'm concerned about is the term \\\"relation alias\\\" seems a\\n>> bit confusing because we already used the term for the alias of the form\\n>> \\\"rN\\\". To avoid that, how about saying \\\"table alias\\\", not \\\"relation alias\\\"?\\n>> (in which case, the comment would be something like \\\"Get the table and\\n>> column identifiers for a given Var node\\\".)\\n\\n> table will be misleading as subquery can represent a join and\\n> corresponding alias would represent the join. Relation is better term.\\n\\nBut the documentation uses the word \\\"table\\\" for a join relation. See \\n7.2.1.2. Table and Column Aliases:\\nhttps://www.postgresql.org/docs/9.6/static/queries-table-expressions.html#QUERIES-TABLE-ALIASES\\n\\n>> I don't think so; in the current version, we essentially deparse from and\\n>> search into the same object, ie, foreignrel->reltarget->exprs, since the\\n>> tlist created by build_tlist_to_deparse is build from the\\n>> foreignrel->reltarget->exprs. Also, since it is just created for deparsing\\n>> the relation as a subquery in deparseRangeTblRef and isn't stored into\\n>> fpinfo or anywhere alse, we would need no concern about creating such\\n>> avenues. IMO I think adding comments would be enough for this.\\n\\n> build_tlist_to_depase() calls pull_var_nodes() before creating the\\n> tlist, whereas the code that searches does not do that. Code-wise\\n> those are not the same things.\\n\\nYou missed the point; the foreignrel->reltarget->exprs doesn't contain \\nany PHVs, so the tlist created by build_tlist_to_depase will be \\nguaranteed to be one-to-one with the foreignrel->reltarget->exprs.\\n\\nBest regards,\\nEtsuro Fujita\\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, 24 Nov 2016 16:57:20 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 24, 2016 at 1:27 PM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> On 2016/11/24 16:46, Ashutosh Bapat wrote:\\n>>>>\\n>>>> Sorry. I think the current version is better than previous one. The\\n>>>> term \\\"subselect alias\\\" is confusing in the previous version. In the\\n>>>> current version, \\\"Get the relation and column alias for a given Var\\n>>>> node,\\\" we need to add word \\\"identifiers\\\" like \\\"Get the relation and\\n>>>> column identifiers for a given Var node\\\".\\n>\\n>\\n> I wrote:\\n>>>\\n>>> OK, but one thing I'm concerned about is the term \\\"relation alias\\\" seems\\n>>> a\\n>>> bit confusing because we already used the term for the alias of the form\\n>>> \\\"rN\\\". To avoid that, how about saying \\\"table alias\\\", not \\\"relation\\n>>> alias\\\"?\\n>>> (in which case, the comment would be something like \\\"Get the table and\\n>>> column identifiers for a given Var node\\\".)\\n>\\n>\\n>> table will be misleading as subquery can represent a join and\\n>> corresponding alias would represent the join. Relation is better term.\\n>\\n>\\n> But the documentation uses the word \\\"table\\\" for a join relation. See\\n> 7.2.1.2. Table and Column Aliases:\\n> https://www.postgresql.org/docs/9.6/static/queries-table-expressions.html#QUERIES-TABLE-ALIASES\\n\\nThe definition there says \\\"A temporary name can be given to tables and\\ncomplex table references to be used for references to the derived\\ntable in the rest of the query. This is called a table alias.\\\", please\\nnote the usage \\\"complex table references\\\". Within the code, we do not\\nrefer to a relation as table. So, the comment in this code should not\\nrefer to a relation as table either.\\n\\n>\\n>>> I don't think so; in the current version, we essentially deparse from and\\n>>> search into the same object, ie, foreignrel->reltarget->exprs, since the\\n>>> tlist created by build_tlist_to_deparse is build from the\\n>>> foreignrel->reltarget->exprs. Also, since it is just created for\\n>>> deparsing\\n>>> the relation as a subquery in deparseRangeTblRef and isn't stored into\\n>>> fpinfo or anywhere alse, we would need no concern about creating such\\n>>> avenues. IMO I think adding comments would be enough for this.\\n>\\n>\\n>> build_tlist_to_depase() calls pull_var_nodes() before creating the\\n>> tlist, whereas the code that searches does not do that. Code-wise\\n>> those are not the same things.\\n>\\n>\\n> You missed the point; the foreignrel->reltarget->exprs doesn't contain any\\n> PHVs, so the tlist created by build_tlist_to_depase will be guaranteed to be\\n> one-to-one with the foreignrel->reltarget->exprs.\\n\\nIt's guaranteed now, but can not be forever. This means anybody who\\nsupports PHV tomorrow, will have to \\\"remember\\\" to update this code as\\nwell. If s/he misses it, a bug will be introduced. That's the avenue\\nfor bug, I am talking about.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 24 Nov 2016 14:09:27 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/11/24 17:39, Ashutosh Bapat wrote:\\n> On Thu, Nov 24, 2016 at 1:27 PM, Etsuro Fujita\\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>> On 2016/11/24 16:46, Ashutosh Bapat wrote:\\n>>> table will be misleading as subquery can represent a join and\\n>>> corresponding alias would represent the join. Relation is better term.\\n\\n>> But the documentation uses the word \\\"table\\\" for a join relation. See\\n>> 7.2.1.2. Table and Column Aliases:\\n>> https://www.postgresql.org/docs/9.6/static/queries-table-expressions.html#QUERIES-TABLE-ALIASES\\n\\n> The definition there says \\\"A temporary name can be given to tables and\\n> complex table references to be used for references to the derived\\n> table in the rest of the query. This is called a table alias.\\\", please\\n> note the usage \\\"complex table references\\\". Within the code, we do not\\n> refer to a relation as table. So, the comment in this code should not\\n> refer to a relation as table either.\\n\\nOK, will keep using the term \\\"relation\\\".\\n\\nI wrote:\\n>>>> I don't think so; in the current version, we essentially deparse from and\\n>>>> search into the same object, ie, foreignrel->reltarget->exprs, since the\\n>>>> tlist created by build_tlist_to_deparse is build from the\\n>>>> foreignrel->reltarget->exprs. Also, since it is just created for\\n>>>> deparsing\\n>>>> the relation as a subquery in deparseRangeTblRef and isn't stored into\\n>>>> fpinfo or anywhere alse, we would need no concern about creating such\\n>>>> avenues. IMO I think adding comments would be enough for this.\\n\\n>>> build_tlist_to_depase() calls pull_var_nodes() before creating the\\n>>> tlist, whereas the code that searches does not do that. Code-wise\\n>>> those are not the same things.\\n\\n>> You missed the point; the foreignrel->reltarget->exprs doesn't contain any\\n>> PHVs, so the tlist created by build_tlist_to_depase will be guaranteed to be\\n>> one-to-one with the foreignrel->reltarget->exprs.\\n\\n> It's guaranteed now, but can not be forever. This means anybody who\\n> supports PHV tomorrow, will have to \\\"remember\\\" to update this code as\\n> well. If s/he misses it, a bug will be introduced. That's the avenue\\n> for bug, I am talking about.\\n\\nIt *can* be guaranteed. See another patch for supporting evaluating \\nPHVs remotely.\\n\\nBest regards,\\nEtsuro Fujita\\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, 24 Nov 2016 17:51:50 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \">\\n>\\n>>>> build_tlist_to_depase() calls pull_var_nodes() before creating the\\n>>>> tlist, whereas the code that searches does not do that. Code-wise\\n>>>> those are not the same things.\\n>\\n>\\n>>> You missed the point; the foreignrel->reltarget->exprs doesn't contain\\n>>> any\\n>>> PHVs, so the tlist created by build_tlist_to_depase will be guaranteed to\\n>>> be\\n>>> one-to-one with the foreignrel->reltarget->exprs.\\n>\\n>\\n>> It's guaranteed now, but can not be forever. This means anybody who\\n>> supports PHV tomorrow, will have to \\\"remember\\\" to update this code as\\n>> well. If s/he misses it, a bug will be introduced. That's the avenue\\n>> for bug, I am talking about.\\n>\\n>\\n> It *can* be guaranteed. See another patch for supporting evaluating PHVs\\n> remotely.\\n\\nWe can't base assumptions in this patch on something in the other\\npatch, esp. when that's not even reviewed once. PHV is just one case,\\nsubqueries involving aggregates is other and there are others. But\\nthat's not really the point. The point is\\nadd_to_flat_tlist(pull_var_clause(rel->reltarget->exprs) can not be\\nproved to be same as rel->reltarget->exprs in general. So, we should\\nbase our code on an assumption that can not be proved.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 24 Nov 2016 14:50:31 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/11/24 18:20, Ashutosh Bapat wrote:\\n\\nI wrote:\\n>>>> You missed the point; the foreignrel->reltarget->exprs doesn't contain\\n>>>> any\\n>>>> PHVs, so the tlist created by build_tlist_to_depase will be guaranteed to\\n>>>> be\\n>>>> one-to-one with the foreignrel->reltarget->exprs.\\n\\nYou wrote:\\n>>> It's guaranteed now, but can not be forever. This means anybody who\\n>>> supports PHV tomorrow, will have to \\\"remember\\\" to update this code as\\n>>> well. If s/he misses it, a bug will be introduced. That's the avenue\\n>>> for bug, I am talking about.\\n\\nI wrote:\\n>> It *can* be guaranteed. See another patch for supporting evaluating PHVs\\n>> remotely.\\n\\n> We can't base assumptions in this patch on something in the other\\n> patch, esp. when that's not even reviewed once. PHV is just one case,\\n> subqueries involving aggregates is other and there are others.\\n\\nLet's discuss this together with the patch for PHVs. That was one of \\nthe reasons why I had merged the two patches into one. I'd like to \\nleave enhancements such as subqueries involving aggregates for future \\nwork, though.\\n\\n> But\\n> that's not really the point. The point is\\n> add_to_flat_tlist(pull_var_clause(rel->reltarget->exprs) can not be\\n> proved to be same as rel->reltarget->exprs in general.\\n\\nReally? As mentioned in comments for RelOptInfo in relation.h shown \\nbelow, the rel->reltarget->exprs for a base/join relation only contains \\nVars and PHVs, so the two things would be proved to be one-to-one. \\n(Note: we don't need to care about the appendrel-child-relation case \\nbecause appendrel child relations wouldn't appear in the main join tree.)\\n\\n * reltarget - Default Path output tlist for this rel; normally \\ncontains\\n * Var and PlaceHolderVar nodes for the values we need to\\n * output from this relation.\\n * List is in no particular order, but all rels of an\\n * appendrel set must use corresponding orders.\\n * NOTE: in an appendrel child relation, may contain\\n * arbitrary expressions pulled up from a subquery!\\n\\nBest regards,\\nEtsuro Fujita\\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, 24 Nov 2016 19:04:46 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/11/22 18:28, Ashutosh Bapat wrote:\\n> The comments should explain why is the assertion true.\\n> + /* Shouldn't be NIL */\\n> + Assert(tlist != NIL);\\n\\nI noticed that I was wrong; in the Assertion the tlist can be empty. An \\nexample for such a case is:\\n\\nSELECT 1 FROM (SELECT c1 FROM ft4 WHERE c1 between 50 and 60) t1 FULL \\nJOIN (SELECT c1 FROM ft5 WHERE c1 between 50 and 60) t2 ON (TRUE);\\n\\nIn this example any columns of the relations ft4 and ft5 wouldn't be \\nneeded for the join or the final output, so both the tlists for the \\nrelations created in deparseRangeTblRef would be empty. Attached is an \\nupdated version fixing this bug. Changes are:\\n\\n* I removed make_outerrel_subquery/make_innerrel_subquery from fpinfo \\nand added a new member \\\"is_subquery_rel\\\" to fpinfo, as a flag on whether \\nto deparse the relation as a subquery.\\n* I modified deparseRangeTblRef, deparseSelectSql, and is_subquery_var \\nto see the is_subquery_rel, not \\nmake_outerrel_subquery/make_innerrel_subquery.\\n* I modified appendSubselectAlias to handle the case where ncols = 0 \\nproperly.\\n* I renamed subquery_rels in fpinfo to lower_subquery_rels, to make it \\neasy to distinguish this from is_subquery_rel clearly.\\n* I added regression tests for that.\\n\\nThe rest of the patch is the same as the previous version, but:\\n\\n> + /* Should be same length */\\n> + Assert(list_length(tlist) ==\\n> list_length(foreignrel->reltarget->exprs));\\n\\nI added comments explaining the reason.\\n\\n> The name get_subselect_alias_id() seems to suggest that the function returns\\n> identifier for subselect alias, which isn't correct. It returns the alias\\n> identifiers for deparsing a Var node. But I guess, we have left this to the\\n> committer's judgement.\\n\\nI agree that the name isn't good, so I changed it to \\nget_relation_column_alias_ids(). Let me know if it may be better.\\n\\nI also revised code and comments a bit, just for consistency.\\n\\nI will update another patch for PHVs on top of the attached one.\\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\": \"Thu, 24 Nov 2016 21:45:40 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/11/24 21:45, Etsuro Fujita wrote:\\n> On 2016/11/22 18:28, Ashutosh Bapat wrote:\\n>> The comments should explain why is the assertion true.\\n>> + /* Shouldn't be NIL */\\n>> + Assert(tlist != NIL);\\n\\n> I noticed that I was wrong; in the Assertion the tlist can be empty. An\\n> example for such a case is:\\n>\\n> SELECT 1 FROM (SELECT c1 FROM ft4 WHERE c1 between 50 and 60) t1 FULL\\n> JOIN (SELECT c1 FROM ft5 WHERE c1 between 50 and 60) t2 ON (TRUE);\\n>\\n> In this example any columns of the relations ft4 and ft5 wouldn't be\\n> needed for the join or the final output, so both the tlists for the\\n> relations created in deparseRangeTblRef would be empty. Attached is an\\n> updated version fixing this bug. Changes are:\\n>\\n> * I removed make_outerrel_subquery/make_innerrel_subquery from fpinfo\\n> and added a new member \\\"is_subquery_rel\\\" to fpinfo, as a flag on whether\\n> to deparse the relation as a subquery.\\n> * I modified deparseRangeTblRef, deparseSelectSql, and is_subquery_var\\n> to see the is_subquery_rel, not\\n> make_outerrel_subquery/make_innerrel_subquery.\\n> * I modified appendSubselectAlias to handle the case where ncols = 0\\n> properly.\\n> * I renamed subquery_rels in fpinfo to lower_subquery_rels, to make it\\n> easy to distinguish this from is_subquery_rel clearly.\\n> * I added regression tests for that.\\n\\nI noticed that the above changes are not adequate; the updated patch \\nbreaks EXPLAIN outputs for EvalPlanQual subplans of foreign joins in \\nwhich foreign tables are full joined and in the remote join query the \\nforeign tables are deparsed as subqueries. Consider:\\n\\npostgres=# explain verbose select * from test, ((select * from ft1 where \\nb between 1 and 3) t1 full join (select * from ft2 where b between 1 and \\n3) t2\\n on (t1.a = t2.a)) ss for update of test;\\n \\n \\nQUERY PLA\\nN\\n-------------------------------------------------------------------------------------------------------------------------------------------------------\\n------------------------------------------------------------------------------------------------------------------------------------------------\\n LockRows (cost=100.00..287.33 rows=6780 width=94)\\n Output: test.a, test.b, ft1.a, ft1.b, ft2.a, ft2.b, test.ctid, \\nft1.*, ft2.*\\n -> Nested Loop (cost=100.00..219.53 rows=6780 width=94)\\n Output: test.a, test.b, ft1.a, ft1.b, ft2.a, ft2.b, test.ctid, \\nft1.*, ft2.*\\n -> Seq Scan on public.test (cost=0.00..32.60 rows=2260 width=14)\\n Output: test.a, test.b, test.ctid\\n -> Materialize (cost=100.00..102.19 rows=3 width=80)\\n Output: ft1.a, ft1.b, ft1.*, ft2.a, ft2.b, ft2.*\\n -> Foreign Scan (cost=100.00..102.17 rows=3 width=80)\\n Output: ft1.a, ft1.b, ft1.*, ft2.a, ft2.b, ft2.*\\n Relations: (public.ft1) FULL JOIN (public.ft2)\\n Remote SQL: SELECT s5.c1, s5.c2, s5.c3, s6.c1, \\ns6.c2, s6.c3 FROM ((SELECT a, b, ROW(a, b) FROM public.t1 WHERE ((b >= \\n1)) AND ((b\\n<= 3))) s5(c1, c2, c3) FULL JOIN (SELECT a, b, ROW(a, b) FROM public.t2 \\nWHERE ((b >= 1)) AND ((b <= 3))) s6(c1, c2, c3) ON (((s5.c1 = s6.c1))))\\n -> Hash Full Join (cost=201.14..202.29 rows=3 \\nwidth=80)\\n Output: ft1.a, ft1.b, ft1.*, ft2.a, ft2.b, ft2.*\\n Hash Cond: (ft1.a = ft2.a)\\n -> Foreign Scan on public.ft1 \\n(cost=100.00..101.11 rows=3 width=40)\\n Output: ft1.a, ft1.b, ft1.*\\n Remote SQL: SELECT NULL FROM public.t1 \\nWHERE ((b >= 1)) AND ((b <= 3))\\n -> Hash (cost=101.11..101.11 rows=3 width=40)\\n Output: ft2.a, ft2.b, ft2.*\\n -> Foreign Scan on public.ft2 \\n(cost=100.00..101.11 rows=3 width=40)\\n Output: ft2.a, ft2.b, ft2.*\\n Remote SQL: SELECT NULL FROM \\npublic.t2 WHERE ((b >= 1)) AND ((b <= 3))\\n(23 rows)\\n\\nThe SELECT clauses of the remote queries for the Foreign Scans in the \\nEPQ subplan are deparsed incorrectly into \\\"SELECT NULL\\\". The reason for \\nthat is that since the foreign tables' fpinfo->is_subquery_rel has been \\nset true at the time those clauses are deparsed (due to that the foreign \\ntables have been deparsed as subqueries in the main remote join query), \\ndeparseSelectSql calls deparseExplicitTargetList with an empty tlist, to \\nconstruct the clauses, so it deparses the tlist into \\\"NULL\\\". This is \\nharmless because we don't use the remote queries for the Foreign Scans \\nduring an EPQ recheck, but that would be confusing for users, so I \\nmodified the patch a bit further to make the EXPLAIN output as-is. \\nAttached is a new version of the patch.\\n\\n> I also revised code and comments a bit, just for consistency.\\n\\nAlso, I renamed appendSubselectAlias to appendSubqueryAlias, because we \\nuse the term \\\"subquery\\\" for other places.\\n\\n> I will update another patch for PHVs on top of the attached one.\\n\\nDone. I also revised some code and comments. I'm also attaching the \\nupdated version of the patch for PHVs.\\n\\nBest regards,\\nEtsuro Fujita\\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, 28 Nov 2016 19:22:03 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 28, 2016 at 3:52 PM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> On 2016/11/24 21:45, Etsuro Fujita wrote:\\n>>\\n>> On 2016/11/22 18:28, Ashutosh Bapat wrote:\\n>>>\\n>>> The comments should explain why is the assertion true.\\n>>> + /* Shouldn't be NIL */\\n>>> + Assert(tlist != NIL);\\n>\\n>\\n>> I noticed that I was wrong; in the Assertion the tlist can be empty. An\\n>> example for such a case is:\\n>>\\n>> SELECT 1 FROM (SELECT c1 FROM ft4 WHERE c1 between 50 and 60) t1 FULL\\n>> JOIN (SELECT c1 FROM ft5 WHERE c1 between 50 and 60) t2 ON (TRUE);\\n>>\\n>> In this example any columns of the relations ft4 and ft5 wouldn't be\\n>> needed for the join or the final output, so both the tlists for the\\n>> relations created in deparseRangeTblRef would be empty. Attached is an\\n>> updated version fixing this bug. Changes are:\\n>>\\n>> * I removed make_outerrel_subquery/make_innerrel_subquery from fpinfo\\n>> and added a new member \\\"is_subquery_rel\\\" to fpinfo, as a flag on whether\\n>> to deparse the relation as a subquery.\\n\\nReplacing make_outerrel_subquery/make_innerrel_subquery with is_subquery_rel\\nseems to be a bad idea. Whether a relation needs to be deparsed as a subquery\\nor not depends upon the relation which joins it with another relation. Take for\\nexample a case when a join ABC can pull up the conditions on AB, but ABD can\\nnot. In this case we will mark that AB in ABD needs to be deparsed as a\\nsubquery but will not mark so in ABC. So, if we choose to join ABCD as (ABC)D\\nwe don't need AB to be deparsed as a subquery. If we choose to join ABCD as\\n(ABD)C, we will need AB to deparsed as a subquery. Regression doesn't have a\\ntestcase like that hence the code seems to be working.\\n\\nSome more comments\\n1. The output of the following query\\n+SELECT 1 FROM (SELECT c1 FROM ft4 WHERE c1 between 50 and 60) t1 FULL\\nJOIN (SELECT c1 FROM ft5 WHERE c1 between 50 and 60) t2 ON (TRUE);\\nproduces 21 rows all containing a \\\"1\\\". That's not quite good use of expected\\noutput space, given that all that the test tests is the empty\\ntargetlists are deparsed\\ncorrectly. You might want to either just test EXPLAIN output or make \\\"between 50\\nand 60\\\" condition more stringent to reduce the number of rows output.\\n\\n2. Got lost here. I guess, all you need to say is \\\"test deparsing FOR UPDATE\\nclause with subqueries.\\\". Anyway, the sentence is very hard to read and needs\\nsimplification.\\n+-- d. test deparsing the remote queries for simple foreign table scans in\\n+-- an EPQ subplan of a foreign join in which the foreign tables are full\\n+-- joined and in the remote join query the foreign tables are deparsed as\\n+-- subqueries\\n\\n3. Why is foreignrel variable changed to rel?\\n-extern void deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root,\\n- RelOptInfo *foreignrel, List *tlist,\\n- List *remote_conds, List *pathkeys,\\n- List **retrieved_attrs, List **params_list);\\n+extern void deparseSelectStmtForRel(StringInfo buf, PlannerInfo\\n*root, RelOptInfo *rel,\\n+ List *tlist, List *remote_conds, List *pathkeys,\\n+ bool is_subquery, List **retrieved_attrs,\\n+ List **params_list);\\n\\n4. I am still not happy with this change\\n+ /*\\n+ * Since (1) the expressions in foreignrel's reltarget doesn't contain\\n+ * any PHVs and (2) foreignrel's local_conds is empty, the tlist\\n+ * created by build_tlist_to_deparse must be one-to-one with the\\n+ * expressions.\\n+ */\\n+ Assert(list_length(tlist) ==\\nlist_length(foreignrel->reltarget->exprs));\\nthe assertion only checks that the number of elements in both the lists are\\nsame but does not check whether those lists are same i.e. they contain the same\\nelements in the same order. This equality is crucial to deparsing logic. If\\nsomehow build_tlist_to_deparse() breaks that assumption in future, we have no\\nway to detect it, unless a regression test fails.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 5 Dec 2016 16:50:45 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 5, 2016 at 6:20 AM, Ashutosh Bapat\\n<ashutosh.bapat@enterprisedb.com> wrote:\\n> 4. I am still not happy with this change\\n> + /*\\n> + * Since (1) the expressions in foreignrel's reltarget doesn't contain\\n> + * any PHVs and (2) foreignrel's local_conds is empty, the tlist\\n> + * created by build_tlist_to_deparse must be one-to-one with the\\n> + * expressions.\\n> + */\\n> + Assert(list_length(tlist) ==\\n> list_length(foreignrel->reltarget->exprs));\\n> the assertion only checks that the number of elements in both the lists are\\n> same but does not check whether those lists are same i.e. they contain the same\\n> elements in the same order. This equality is crucial to deparsing logic. If\\n> somehow build_tlist_to_deparse() breaks that assumption in future, we have no\\n> way to detect it, unless a regression test fails.\\n\\nIf there's an easy way to do a more exact comparison, great. But if\\nwe can't get an awesome Assert(), a helpful Assert() is still better\\nthan a kick in the head.\\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, 6 Dec 2016 15:27:06 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/12/07 5:27, Robert Haas wrote:\\n> On Mon, Dec 5, 2016 at 6:20 AM, Ashutosh Bapat\\n> <ashutosh.bapat@enterprisedb.com> wrote:\\n>> 4. I am still not happy with this change\\n>> + /*\\n>> + * Since (1) the expressions in foreignrel's reltarget doesn't contain\\n>> + * any PHVs and (2) foreignrel's local_conds is empty, the tlist\\n>> + * created by build_tlist_to_deparse must be one-to-one with the\\n>> + * expressions.\\n>> + */\\n>> + Assert(list_length(tlist) ==\\n>> list_length(foreignrel->reltarget->exprs));\\n>> the assertion only checks that the number of elements in both the lists are\\n>> same but does not check whether those lists are same i.e. they contain the same\\n>> elements in the same order. This equality is crucial to deparsing logic. If\\n>> somehow build_tlist_to_deparse() breaks that assumption in future, we have no\\n>> way to detect it, unless a regression test fails.\\n\\n> If there's an easy way to do a more exact comparison, great.\\n\\nAshutosh proposed this to do the comparison:\\n\\nOn 2016/11/22 18:28, Ashutosh Bapat wrote:\\n> I guess, the reason why you are doing it this way, is SELECT clause for the\\n> outermost query gets deparsed before FROM clause. For later we call\\n> deparseRangeTblRef(), which builds the tlist. So, while deparsing SELECT\\n> clause, we do not have tlist to build from. In that case, I guess, we have to\\n> build the tlist in get_subselect_alias_id() if it's not available and stick it\\n> in fpinfo. Subsequent calls to get_subselect_alias_id() should find tlist\\n> there. Then in deparseRangeTblRef() assert that there's a tlist in fpinfo\\n> and use it to build the SELECT clause of subquery. That way, we don't build\\n> tlist unless it's needed and also use the same tlist for all searches. Please\\n> use tlist_member() to search into the tlist.\\n\\nThis would probably work, but seems to me a bit complicated. Instead, \\nI'd like to propose that we build the tlist for each relation being \\ndeparsed as a subquery in a given join tree, right before deparsing the \\nSELECT clause in deparseSelectStmtForRel, if is_subquery is false and \\nlower_subquery_rels isn't NULL, and store the tlist into the relation's \\nfpinfo. That would allow us to build the tlist only when we need it, \\nand to use tlist_member for the exact comparison. I think it would be \\nmuch easier to implement that.\\n\\nBest regards,\\nEtsuro Fujita\\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\": \"Wed, 7 Dec 2016 12:15:41 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 7, 2016 at 1:57 AM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Mon, Dec 5, 2016 at 6:20 AM, Ashutosh Bapat\\n> <ashutosh.bapat@enterprisedb.com> wrote:\\n>> 4. I am still not happy with this change\\n>> + /*\\n>> + * Since (1) the expressions in foreignrel's reltarget doesn't contain\\n>> + * any PHVs and (2) foreignrel's local_conds is empty, the tlist\\n>> + * created by build_tlist_to_deparse must be one-to-one with the\\n>> + * expressions.\\n>> + */\\n>> + Assert(list_length(tlist) ==\\n>> list_length(foreignrel->reltarget->exprs));\\n>> the assertion only checks that the number of elements in both the lists are\\n>> same but does not check whether those lists are same i.e. they contain the same\\n>> elements in the same order. This equality is crucial to deparsing logic. If\\n>> somehow build_tlist_to_deparse() breaks that assumption in future, we have no\\n>> way to detect it, unless a regression test fails.\\n>\\n> If there's an easy way to do a more exact comparison, great. But if\\n> we can't get an awesome Assert(), a helpful Assert() is still better\\n> than a kick in the head.\\n\\nThe assert is not a problem in itself, but the reason we have to add\\nthe assert. The problem is explained in [1], point #9.\\n\\n[1] https://www.postgresql.org/message-id/CAFjFpRfwoSsJr9418b2jA7g0nwagjZSWhPQPUmK9M6z5XSOAqQ%40mail.gmail.com\\n\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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 Dec 2016 12:02:35 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \">\\n>\\n> Ashutosh proposed this to do the comparison:\\n>\\n> On 2016/11/22 18:28, Ashutosh Bapat wrote:\\n>>\\n>> I guess, the reason why you are doing it this way, is SELECT clause for\\n>> the\\n>> outermost query gets deparsed before FROM clause. For later we call\\n>> deparseRangeTblRef(), which builds the tlist. So, while deparsing SELECT\\n>> clause, we do not have tlist to build from. In that case, I guess, we have\\n>> to\\n>> build the tlist in get_subselect_alias_id() if it's not available and\\n>> stick it\\n>> in fpinfo. Subsequent calls to get_subselect_alias_id() should find tlist\\n>> there. Then in deparseRangeTblRef() assert that there's a tlist in fpinfo\\n>> and use it to build the SELECT clause of subquery. That way, we don't\\n>> build\\n>> tlist unless it's needed and also use the same tlist for all searches.\\n>> Please\\n>> use tlist_member() to search into the tlist.\\n>\\n>\\n> This would probably work, but seems to me a bit complicated. Instead, I'd\\n> like to propose that we build the tlist for each relation being deparsed as\\n> a subquery in a given join tree, right before deparsing the SELECT clause in\\n> deparseSelectStmtForRel, if is_subquery is false and lower_subquery_rels\\n> isn't NULL, and store the tlist into the relation's fpinfo. That would\\n> allow us to build the tlist only when we need it, and to use tlist_member\\n> for the exact comparison. I think it would be much easier to implement\\n> that.\\n>\\n\\nIIRC, this is inline with my original proposal of creating tlists\\nbefore deparsing anything. Along-with that I also proposed to bubble\\nthis array of tlist up the join hierarchy to avoid recursion [1] point\\n#15, further elaborated in [2]\\n\\n[1] https://www.postgresql.org/message-id/ad449b25-66ee-4c06-568c-0eb2e1bef9f9%40lab.ntt.co.jp\\n[2] https://www.postgresql.org/message-id/CAFjFpRcn7%3DDNOXm-PJ_jVDqAmghKVf6tApQC%2B_RgMZ8tOPExcA%40mail.gmail.com\\n\\nWe seem to be moving towards that solution, but as discussed we have\\nleft it to committer's judgement.\\n--\\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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 Dec 2016 12:09:01 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/12/05 20:20, Ashutosh Bapat wrote:\\n> On Mon, Nov 28, 2016 at 3:52 PM, Etsuro Fujita\\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>> On 2016/11/24 21:45, Etsuro Fujita wrote:\\n>>> * I removed make_outerrel_subquery/make_innerrel_subquery from fpinfo\\n>>> and added a new member \\\"is_subquery_rel\\\" to fpinfo, as a flag on whether\\n>>> to deparse the relation as a subquery.\\n\\n> Replacing make_outerrel_subquery/make_innerrel_subquery with is_subquery_rel\\n> seems to be a bad idea. Whether a relation needs to be deparsed as a subquery\\n> or not depends upon the relation which joins it with another relation. Take for\\n> example a case when a join ABC can pull up the conditions on AB, but ABD can\\n> not. In this case we will mark that AB in ABD needs to be deparsed as a\\n> subquery but will not mark so in ABC. So, if we choose to join ABCD as (ABC)D\\n> we don't need AB to be deparsed as a subquery. If we choose to join ABCD as\\n> (ABD)C, we will need AB to deparsed as a subquery.\\n\\nThis is not true; since (1) currently, a relation needs to be deparsed \\nas a subquery only when the relation is full-joined with any other \\nrelation, and (2) the join ordering for full joins is preserved, we \\nwouldn't have such an example. (You might think we would have that when \\nextending the patch to handle PHVs, but the answer would be no, because \\nthe patch for PHVs would determine whether a relation emitting PHVs \\nneeds to be deparsed as a subquery, depending on the relation itself. \\nSee the patch for PHVs.) I like is_subquery_rel more than \\nmake_outerrel_subquery/make_innerrel_subquery, because it reduces the \\nnumber of members in fpinfo. But the choice would be arbitrary, so I \\nwouldn't object to going back to \\nmake_outerrel_subquery/make_innerrel_subquery.\\n\\n> Some more comments\\n> 1. The output of the following query\\n> +SELECT 1 FROM (SELECT c1 FROM ft4 WHERE c1 between 50 and 60) t1 FULL\\n> JOIN (SELECT c1 FROM ft5 WHERE c1 between 50 and 60) t2 ON (TRUE);\\n> produces 21 rows all containing a \\\"1\\\". That's not quite good use of expected\\n> output space, given that all that the test tests is the empty\\n> targetlists are deparsed\\n> correctly. You might want to either just test EXPLAIN output or make \\\"between 50\\n> and 60\\\" condition more stringent to reduce the number of rows output.\\n\\nWill do.\\n\\n> 2. Got lost here. I guess, all you need to say is \\\"test deparsing FOR UPDATE\\n> clause with subqueries.\\\". Anyway, the sentence is very hard to read and needs\\n> simplification.\\n> +-- d. test deparsing the remote queries for simple foreign table scans in\\n> +-- an EPQ subplan of a foreign join in which the foreign tables are full\\n> +-- joined and in the remote join query the foreign tables are deparsed as\\n> +-- subqueries\\n\\nWill do.\\n\\n> 3. Why is foreignrel variable changed to rel?\\n> -extern void deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root,\\n> - RelOptInfo *foreignrel, List *tlist,\\n> - List *remote_conds, List *pathkeys,\\n> - List **retrieved_attrs, List **params_list);\\n> +extern void deparseSelectStmtForRel(StringInfo buf, PlannerInfo\\n> *root, RelOptInfo *rel,\\n> + List *tlist, List *remote_conds, List *pathkeys,\\n> + bool is_subquery, List **retrieved_attrs,\\n> + List **params_list);\\n\\nI changed the name that way to match with the function definition in \\ndeparse.c.\\n\\nThanks for the comments!\\n\\nBest regards,\\nEtsuro Fujita\\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\": \"Wed, 7 Dec 2016 20:21:58 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/12/07 15:39, Ashutosh Bapat wrote:\\n\\n>> On 2016/11/22 18:28, Ashutosh Bapat wrote:\\n>>>\\n>>> I guess, the reason why you are doing it this way, is SELECT clause for\\n>>> the\\n>>> outermost query gets deparsed before FROM clause. For later we call\\n>>> deparseRangeTblRef(), which builds the tlist. So, while deparsing SELECT\\n>>> clause, we do not have tlist to build from. In that case, I guess, we have\\n>>> to\\n>>> build the tlist in get_subselect_alias_id() if it's not available and\\n>>> stick it\\n>>> in fpinfo. Subsequent calls to get_subselect_alias_id() should find tlist\\n>>> there. Then in deparseRangeTblRef() assert that there's a tlist in fpinfo\\n>>> and use it to build the SELECT clause of subquery. That way, we don't\\n>>> build\\n>>> tlist unless it's needed and also use the same tlist for all searches.\\n>>> Please\\n>>> use tlist_member() to search into the tlist.\\n\\nI wrote:\\n>> This would probably work, but seems to me a bit complicated. Instead, I'd\\n>> like to propose that we build the tlist for each relation being deparsed as\\n>> a subquery in a given join tree, right before deparsing the SELECT clause in\\n>> deparseSelectStmtForRel, if is_subquery is false and lower_subquery_rels\\n>> isn't NULL, and store the tlist into the relation's fpinfo. That would\\n>> allow us to build the tlist only when we need it, and to use tlist_member\\n>> for the exact comparison. I think it would be much easier to implement\\n>> that.\\n\\n> IIRC, this is inline with my original proposal of creating tlists\\n> before deparsing anything. Along-with that I also proposed to bubble\\n> this array of tlist up the join hierarchy to avoid recursion [1] point\\n> #15, further elaborated in [2]\\n>\\n> [1] https://www.postgresql.org/message-id/ad449b25-66ee-4c06-568c-0eb2e1bef9f9%40lab.ntt.co.jp\\n> [2] https://www.postgresql.org/message-id/CAFjFpRcn7%3DDNOXm-PJ_jVDqAmghKVf6tApQC%2B_RgMZ8tOPExcA%40mail.gmail.com\\n>\\n> We seem to be moving towards that solution, but as discussed we have\\n> left it to committer's judgement.\\n\\nMy proposal here would be a bit different from what you proposed; right \\nbefore deparseSelectSql in deparseSelectStmtForRel, build the tlists for \\nrelations present in the given jointree that will be deparsed as \\nsubqueries, by (1) traversing the given jointree and (2) applying \\nbuild_tlist_to_deparse to each relation to be deparsed as a subquery and \\nsaving the result in that relation's fpinfo. I think that would be \\nimplemented easily, and the overhead would be small.\\n\\nBest regards,\\nEtsuro Fujita\\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\": \"Wed, 7 Dec 2016 20:23:41 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 7, 2016 at 4:51 PM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> On 2016/12/05 20:20, Ashutosh Bapat wrote:\\n>>\\n>> On Mon, Nov 28, 2016 at 3:52 PM, Etsuro Fujita\\n>> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>>>\\n>>> On 2016/11/24 21:45, Etsuro Fujita wrote:\\n>>>>\\n>>>> * I removed make_outerrel_subquery/make_innerrel_subquery from fpinfo\\n>>>> and added a new member \\\"is_subquery_rel\\\" to fpinfo, as a flag on whether\\n>>>> to deparse the relation as a subquery.\\n>\\n>\\n>> Replacing make_outerrel_subquery/make_innerrel_subquery with\\n>> is_subquery_rel\\n>> seems to be a bad idea. Whether a relation needs to be deparsed as a\\n>> subquery\\n>> or not depends upon the relation which joins it with another relation.\\n>> Take for\\n>> example a case when a join ABC can pull up the conditions on AB, but ABD\\n>> can\\n>> not. In this case we will mark that AB in ABD needs to be deparsed as a\\n>> subquery but will not mark so in ABC. So, if we choose to join ABCD as\\n>> (ABC)D\\n>> we don't need AB to be deparsed as a subquery. If we choose to join ABCD\\n>> as\\n>> (ABD)C, we will need AB to deparsed as a subquery.\\n>\\n>\\n> This is not true; since (1) currently, a relation needs to be deparsed as a\\n> subquery only when the relation is full-joined with any other relation, and\\n> (2) the join ordering for full joins is preserved, we wouldn't have such an\\n> example. (You might think we would have that when extending the patch to\\n> handle PHVs, but the answer would be no, because the patch for PHVs would\\n> determine whether a relation emitting PHVs needs to be deparsed as a\\n> subquery, depending on the relation itself. See the patch for PHVs.) I like\\n> is_subquery_rel more than make_outerrel_subquery/make_innerrel_subquery,\\n> because it reduces the number of members in fpinfo. But the choice would be\\n> arbitrary, so I wouldn't object to going back to\\n> make_outerrel_subquery/make_innerrel_subquery.\\n\\nI think you are right. And even if we were to deparse a relation as\\nsubquery, when it shouldn't be, we will only loose a syntactic\\noptimization. So, it shouldn't result in a bug.\\n\\n\\n>> 3. Why is foreignrel variable changed to rel?\\n>> -extern void deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root,\\n>> - RelOptInfo *foreignrel, List *tlist,\\n>> - List *remote_conds, List *pathkeys,\\n>> - List **retrieved_attrs, List **params_list);\\n>> +extern void deparseSelectStmtForRel(StringInfo buf, PlannerInfo\\n>> *root, RelOptInfo *rel,\\n>> + List *tlist, List *remote_conds, List *pathkeys,\\n>> + bool is_subquery, List **retrieved_attrs,\\n>> + List **params_list);\\n>\\n>\\n> I changed the name that way to match with the function definition in\\n> deparse.c.\\n>\\n\\nThat's something good to do but it's unrelated to this patch. I have\\nrepeated this line several times in this thread :)\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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 Dec 2016 17:41:02 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/12/07 21:11, Ashutosh Bapat wrote:\\n> On Wed, Dec 7, 2016 at 4:51 PM, Etsuro Fujita\\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>> On 2016/12/05 20:20, Ashutosh Bapat wrote:\\n\\n>>> 3. Why is foreignrel variable changed to rel?\\n>>> -extern void deparseSelectStmtForRel(StringInfo buf, PlannerInfo *root,\\n>>> - RelOptInfo *foreignrel, List *tlist,\\n>>> - List *remote_conds, List *pathkeys,\\n>>> - List **retrieved_attrs, List **params_list);\\n>>> +extern void deparseSelectStmtForRel(StringInfo buf, PlannerInfo\\n>>> *root, RelOptInfo *rel,\\n>>> + List *tlist, List *remote_conds, List *pathkeys,\\n>>> + bool is_subquery, List **retrieved_attrs,\\n>>> + List **params_list);\\n\\n>> I changed the name that way to match with the function definition in\\n>> deparse.c.\\n\\n> That's something good to do but it's unrelated to this patch. I have\\n> repeated this line several times in this thread :)\\n\\nOK.\\n\\nBest regards,\\nEtsuro Fujita\\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, 8 Dec 2016 21:04:57 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/12/07 20:23, Etsuro Fujita wrote:\\n> On 2016/12/07 15:39, Ashutosh Bapat wrote:\\n\\n>>> On 2016/11/22 18:28, Ashutosh Bapat wrote:\\n>>>>\\n>>>> I guess, the reason why you are doing it this way, is SELECT clause for\\n>>>> the\\n>>>> outermost query gets deparsed before FROM clause. For later we call\\n>>>> deparseRangeTblRef(), which builds the tlist. So, while deparsing\\n>>>> SELECT\\n>>>> clause, we do not have tlist to build from. In that case, I guess,\\n>>>> we have\\n>>>> to\\n>>>> build the tlist in get_subselect_alias_id() if it's not available and\\n>>>> stick it\\n>>>> in fpinfo. Subsequent calls to get_subselect_alias_id() should find\\n>>>> tlist\\n>>>> there. Then in deparseRangeTblRef() assert that there's a tlist in\\n>>>> fpinfo\\n>>>> and use it to build the SELECT clause of subquery. That way, we don't\\n>>>> build\\n>>>> tlist unless it's needed and also use the same tlist for all searches.\\n>>>> Please\\n>>>> use tlist_member() to search into the tlist.\\n\\n> I wrote:\\n>>> This would probably work, but seems to me a bit complicated.\\n>>> Instead, I'd\\n>>> like to propose that we build the tlist for each relation being\\n>>> deparsed as\\n>>> a subquery in a given join tree, right before deparsing the SELECT\\n>>> clause in\\n>>> deparseSelectStmtForRel, if is_subquery is false and lower_subquery_rels\\n>>> isn't NULL, and store the tlist into the relation's fpinfo. That would\\n>>> allow us to build the tlist only when we need it, and to use\\n>>> tlist_member\\n>>> for the exact comparison. I think it would be much easier to implement\\n>>> that.\\n\\n>> IIRC, this is inline with my original proposal of creating tlists\\n>> before deparsing anything. Along-with that I also proposed to bubble\\n>> this array of tlist up the join hierarchy to avoid recursion [1] point\\n>> #15, further elaborated in [2]\\n>>\\n>> [1]\\n>> https://www.postgresql.org/message-id/ad449b25-66ee-4c06-568c-0eb2e1bef9f9%40lab.ntt.co.jp\\n>>\\n>> [2]\\n>> https://www.postgresql.org/message-id/CAFjFpRcn7%3DDNOXm-PJ_jVDqAmghKVf6tApQC%2B_RgMZ8tOPExcA%40mail.gmail.com\\n\\n> My proposal here would be a bit different from what you proposed; right\\n> before deparseSelectSql in deparseSelectStmtForRel, build the tlists for\\n> relations present in the given jointree that will be deparsed as\\n> subqueries, by (1) traversing the given jointree and (2) applying\\n> build_tlist_to_deparse to each relation to be deparsed as a subquery and\\n> saving the result in that relation's fpinfo. I think that would be\\n> implemented easily, and the overhead would be small.\\n\\nI implemented that to address your concern:\\n* deparseRangeTblRef uses the tlist created as above, to build the \\nSELECT clause of the subquery. (You said \\\"Then in deparseRangeTblRef() \\nassert that there's a tlist in fpinfo\\\", but the tlist may be empty, so I \\ndidn't add any assertion to that function.)\\n* get_relation_column_alias_ids uses tlist_member with the tlist.\\n\\nI also addressed the comments #1, #2 and #3 in [1]. For #1, I added \\n\\\"LIMIT 10\\\" to the query. Attached is the new version of the patch.\\n\\nOther changes:\\n* As discussed before, renamed grouped_tlist in fpinfo to \\\"tlist\\\" and \\nused it to store the tlist created as above.\\n* Also, renamed SS_REL_ALIAS_PREFIX to SUBQUERY_REL_ALIAS_PREFIX \\n(Likewise for SS_COL_ALIAS_PREFIX).\\n* Relocated some functions.\\n* Revised comments a little bit.\\n\\nBest regards,\\nEtsuro Fujita\\n\\n[1] \\nhttps://www.postgresql.org/message-id/CAFjFpRfU4-gxqZ8ahoKM1ZtDJEThe3K8Lb_6beRKa8mmP%3Dv%3DfA%40mail.gmail.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, 8 Dec 2016 21:08:57 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2016/12/08 21:08, Etsuro Fujita wrote:\\n> On 2016/12/07 20:23, Etsuro Fujita wrote:\\n>> My proposal here would be a bit different from what you proposed; right\\n>> before deparseSelectSql in deparseSelectStmtForRel, build the tlists for\\n>> relations present in the given jointree that will be deparsed as\\n>> subqueries, by (1) traversing the given jointree and (2) applying\\n>> build_tlist_to_deparse to each relation to be deparsed as a subquery and\\n>> saving the result in that relation's fpinfo. I think that would be\\n>> implemented easily, and the overhead would be small.\\n\\n> I implemented that to address your concern:\\n> * deparseRangeTblRef uses the tlist created as above, to build the\\n> SELECT clause of the subquery. (You said \\\"Then in deparseRangeTblRef()\\n> assert that there's a tlist in fpinfo\\\", but the tlist may be empty, so I\\n> didn't add any assertion to that function.)\\n> * get_relation_column_alias_ids uses tlist_member with the tlist.\\n>\\n> I also addressed the comments #1, #2 and #3 in [1]. For #1, I added\\n> \\\"LIMIT 10\\\" to the query. Attached is the new version of the patch.\\n>\\n> Other changes:\\n> * As discussed before, renamed grouped_tlist in fpinfo to \\\"tlist\\\" and\\n> used it to store the tlist created as above.\\n> * Also, renamed SS_REL_ALIAS_PREFIX to SUBQUERY_REL_ALIAS_PREFIX\\n> (Likewise for SS_COL_ALIAS_PREFIX).\\n> * Relocated some functions.\\n> * Revised comments a little bit.\\n\\nI updated the patch a bit further: simplified the function name \\n(s/build_subquery_rel_tlists/build_subquery_tlists/), and revised \\ncomments a little bit. Attached is an updated version \\n(postgres-fdw-subquery-support-v14.patch). And I rebased another patch \\nfor PHVs against that patch, which is also attached \\n(postgres-fdw-phv-pushdown-v14.patch).\\n\\nBest regards,\\nEtsuro Fujita\\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, 27 Dec 2016 21:51:06 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \">\\n> I updated the patch a bit further: simplified the function name\\n> (s/build_subquery_rel_tlists/build_subquery_tlists/), and revised comments a\\n> little bit. Attached is an updated version\\n> (postgres-fdw-subquery-support-v14.patch). And I rebased another patch for\\n> PHVs against that patch, which is also attached\\n> (postgres-fdw-phv-pushdown-v14.patch).\\n>\\nFew comments\\n\\nIn build_subquery_tlists(), why don't we handle base relations?\\n+ if (foreignrel->reloptkind != RELOPT_JOINREL)\\n+ return;\\n\\nAlso, in this function, if fpinfo->tlist is already set, why do we want to\\nbuild it again?\\n\\nIn build_tlist_to_deparse(), if fpinfo->tlist for the given relation is set, we\\nshould just return it rather than constructing it again.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 3 Jan 2017 13:58:23 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2017/01/03 17:28, Ashutosh Bapat wrote:\\n\\nI wrote:\\n>> I updated the patch a bit further: simplified the function name\\n>> (s/build_subquery_rel_tlists/build_subquery_tlists/), and revised comments a\\n>> little bit. Attached is an updated version\\n>> (postgres-fdw-subquery-support-v14.patch).\\n\\n> Few comments\\n\\nThanks for the comments!\\n\\n> In build_subquery_tlists(), why don't we handle base relations?\\n> + if (foreignrel->reloptkind != RELOPT_JOINREL)\\n> + return;\\n\\nThe reason for that is we don't need to handle the baserel cases; the \\ntlist for a base relation, if needed, would be created while recursing \\ninto a join relation that joins the base relation to other base/join \\nrelation.\\n\\n> Also, in this function, if fpinfo->tlist is already set, why do we want to\\n> build it again?\\n\\nWhen this function gets called, fpinfo->tlist isn't set for any base or \\njoin relation that needs to build the tlist, so we always need to build \\nit for each such relation.\\n\\n> In build_tlist_to_deparse(), if fpinfo->tlist for the given relation is set, we\\n> should just return it rather than constructing it again.\\n\\nIn that function we wouldn't have such cases for base or join relations \\nneeding the tlist.\\n\\nBest regards,\\nEtsuro Fujita\\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, 5 Jan 2017 20:44:51 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 5, 2017 at 5:14 PM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> On 2017/01/03 17:28, Ashutosh Bapat wrote:\\n>\\n> I wrote:\\n>>>\\n>>> I updated the patch a bit further: simplified the function name\\n>>> (s/build_subquery_rel_tlists/build_subquery_tlists/), and revised\\n>>> comments a\\n>>> little bit. Attached is an updated version\\n>>> (postgres-fdw-subquery-support-v14.patch).\\n>\\n>\\n>> Few comments\\n>\\n>\\n> Thanks for the comments!\\n>\\n>> In build_subquery_tlists(), why don't we handle base relations?\\n>> + if (foreignrel->reloptkind != RELOPT_JOINREL)\\n>> + return;\\n>\\n>\\n> The reason for that is we don't need to handle the baserel cases; the tlist\\n> for a base relation, if needed, would be created while recursing into a join\\n> relation that joins the base relation to other base/join relation.\\n\\nRight. Sorry, I misunderstood the code. May be a comment would help.\\n\\n>\\n>> Also, in this function, if fpinfo->tlist is already set, why do we want to\\n>> build it again?\\n>\\n>\\n> When this function gets called, fpinfo->tlist isn't set for any base or join\\n> relation that needs to build the tlist, so we always need to build it for\\n> each such relation.\\n\\nIIUC, for a relation with use_remote_estimates we will deparse the\\nquery twice and will build the targetlist twice.\\n\\n>\\n>> In build_tlist_to_deparse(), if fpinfo->tlist for the given relation is\\n>> set, we\\n>> should just return it rather than constructing it again.\\n>\\n>\\n> In that function we wouldn't have such cases for base or join relations\\n> needing the tlist.\\n\\nSame explanation as above.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 5 Jan 2017 17:41:26 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2017/01/05 21:11, Ashutosh Bapat wrote:\\n> On Thu, Jan 5, 2017 at 5:14 PM, Etsuro Fujita\\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>> On 2017/01/03 17:28, Ashutosh Bapat wrote:\\n>>> In build_subquery_tlists(), why don't we handle base relations?\\n>>> + if (foreignrel->reloptkind != RELOPT_JOINREL)\\n>>> + return;\\n\\n>> The reason for that is we don't need to handle the baserel cases; the tlist\\n>> for a base relation, if needed, would be created while recursing into a join\\n>> relation that joins the base relation to other base/join relation.\\n\\n> Right. Sorry, I misunderstood the code. May be a comment would help.\\n\\nWill add the comment.\\n\\n>>> Also, in this function, if fpinfo->tlist is already set, why do we want to\\n>>> build it again?\\n\\n>> When this function gets called, fpinfo->tlist isn't set for any base or join\\n>> relation that needs to build the tlist, so we always need to build it for\\n>> each such relation.\\n\\n> IIUC, for a relation with use_remote_estimates we will deparse the\\n> query twice and will build the targetlist twice.\\n\\nThat's right. We could avoid the duplicate work the way you proposed, \\nbut I was thinking to leave that for another patch. Should we do that \\nin this patch?\\n\\n>>> In build_tlist_to_deparse(), if fpinfo->tlist for the given relation is\\n>>> set, we\\n>>> should just return it rather than constructing it again.\\n\\n>> In that function we wouldn't have such cases for base or join relations\\n>> needing the tlist.\\n\\n> Same explanation as above.\\n\\nWill revise if it's better to do that in this patch.\\n\\nBest regards,\\nEtsuro Fujita\\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, 5 Jan 2017 21:21:56 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 5, 2017 at 5:51 PM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> On 2017/01/05 21:11, Ashutosh Bapat wrote:\\n>>\\n>> On Thu, Jan 5, 2017 at 5:14 PM, Etsuro Fujita\\n>> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>>>\\n>>> On 2017/01/03 17:28, Ashutosh Bapat wrote:\\n>>>>\\n>>>> In build_subquery_tlists(), why don't we handle base relations?\\n>>>> + if (foreignrel->reloptkind != RELOPT_JOINREL)\\n>>>> + return;\\n>\\n>\\n>>> The reason for that is we don't need to handle the baserel cases; the\\n>>> tlist\\n>>> for a base relation, if needed, would be created while recursing into a\\n>>> join\\n>>> relation that joins the base relation to other base/join relation.\\n>\\n>\\n>> Right. Sorry, I misunderstood the code. May be a comment would help.\\n>\\n>\\n> Will add the comment.\\n>\\n>>>> Also, in this function, if fpinfo->tlist is already set, why do we want\\n>>>> to\\n>>>> build it again?\\n>\\n>\\n>>> When this function gets called, fpinfo->tlist isn't set for any base or\\n>>> join\\n>>> relation that needs to build the tlist, so we always need to build it for\\n>>> each such relation.\\n>\\n>\\n>> IIUC, for a relation with use_remote_estimates we will deparse the\\n>> query twice and will build the targetlist twice.\\n>\\n>\\n> That's right. We could avoid the duplicate work the way you proposed, but I\\n> was thinking to leave that for another patch. Should we do that in this\\n> patch?\\n\\nIf you are agree that the change is needed, it's better to do it in\\nthis patch itself if we can, instead of a one liner patch.\\n>\\n>>>> In build_tlist_to_deparse(), if fpinfo->tlist for the given relation is\\n>>>> set, we\\n>>>> should just return it rather than constructing it again.\\n>\\n>\\n>>> In that function we wouldn't have such cases for base or join relations\\n>>> needing the tlist.\\n>\\n>\\n>> Same explanation as above.\\n>\\n>\\n> Will revise if it's better to do that in this patch.\\n\\nThanks.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 5 Jan 2017 18:08:17 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2017/01/05 21:38, Ashutosh Bapat wrote:\\n> On Thu, Jan 5, 2017 at 5:51 PM, Etsuro Fujita\\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>> On 2017/01/05 21:11, Ashutosh Bapat wrote:\\n\\n>>>> On 2017/01/03 17:28, Ashutosh Bapat wrote:\\n>>>>> Also, in this function, if fpinfo->tlist is already set, why do we want\\n>>>>> to\\n>>>>> build it again?\\n\\n>>> IIUC, for a relation with use_remote_estimates we will deparse the\\n>>> query twice and will build the targetlist twice.\\n\\n>> That's right. We could avoid the duplicate work the way you proposed, but I\\n>> was thinking to leave that for another patch. Should we do that in this\\n>> patch?\\n\\n> If you are agree that the change is needed, it's better to do it in\\n> this patch itself if we can, instead of a one liner patch.\\n\\nWhile working on this, I noticed a weird case. Consider:\\n\\npostgres=# explain verbose select 1 from ft1 left join ft2 on (ft1.a = \\nft2.a) inner join test on (true);\\n QUERY PLAN\\n-------------------------------------------------------------------------------------------------\\n Nested Loop (cost=100.00..103.06 rows=1 width=4)\\n Output: 1\\n -> Foreign Scan (cost=100.00..102.04 rows=1 width=0)\\n Relations: (public.ft1) LEFT JOIN (public.ft2)\\n Remote SQL: SELECT NULL FROM (public.t1 r1 LEFT JOIN public.t2 \\nr2 ON (((r1.a = r2.a))))\\n -> Seq Scan on public.test (cost=0.00..1.01 rows=1 width=0)\\n Output: test.a, test.b\\n(7 rows)\\n\\nIn this case the fpinfo->tlist of the foreign join is NIL, so whether or \\nnot the tlist is already built cannot be discriminated by the \\nfpinfo->tlist. We might need another flag to show that the tlist has \\nbeen built already. Since this is an existing issue and we would need \\nto give careful thought to this, so I'd like to leave this for another \\npatch.\\n\\nBest regards,\\nEtsuro Fujita\\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\": \"Wed, 11 Jan 2017 21:15:34 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 11, 2017 at 5:45 PM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> On 2017/01/05 21:38, Ashutosh Bapat wrote:\\n>>\\n>> On Thu, Jan 5, 2017 at 5:51 PM, Etsuro Fujita\\n>> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>>>\\n>>> On 2017/01/05 21:11, Ashutosh Bapat wrote:\\n>\\n>\\n>>>>> On 2017/01/03 17:28, Ashutosh Bapat wrote:\\n>>>>>>\\n>>>>>> Also, in this function, if fpinfo->tlist is already set, why do we\\n>>>>>> want\\n>>>>>> to\\n>>>>>> build it again?\\n>\\n>\\n>>>> IIUC, for a relation with use_remote_estimates we will deparse the\\n>>>> query twice and will build the targetlist twice.\\n>\\n>\\n>>> That's right. We could avoid the duplicate work the way you proposed,\\n>>> but I\\n>>> was thinking to leave that for another patch. Should we do that in this\\n>>> patch?\\n>\\n>\\n>> If you are agree that the change is needed, it's better to do it in\\n>> this patch itself if we can, instead of a one liner patch.\\n>\\n>\\n> While working on this, I noticed a weird case. Consider:\\n>\\n> postgres=# explain verbose select 1 from ft1 left join ft2 on (ft1.a =\\n> ft2.a) inner join test on (true);\\n> QUERY PLAN\\n> -------------------------------------------------------------------------------------------------\\n> Nested Loop (cost=100.00..103.06 rows=1 width=4)\\n> Output: 1\\n> -> Foreign Scan (cost=100.00..102.04 rows=1 width=0)\\n> Relations: (public.ft1) LEFT JOIN (public.ft2)\\n> Remote SQL: SELECT NULL FROM (public.t1 r1 LEFT JOIN public.t2 r2\\n> ON (((r1.a = r2.a))))\\n> -> Seq Scan on public.test (cost=0.00..1.01 rows=1 width=0)\\n> Output: test.a, test.b\\n> (7 rows)\\n>\\n> In this case the fpinfo->tlist of the foreign join is NIL, so whether or not\\n> the tlist is already built cannot be discriminated by the fpinfo->tlist. We\\n> might need another flag to show that the tlist has been built already.\\n> Since this is an existing issue and we would need to give careful thought to\\n> this, so I'd like to leave this for another patch.\\n\\nI think in that case, relation's targetlist will also be NIL or\\ncontain no Var node. It wouldn't be expensive to build it again and\\nagain. That's better than maintaining a new flag. This is just a\\nsuggestion, for an additional check, you might want to check\\nrel->reltarget->exprs for NIL. But I think we don't need it.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 12 Jan 2017 14:55:01 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2017/01/12 18:25, Ashutosh Bapat wrote:\\n> On Wed, Jan 11, 2017 at 5:45 PM, Etsuro Fujita\\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n\\n>>>> On 2017/01/05 21:11, Ashutosh Bapat wrote:\\n>>>>> IIUC, for a relation with use_remote_estimates we will deparse the\\n>>>>> query twice and will build the targetlist twice.\\n\\n>> While working on this, I noticed a weird case. Consider:\\n>>\\n>> postgres=# explain verbose select 1 from ft1 left join ft2 on (ft1.a =\\n>> ft2.a) inner join test on (true);\\n>> QUERY PLAN\\n>> -------------------------------------------------------------------------------------------------\\n>> Nested Loop (cost=100.00..103.06 rows=1 width=4)\\n>> Output: 1\\n>> -> Foreign Scan (cost=100.00..102.04 rows=1 width=0)\\n>> Relations: (public.ft1) LEFT JOIN (public.ft2)\\n>> Remote SQL: SELECT NULL FROM (public.t1 r1 LEFT JOIN public.t2 r2\\n>> ON (((r1.a = r2.a))))\\n>> -> Seq Scan on public.test (cost=0.00..1.01 rows=1 width=0)\\n>> Output: test.a, test.b\\n>> (7 rows)\\n>>\\n>> In this case the fpinfo->tlist of the foreign join is NIL, so whether or not\\n>> the tlist is already built cannot be discriminated by the fpinfo->tlist. We\\n>> might need another flag to show that the tlist has been built already.\\n>> Since this is an existing issue and we would need to give careful thought to\\n>> this, so I'd like to leave this for another patch.\\n\\n> I think in that case, relation's targetlist will also be NIL or\\n> contain no Var node. It wouldn't be expensive to build it again and\\n> again. That's better than maintaining a new flag.\\n\\nI think that's ugly. A more clean way I'm thinking is: (1) in \\npostgresGetForeignJoinPaths(), create a tlist by \\nbuild_tlist_to_deparse() and save it in fpinfo->tlist before \\nestimate_path_cost_size() if use_remote_estimates=true, (2) in \\nestimate_path_cost_size(), use the fpinfo->tlist if \\nuse_remote_estimates=true, and (3) in postgresGetForeignPlan(), use the \\nfpinfo->tlist as the fdw_scan_tlist if use_remote_estimates=true, \\notherwise create a tlist as the fdw_scan_tlist by \\nbuild_tlist_to_deparse(), like the attached.\\n\\nWhat do you think about that?\\n\\nAnother change is: I simplified build_tlist_to_deparse() a bit and put \\nthat in postgres_fdw.c because that is used only in postgres_fdw.c.\\n\\nI still think we should discuss this separately because this is an \\nexisting issue and that makes it easy to review the patch. If the \\nattached is the right way to go, I'll update the join-pushdown patch on \\ntop of the patch.\\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\": \"Fri, 13 Jan 2017 16:00:19 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 13, 2017 at 12:30 PM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> On 2017/01/12 18:25, Ashutosh Bapat wrote:\\n>>\\n>> On Wed, Jan 11, 2017 at 5:45 PM, Etsuro Fujita\\n>> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>\\n>\\n>>>>> On 2017/01/05 21:11, Ashutosh Bapat wrote:\\n>>>>>>\\n>>>>>> IIUC, for a relation with use_remote_estimates we will deparse the\\n>>>>>> query twice and will build the targetlist twice.\\n>\\n>\\n>>> While working on this, I noticed a weird case. Consider:\\n>>>\\n>>> postgres=# explain verbose select 1 from ft1 left join ft2 on (ft1.a =\\n>>> ft2.a) inner join test on (true);\\n>>> QUERY PLAN\\n>>>\\n>>> -------------------------------------------------------------------------------------------------\\n>>> Nested Loop (cost=100.00..103.06 rows=1 width=4)\\n>>> Output: 1\\n>>> -> Foreign Scan (cost=100.00..102.04 rows=1 width=0)\\n>>> Relations: (public.ft1) LEFT JOIN (public.ft2)\\n>>> Remote SQL: SELECT NULL FROM (public.t1 r1 LEFT JOIN public.t2\\n>>> r2\\n>>> ON (((r1.a = r2.a))))\\n>>> -> Seq Scan on public.test (cost=0.00..1.01 rows=1 width=0)\\n>>> Output: test.a, test.b\\n>>> (7 rows)\\n>>>\\n>>> In this case the fpinfo->tlist of the foreign join is NIL, so whether or\\n>>> not\\n>>> the tlist is already built cannot be discriminated by the fpinfo->tlist.\\n>>> We\\n>>> might need another flag to show that the tlist has been built already.\\n>>> Since this is an existing issue and we would need to give careful thought\\n>>> to\\n>>> this, so I'd like to leave this for another patch.\\n>\\n>\\n>> I think in that case, relation's targetlist will also be NIL or\\n>> contain no Var node. It wouldn't be expensive to build it again and\\n>> again. That's better than maintaining a new flag.\\n>\\n>\\n> I think that's ugly. A more clean way I'm thinking is: (1) in\\n> postgresGetForeignJoinPaths(), create a tlist by build_tlist_to_deparse()\\n> and save it in fpinfo->tlist before estimate_path_cost_size() if\\n> use_remote_estimates=true, (2) in estimate_path_cost_size(), use the\\n> fpinfo->tlist if use_remote_estimates=true, and (3) in\\n> postgresGetForeignPlan(), use the fpinfo->tlist as the fdw_scan_tlist if\\n> use_remote_estimates=true, otherwise create a tlist as the fdw_scan_tlist by\\n> build_tlist_to_deparse(), like the attached.\\n>\\n> What do you think about that?\\n>\\n> Another change is: I simplified build_tlist_to_deparse() a bit and put that\\n> in postgres_fdw.c because that is used only in postgres_fdw.c.\\n>\\n> I still think we should discuss this separately because this is an existing\\n> issue and that makes it easy to review the patch. If the attached is the\\n> right way to go, I'll update the join-pushdown patch on top of the patch.\\n\\nI don't think it's right to assume that the targetlist will be\\navailable only when use_remote_estimate is true; for grouping it's\\ncertainly not.\\n\\nBut I don't see this discussion getting anywhere. I will leave it to\\nthe committer's judgement. I think we should pick up your patch on\\n27th December, update the comment per your mail on 5th Jan. I will\\nreview it once and list down the things left to committer's judgement.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 27 Jan 2017 16:34:06 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2017/01/27 20:04, Ashutosh Bapat wrote:\\n> On Fri, Jan 13, 2017 at 12:30 PM, Etsuro Fujita\\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>> A more clean way I'm thinking is: (1) in\\n>> postgresGetForeignJoinPaths(), create a tlist by build_tlist_to_deparse()\\n>> and save it in fpinfo->tlist before estimate_path_cost_size() if\\n>> use_remote_estimates=true, (2) in estimate_path_cost_size(), use the\\n>> fpinfo->tlist if use_remote_estimates=true, and (3) in\\n>> postgresGetForeignPlan(), use the fpinfo->tlist as the fdw_scan_tlist if\\n>> use_remote_estimates=true, otherwise create a tlist as the fdw_scan_tlist by\\n>> build_tlist_to_deparse(), like the attached.\\n>>\\n>> What do you think about that?\\n>>\\n>> Another change is: I simplified build_tlist_to_deparse() a bit and put that\\n>> in postgres_fdw.c because that is used only in postgres_fdw.c.\\n>>\\n>> I still think we should discuss this separately because this is an existing\\n>> issue and that makes it easy to review the patch. If the attached is the\\n>> right way to go, I'll update the join-pushdown patch on top of the patch.\\n\\n> I don't think it's right to assume that the targetlist will be\\n> available only when use_remote_estimate is true; for grouping it's\\n> certainly not.\\n\\nMy explanation was not enough. Sorry about that. My proposal described \\nabove was for join relations, not upper relations. We build the \\ntargetlist of an upper relation during postgresGetForeignUpperPaths, so \\nfor grouping I think we should use that targetlist in \\nestimate_path_cost_size and postgresGetForeignPlan. The patch was \\ncreated that way.\\n\\n> But I don't see this discussion getting anywhere. I will leave it to\\n> the committer's judgement.\\n\\nI'm fine with that.\\n\\n> I think we should pick up your patch on\\n> 27th December, update the comment per your mail on 5th Jan. I will\\n> review it once and list down the things left to committer's judgement.\\n\\nSorry, I started thinking we went in the wrong direction. I added to \\ndeparseSelectStmtForRel build_subquery_tlists, which creates a tlist for \\neach subquery present in a given join tree prior to deparsing a whole \\nremote query. But that's nothing but an overhead; we need to create a \\ntlist for the top-level query because we use it as fdw_scan_tlist, but \\nnot for subqueries, and we need to create retrieved_attrs for the \\ntop-level query while deparsing the targetlist in \\ndeparseExplicitTargetList, but not for subqueries. So, we should need \\nsome work to avoid such a useless overhead.\\n\\nBest regards,\\nEtsuro Fujita\\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, 27 Jan 2017 21:25:49 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2017/01/27 21:25, Etsuro Fujita wrote:\\n> On 2017/01/27 20:04, Ashutosh Bapat wrote:\\n>> I think we should pick up your patch on\\n>> 27th December, update the comment per your mail on 5th Jan. I will\\n>> review it once and list down the things left to committer's judgement.\\n\\n> Sorry, I started thinking we went in the wrong direction. I added to\\n> deparseSelectStmtForRel build_subquery_tlists, which creates a tlist for\\n> each subquery present in a given join tree prior to deparsing a whole\\n> remote query. But that's nothing but an overhead; we need to create a\\n> tlist for the top-level query because we use it as fdw_scan_tlist, but\\n> not for subqueries, and we need to create retrieved_attrs for the\\n> top-level query while deparsing the targetlist in\\n> deparseExplicitTargetList, but not for subqueries. So, we should need\\n> some work to avoid such a useless overhead.\\n\\nI think we can avoid the useless overhead by adding a new function, \\ndeparseSubqueryTargetList, that deparses expressions in the given \\nrelation's reltarget, not the tlist, as a SELECT clause of the subquery \\nrepresenting the given relation. That would also allow us to make the \\n1-to-1 relationship between the subquery output columns and their \\naliases more explicit, which was your original comment. Please find \\nattached the new version. (The patch doesn't need the patch to avoid \\nduplicate construction of the tlist, discussed upthread.)\\n\\nOther changes:\\n* I went back to make_outerrel_subquery and make_innerrel_subquery, \\nwhich are flags to indicate whether to deparse the input relations as \\nsubqueries. is_subquery_rel would work well for handling the cases of \\nfull joins with restrictions on the input relations, but I noticed that \\nthat wouldn't work well when extending to handle the cases where we \\ndeparse the input relations as subqueries to evaluate PHVs remotely.\\n* Since appendSubqueryAlias in the previous patch is pretty small, I \\nincluded the code into deparseRangeTblRef.\\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, 30 Jan 2017 20:30:51 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 30, 2017 at 5:00 PM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> On 2017/01/27 21:25, Etsuro Fujita wrote:\\n>>\\n>> On 2017/01/27 20:04, Ashutosh Bapat wrote:\\n>>>\\n>>> I think we should pick up your patch on\\n>>> 27th December, update the comment per your mail on 5th Jan. I will\\n>>> review it once and list down the things left to committer's judgement.\\n>\\n>\\n>> Sorry, I started thinking we went in the wrong direction. I added to\\n>> deparseSelectStmtForRel build_subquery_tlists, which creates a tlist for\\n>> each subquery present in a given join tree prior to deparsing a whole\\n>> remote query. But that's nothing but an overhead; we need to create a\\n>> tlist for the top-level query because we use it as fdw_scan_tlist, but\\n>> not for subqueries, and we need to create retrieved_attrs for the\\n>> top-level query while deparsing the targetlist in\\n>> deparseExplicitTargetList, but not for subqueries. So, we should need\\n>> some work to avoid such a useless overhead.\\n>\\n>\\n> I think we can avoid the useless overhead by adding a new function,\\n> deparseSubqueryTargetList, that deparses expressions in the given relation's\\n> reltarget, not the tlist, as a SELECT clause of the subquery representing\\n> the given relation. That would also allow us to make the 1-to-1\\n> relationship between the subquery output columns and their aliases more\\n> explicit, which was your original comment. Please find attached the new\\n> version. (The patch doesn't need the patch to avoid duplicate construction\\n> of the tlist, discussed upthread.)\\n\\nI have not looked at the patch, but the reason we use a tlist instead\\nof list of expressions is because fdw_scan_tlist is expected in that\\nform and we don't want two different representations one to deparse\\nSELECT and one to interpret results from the foreign server. What you\\ndescribe above seems to introduce exactly that hazard.\\n\\n>\\n> Other changes:\\n> * I went back to make_outerrel_subquery and make_innerrel_subquery, which\\n> are flags to indicate whether to deparse the input relations as subqueries.\\n> is_subquery_rel would work well for handling the cases of full joins with\\n> restrictions on the input relations, but I noticed that that wouldn't work\\n> well when extending to handle the cases where we deparse the input relations\\n> as subqueries to evaluate PHVs remotely.\\n\\nI had objected to using a single variable instead of two previously\\nand you had argued against it in [1]. There you had mentioned that PHV\\ndoesn't need two variables, but now you are taking the other stand,\\nwithout any apparent reason. Can you please clarify it?\\n\\n[1] https://www.postgresql.org/message-id/1eb58ee4-8ffa-7c40-1229-c8973e6498ea@lab.ntt.co.jp\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database 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, 30 Jan 2017 17:35:18 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2017/01/30 21:05, Ashutosh Bapat wrote:\\n> On Mon, Jan 30, 2017 at 5:00 PM, Etsuro Fujita\\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>> On 2017/01/27 21:25, Etsuro Fujita wrote:\\n>>> Sorry, I started thinking we went in the wrong direction. I added to\\n>>> deparseSelectStmtForRel build_subquery_tlists, which creates a tlist for\\n>>> each subquery present in a given join tree prior to deparsing a whole\\n>>> remote query. But that's nothing but an overhead; we need to create a\\n>>> tlist for the top-level query because we use it as fdw_scan_tlist, but\\n>>> not for subqueries, and we need to create retrieved_attrs for the\\n>>> top-level query while deparsing the targetlist in\\n>>> deparseExplicitTargetList, but not for subqueries. So, we should need\\n>>> some work to avoid such a useless overhead.\\n\\n>> I think we can avoid the useless overhead by adding a new function,\\n>> deparseSubqueryTargetList, that deparses expressions in the given relation's\\n>> reltarget, not the tlist, as a SELECT clause of the subquery representing\\n>> the given relation. That would also allow us to make the 1-to-1\\n>> relationship between the subquery output columns and their aliases more\\n>> explicit, which was your original comment. Please find attached the new\\n>> version. (The patch doesn't need the patch to avoid duplicate construction\\n>> of the tlist, discussed upthread.)\\n\\n> I have not looked at the patch, but the reason we use a tlist instead\\n> of list of expressions is because fdw_scan_tlist is expected in that\\n> form\\n\\nActually, we wouldn't need that to deparse a remote join query; a list \\nof expressions would be enough.\\n\\n> and we don't want two different representations one to deparse\\n> SELECT and one to interpret results from the foreign server. What you\\n> describe above seems to introduce exactly that hazard.\\n\\nI agree with you to some extent, BUT:\\n* I don't think it's a good idea to create a tlist for each base/join \\nrelation that is deparsed as a subquery, to just avoid that hazard. As \\nI said above, that's nothing but an overhead.\\n* I think we would need to have two different representations for at \\nleast base relations; we use fpinfo->attrs_used to deparse a simple \\nforeign table scan query for a base relation, but when deparsing the \\nbase relation as a subquery, we would need to use the list of \\nexpressions in the base relation's reltarget, to deparse a SELECT clause \\nof the subquery, because we need to deparse a whole-row reference to the \\nbase relation as ROW, not all the user columns expanded, as shown in \\nthis extracted from the regression tests in the patch:\\n\\n+ EXPLAIN (VERBOSE, COSTS OFF)\\n+ SELECT t1.c1, ss.a, ss.b FROM (SELECT c1 FROM \\\"S 1\\\".\\\"T 3\\\" WHERE c1 = \\n50) t1 INNER JOIN (SELECT t2.c1, t3.c1 FROM (SELECT c1 FROM ft4 WHERE c1 \\nbetween 50 and 60) t2 FULL JOIN (SELECT c1 FROM ft5 WHERE c1 between 50 \\nand 60) t3 ON (t2.c1 = t3.c1) WHERE t2.c1 IS NULL OR t2.c1 IS NOT NULL) \\nss(a, b) ON (TRUE) ORDER BY t1.c1, ss.a, ss.b FOR UPDATE OF t1;\\n+ \\n \\n QUERY PLAN \\n \\n \\n\\n+ \\n----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------\\n+ LockRows\\n+ Output: \\\"T 3\\\".c1, ft4.c1, ft5.c1, \\\"T 3\\\".ctid, ft4.*, ft5.*\\n+ -> Nested Loop\\n+ Output: \\\"T 3\\\".c1, ft4.c1, ft5.c1, \\\"T 3\\\".ctid, ft4.*, ft5.*\\n+ -> Foreign Scan\\n+ Output: ft4.c1, ft4.*, ft5.c1, ft5.*\\n+ Relations: (public.ft4) FULL JOIN (public.ft5)\\n+ Remote SQL: SELECT s8.c1, s8.c2, s9.c1, s9.c2 FROM \\n((SELECT c1, ROW(c1, c2, c3) FROM \\\"S 1\\\".\\\"T 3\\\" WHERE ((c1 >= 50)) AND \\n((c1 <= 60))) s8(c1, c2) FULL JOIN (SELECT c1, ROW(c1, c2, c3) FROM \\\"S \\n1\\\".\\\"T 4\\\" WHERE ((c1 >= 50)) AND ((c1 <= 60))) s9(c1, c2) ON (((s8.c1 = \\ns9.c1)))) WHERE (((s8.c1 IS NULL) OR (s8.c1 IS NOT NULL))) ORDER BY \\ns8.c1 ASC NULLS LAST, s9.c1 ASC NULLS LAST\\n+ -> Hash Full Join\\n+ Output: ft4.c1, ft4.*, ft5.c1, ft5.*\\n+ Hash Cond: (ft4.c1 = ft5.c1)\\n+ Filter: ((ft4.c1 IS NULL) OR (ft4.c1 IS NOT NULL))\\n+ -> Foreign Scan on public.ft4\\n+ Output: ft4.c1, ft4.*\\n+ Remote SQL: SELECT c1, c2, c3 FROM \\\"S 1\\\".\\\"T \\n3\\\" WHERE ((c1 >= 50)) AND ((c1 <= 60))\\n+ -> Hash\\n+ Output: ft5.c1, ft5.*\\n+ -> Foreign Scan on public.ft5\\n+ Output: ft5.c1, ft5.*\\n+ Remote SQL: SELECT c1, c2, c3 FROM \\\"S \\n1\\\".\\\"T 4\\\" WHERE ((c1 >= 50)) AND ((c1 <= 60))\\n+ -> Materialize\\n+ Output: \\\"T 3\\\".c1, \\\"T 3\\\".ctid\\n+ -> Seq Scan on \\\"S 1\\\".\\\"T 3\\\"\\n+ Output: \\\"T 3\\\".c1, \\\"T 3\\\".ctid\\n+ Filter: (\\\"T 3\\\".c1 = 50)\\n+ (25 rows)\\n\\n>> Other changes:\\n>> * I went back to make_outerrel_subquery and make_innerrel_subquery, which\\n>> are flags to indicate whether to deparse the input relations as subqueries.\\n>> is_subquery_rel would work well for handling the cases of full joins with\\n>> restrictions on the input relations, but I noticed that that wouldn't work\\n>> well when extending to handle the cases where we deparse the input relations\\n>> as subqueries to evaluate PHVs remotely.\\n\\n> I had objected to using a single variable instead of two previously\\n> and you had argued against it in [1]. There you had mentioned that PHV\\n> doesn't need two variables, but now you are taking the other stand,\\n> without any apparent reason. Can you please clarify it?\\n\\nSorry, I missed some cases. Consider the join {A, B, C} satisfying the \\nidentity 3 in optimizer/README, ie,\\n\\n (A leftjoin B on (Pab)) leftjoin C on (Pbc)\\n = A leftjoin (B leftjoin C on (Pbc)) on (Pab)\\n\\nwhere predicate Pbc fails for all-null B rows. Assume that B has a PHV \\nin the relation's reltarget. While considering 2-way joins, \\nforeign_join_ok would determine to deparse B as a subquery and hence set \\nthe B's is_subquery_rel to true for the join relation {A, B}. However, \\nif the planner selects the join order {A leftjoin (B leftjoin C on \\n(Pbc)) on (Pab)} as the cheapest path for the join {A, B, C}, we would \\ndeparse B as a subquery according to the is_subquery_rel flag while \\ncreating the FROM clause entry for the join relation {B, C}. That \\nwouldn't look good. That would be logically correct, though. To avoid \\nthis, I'd like to go back to the two variables previously proposed.\\n\\nBest regards,\\nEtsuro Fujita\\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, 31 Jan 2017 17:50:51 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 30, 2017 at 8:30 PM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> Other changes:\\n> * I went back to make_outerrel_subquery and make_innerrel_subquery, which\\n> are flags to indicate whether to deparse the input relations as subqueries.\\n> is_subquery_rel would work well for handling the cases of full joins with\\n> restrictions on the input relations, but I noticed that that wouldn't work\\n> well when extending to handle the cases where we deparse the input relations\\n> as subqueries to evaluate PHVs remotely.\\n> * Since appendSubqueryAlias in the previous patch is pretty small, I\\n> included the code into deparseRangeTblRef.\\n\\nThe patch is very fresh, so moved to CF 2017-03.\\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, 1 Feb 2017 13:51:03 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 1/30/17 6:30 AM, Etsuro Fujita wrote:\\n> On 2017/01/27 21:25, Etsuro Fujita wrote:\\n>> On 2017/01/27 20:04, Ashutosh Bapat wrote:\\n>>> I think we should pick up your patch on\\n>>> 27th December, update the comment per your mail on 5th Jan. I will\\n>>> review it once and list down the things left to committer's judgement.\\n> \\n>> Sorry, I started thinking we went in the wrong direction. I added to\\n>> deparseSelectStmtForRel build_subquery_tlists, which creates a tlist for\\n>> each subquery present in a given join tree prior to deparsing a whole\\n>> remote query. But that's nothing but an overhead; we need to create a\\n>> tlist for the top-level query because we use it as fdw_scan_tlist, but\\n>> not for subqueries, and we need to create retrieved_attrs for the\\n>> top-level query while deparsing the targetlist in\\n>> deparseExplicitTargetList, but not for subqueries. So, we should need\\n>> some work to avoid such a useless overhead.\\n> \\n> I think we can avoid the useless overhead by adding a new function,\\n> deparseSubqueryTargetList, that deparses expressions in the given\\n> relation's reltarget, not the tlist, as a SELECT clause of the subquery\\n> representing the given relation. That would also allow us to make the\\n> 1-to-1 relationship between the subquery output columns and their\\n> aliases more explicit, which was your original comment. Please find\\n> attached the new version. (The patch doesn't need the patch to avoid\\n> duplicate construction of the tlist, discussed upthread.)\\n> \\n> Other changes:\\n> * I went back to make_outerrel_subquery and make_innerrel_subquery,\\n> which are flags to indicate whether to deparse the input relations as\\n> subqueries. is_subquery_rel would work well for handling the cases of\\n> full joins with restrictions on the input relations, but I noticed that\\n> that wouldn't work well when extending to handle the cases where we\\n> deparse the input relations as subqueries to evaluate PHVs remotely.\\n> * Since appendSubqueryAlias in the previous patch is pretty small, I\\n> included the code into deparseRangeTblRef.\\n\\nThis patch does not apply cleanly at cccbdde:\\n\\n$ patch -p1 < ../other/postgres-fdw-subquery-support-v15.patch\\npatching file contrib/postgres_fdw/deparse.c\\nHunk #11 succeeded at 1371 (offset -3 lines).\\nHunk #12 succeeded at 1419 (offset -3 lines).\\nHunk #13 succeeded at 1486 (offset -3 lines).\\nHunk #14 succeeded at 2186 (offset -3 lines).\\nHunk #15 succeeded at 3082 (offset -3 lines).\\npatching file contrib/postgres_fdw/expected/postgres_fdw.out\\npatching file contrib/postgres_fdw/postgres_fdw.c\\nHunk #1 succeeded at 669 (offset 1 line).\\nHunk #2 succeeded at 1245 (offset -1 lines).\\nHunk #3 succeeded at 2557 (offset -1 lines).\\nHunk #4 succeeded at 4157 (offset 3 lines).\\nHunk #5 succeeded at 4183 (offset 3 lines).\\nHunk #6 succeeded at 4212 (offset 3 lines).\\nHunk #7 succeeded at 4315 (offset 3 lines).\\npatching file contrib/postgres_fdw/postgres_fdw.h\\npatching file contrib/postgres_fdw/sql/postgres_fdw.sql\\n\\nSince these are just offsets I'll leave the patch as \\\"Needs review\\\" but\\nan updated patch would be appreciated.\\n\\nThanks,\\n-- \\n-David\\ndavid@pgmasters.net\\n\\n\\n-- \\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, 16 Mar 2017 11:46:22 -0400\",\n \"msg_from\": \"David Steele <david@pgmasters.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 16, 2017 at 11:46 AM, David Steele <david@pgmasters.net> wrote:\\n> $ patch -p1 < ../other/postgres-fdw-subquery-support-v15.patch\\n> patching file contrib/postgres_fdw/deparse.c\\n> Hunk #11 succeeded at 1371 (offset -3 lines).\\n> Hunk #12 succeeded at 1419 (offset -3 lines).\\n> Hunk #13 succeeded at 1486 (offset -3 lines).\\n> Hunk #14 succeeded at 2186 (offset -3 lines).\\n> Hunk #15 succeeded at 3082 (offset -3 lines).\\n> patching file contrib/postgres_fdw/expected/postgres_fdw.out\\n> patching file contrib/postgres_fdw/postgres_fdw.c\\n> Hunk #1 succeeded at 669 (offset 1 line).\\n> Hunk #2 succeeded at 1245 (offset -1 lines).\\n> Hunk #3 succeeded at 2557 (offset -1 lines).\\n> Hunk #4 succeeded at 4157 (offset 3 lines).\\n> Hunk #5 succeeded at 4183 (offset 3 lines).\\n> Hunk #6 succeeded at 4212 (offset 3 lines).\\n> Hunk #7 succeeded at 4315 (offset 3 lines).\\n> patching file contrib/postgres_fdw/postgres_fdw.h\\n> patching file contrib/postgres_fdw/sql/postgres_fdw.sql\\n>\\n> Since these are just offsets I'll leave the patch as \\\"Needs review\\\" but\\n> an updated patch would be appreciated.\\n\\nI don't think that's really needed. Offsets don't hurt anything.\\nEven fuzz is OK. As long as the hunks are applying, I think it's\\nfine.\\n\\nIncidentally, I'm reading through this one now.\\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, 16 Mar 2017 12:48:43 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 16, 2017 at 12:48 PM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Thu, Mar 16, 2017 at 11:46 AM, David Steele <david@pgmasters.net> wrote:\\n>> $ patch -p1 < ../other/postgres-fdw-subquery-support-v15.patch\\n>> patching file contrib/postgres_fdw/deparse.c\\n>> Hunk #11 succeeded at 1371 (offset -3 lines).\\n>> Hunk #12 succeeded at 1419 (offset -3 lines).\\n>> Hunk #13 succeeded at 1486 (offset -3 lines).\\n>> Hunk #14 succeeded at 2186 (offset -3 lines).\\n>> Hunk #15 succeeded at 3082 (offset -3 lines).\\n>> patching file contrib/postgres_fdw/expected/postgres_fdw.out\\n>> patching file contrib/postgres_fdw/postgres_fdw.c\\n>> Hunk #1 succeeded at 669 (offset 1 line).\\n>> Hunk #2 succeeded at 1245 (offset -1 lines).\\n>> Hunk #3 succeeded at 2557 (offset -1 lines).\\n>> Hunk #4 succeeded at 4157 (offset 3 lines).\\n>> Hunk #5 succeeded at 4183 (offset 3 lines).\\n>> Hunk #6 succeeded at 4212 (offset 3 lines).\\n>> Hunk #7 succeeded at 4315 (offset 3 lines).\\n>> patching file contrib/postgres_fdw/postgres_fdw.h\\n>> patching file contrib/postgres_fdw/sql/postgres_fdw.sql\\n>>\\n>> Since these are just offsets I'll leave the patch as \\\"Needs review\\\" but\\n>> an updated patch would be appreciated.\\n>\\n> I don't think that's really needed. Offsets don't hurt anything.\\n> Even fuzz is OK. As long as the hunks are applying, I think it's\\n> fine.\\n>\\n> Incidentally, I'm reading through this one now.\\n\\nAnd ... I don't see anything to complain about, so, committed.\\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, 16 Mar 2017 13:35:23 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n },\n {\n \"msg_contents\": \"On 2017/03/17 2:35, Robert Haas wrote:\\n> And ... I don't see anything to complain about, so, committed.\\n\\nThanks for committing, Robert! Thanks for reviewing, Ashutosh and David!\\n\\nBest regards,\\nEtsuro Fujita\\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 Mar 2017 18:42:49 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Push down more full joins in postgres_fdw\"\n }\n]","isConversational":false}}},{"rowIdx":42987,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"plpgsql.h defines a number of enums, but most of the code passes them\naround as ints. The attached patch updates structs and function\nprototypes to take enum types instead. This clarifies the struct\ndefinitions in plpgsql.h in particular.\n\nI didn't deal with the PLPGSQL_RC_* symbols, since they are only used in\npl_exec.c (could be moved there?), and it would bloat this patch.\n\n-- \nPeter Eisentraut 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":"Fri, 19 Aug 2016 10:56:20 -0400","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Make better use of existing enums in plpgsql"},{"msg_contents":"Hi\n\n2016-08-19 16:56 GMT+02:00 Peter Eisentraut <\npeter.eisentraut@2ndquadrant.com>:\n\n> plpgsql.h defines a number of enums, but most of the code passes them\n> around as ints. The attached patch updates structs and function\n> prototypes to take enum types instead. This clarifies the struct\n> definitions in plpgsql.h in particular.\n>\n\nI am sending review of this trivial patch.\n\n1. No problems with patching, compiling\n2. all regress tests passed\n3. There are not any visible change, so tests, docs are not necessary\n4. Using enum instead int is generally good idea\n\nI will mark this patch as ready for commiters.\n\n\n>\n> I didn't deal with the PLPGSQL_RC_* symbols, since they are only used in\n> pl_exec.c (could be moved there?), and it would bloat this patch.\n>\n\nIt is good idea. PLPGSQL_RC_* should be private in pl_exec.c\n\nRegards\n\nPavel\n\n\n>\n> --\n> Peter Eisentraut http://www.2ndQuadrant.com/\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\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>\n\nHi2016-08-19 16:56 GMT+02:00 Peter Eisentraut <peter.eisentraut@2ndquadrant.com>:plpgsql.h defines a number of enums, but most of the code passes them\naround as ints. The attached patch updates structs and function\nprototypes to take enum types instead. This clarifies the struct\ndefinitions in plpgsql.h in particular.I am sending review of this trivial patch.1. No problems with patching, compiling2. all regress tests passed3. There are not any visible change, so tests, docs are not necessary4. Using enum instead int is generally good ideaI will mark this patch as ready for commiters. \n\nI didn't deal with the PLPGSQL_RC_* symbols, since they are only used in\npl_exec.c (could be moved there?), and it would bloat this patch.It is good idea. PLPGSQL_RC_* should be private in pl_exec.cRegardsPavel \n\n--\nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\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, 4 Sep 2016 18:26:26 +0200","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":false,"msg_subject":"Re: Make better use of existing enums in plpgsql"},{"msg_contents":"On 9/4/16 12:26 PM, Pavel Stehule wrote:\n> I am sending review of this trivial patch.\n> \n> 1. No problems with patching, compiling\n> 2. all regress tests passed\n> 3. There are not any visible change, so tests, docs are not necessary\n> 4. Using enum instead int is generally good idea\n> \n> I will mark this patch as ready for commiters.\n\nCommitted, thanks.\n\n-- \nPeter Eisentraut 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":"Fri, 9 Sep 2016 10:23:59 -0400","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Make better use of existing enums in plpgsql"}],"string":"[\n {\n \"msg_contents\": \"plpgsql.h defines a number of enums, but most of the code passes them\\naround as ints. The attached patch updates structs and function\\nprototypes to take enum types instead. This clarifies the struct\\ndefinitions in plpgsql.h in particular.\\n\\nI didn't deal with the PLPGSQL_RC_* symbols, since they are only used in\\npl_exec.c (could be moved there?), and it would bloat this patch.\\n\\n-- \\nPeter Eisentraut 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\": \"Fri, 19 Aug 2016 10:56:20 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Make better use of existing enums in plpgsql\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2016-08-19 16:56 GMT+02:00 Peter Eisentraut <\\npeter.eisentraut@2ndquadrant.com>:\\n\\n> plpgsql.h defines a number of enums, but most of the code passes them\\n> around as ints. The attached patch updates structs and function\\n> prototypes to take enum types instead. This clarifies the struct\\n> definitions in plpgsql.h in particular.\\n>\\n\\nI am sending review of this trivial patch.\\n\\n1. No problems with patching, compiling\\n2. all regress tests passed\\n3. There are not any visible change, so tests, docs are not necessary\\n4. Using enum instead int is generally good idea\\n\\nI will mark this patch as ready for commiters.\\n\\n\\n>\\n> I didn't deal with the PLPGSQL_RC_* symbols, since they are only used in\\n> pl_exec.c (could be moved there?), and it would bloat this patch.\\n>\\n\\nIt is good idea. PLPGSQL_RC_* should be private in pl_exec.c\\n\\nRegards\\n\\nPavel\\n\\n\\n>\\n> --\\n> Peter Eisentraut http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\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>\\n\\nHi2016-08-19 16:56 GMT+02:00 Peter Eisentraut <peter.eisentraut@2ndquadrant.com>:plpgsql.h defines a number of enums, but most of the code passes them\\naround as ints. The attached patch updates structs and function\\nprototypes to take enum types instead. This clarifies the struct\\ndefinitions in plpgsql.h in particular.I am sending review of this trivial patch.1. No problems with patching, compiling2. all regress tests passed3. There are not any visible change, so tests, docs are not necessary4. Using enum instead int is generally good ideaI will mark this patch as ready for commiters. \\n\\nI didn't deal with the PLPGSQL_RC_* symbols, since they are only used in\\npl_exec.c (could be moved there?), and it would bloat this patch.It is good idea. PLPGSQL_RC_* should be private in pl_exec.cRegardsPavel \\n\\n--\\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n--\\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 Sep 2016 18:26:26 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Make better use of existing enums in plpgsql\"\n },\n {\n \"msg_contents\": \"On 9/4/16 12:26 PM, Pavel Stehule wrote:\\n> I am sending review of this trivial patch.\\n> \\n> 1. No problems with patching, compiling\\n> 2. all regress tests passed\\n> 3. There are not any visible change, so tests, docs are not necessary\\n> 4. Using enum instead int is generally good idea\\n> \\n> I will mark this patch as ready for commiters.\\n\\nCommitted, thanks.\\n\\n-- \\nPeter Eisentraut 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\": \"Fri, 9 Sep 2016 10:23:59 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Make better use of existing enums in plpgsql\"\n }\n]","isConversational":false}}},{"rowIdx":42988,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Use LEFT JOINs in some system views in case referenced row doesn't exist.\n\nIn particular, left join to pg_authid so that rows in pg_stat_activity\ndon't disappear if the session's owning user has been dropped.\nAlso convert a few joins to pg_database to left joins, in the same spirit,\nthough that case might be harder to hit. We were doing this in other\nviews already, so it was a bit inconsistent that these views didn't.\n\nOskari Saarenmaa, with some further tweaking by me\n\nDiscussion: <56E87CD8.60007@ohmu.fi>\n\nBranch\n------\nmaster\n\nDetails\n-------\nhttp://git.postgresql.org/pg/commitdiff/8299471c37fff0b0f5a777a12f920125310c0efe\n\nModified Files\n--------------\nsrc/backend/catalog/system_views.sql | 20 +++++++++-----------\nsrc/include/catalog/catversion.h | 2 +-\nsrc/test/regress/expected/rules.out | 20 +++++++++-----------\n3 files changed, 19 insertions(+), 23 deletions(-)\n\n\n-- \nSent via pgsql-committers mailing list (pgsql-committers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-committers\n","msg_date":"Fri, 19 Aug 2016 21:13:52 +0000","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"pgsql: Use LEFT JOINs in some system views in case referenced row\n doesn"},{"msg_contents":"On Sat, Aug 20, 2016 at 6:13 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Use LEFT JOINs in some system views in case referenced row doesn't exist.\n>\n> In particular, left join to pg_authid so that rows in pg_stat_activity\n> don't disappear if the session's owning user has been dropped.\n> Also convert a few joins to pg_database to left joins, in the same spirit,\n> though that case might be harder to hit. We were doing this in other\n> views already, so it was a bit inconsistent that these views didn't.\n\nThis change causes pg_stat_activity to report the \"bogus\" information about\nwalsenders as follows.\n\n$ ps aux | grep 75076\npostgres 75076 0.0 0.0 2608252 1364 ?? Ss 2:41AM\n0:00.07 postgres: wal sender process postgres [local] streaming\n0/3000140\n\n=# SELECT * FROM pg_stat_activity;\n-[ RECORD 1 ]----+--------------------------------\ndatid | 0\ndatname | (null)\npid | 75076\nusesysid | 10\nusename | postgres\napplication_name | sby1\nclient_addr | (null)\nclient_hostname | (null)\nclient_port | -1\nbackend_start | 2016-09-10 02:41:24.568313+09\nxact_start | (null)\nquery_start | (null)\nstate_change | 2016-09-10 02:41:24.570135+09\nwait_event_type | (null)\nwait_event | (null)\nstate | idle\nbackend_xid | (null)\nbackend_xmin | (null)\nquery |\n\nI think that this information is confusing and should not exist in\npg_stat_activity.\n\nRegards,\n\n-- \nFujii Masao\n\n\n-- \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, 10 Sep 2016 02:51:41 +0900","msg_from":"Fujii Masao <masao.fujii@gmail.com>","msg_from_op":false,"msg_subject":"Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system views in case\n referenced row doesn"},{"msg_contents":"Fujii Masao <masao.fujii@gmail.com> writes:\n> On Sat, Aug 20, 2016 at 6:13 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> Use LEFT JOINs in some system views in case referenced row doesn't exist.\n\n> This change causes pg_stat_activity to report the \"bogus\" information about\n> walsenders as follows.\n\nHmm ... but if we want to exclude walsenders from that view, surely we\nshould do so explicitly, not have it happen as a hidden side-effect\nof not being able to join them to pg_database.\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, 09 Sep 2016 13:59:24 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system views in\n case referenced row doesn"},{"msg_contents":"I wrote:\n> Fujii Masao <masao.fujii@gmail.com> writes:\n>> On Sat, Aug 20, 2016 at 6:13 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>>> Use LEFT JOINs in some system views in case referenced row doesn't exist.\n\n>> This change causes pg_stat_activity to report the \"bogus\" information about\n>> walsenders as follows.\n\n> Hmm ... but if we want to exclude walsenders from that view, surely we\n> should do so explicitly, not have it happen as a hidden side-effect\n> of not being able to join them to pg_database.\n\nI looked into this a little. There are at least three things we could\ndo here:\n\n1. Decide that showing walsenders is a good thing. I'm not really\nsure why it isn't -- for example, we could take the trouble to display\nthe current replication command as the walsender's activity.\n\n2. Suppress walsenders by the expedient of ignoring PGPROCs that have\nzero database ID. This would also ignore other process types that aren't\nconnected to a specific database. This is pretty backwards-compatible\nbecause it's equivalent to what used to happen implicitly via the inner\njoin in the view.\n\n3. Suppress walsenders by adding some new field to PgBackendStatus that\nidentifies them, and having pg_stat_get_activity() use that to filter.\n\n\nIt looks to me like even before the change Fujii-san complains of, it was\nonly generic walsenders that were hidden in the view. Database-specific\nwalsenders (am_db_walsender) would have a valid database OID in\nPgBackendStatus and therefore would pass the join. This makes me think\nthat option #1 is really the sanest alternative. While option #2 is\nthe most backwards-compatible, it's pretty hard to see why we would think\nthat showing only database-specific walsenders is a desirable behavior.\nIf we want to exclude walsenders altogether, we need option #3.\n\nThoughts?\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, 10 Sep 2016 12:51:02 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system views in\n case referenced row doesn"},{"msg_contents":"I wrote:\n> I looked into this a little. There are at least three things we could\n> do here:\n> 1. Decide that showing walsenders is a good thing. I'm not really\n> sure why it isn't -- for example, we could take the trouble to display\n> the current replication command as the walsender's activity.\n> 2. Suppress walsenders by the expedient of ignoring PGPROCs that have\n> zero database ID. This would also ignore other process types that aren't\n> connected to a specific database. This is pretty backwards-compatible\n> because it's equivalent to what used to happen implicitly via the inner\n> join in the view.\n> 3. Suppress walsenders by adding some new field to PgBackendStatus that\n> identifies them, and having pg_stat_get_activity() use that to filter.\n\nBTW, I had been thinking we could implement #2 or #3 relatively painlessly\nby adding filter logic inside pg_stat_get_activity(). However, I now\nnotice that that function is also used by the pg_stat_replication view,\nwhich joins its output to pg_stat_get_wal_senders(). That means we can\n*not* simply suppress walsenders in the function's output; if we want to\nfilter them away in pg_stat_activity, we have to change the view\ndefinition somehow, and probably the function's API as well. And that\nmeans we can't change the behavior without an initdb, which is kind of\nannoying to have to do after rc1.\n\nThe fact that the pg_stat_replication view does show walsender processes\nseems like possibly a reasonable argument for not showing them in\npg_stat_activity. But we'd have to do some rejiggering of the view\ndefinition to make that happen.\n\nDon't have a strong opinion which way it ought to work --- I could go with\neither the \"walsenders are a perfectly good activity\" position or the\n\"they should appear only in pg_stat_replication\" one. But the latter will\ntake an initdb to do, and I'm inclined to the position that it's too late\nfor that in 9.6. As I noted, it was already inconsistent for per-db\nwalsenders long before 9.6, without field complaints, so I'm having a hard\ntime seeing this as a critical 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, 10 Sep 2016 16:21:50 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system views in\n case referenced row doesn"},{"msg_contents":"On Sun, Sep 11, 2016 at 5:21 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> The fact that the pg_stat_replication view does show walsender processes\n> seems like possibly a reasonable argument for not showing them in\n> pg_stat_activity. But we'd have to do some rejiggering of the view\n> definition to make that happen.\n\nWe may actually had better show WAL sender processes in\npg_stat_activity. An argument in favor of that is the tracking of\nWaitEventSet events (or latches if you want).\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, 12 Sep 2016 10:00:50 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\n views in case referenced row doesn"},{"msg_contents":"Michael Paquier <michael.paquier@gmail.com> writes:\n> On Sun, Sep 11, 2016 at 5:21 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> The fact that the pg_stat_replication view does show walsender processes\n>> seems like possibly a reasonable argument for not showing them in\n>> pg_stat_activity. But we'd have to do some rejiggering of the view\n>> definition to make that happen.\n\n> We may actually had better show WAL sender processes in\n> pg_stat_activity. An argument in favor of that is the tracking of\n> WaitEventSet events (or latches if you want).\n\nAlso, walsenders count against MaxBackends don't they? So not showing\nthem could contribute to confusion about why an installation is hitting\nthe connection limit.\n\nIf we do keep them in the view, I would definitely vote for having them\nset their \"query\" fields to something that shows they're walsenders.\nIt's awfully late to be doing anything complicated there for 9.6,\nbut we could just set the query to \"walsender\" and plan to improve\non that in future releases.\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, 11 Sep 2016 21:13:05 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system views in\n case referenced row doesn"},{"msg_contents":"* Tom Lane (tgl@sss.pgh.pa.us) wrote:\n> Michael Paquier <michael.paquier@gmail.com> writes:\n> > On Sun, Sep 11, 2016 at 5:21 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> >> The fact that the pg_stat_replication view does show walsender processes\n> >> seems like possibly a reasonable argument for not showing them in\n> >> pg_stat_activity. But we'd have to do some rejiggering of the view\n> >> definition to make that happen.\n> \n> > We may actually had better show WAL sender processes in\n> > pg_stat_activity. An argument in favor of that is the tracking of\n> > WaitEventSet events (or latches if you want).\n> \n> Also, walsenders count against MaxBackends don't they? So not showing\n> them could contribute to confusion about why an installation is hitting\n> the connection limit.\n> \n> If we do keep them in the view, I would definitely vote for having them\n> set their \"query\" fields to something that shows they're walsenders.\n> It's awfully late to be doing anything complicated there for 9.6,\n> but we could just set the query to \"walsender\" and plan to improve\n> on that in future releases.\n\n+1\n\nThanks!\n\nStephen","msg_date":"Sun, 11 Sep 2016 21:16:46 -0400","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\n views in case referenced row doesn"},{"msg_contents":"On Mon, Sep 12, 2016 at 10:16 AM, Stephen Frost <sfrost@snowman.net> wrote:\n> * Tom Lane (tgl@sss.pgh.pa.us) wrote:\n>> Michael Paquier <michael.paquier@gmail.com> writes:\n>> > On Sun, Sep 11, 2016 at 5:21 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> >> The fact that the pg_stat_replication view does show walsender processes\n>> >> seems like possibly a reasonable argument for not showing them in\n>> >> pg_stat_activity. But we'd have to do some rejiggering of the view\n>> >> definition to make that happen.\n>>\n>> > We may actually had better show WAL sender processes in\n>> > pg_stat_activity. An argument in favor of that is the tracking of\n>> > WaitEventSet events (or latches if you want).\n>>\n>> Also, walsenders count against MaxBackends don't they? So not showing\n>> them could contribute to confusion about why an installation is hitting\n>> the connection limit.\n\nYes they are counted in MaxBackends. So that would help as well in\nlooking at connection limit issues.\n\n>> If we do keep them in the view, I would definitely vote for having them\n>> set their \"query\" fields to something that shows they're walsenders.\n>> It's awfully late to be doing anything complicated there for 9.6,\n>> but we could just set the query to \"walsender\" and plan to improve\n>> on that in future releases.\n>\n> +1\n\nIndeed, and the query field does not have much more meaning for a WAL\nsender. So I'd propose the attached for 9.6 and HEAD. I have thought\nabout reporting that to pgstat in StartReplication(), but as there is\nsome error handling there I'd think that WalSndLoop() is a better\nplace to call pgstat_report_activity, as per the attached.\n\nWe could have far more fancy verbose information to offer to the user,\nbut that's another story..\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":"Mon, 12 Sep 2016 11:27:20 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\n views in case referenced row doesn"},{"msg_contents":"Michael Paquier <michael.paquier@gmail.com> writes:\n> Indeed, and the query field does not have much more meaning for a WAL\n> sender. So I'd propose the attached for 9.6 and HEAD. I have thought\n> about reporting that to pgstat in StartReplication(), but as there is\n> some error handling there I'd think that WalSndLoop() is a better\n> place to call pgstat_report_activity, as per the attached.\n\nAs long as it's a fixed string there's no reason to set it repeatedly,\nso this placement looks fine for now. We can reconsider when/if we\nmake it variable and decide what is going to drive it.\n\nOn reflection, maybe s/walsender/WAL sender/? It doesn't look like\nwe really use the word \"walsender\" in user-facing docs.\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, 11 Sep 2016 22:38:23 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system views in\n case referenced row doesn"},{"msg_contents":"On Mon, Sep 12, 2016 at 11:38 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Michael Paquier <michael.paquier@gmail.com> writes:\n>> Indeed, and the query field does not have much more meaning for a WAL\n>> sender. So I'd propose the attached for 9.6 and HEAD. I have thought\n>> about reporting that to pgstat in StartReplication(), but as there is\n>> some error handling there I'd think that WalSndLoop() is a better\n>> place to call pgstat_report_activity, as per the attached.\n>\n> As long as it's a fixed string there's no reason to set it repeatedly,\n> so this placement looks fine for now. We can reconsider when/if we\n> make it variable and decide what is going to drive it.\n>\n> On reflection, maybe s/walsender/WAL sender/? It doesn't look like\n> we really use the word \"walsender\" in user-facing docs.\n\nIndeed, that may be better for clarity.\n\nExcept from the release notes, walsender is mentioned a couple of\ntimes in the protocol docs, as *walsender mode*:\nsrc/sgml/high-availability.sgml: a corresponding walsender process\nin the primary.\nsrc/sgml/protocol.sgml: Copy-both mode is initiated when a backend\nin walsender mode\nsrc/sgml/protocol.sgml:of <literal>true</> tells the backend to go\ninto walsender mode, wherein a\nsrc/sgml/protocol.sgml:the simple query protocol can be used in walsender mode.\nsrc/sgml/protocol.sgml:Passing <literal>database</> as the value\ninstructs walsender to connect to\nsrc/sgml/protocol.sgml:The commands accepted in walsender mode are:\nSo that looked adapted to me at first sight. Actually, the text had\nbetter be \"WAL sender\" and not \"walsender\" in high-availability, no?\nThat refers to the process, and not the backend mode.\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, 12 Sep 2016 11:46:03 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\n views in case referenced row doesn"},{"msg_contents":"On 12 September 2016 at 03:38, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n\n> On reflection, maybe s/walsender/WAL sender/? It doesn't look like\n> we really use the word \"walsender\" in user-facing docs.\n\nThere are already\n* 3 user messages referring to walsender and 2 referring to walreceiver\n* multiple references in the docs and release notes referring to walsender\n\nso usage has already been set and I vote to just leave things that way.\n\n-- \nSimon Riggs 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":"Mon, 12 Sep 2016 07:30:03 +0100","msg_from":"Simon Riggs <simon@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\n views in case referenced row doesn"},{"msg_contents":"On 12 September 2016 at 03:27, Michael Paquier\n<michael.paquier@gmail.com> wrote:\n\n> So I'd propose the attached for 9.6 and HEAD.\n\nThe $OP commit was against HEAD only, not against 9.6\n\nWhy would we need to backpatch this commit?\n\n-- \nSimon Riggs 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":"Mon, 12 Sep 2016 08:19:32 +0100","msg_from":"Simon Riggs <simon@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\n views in case referenced row doesn"},{"msg_contents":"On Mon, Sep 12, 2016 at 4:19 PM, Simon Riggs <simon@2ndquadrant.com> wrote:\n> On 12 September 2016 at 03:27, Michael Paquier\n> <michael.paquier@gmail.com> wrote:\n>\n>> So I'd propose the attached for 9.6 and HEAD.\n>\n> The $OP commit was against HEAD only, not against 9.6\n>\n> Why would we need to backpatch this commit?\n\nYou are right there is no need to get that into 9.6. Sorry for the mistake.\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, 12 Sep 2016 16:28:41 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\n views in case referenced row doesn"},{"msg_contents":"On 12 September 2016 at 08:28, Michael Paquier\n<michael.paquier@gmail.com> wrote:\n> On Mon, Sep 12, 2016 at 4:19 PM, Simon Riggs <simon@2ndquadrant.com> wrote:\n>> On 12 September 2016 at 03:27, Michael Paquier\n>> <michael.paquier@gmail.com> wrote:\n>>\n>>> So I'd propose the attached for 9.6 and HEAD.\n>>\n>> The $OP commit was against HEAD only, not against 9.6\n>>\n>> Why would we need to backpatch this commit?\n>\n> You are right there is no need to get that into 9.6. Sorry for the mistake.\n\nCommitted to HEAD only.\n\n\n[No longer a 9.6 issue.]\n\n-- \nSimon Riggs 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":"Mon, 12 Sep 2016 08:59:27 +0100","msg_from":"Simon Riggs <simon@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\n views in case referenced row doesn"},{"msg_contents":"On Mon, Sep 12, 2016 at 4:59 PM, Simon Riggs <simon@2ndquadrant.com> wrote:\n> On 12 September 2016 at 08:28, Michael Paquier\n> <michael.paquier@gmail.com> wrote:\n>> On Mon, Sep 12, 2016 at 4:19 PM, Simon Riggs <simon@2ndquadrant.com> wrote:\n>>> On 12 September 2016 at 03:27, Michael Paquier\n>>> <michael.paquier@gmail.com> wrote:\n>>>\n>>>> So I'd propose the attached for 9.6 and HEAD.\n>>>\n>>> The $OP commit was against HEAD only, not against 9.6\n>>>\n>>> Why would we need to backpatch this commit?\n>>\n>> You are right there is no need to get that into 9.6. Sorry for the mistake.\n>\n> Committed to HEAD only.\n\nThanks. Using walsender here is fine for me.\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, 12 Sep 2016 17:10:22 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\n views in case referenced row doesn"},{"msg_contents":"Michael Paquier <michael.paquier@gmail.com> writes:\n> On Mon, Sep 12, 2016 at 4:19 PM, Simon Riggs <simon@2ndquadrant.com> wrote:\n>> Why would we need to backpatch this commit?\n\n> You are right there is no need to get that into 9.6. Sorry for the mistake.\n\nOh, that's my fault, I'd incorrectly remembered this commit as having been\nfurther back than it was. HEAD-only is correct so far as fixing\nFujii-san's original complaint is concerned. However, don't we have the\nproblem that am_db_walsender processes will show up in pg_stat_activity\nanyway? Do we want to do something about those further back?\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, 12 Sep 2016 08:08:43 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system views in\n case referenced row doesn"},{"msg_contents":"On Mon, Sep 12, 2016 at 9:08 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Michael Paquier <michael.paquier@gmail.com> writes:\n>> On Mon, Sep 12, 2016 at 4:19 PM, Simon Riggs <simon@2ndquadrant.com> wrote:\n>>> Why would we need to backpatch this commit?\n>\n>> You are right there is no need to get that into 9.6. Sorry for the mistake.\n>\n> Oh, that's my fault, I'd incorrectly remembered this commit as having been\n> further back than it was. HEAD-only is correct so far as fixing\n> Fujii-san's original complaint is concerned. However, don't we have the\n> problem that am_db_walsender processes will show up in pg_stat_activity\n> anyway?\n\nYes, those show up..\n\n> Do we want to do something about those further back?\n\nHiding them is not something that we should do, and changing the query\nfield to show something that we think is helpful may impact existing\napplications that rely on the fact that this field is NULL. So I'd\nvote for doing nothing.\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, 12 Sep 2016 21:53:05 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\n views in case referenced row doesn"}],"string":"[\n {\n \"msg_contents\": \"Use LEFT JOINs in some system views in case referenced row doesn't exist.\\n\\nIn particular, left join to pg_authid so that rows in pg_stat_activity\\ndon't disappear if the session's owning user has been dropped.\\nAlso convert a few joins to pg_database to left joins, in the same spirit,\\nthough that case might be harder to hit. We were doing this in other\\nviews already, so it was a bit inconsistent that these views didn't.\\n\\nOskari Saarenmaa, with some further tweaking by me\\n\\nDiscussion: <56E87CD8.60007@ohmu.fi>\\n\\nBranch\\n------\\nmaster\\n\\nDetails\\n-------\\nhttp://git.postgresql.org/pg/commitdiff/8299471c37fff0b0f5a777a12f920125310c0efe\\n\\nModified Files\\n--------------\\nsrc/backend/catalog/system_views.sql | 20 +++++++++-----------\\nsrc/include/catalog/catversion.h | 2 +-\\nsrc/test/regress/expected/rules.out | 20 +++++++++-----------\\n3 files changed, 19 insertions(+), 23 deletions(-)\\n\\n\\n-- \\nSent via pgsql-committers mailing list (pgsql-committers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-committers\\n\",\n \"msg_date\": \"Fri, 19 Aug 2016 21:13:52 +0000\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"pgsql: Use LEFT JOINs in some system views in case referenced row\\n doesn\"\n },\n {\n \"msg_contents\": \"On Sat, Aug 20, 2016 at 6:13 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Use LEFT JOINs in some system views in case referenced row doesn't exist.\\n>\\n> In particular, left join to pg_authid so that rows in pg_stat_activity\\n> don't disappear if the session's owning user has been dropped.\\n> Also convert a few joins to pg_database to left joins, in the same spirit,\\n> though that case might be harder to hit. We were doing this in other\\n> views already, so it was a bit inconsistent that these views didn't.\\n\\nThis change causes pg_stat_activity to report the \\\"bogus\\\" information about\\nwalsenders as follows.\\n\\n$ ps aux | grep 75076\\npostgres 75076 0.0 0.0 2608252 1364 ?? Ss 2:41AM\\n0:00.07 postgres: wal sender process postgres [local] streaming\\n0/3000140\\n\\n=# SELECT * FROM pg_stat_activity;\\n-[ RECORD 1 ]----+--------------------------------\\ndatid | 0\\ndatname | (null)\\npid | 75076\\nusesysid | 10\\nusename | postgres\\napplication_name | sby1\\nclient_addr | (null)\\nclient_hostname | (null)\\nclient_port | -1\\nbackend_start | 2016-09-10 02:41:24.568313+09\\nxact_start | (null)\\nquery_start | (null)\\nstate_change | 2016-09-10 02:41:24.570135+09\\nwait_event_type | (null)\\nwait_event | (null)\\nstate | idle\\nbackend_xid | (null)\\nbackend_xmin | (null)\\nquery |\\n\\nI think that this information is confusing and should not exist in\\npg_stat_activity.\\n\\nRegards,\\n\\n-- \\nFujii Masao\\n\\n\\n-- \\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, 10 Sep 2016 02:51:41 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system views in case\\n referenced row doesn\"\n },\n {\n \"msg_contents\": \"Fujii Masao <masao.fujii@gmail.com> writes:\\n> On Sat, Aug 20, 2016 at 6:13 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> Use LEFT JOINs in some system views in case referenced row doesn't exist.\\n\\n> This change causes pg_stat_activity to report the \\\"bogus\\\" information about\\n> walsenders as follows.\\n\\nHmm ... but if we want to exclude walsenders from that view, surely we\\nshould do so explicitly, not have it happen as a hidden side-effect\\nof not being able to join them to pg_database.\\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, 09 Sep 2016 13:59:24 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system views in\\n case referenced row doesn\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> Fujii Masao <masao.fujii@gmail.com> writes:\\n>> On Sat, Aug 20, 2016 at 6:13 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>>> Use LEFT JOINs in some system views in case referenced row doesn't exist.\\n\\n>> This change causes pg_stat_activity to report the \\\"bogus\\\" information about\\n>> walsenders as follows.\\n\\n> Hmm ... but if we want to exclude walsenders from that view, surely we\\n> should do so explicitly, not have it happen as a hidden side-effect\\n> of not being able to join them to pg_database.\\n\\nI looked into this a little. There are at least three things we could\\ndo here:\\n\\n1. Decide that showing walsenders is a good thing. I'm not really\\nsure why it isn't -- for example, we could take the trouble to display\\nthe current replication command as the walsender's activity.\\n\\n2. Suppress walsenders by the expedient of ignoring PGPROCs that have\\nzero database ID. This would also ignore other process types that aren't\\nconnected to a specific database. This is pretty backwards-compatible\\nbecause it's equivalent to what used to happen implicitly via the inner\\njoin in the view.\\n\\n3. Suppress walsenders by adding some new field to PgBackendStatus that\\nidentifies them, and having pg_stat_get_activity() use that to filter.\\n\\n\\nIt looks to me like even before the change Fujii-san complains of, it was\\nonly generic walsenders that were hidden in the view. Database-specific\\nwalsenders (am_db_walsender) would have a valid database OID in\\nPgBackendStatus and therefore would pass the join. This makes me think\\nthat option #1 is really the sanest alternative. While option #2 is\\nthe most backwards-compatible, it's pretty hard to see why we would think\\nthat showing only database-specific walsenders is a desirable behavior.\\nIf we want to exclude walsenders altogether, we need option #3.\\n\\nThoughts?\\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, 10 Sep 2016 12:51:02 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system views in\\n case referenced row doesn\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> I looked into this a little. There are at least three things we could\\n> do here:\\n> 1. Decide that showing walsenders is a good thing. I'm not really\\n> sure why it isn't -- for example, we could take the trouble to display\\n> the current replication command as the walsender's activity.\\n> 2. Suppress walsenders by the expedient of ignoring PGPROCs that have\\n> zero database ID. This would also ignore other process types that aren't\\n> connected to a specific database. This is pretty backwards-compatible\\n> because it's equivalent to what used to happen implicitly via the inner\\n> join in the view.\\n> 3. Suppress walsenders by adding some new field to PgBackendStatus that\\n> identifies them, and having pg_stat_get_activity() use that to filter.\\n\\nBTW, I had been thinking we could implement #2 or #3 relatively painlessly\\nby adding filter logic inside pg_stat_get_activity(). However, I now\\nnotice that that function is also used by the pg_stat_replication view,\\nwhich joins its output to pg_stat_get_wal_senders(). That means we can\\n*not* simply suppress walsenders in the function's output; if we want to\\nfilter them away in pg_stat_activity, we have to change the view\\ndefinition somehow, and probably the function's API as well. And that\\nmeans we can't change the behavior without an initdb, which is kind of\\nannoying to have to do after rc1.\\n\\nThe fact that the pg_stat_replication view does show walsender processes\\nseems like possibly a reasonable argument for not showing them in\\npg_stat_activity. But we'd have to do some rejiggering of the view\\ndefinition to make that happen.\\n\\nDon't have a strong opinion which way it ought to work --- I could go with\\neither the \\\"walsenders are a perfectly good activity\\\" position or the\\n\\\"they should appear only in pg_stat_replication\\\" one. But the latter will\\ntake an initdb to do, and I'm inclined to the position that it's too late\\nfor that in 9.6. As I noted, it was already inconsistent for per-db\\nwalsenders long before 9.6, without field complaints, so I'm having a hard\\ntime seeing this as a critical 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, 10 Sep 2016 16:21:50 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system views in\\n case referenced row doesn\"\n },\n {\n \"msg_contents\": \"On Sun, Sep 11, 2016 at 5:21 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> The fact that the pg_stat_replication view does show walsender processes\\n> seems like possibly a reasonable argument for not showing them in\\n> pg_stat_activity. But we'd have to do some rejiggering of the view\\n> definition to make that happen.\\n\\nWe may actually had better show WAL sender processes in\\npg_stat_activity. An argument in favor of that is the tracking of\\nWaitEventSet events (or latches if you want).\\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, 12 Sep 2016 10:00:50 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\\n views in case referenced row doesn\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael.paquier@gmail.com> writes:\\n> On Sun, Sep 11, 2016 at 5:21 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> The fact that the pg_stat_replication view does show walsender processes\\n>> seems like possibly a reasonable argument for not showing them in\\n>> pg_stat_activity. But we'd have to do some rejiggering of the view\\n>> definition to make that happen.\\n\\n> We may actually had better show WAL sender processes in\\n> pg_stat_activity. An argument in favor of that is the tracking of\\n> WaitEventSet events (or latches if you want).\\n\\nAlso, walsenders count against MaxBackends don't they? So not showing\\nthem could contribute to confusion about why an installation is hitting\\nthe connection limit.\\n\\nIf we do keep them in the view, I would definitely vote for having them\\nset their \\\"query\\\" fields to something that shows they're walsenders.\\nIt's awfully late to be doing anything complicated there for 9.6,\\nbut we could just set the query to \\\"walsender\\\" and plan to improve\\non that in future releases.\\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, 11 Sep 2016 21:13:05 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system views in\\n case referenced row doesn\"\n },\n {\n \"msg_contents\": \"* Tom Lane (tgl@sss.pgh.pa.us) wrote:\\n> Michael Paquier <michael.paquier@gmail.com> writes:\\n> > On Sun, Sep 11, 2016 at 5:21 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> >> The fact that the pg_stat_replication view does show walsender processes\\n> >> seems like possibly a reasonable argument for not showing them in\\n> >> pg_stat_activity. But we'd have to do some rejiggering of the view\\n> >> definition to make that happen.\\n> \\n> > We may actually had better show WAL sender processes in\\n> > pg_stat_activity. An argument in favor of that is the tracking of\\n> > WaitEventSet events (or latches if you want).\\n> \\n> Also, walsenders count against MaxBackends don't they? So not showing\\n> them could contribute to confusion about why an installation is hitting\\n> the connection limit.\\n> \\n> If we do keep them in the view, I would definitely vote for having them\\n> set their \\\"query\\\" fields to something that shows they're walsenders.\\n> It's awfully late to be doing anything complicated there for 9.6,\\n> but we could just set the query to \\\"walsender\\\" and plan to improve\\n> on that in future releases.\\n\\n+1\\n\\nThanks!\\n\\nStephen\",\n \"msg_date\": \"Sun, 11 Sep 2016 21:16:46 -0400\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\\n views in case referenced row doesn\"\n },\n {\n \"msg_contents\": \"On Mon, Sep 12, 2016 at 10:16 AM, Stephen Frost <sfrost@snowman.net> wrote:\\n> * Tom Lane (tgl@sss.pgh.pa.us) wrote:\\n>> Michael Paquier <michael.paquier@gmail.com> writes:\\n>> > On Sun, Sep 11, 2016 at 5:21 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> >> The fact that the pg_stat_replication view does show walsender processes\\n>> >> seems like possibly a reasonable argument for not showing them in\\n>> >> pg_stat_activity. But we'd have to do some rejiggering of the view\\n>> >> definition to make that happen.\\n>>\\n>> > We may actually had better show WAL sender processes in\\n>> > pg_stat_activity. An argument in favor of that is the tracking of\\n>> > WaitEventSet events (or latches if you want).\\n>>\\n>> Also, walsenders count against MaxBackends don't they? So not showing\\n>> them could contribute to confusion about why an installation is hitting\\n>> the connection limit.\\n\\nYes they are counted in MaxBackends. So that would help as well in\\nlooking at connection limit issues.\\n\\n>> If we do keep them in the view, I would definitely vote for having them\\n>> set their \\\"query\\\" fields to something that shows they're walsenders.\\n>> It's awfully late to be doing anything complicated there for 9.6,\\n>> but we could just set the query to \\\"walsender\\\" and plan to improve\\n>> on that in future releases.\\n>\\n> +1\\n\\nIndeed, and the query field does not have much more meaning for a WAL\\nsender. So I'd propose the attached for 9.6 and HEAD. I have thought\\nabout reporting that to pgstat in StartReplication(), but as there is\\nsome error handling there I'd think that WalSndLoop() is a better\\nplace to call pgstat_report_activity, as per the attached.\\n\\nWe could have far more fancy verbose information to offer to the user,\\nbut that's another story..\\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\": \"Mon, 12 Sep 2016 11:27:20 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\\n views in case referenced row doesn\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael.paquier@gmail.com> writes:\\n> Indeed, and the query field does not have much more meaning for a WAL\\n> sender. So I'd propose the attached for 9.6 and HEAD. I have thought\\n> about reporting that to pgstat in StartReplication(), but as there is\\n> some error handling there I'd think that WalSndLoop() is a better\\n> place to call pgstat_report_activity, as per the attached.\\n\\nAs long as it's a fixed string there's no reason to set it repeatedly,\\nso this placement looks fine for now. We can reconsider when/if we\\nmake it variable and decide what is going to drive it.\\n\\nOn reflection, maybe s/walsender/WAL sender/? It doesn't look like\\nwe really use the word \\\"walsender\\\" in user-facing docs.\\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, 11 Sep 2016 22:38:23 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system views in\\n case referenced row doesn\"\n },\n {\n \"msg_contents\": \"On Mon, Sep 12, 2016 at 11:38 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Michael Paquier <michael.paquier@gmail.com> writes:\\n>> Indeed, and the query field does not have much more meaning for a WAL\\n>> sender. So I'd propose the attached for 9.6 and HEAD. I have thought\\n>> about reporting that to pgstat in StartReplication(), but as there is\\n>> some error handling there I'd think that WalSndLoop() is a better\\n>> place to call pgstat_report_activity, as per the attached.\\n>\\n> As long as it's a fixed string there's no reason to set it repeatedly,\\n> so this placement looks fine for now. We can reconsider when/if we\\n> make it variable and decide what is going to drive it.\\n>\\n> On reflection, maybe s/walsender/WAL sender/? It doesn't look like\\n> we really use the word \\\"walsender\\\" in user-facing docs.\\n\\nIndeed, that may be better for clarity.\\n\\nExcept from the release notes, walsender is mentioned a couple of\\ntimes in the protocol docs, as *walsender mode*:\\nsrc/sgml/high-availability.sgml: a corresponding walsender process\\nin the primary.\\nsrc/sgml/protocol.sgml: Copy-both mode is initiated when a backend\\nin walsender mode\\nsrc/sgml/protocol.sgml:of <literal>true</> tells the backend to go\\ninto walsender mode, wherein a\\nsrc/sgml/protocol.sgml:the simple query protocol can be used in walsender mode.\\nsrc/sgml/protocol.sgml:Passing <literal>database</> as the value\\ninstructs walsender to connect to\\nsrc/sgml/protocol.sgml:The commands accepted in walsender mode are:\\nSo that looked adapted to me at first sight. Actually, the text had\\nbetter be \\\"WAL sender\\\" and not \\\"walsender\\\" in high-availability, no?\\nThat refers to the process, and not the backend mode.\\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, 12 Sep 2016 11:46:03 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\\n views in case referenced row doesn\"\n },\n {\n \"msg_contents\": \"On 12 September 2016 at 03:38, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n\\n> On reflection, maybe s/walsender/WAL sender/? It doesn't look like\\n> we really use the word \\\"walsender\\\" in user-facing docs.\\n\\nThere are already\\n* 3 user messages referring to walsender and 2 referring to walreceiver\\n* multiple references in the docs and release notes referring to walsender\\n\\nso usage has already been set and I vote to just leave things that way.\\n\\n-- \\nSimon Riggs 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\": \"Mon, 12 Sep 2016 07:30:03 +0100\",\n \"msg_from\": \"Simon Riggs <simon@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\\n views in case referenced row doesn\"\n },\n {\n \"msg_contents\": \"On 12 September 2016 at 03:27, Michael Paquier\\n<michael.paquier@gmail.com> wrote:\\n\\n> So I'd propose the attached for 9.6 and HEAD.\\n\\nThe $OP commit was against HEAD only, not against 9.6\\n\\nWhy would we need to backpatch this commit?\\n\\n-- \\nSimon Riggs 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\": \"Mon, 12 Sep 2016 08:19:32 +0100\",\n \"msg_from\": \"Simon Riggs <simon@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\\n views in case referenced row doesn\"\n },\n {\n \"msg_contents\": \"On Mon, Sep 12, 2016 at 4:19 PM, Simon Riggs <simon@2ndquadrant.com> wrote:\\n> On 12 September 2016 at 03:27, Michael Paquier\\n> <michael.paquier@gmail.com> wrote:\\n>\\n>> So I'd propose the attached for 9.6 and HEAD.\\n>\\n> The $OP commit was against HEAD only, not against 9.6\\n>\\n> Why would we need to backpatch this commit?\\n\\nYou are right there is no need to get that into 9.6. Sorry for the mistake.\\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, 12 Sep 2016 16:28:41 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\\n views in case referenced row doesn\"\n },\n {\n \"msg_contents\": \"On 12 September 2016 at 08:28, Michael Paquier\\n<michael.paquier@gmail.com> wrote:\\n> On Mon, Sep 12, 2016 at 4:19 PM, Simon Riggs <simon@2ndquadrant.com> wrote:\\n>> On 12 September 2016 at 03:27, Michael Paquier\\n>> <michael.paquier@gmail.com> wrote:\\n>>\\n>>> So I'd propose the attached for 9.6 and HEAD.\\n>>\\n>> The $OP commit was against HEAD only, not against 9.6\\n>>\\n>> Why would we need to backpatch this commit?\\n>\\n> You are right there is no need to get that into 9.6. Sorry for the mistake.\\n\\nCommitted to HEAD only.\\n\\n\\n[No longer a 9.6 issue.]\\n\\n-- \\nSimon Riggs 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\": \"Mon, 12 Sep 2016 08:59:27 +0100\",\n \"msg_from\": \"Simon Riggs <simon@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\\n views in case referenced row doesn\"\n },\n {\n \"msg_contents\": \"On Mon, Sep 12, 2016 at 4:59 PM, Simon Riggs <simon@2ndquadrant.com> wrote:\\n> On 12 September 2016 at 08:28, Michael Paquier\\n> <michael.paquier@gmail.com> wrote:\\n>> On Mon, Sep 12, 2016 at 4:19 PM, Simon Riggs <simon@2ndquadrant.com> wrote:\\n>>> On 12 September 2016 at 03:27, Michael Paquier\\n>>> <michael.paquier@gmail.com> wrote:\\n>>>\\n>>>> So I'd propose the attached for 9.6 and HEAD.\\n>>>\\n>>> The $OP commit was against HEAD only, not against 9.6\\n>>>\\n>>> Why would we need to backpatch this commit?\\n>>\\n>> You are right there is no need to get that into 9.6. Sorry for the mistake.\\n>\\n> Committed to HEAD only.\\n\\nThanks. Using walsender here is fine for me.\\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, 12 Sep 2016 17:10:22 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\\n views in case referenced row doesn\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael.paquier@gmail.com> writes:\\n> On Mon, Sep 12, 2016 at 4:19 PM, Simon Riggs <simon@2ndquadrant.com> wrote:\\n>> Why would we need to backpatch this commit?\\n\\n> You are right there is no need to get that into 9.6. Sorry for the mistake.\\n\\nOh, that's my fault, I'd incorrectly remembered this commit as having been\\nfurther back than it was. HEAD-only is correct so far as fixing\\nFujii-san's original complaint is concerned. However, don't we have the\\nproblem that am_db_walsender processes will show up in pg_stat_activity\\nanyway? Do we want to do something about those further back?\\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, 12 Sep 2016 08:08:43 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system views in\\n case referenced row doesn\"\n },\n {\n \"msg_contents\": \"On Mon, Sep 12, 2016 at 9:08 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Michael Paquier <michael.paquier@gmail.com> writes:\\n>> On Mon, Sep 12, 2016 at 4:19 PM, Simon Riggs <simon@2ndquadrant.com> wrote:\\n>>> Why would we need to backpatch this commit?\\n>\\n>> You are right there is no need to get that into 9.6. Sorry for the mistake.\\n>\\n> Oh, that's my fault, I'd incorrectly remembered this commit as having been\\n> further back than it was. HEAD-only is correct so far as fixing\\n> Fujii-san's original complaint is concerned. However, don't we have the\\n> problem that am_db_walsender processes will show up in pg_stat_activity\\n> anyway?\\n\\nYes, those show up..\\n\\n> Do we want to do something about those further back?\\n\\nHiding them is not something that we should do, and changing the query\\nfield to show something that we think is helpful may impact existing\\napplications that rely on the fact that this field is NULL. So I'd\\nvote for doing nothing.\\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, 12 Sep 2016 21:53:05 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Use LEFT JOINs in some system\\n views in case referenced row doesn\"\n }\n]","isConversational":false}}},{"rowIdx":42989,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"I found another bug as a result of using amcheck on Heroku customer\ndatabases. This time, the bug is in core Postgres. It's one of mine.\n\nThere was a thinko in tuplesort's abbreviation abort logic, causing\ncertain SortTuples to be spuriously marked NULL (and so, subsequently\nsorted as a NULL tuple, despite not actually changing anything about\nthe representation of caller tuples). The attached patch fixes this\nbug.\n\nI noticed this following a complaint by amcheck about a tuple in the\nwrong order on a leaf page in some random text index. The leaf page\nwas entirely full of NULL values, aside from this one tuple at some\nseemingly random position. All non-NULL index tuples were of the kind\nthat you'd expect to trigger abbreviation to abort (many distinct\nvalues, but with little entropy at the beginning).\n\nI believe that this particular problem has been observed on a tiny\nfraction of all databases tested, so I don't think it's very common in\nthe wild.\n\nI'd be surprised if amcheck does not bring more bugs like this to my\nattention before too long. We should work on improving it, so that we\nhave greater visibility into problems that occur in the field.\n-- \nPeter Geoghegan\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, 19 Aug 2016 15:07:02 -0700","msg_from":"Peter Geoghegan <pg@heroku.com>","msg_from_op":true,"msg_subject":"Bug in abbreviated keys abort handling (found with amcheck)"},{"msg_contents":"On Fri, Aug 19, 2016 at 6:07 PM, Peter Geoghegan <pg@heroku.com> wrote:\n> I found another bug as a result of using amcheck on Heroku customer\n> databases. This time, the bug is in core Postgres. It's one of mine.\n>\n> There was a thinko in tuplesort's abbreviation abort logic, causing\n> certain SortTuples to be spuriously marked NULL (and so, subsequently\n> sorted as a NULL tuple, despite not actually changing anything about\n> the representation of caller tuples). The attached patch fixes this\n> bug.\n\nUgh, that sucks. Thanks for the report and patch. Committed and\nback-patched to 9.5.\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, 22 Aug 2016 15:34:23 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Bug in abbreviated keys abort handling (found with amcheck)"},{"msg_contents":"On Mon, Aug 22, 2016 at 12:34 PM, Robert Haas <robertmhaas@gmail.com> wrote:\n> Ugh, that sucks. Thanks for the report and patch. Committed and\n> back-patched to 9.5.\n\nThanks.\n\nWithin Heroku, there is a lot of enthusiasm for the idea of sharing\nhard data about the prevalence of problems like this. I hope to be\nable to share figures in the next few weeks, when I finish working\nthrough the backlog.\n\nSeparately, I would like amcheck to play a role in how we direct users\nto REINDEX, as issues like this come to light. It would be much more\nhelpful if we didn't have to be so conservative. I hesitate to say\nthat amcheck will detect cases where this bug led to corruption with\n100% reliability, but I think that any case that one can imagine in\nwhich amcheck fails here is unlikely in the extreme. The same applies\nto the glibc abbreviated keys issue.\n\nI actually didn't find any glibc strxfrm() issues yet, even though any\ninstances of corruption of text indexes I've seen originated before\nthe point release in which strxfrm() became distrusted. I guess that\nnot that many Heroku users use the \"C\" locale, which would still be\naffected with the latest point release.\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":"Mon, 22 Aug 2016 13:22:56 -0700","msg_from":"Peter Geoghegan <pg@heroku.com>","msg_from_op":true,"msg_subject":"Re: Bug in abbreviated keys abort handling (found with amcheck)"}],"string":"[\n {\n \"msg_contents\": \"I found another bug as a result of using amcheck on Heroku customer\\ndatabases. This time, the bug is in core Postgres. It's one of mine.\\n\\nThere was a thinko in tuplesort's abbreviation abort logic, causing\\ncertain SortTuples to be spuriously marked NULL (and so, subsequently\\nsorted as a NULL tuple, despite not actually changing anything about\\nthe representation of caller tuples). The attached patch fixes this\\nbug.\\n\\nI noticed this following a complaint by amcheck about a tuple in the\\nwrong order on a leaf page in some random text index. The leaf page\\nwas entirely full of NULL values, aside from this one tuple at some\\nseemingly random position. All non-NULL index tuples were of the kind\\nthat you'd expect to trigger abbreviation to abort (many distinct\\nvalues, but with little entropy at the beginning).\\n\\nI believe that this particular problem has been observed on a tiny\\nfraction of all databases tested, so I don't think it's very common in\\nthe wild.\\n\\nI'd be surprised if amcheck does not bring more bugs like this to my\\nattention before too long. We should work on improving it, so that we\\nhave greater visibility into problems that occur in the field.\\n-- \\nPeter Geoghegan\\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, 19 Aug 2016 15:07:02 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@heroku.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Bug in abbreviated keys abort handling (found with amcheck)\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 19, 2016 at 6:07 PM, Peter Geoghegan <pg@heroku.com> wrote:\\n> I found another bug as a result of using amcheck on Heroku customer\\n> databases. This time, the bug is in core Postgres. It's one of mine.\\n>\\n> There was a thinko in tuplesort's abbreviation abort logic, causing\\n> certain SortTuples to be spuriously marked NULL (and so, subsequently\\n> sorted as a NULL tuple, despite not actually changing anything about\\n> the representation of caller tuples). The attached patch fixes this\\n> bug.\\n\\nUgh, that sucks. Thanks for the report and patch. Committed and\\nback-patched to 9.5.\\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, 22 Aug 2016 15:34:23 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Bug in abbreviated keys abort handling (found with amcheck)\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 22, 2016 at 12:34 PM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> Ugh, that sucks. Thanks for the report and patch. Committed and\\n> back-patched to 9.5.\\n\\nThanks.\\n\\nWithin Heroku, there is a lot of enthusiasm for the idea of sharing\\nhard data about the prevalence of problems like this. I hope to be\\nable to share figures in the next few weeks, when I finish working\\nthrough the backlog.\\n\\nSeparately, I would like amcheck to play a role in how we direct users\\nto REINDEX, as issues like this come to light. It would be much more\\nhelpful if we didn't have to be so conservative. I hesitate to say\\nthat amcheck will detect cases where this bug led to corruption with\\n100% reliability, but I think that any case that one can imagine in\\nwhich amcheck fails here is unlikely in the extreme. The same applies\\nto the glibc abbreviated keys issue.\\n\\nI actually didn't find any glibc strxfrm() issues yet, even though any\\ninstances of corruption of text indexes I've seen originated before\\nthe point release in which strxfrm() became distrusted. I guess that\\nnot that many Heroku users use the \\\"C\\\" locale, which would still be\\naffected with the latest point release.\\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\": \"Mon, 22 Aug 2016 13:22:56 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@heroku.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Bug in abbreviated keys abort handling (found with amcheck)\"\n }\n]","isConversational":false}}},{"rowIdx":42990,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hello,\ni'm testing the new row security level functionality in postgresql 9.5.\nTo do that i run this script:\n\n-----------cut here ----------------------\n\nCREATE TABLE public.policy_tab\n(\n id bigint NOT NULL,\n description character varying(160) NOT NULL,\n usr name NOT NULL,\n CONSTRAINT policy_tab_pk PRIMARY KEY (id)\n);\n\nALTER TABLE public.policy_tab OWNER TO postgres;\n\nGRANT ALL ON TABLE public.policy_tab TO public;\n\nCREATE OR REPLACE VIEW public.policy_view AS\n SELECT id,\n description,\n usr\n FROM public.policy_tab;\n\nALTER TABLE public.policy_view\n OWNER TO postgres;\nGRANT ALL ON TABLE public.policy_view TO public;\n\nALTER TABLE public.policy_tab ENABLE ROW LEVEL SECURITY;\n\nCREATE POLICY standard ON public.policy_tab\n FOR ALL\n TO PUBLIC\n USING (usr = current_user);\n\nINSERT INTO public.policy_tab (id, description, usr) VALUES (1,'uno','\nmanager@scuola247.it');\nINSERT INTO public.policy_tab (id, description, usr) VALUES (2,'due','\nmanager@scuola247.it');\nINSERT INTO public.policy_tab (id, description, usr) VALUES (3,'tre','\nmanager@scuola247.it');\nINSERT INTO public.policy_tab (id, description, usr) VALUES (4,'quattro','\nteacher@scuola247.it');\nINSERT INTO public.policy_tab (id, description, usr) VALUES (5,'cinque','\nteacher@scuola247.it');\n\n-----------cut here ----------------------\n\nafter that i run the query: \"select * from public.policy_tab\"\n\nand the the oupt was what i excpected:\n\nrows 1,2,3 for user: manager@scuola247.it\nrows 4,5 for user: teacher@scuola247.it\nrows 1,2,3,4,5 for user: postgres (the policy doesn't work for him)\n\nbut when i run the query: \"select * from public.policy_view\"\n\nthe ouput is the same (all rows) for all users\n\ni'm doing some mistakes or this is a bug ?\n\nthank you in advance for the time you would like dedicate to me.\n\nAndrea Adami\n\nHello,i'm testing the new row security level functionality in postgresql 9.5.To do that i run this script:-----------cut here ----------------------CREATE TABLE public.policy_tab( id bigint NOT NULL, description character varying(160) NOT NULL, usr name NOT NULL, CONSTRAINT policy_tab_pk PRIMARY KEY (id));ALTER TABLE public.policy_tab OWNER TO postgres;GRANT ALL ON TABLE public.policy_tab TO public;CREATE OR REPLACE VIEW public.policy_view AS SELECT id, description, usr FROM public.policy_tab;ALTER TABLE public.policy_view OWNER TO postgres;GRANT ALL ON TABLE public.policy_view TO public;ALTER TABLE public.policy_tab ENABLE ROW LEVEL SECURITY;CREATE POLICY standard ON public.policy_tab FOR ALL TO PUBLIC USING (usr = current_user);INSERT INTO public.policy_tab (id, description, usr) VALUES (1,'uno','manager@scuola247.it');INSERT INTO public.policy_tab (id, description, usr) VALUES (2,'due','manager@scuola247.it');INSERT INTO public.policy_tab (id, description, usr) VALUES (3,'tre','manager@scuola247.it');INSERT INTO public.policy_tab (id, description, usr) VALUES (4,'quattro','teacher@scuola247.it');INSERT INTO public.policy_tab (id, description, usr) VALUES (5,'cinque','teacher@scuola247.it');-----------cut here ----------------------after that i run the query: \"select * from public.policy_tab\"and the the oupt was what i excpected:rows 1,2,3 for user: manager@scuola247.itrows 4,5 for user: teacher@scuola247.itrows 1,2,3,4,5 for user: postgres (the policy doesn't work for him)but when i run the query: \"select * from public.policy_view\"the ouput is the same (all rows) for all users i'm doing some mistakes or this is a bug ?thank you in advance for the time you would like dedicate to me.Andrea Adami","msg_date":"Sat, 20 Aug 2016 04:15:10 +0200","msg_from":"Andrea Adami <fol@fulcro.net>","msg_from_op":true,"msg_subject":"CREATE POLICY bug ?"},{"msg_contents":"On 20 August 2016 at 03:15, Andrea Adami <fol@fulcro.net> wrote:\n> when i run the query: \"select * from public.policy_view\"\n> the ouput is the same (all rows) for all users\n> i'm doing some mistakes or this is a bug ?\n>\n\nNo, it looks correct to me. When going through a view, the policies\nand permission checks that apply are those that would apply to the\nview's owner, which in this case is postgres, so no policies are\napplied.\n\nOr, quoting from the notes in the CREATE POLICY documentation:\n\n As with normal queries and views, permission checks and policies for\n the tables which are referenced by a view will use the view owner's\n rights and any policies which apply to the view owner.\n\nRegards,\nDean\n\n\n-- \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, 20 Aug 2016 15:01:40 +0100","msg_from":"Dean Rasheed <dean.a.rasheed@gmail.com>","msg_from_op":false,"msg_subject":"Re: CREATE POLICY bug ?"},{"msg_contents":"[Please reply to the list, not just to me, so that others can benefit\nfrom and contribute to the discussion]\n\nOn 31 August 2016 at 11:52, Andrea Adami <fol@fulcro.net> wrote:\n> Thnaks Dean, i did further investigations:\n> i set the owner of the view to: \"manager@scuola247.it\" with:\n> ALTER TABLE public.policy_view OWNER TO \"manager@scuola247.it\";\n> and i thinking to see from the select:\n> select * from policy_view\n> the rows: 1,2,3\n> then\n> set role 'manager@scuola247.it';\n> select * from policy_view;\n> return rows 1,2,3 as expected but:\n> set role 'teacher@scuola247.it';\n> select * from policy_view;\n> returns rows 4,5 and\n> set role 'postgres'\n> select * from policy_view\n> return nothing ...\n> what you thinking about ?\n>\n> Andrea\n\nThat's correct. With the table owned by postgres and the view owned by\n\"manager@scuola247.it\", access to the table via the view is subject to\nthe policies that apply to \"manager@scuola247.it\". So regardless of\nwho the current user is, when selecting from the view, the policy\n\"standard\" will be applied, and that will limit the visible rows to\nthose for which usr = current_user.\n\nRegards,\nDean\n\n\n-- \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, 1 Sep 2016 10:53:51 +0100","msg_from":"Dean Rasheed <dean.a.rasheed@gmail.com>","msg_from_op":false,"msg_subject":"Re: CREATE POLICY bug ?"}],"string":"[\n {\n \"msg_contents\": \"Hello,\\ni'm testing the new row security level functionality in postgresql 9.5.\\nTo do that i run this script:\\n\\n-----------cut here ----------------------\\n\\nCREATE TABLE public.policy_tab\\n(\\n id bigint NOT NULL,\\n description character varying(160) NOT NULL,\\n usr name NOT NULL,\\n CONSTRAINT policy_tab_pk PRIMARY KEY (id)\\n);\\n\\nALTER TABLE public.policy_tab OWNER TO postgres;\\n\\nGRANT ALL ON TABLE public.policy_tab TO public;\\n\\nCREATE OR REPLACE VIEW public.policy_view AS\\n SELECT id,\\n description,\\n usr\\n FROM public.policy_tab;\\n\\nALTER TABLE public.policy_view\\n OWNER TO postgres;\\nGRANT ALL ON TABLE public.policy_view TO public;\\n\\nALTER TABLE public.policy_tab ENABLE ROW LEVEL SECURITY;\\n\\nCREATE POLICY standard ON public.policy_tab\\n FOR ALL\\n TO PUBLIC\\n USING (usr = current_user);\\n\\nINSERT INTO public.policy_tab (id, description, usr) VALUES (1,'uno','\\nmanager@scuola247.it');\\nINSERT INTO public.policy_tab (id, description, usr) VALUES (2,'due','\\nmanager@scuola247.it');\\nINSERT INTO public.policy_tab (id, description, usr) VALUES (3,'tre','\\nmanager@scuola247.it');\\nINSERT INTO public.policy_tab (id, description, usr) VALUES (4,'quattro','\\nteacher@scuola247.it');\\nINSERT INTO public.policy_tab (id, description, usr) VALUES (5,'cinque','\\nteacher@scuola247.it');\\n\\n-----------cut here ----------------------\\n\\nafter that i run the query: \\\"select * from public.policy_tab\\\"\\n\\nand the the oupt was what i excpected:\\n\\nrows 1,2,3 for user: manager@scuola247.it\\nrows 4,5 for user: teacher@scuola247.it\\nrows 1,2,3,4,5 for user: postgres (the policy doesn't work for him)\\n\\nbut when i run the query: \\\"select * from public.policy_view\\\"\\n\\nthe ouput is the same (all rows) for all users\\n\\ni'm doing some mistakes or this is a bug ?\\n\\nthank you in advance for the time you would like dedicate to me.\\n\\nAndrea Adami\\n\\nHello,i'm testing the new row security level functionality in postgresql 9.5.To do that i run this script:-----------cut here ----------------------CREATE TABLE public.policy_tab( id bigint NOT NULL, description character varying(160) NOT NULL, usr name NOT NULL, CONSTRAINT policy_tab_pk PRIMARY KEY (id));ALTER TABLE public.policy_tab OWNER TO postgres;GRANT ALL ON TABLE public.policy_tab TO public;CREATE OR REPLACE VIEW public.policy_view AS SELECT id, description, usr FROM public.policy_tab;ALTER TABLE public.policy_view OWNER TO postgres;GRANT ALL ON TABLE public.policy_view TO public;ALTER TABLE public.policy_tab ENABLE ROW LEVEL SECURITY;CREATE POLICY standard ON public.policy_tab FOR ALL TO PUBLIC USING (usr = current_user);INSERT INTO public.policy_tab (id, description, usr) VALUES (1,'uno','manager@scuola247.it');INSERT INTO public.policy_tab (id, description, usr) VALUES (2,'due','manager@scuola247.it');INSERT INTO public.policy_tab (id, description, usr) VALUES (3,'tre','manager@scuola247.it');INSERT INTO public.policy_tab (id, description, usr) VALUES (4,'quattro','teacher@scuola247.it');INSERT INTO public.policy_tab (id, description, usr) VALUES (5,'cinque','teacher@scuola247.it');-----------cut here ----------------------after that i run the query: \\\"select * from public.policy_tab\\\"and the the oupt was what i excpected:rows 1,2,3 for user: manager@scuola247.itrows 4,5 for user: teacher@scuola247.itrows 1,2,3,4,5 for user: postgres (the policy doesn't work for him)but when i run the query: \\\"select * from public.policy_view\\\"the ouput is the same (all rows) for all users i'm doing some mistakes or this is a bug ?thank you in advance for the time you would like dedicate to me.Andrea Adami\",\n \"msg_date\": \"Sat, 20 Aug 2016 04:15:10 +0200\",\n \"msg_from\": \"Andrea Adami <fol@fulcro.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"CREATE POLICY bug ?\"\n },\n {\n \"msg_contents\": \"On 20 August 2016 at 03:15, Andrea Adami <fol@fulcro.net> wrote:\\n> when i run the query: \\\"select * from public.policy_view\\\"\\n> the ouput is the same (all rows) for all users\\n> i'm doing some mistakes or this is a bug ?\\n>\\n\\nNo, it looks correct to me. When going through a view, the policies\\nand permission checks that apply are those that would apply to the\\nview's owner, which in this case is postgres, so no policies are\\napplied.\\n\\nOr, quoting from the notes in the CREATE POLICY documentation:\\n\\n As with normal queries and views, permission checks and policies for\\n the tables which are referenced by a view will use the view owner's\\n rights and any policies which apply to the view owner.\\n\\nRegards,\\nDean\\n\\n\\n-- \\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, 20 Aug 2016 15:01:40 +0100\",\n \"msg_from\": \"Dean Rasheed <dean.a.rasheed@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: CREATE POLICY bug ?\"\n },\n {\n \"msg_contents\": \"[Please reply to the list, not just to me, so that others can benefit\\nfrom and contribute to the discussion]\\n\\nOn 31 August 2016 at 11:52, Andrea Adami <fol@fulcro.net> wrote:\\n> Thnaks Dean, i did further investigations:\\n> i set the owner of the view to: \\\"manager@scuola247.it\\\" with:\\n> ALTER TABLE public.policy_view OWNER TO \\\"manager@scuola247.it\\\";\\n> and i thinking to see from the select:\\n> select * from policy_view\\n> the rows: 1,2,3\\n> then\\n> set role 'manager@scuola247.it';\\n> select * from policy_view;\\n> return rows 1,2,3 as expected but:\\n> set role 'teacher@scuola247.it';\\n> select * from policy_view;\\n> returns rows 4,5 and\\n> set role 'postgres'\\n> select * from policy_view\\n> return nothing ...\\n> what you thinking about ?\\n>\\n> Andrea\\n\\nThat's correct. With the table owned by postgres and the view owned by\\n\\\"manager@scuola247.it\\\", access to the table via the view is subject to\\nthe policies that apply to \\\"manager@scuola247.it\\\". So regardless of\\nwho the current user is, when selecting from the view, the policy\\n\\\"standard\\\" will be applied, and that will limit the visible rows to\\nthose for which usr = current_user.\\n\\nRegards,\\nDean\\n\\n\\n-- \\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, 1 Sep 2016 10:53:51 +0100\",\n \"msg_from\": \"Dean Rasheed <dean.a.rasheed@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: CREATE POLICY bug ?\"\n }\n]","isConversational":false}}},{"rowIdx":42991,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"I have done my yearly TODO list cleanup, and I did more extensive item\nremoval this time. There were a number of entries I could not figure out\nso if people want to review what is left and remove items that are\nundesired or done, please do that. Thanks.\n\n\thttps://wiki.postgresql.org/wiki/Todo\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n\n-- \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, 20 Aug 2016 00:31:24 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":true,"msg_subject":"TODO list updated for PG 10"}],"string":"[\n {\n \"msg_contents\": \"I have done my yearly TODO list cleanup, and I did more extensive item\\nremoval this time. There were a number of entries I could not figure out\\nso if people want to review what is left and remove items that are\\nundesired or done, please do that. Thanks.\\n\\n\\thttps://wiki.postgresql.org/wiki/Todo\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\\n-- \\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, 20 Aug 2016 00:31:24 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"TODO list updated for PG 10\"\n }\n]","isConversational":false}}},{"rowIdx":42992,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Someone reported that a replication slot that existed at the time a base\nbackup was done on the master was copied to the standby. Because they\ndidn't realize it, their WAL was not being recycled on the standby.\n\nIs that possible? Is it a known behavior? I don't see it documented.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n\n-- \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, 20 Aug 2016 00:39:39 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":true,"msg_subject":"replication slots replicated to standbys?"},{"msg_contents":"On Sat, Aug 20, 2016 at 1:39 PM, Bruce Momjian <bruce@momjian.us> wrote:\n> Someone reported that a replication slot that existed at the time a base\n> backup was done on the master was copied to the standby. Because they\n> didn't realize it, their WAL was not being recycled on the standby.\n>\n> Is that possible? Is it a known behavior? I don't see it documented.\n\n From backup.sgml:\n <para>\n It is often a good idea to also omit from the backup the files\n within the cluster's <filename>pg_replslot/</> directory, so that\n replication slots that exist on the master do not become part of the\n backup. Otherwise, the subsequent use of the backup to create a standby\n may result in indefinite retention of WAL files on the standby, and\n possibly bloat on the master if hot standby feedback is enabled, because\n the clients that are using those replication slots will still be connecting\n to and updating the slots on the master, not the standby. Even if the\n backup is only intended for use in creating a new master, copying the\n replication slots isn't expected to be particularly useful, since the\n contents of those slots will likely be badly out of date by the time\n the new master comes on line.\n </para>\n\nNote as well that pg_basebackup omits its content and creates an empty\ndirectory.\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":"Sat, 20 Aug 2016 13:43:42 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: replication slots replicated to standbys?"},{"msg_contents":"On Sat, Aug 20, 2016 at 01:43:42PM +0900, Michael Paquier wrote:\n> On Sat, Aug 20, 2016 at 1:39 PM, Bruce Momjian <bruce@momjian.us> wrote:\n> > Someone reported that a replication slot that existed at the time a base\n> > backup was done on the master was copied to the standby. Because they\n> > didn't realize it, their WAL was not being recycled on the standby.\n> >\n> > Is that possible? Is it a known behavior? I don't see it documented.\n> \n> >From backup.sgml:\n> <para>\n> It is often a good idea to also omit from the backup the files\n> within the cluster's <filename>pg_replslot/</> directory, so that\n> replication slots that exist on the master do not become part of the\n> backup. Otherwise, the subsequent use of the backup to create a standby\n> may result in indefinite retention of WAL files on the standby, and\n> possibly bloat on the master if hot standby feedback is enabled, because\n> the clients that are using those replication slots will still be connecting\n> to and updating the slots on the master, not the standby. Even if the\n> backup is only intended for use in creating a new master, copying the\n> replication slots isn't expected to be particularly useful, since the\n> contents of those slots will likely be badly out of date by the time\n> the new master comes on line.\n> </para>\n> \n> Note as well that pg_basebackup omits its content and creates an empty\n> directory.\n\nSeems like another good idea to use pg_basebackup rather than manually\ndoing base backups; Magnus has been saying this for a while.\n\nI supposed there is no way we could remove this error-prone behavior\nbecause replication slots must survive server restarts. Is there no way\nto know if we are starting a standby from a fresh base backup vs.\nrestarting a standby? In that case we could clear the replication\nslots. Are there any other error-prone things copied from the master?\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n\n-- \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, 20 Aug 2016 12:35:36 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":true,"msg_subject":"Re: replication slots replicated to standbys?"},{"msg_contents":"On 21 Aug 2016 12:36 AM, \"Bruce Momjian\" <bruce@momjian.us> wrote:\n>\n> On Sat, Aug 20, 2016 at 01:43:42PM +0900, Michael Paquier wrote:\n> > On Sat, Aug 20, 2016 at 1:39 PM, Bruce Momjian <bruce@momjian.us> wrote:\n> > > Someone reported that a replication slot that existed at the time a\nbase\n> > > backup was done on the master was copied to the standby. Because they\n> > > didn't realize it, their WAL was not being recycled on the standby.\n> > >\n> > > Is that possible? Is it a known behavior? I don't see it documented.\n> >\n> > >From backup.sgml:\n> > <para>\n> > It is often a good idea to also omit from the backup the files\n> > within the cluster's <filename>pg_replslot/</> directory, so that\n> > replication slots that exist on the master do not become part of the\n> > backup. Otherwise, the subsequent use of the backup to create a\nstandby\n> > may result in indefinite retention of WAL files on the standby, and\n> > possibly bloat on the master if hot standby feedback is enabled,\nbecause\n> > the clients that are using those replication slots will still be\nconnecting\n> > to and updating the slots on the master, not the standby. Even if\nthe\n> > backup is only intended for use in creating a new master, copying\nthe\n> > replication slots isn't expected to be particularly useful, since\nthe\n> > contents of those slots will likely be badly out of date by the time\n> > the new master comes on line.\n> > </para>\n> >\n> > Note as well that pg_basebackup omits its content and creates an empty\n> > directory.\n>\n> Seems like another good idea to use pg_basebackup rather than manually\n> doing base backups; Magnus has been saying this for a while.\n\nThe main time that's an issue is when you're rsync'ing to save bandwidth,\nusing CoW volume snapshots, etc. pg_basebackup becomes totally impractical\non big systems.\n\n> I supposed there is no way we could remove this error-prone behavior\n> because replication slots must survive server restarts. Is there no way\n> to know if we are starting a standby from a fresh base backup vs.\n> restarting a standby? In that case we could clear the replication\n> slots. Are there any other error-prone things copied from the master?\n\nWe could remove slots when we enter archive recovery. But I've recently\nimplememted support for logical decoding from standbys, which needs slots.\nPhysical slot use on standby is also handy. We cannot tell whether a slot\nwas created on the replica or created on the master and copied in the base\nbackup and don't want to drop slots created on the replica.\n\nI also have use cases for slots being retained in restore from snapshot,\nfor re-integrating restored nodes into an MM mesh.\n\nI think a recovery.conf option to remove all slots during archive recovery\ncould be handy. But mostly it comes down to tools not copying them.\n\n\nOn 21 Aug 2016 12:36 AM, \"Bruce Momjian\" <bruce@momjian.us> wrote:\n>\n> On Sat, Aug 20, 2016 at 01:43:42PM +0900, Michael Paquier wrote:\n> > On Sat, Aug 20, 2016 at 1:39 PM, Bruce Momjian <bruce@momjian.us> wrote:\n> > > Someone reported that a replication slot that existed at the time a base\n> > > backup was done on the master was copied to the standby. Because they\n> > > didn't realize it, their WAL was not being recycled on the standby.\n> > >\n> > > Is that possible? Is it a known behavior? I don't see it documented.\n> >\n> > >From backup.sgml:\n> > <para>\n> > It is often a good idea to also omit from the backup the files\n> > within the cluster's <filename>pg_replslot/</> directory, so that\n> > replication slots that exist on the master do not become part of the\n> > backup. Otherwise, the subsequent use of the backup to create a standby\n> > may result in indefinite retention of WAL files on the standby, and\n> > possibly bloat on the master if hot standby feedback is enabled, because\n> > the clients that are using those replication slots will still be connecting\n> > to and updating the slots on the master, not the standby. Even if the\n> > backup is only intended for use in creating a new master, copying the\n> > replication slots isn't expected to be particularly useful, since the\n> > contents of those slots will likely be badly out of date by the time\n> > the new master comes on line.\n> > </para>\n> >\n> > Note as well that pg_basebackup omits its content and creates an empty\n> > directory.\n>\n> Seems like another good idea to use pg_basebackup rather than manually\n> doing base backups; Magnus has been saying this for a while.\nThe main time that's an issue is when you're rsync'ing to save bandwidth, using CoW volume snapshots, etc. pg_basebackup becomes totally impractical on big systems.\n> I supposed there is no way we could remove this error-prone behavior\n> because replication slots must survive server restarts. Is there no way\n> to know if we are starting a standby from a fresh base backup vs.\n> restarting a standby? In that case we could clear the replication\n> slots. Are there any other error-prone things copied from the master?\nWe could remove slots when we enter archive recovery. But I've recently implememted support for logical decoding from standbys, which needs slots. Physical slot use on standby is also handy. We cannot tell whether a slot was created on the replica or created on the master and copied in the base backup and don't want to drop slots created on the replica.\nI also have use cases for slots being retained in restore from snapshot, for re-integrating restored nodes into an MM mesh.\nI think a recovery.conf option to remove all slots during archive recovery could be handy. But mostly it comes down to tools not copying them.","msg_date":"Sun, 21 Aug 2016 12:24:50 +0800","msg_from":"Craig Ringer <craig.ringer@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: replication slots replicated to standbys?"},{"msg_contents":"On Sun, Aug 21, 2016 at 1:24 PM, Craig Ringer\n<craig.ringer@2ndquadrant.com> wrote:\n> On 21 Aug 2016 12:36 AM, \"Bruce Momjian\" <bruce@momjian.us> wrote:\n>> Seems like another good idea to use pg_basebackup rather than manually\n>> doing base backups; Magnus has been saying this for a while.\n>\n> The main time that's an issue is when you're rsync'ing to save bandwidth,\n> using CoW volume snapshots, etc. pg_basebackup becomes totally impractical\n> on big systems.\n\nYes, and that's not fun. Particularly when the backup takes so long\nthat WAL has already been recycled... Replication slots help here but\nthe partitions dedicated to pg_xlog have their limit as well.\n\n>> I supposed there is no way we could remove this error-prone behavior\n>> because replication slots must survive server restarts. Is there no way\n>> to know if we are starting a standby from a fresh base backup vs.\n>> restarting a standby? In that case we could clear the replication\n>> slots. Are there any other error-prone things copied from the master?\n>\n> We could remove slots when we enter archive recovery. But I've recently\n> implemented support for logical decoding from standbys, which needs slots.\n> Physical slot use on standby is also handy. We cannot tell whether a slot\n> was created on the replica or created on the master and copied in the base\n> backup and don't want to drop slots created on the replica.\n>\n> I also have use cases for slots being retained in restore from snapshot, for\n> re-integrating restored nodes into an MM mesh.\n>\n> I think a recovery.conf option to remove all slots during archive recovery\n> could be handy. But mostly it comes down to tools not copying them.\n\nYes, I'd personally let recovery.conf out of that, as well as the\nremoval of replication slot data when archive recovery begins to keep\nthe configuration simple. The decision-making of the data included in\nany backup will be done by the tool itself anyway..\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, 22 Aug 2016 11:31:42 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: replication slots replicated to standbys?"},{"msg_contents":"On Sun, Aug 21, 2016 at 1:35 AM, Bruce Momjian <bruce@momjian.us> wrote:\n> On Sat, Aug 20, 2016 at 01:43:42PM +0900, Michael Paquier wrote:\n>> Note as well that pg_basebackup omits its content and creates an empty\n>> directory.\n>\n> Are there any other error-prone things copied from the master?\n\nThe contents of pg_snapshots get copied by pg_basebackup. Those are\nuseless in a backup, but harmless.\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, 22 Aug 2016 11:49:02 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: replication slots replicated to standbys?"},{"msg_contents":"On 22 August 2016 at 10:31, Michael Paquier <michael.paquier@gmail.com>\nwrote:\n\n> On Sun, Aug 21, 2016 at 1:24 PM, Craig Ringer\n> <craig.ringer@2ndquadrant.com> wrote:\n> > On 21 Aug 2016 12:36 AM, \"Bruce Momjian\" <bruce@momjian.us> wrote:\n> >> Seems like another good idea to use pg_basebackup rather than manually\n> >> doing base backups; Magnus has been saying this for a while.\n> >\n> > The main time that's an issue is when you're rsync'ing to save bandwidth,\n> > using CoW volume snapshots, etc. pg_basebackup becomes totally\n> impractical\n> > on big systems.\n>\n> Yes, and that's not fun. Particularly when the backup takes so long\n> that WAL has already been recycled... Replication slots help here but\n> the partitions dedicated to pg_xlog have their limit as well.\n>\n\nWe can and probably should allow XLogReader to invoke restore_command to\nfetch WAL, read it, and discard/recycle it again. This would greatly\nalleviate the pain of indefinite xlog retention.\n\nIt's a pain to do so while recovery.conf is its own separate magic though,\nnot part of postgresql.conf.\n\nI have no plans to work on this at this time.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\nOn 22 August 2016 at 10:31, Michael Paquier <michael.paquier@gmail.com> wrote:On Sun, Aug 21, 2016 at 1:24 PM, Craig Ringer\n<craig.ringer@2ndquadrant.com> wrote:\n> On 21 Aug 2016 12:36 AM, \"Bruce Momjian\" <bruce@momjian.us> wrote:\n>> Seems like another good idea to use pg_basebackup rather than manually\n>> doing base backups; Magnus has been saying this for a while.\n>\n> The main time that's an issue is when you're rsync'ing to save bandwidth,\n> using CoW volume snapshots, etc. pg_basebackup becomes totally impractical\n> on big systems.\n\nYes, and that's not fun. Particularly when the backup takes so long\nthat WAL has already been recycled... Replication slots help here but\nthe partitions dedicated to pg_xlog have their limit as well.We can and probably should allow XLogReader to invoke restore_command to fetch WAL, read it, and discard/recycle it again. This would greatly alleviate the pain of indefinite xlog retention.It's a pain to do so while recovery.conf is its own separate magic though, not part of postgresql.conf. I have no plans to work on this at this time.-- Craig Ringer http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Mon, 22 Aug 2016 11:17:25 +0800","msg_from":"Craig Ringer <craig.ringer@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: replication slots replicated to standbys?"}],"string":"[\n {\n \"msg_contents\": \"Someone reported that a replication slot that existed at the time a base\\nbackup was done on the master was copied to the standby. Because they\\ndidn't realize it, their WAL was not being recycled on the standby.\\n\\nIs that possible? Is it a known behavior? I don't see it documented.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\\n-- \\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, 20 Aug 2016 00:39:39 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"replication slots replicated to standbys?\"\n },\n {\n \"msg_contents\": \"On Sat, Aug 20, 2016 at 1:39 PM, Bruce Momjian <bruce@momjian.us> wrote:\\n> Someone reported that a replication slot that existed at the time a base\\n> backup was done on the master was copied to the standby. Because they\\n> didn't realize it, their WAL was not being recycled on the standby.\\n>\\n> Is that possible? Is it a known behavior? I don't see it documented.\\n\\n From backup.sgml:\\n <para>\\n It is often a good idea to also omit from the backup the files\\n within the cluster's <filename>pg_replslot/</> directory, so that\\n replication slots that exist on the master do not become part of the\\n backup. Otherwise, the subsequent use of the backup to create a standby\\n may result in indefinite retention of WAL files on the standby, and\\n possibly bloat on the master if hot standby feedback is enabled, because\\n the clients that are using those replication slots will still be connecting\\n to and updating the slots on the master, not the standby. Even if the\\n backup is only intended for use in creating a new master, copying the\\n replication slots isn't expected to be particularly useful, since the\\n contents of those slots will likely be badly out of date by the time\\n the new master comes on line.\\n </para>\\n\\nNote as well that pg_basebackup omits its content and creates an empty\\ndirectory.\\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\": \"Sat, 20 Aug 2016 13:43:42 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: replication slots replicated to standbys?\"\n },\n {\n \"msg_contents\": \"On Sat, Aug 20, 2016 at 01:43:42PM +0900, Michael Paquier wrote:\\n> On Sat, Aug 20, 2016 at 1:39 PM, Bruce Momjian <bruce@momjian.us> wrote:\\n> > Someone reported that a replication slot that existed at the time a base\\n> > backup was done on the master was copied to the standby. Because they\\n> > didn't realize it, their WAL was not being recycled on the standby.\\n> >\\n> > Is that possible? Is it a known behavior? I don't see it documented.\\n> \\n> >From backup.sgml:\\n> <para>\\n> It is often a good idea to also omit from the backup the files\\n> within the cluster's <filename>pg_replslot/</> directory, so that\\n> replication slots that exist on the master do not become part of the\\n> backup. Otherwise, the subsequent use of the backup to create a standby\\n> may result in indefinite retention of WAL files on the standby, and\\n> possibly bloat on the master if hot standby feedback is enabled, because\\n> the clients that are using those replication slots will still be connecting\\n> to and updating the slots on the master, not the standby. Even if the\\n> backup is only intended for use in creating a new master, copying the\\n> replication slots isn't expected to be particularly useful, since the\\n> contents of those slots will likely be badly out of date by the time\\n> the new master comes on line.\\n> </para>\\n> \\n> Note as well that pg_basebackup omits its content and creates an empty\\n> directory.\\n\\nSeems like another good idea to use pg_basebackup rather than manually\\ndoing base backups; Magnus has been saying this for a while.\\n\\nI supposed there is no way we could remove this error-prone behavior\\nbecause replication slots must survive server restarts. Is there no way\\nto know if we are starting a standby from a fresh base backup vs.\\nrestarting a standby? In that case we could clear the replication\\nslots. Are there any other error-prone things copied from the master?\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\\n-- \\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, 20 Aug 2016 12:35:36 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: replication slots replicated to standbys?\"\n },\n {\n \"msg_contents\": \"On 21 Aug 2016 12:36 AM, \\\"Bruce Momjian\\\" <bruce@momjian.us> wrote:\\n>\\n> On Sat, Aug 20, 2016 at 01:43:42PM +0900, Michael Paquier wrote:\\n> > On Sat, Aug 20, 2016 at 1:39 PM, Bruce Momjian <bruce@momjian.us> wrote:\\n> > > Someone reported that a replication slot that existed at the time a\\nbase\\n> > > backup was done on the master was copied to the standby. Because they\\n> > > didn't realize it, their WAL was not being recycled on the standby.\\n> > >\\n> > > Is that possible? Is it a known behavior? I don't see it documented.\\n> >\\n> > >From backup.sgml:\\n> > <para>\\n> > It is often a good idea to also omit from the backup the files\\n> > within the cluster's <filename>pg_replslot/</> directory, so that\\n> > replication slots that exist on the master do not become part of the\\n> > backup. Otherwise, the subsequent use of the backup to create a\\nstandby\\n> > may result in indefinite retention of WAL files on the standby, and\\n> > possibly bloat on the master if hot standby feedback is enabled,\\nbecause\\n> > the clients that are using those replication slots will still be\\nconnecting\\n> > to and updating the slots on the master, not the standby. Even if\\nthe\\n> > backup is only intended for use in creating a new master, copying\\nthe\\n> > replication slots isn't expected to be particularly useful, since\\nthe\\n> > contents of those slots will likely be badly out of date by the time\\n> > the new master comes on line.\\n> > </para>\\n> >\\n> > Note as well that pg_basebackup omits its content and creates an empty\\n> > directory.\\n>\\n> Seems like another good idea to use pg_basebackup rather than manually\\n> doing base backups; Magnus has been saying this for a while.\\n\\nThe main time that's an issue is when you're rsync'ing to save bandwidth,\\nusing CoW volume snapshots, etc. pg_basebackup becomes totally impractical\\non big systems.\\n\\n> I supposed there is no way we could remove this error-prone behavior\\n> because replication slots must survive server restarts. Is there no way\\n> to know if we are starting a standby from a fresh base backup vs.\\n> restarting a standby? In that case we could clear the replication\\n> slots. Are there any other error-prone things copied from the master?\\n\\nWe could remove slots when we enter archive recovery. But I've recently\\nimplememted support for logical decoding from standbys, which needs slots.\\nPhysical slot use on standby is also handy. We cannot tell whether a slot\\nwas created on the replica or created on the master and copied in the base\\nbackup and don't want to drop slots created on the replica.\\n\\nI also have use cases for slots being retained in restore from snapshot,\\nfor re-integrating restored nodes into an MM mesh.\\n\\nI think a recovery.conf option to remove all slots during archive recovery\\ncould be handy. But mostly it comes down to tools not copying them.\\n\\n\\nOn 21 Aug 2016 12:36 AM, \\\"Bruce Momjian\\\" <bruce@momjian.us> wrote:\\n>\\n> On Sat, Aug 20, 2016 at 01:43:42PM +0900, Michael Paquier wrote:\\n> > On Sat, Aug 20, 2016 at 1:39 PM, Bruce Momjian <bruce@momjian.us> wrote:\\n> > > Someone reported that a replication slot that existed at the time a base\\n> > > backup was done on the master was copied to the standby. Because they\\n> > > didn't realize it, their WAL was not being recycled on the standby.\\n> > >\\n> > > Is that possible? Is it a known behavior? I don't see it documented.\\n> >\\n> > >From backup.sgml:\\n> > <para>\\n> > It is often a good idea to also omit from the backup the files\\n> > within the cluster's <filename>pg_replslot/</> directory, so that\\n> > replication slots that exist on the master do not become part of the\\n> > backup. Otherwise, the subsequent use of the backup to create a standby\\n> > may result in indefinite retention of WAL files on the standby, and\\n> > possibly bloat on the master if hot standby feedback is enabled, because\\n> > the clients that are using those replication slots will still be connecting\\n> > to and updating the slots on the master, not the standby. Even if the\\n> > backup is only intended for use in creating a new master, copying the\\n> > replication slots isn't expected to be particularly useful, since the\\n> > contents of those slots will likely be badly out of date by the time\\n> > the new master comes on line.\\n> > </para>\\n> >\\n> > Note as well that pg_basebackup omits its content and creates an empty\\n> > directory.\\n>\\n> Seems like another good idea to use pg_basebackup rather than manually\\n> doing base backups; Magnus has been saying this for a while.\\nThe main time that's an issue is when you're rsync'ing to save bandwidth, using CoW volume snapshots, etc. pg_basebackup becomes totally impractical on big systems.\\n> I supposed there is no way we could remove this error-prone behavior\\n> because replication slots must survive server restarts. Is there no way\\n> to know if we are starting a standby from a fresh base backup vs.\\n> restarting a standby? In that case we could clear the replication\\n> slots. Are there any other error-prone things copied from the master?\\nWe could remove slots when we enter archive recovery. But I've recently implememted support for logical decoding from standbys, which needs slots. Physical slot use on standby is also handy. We cannot tell whether a slot was created on the replica or created on the master and copied in the base backup and don't want to drop slots created on the replica.\\nI also have use cases for slots being retained in restore from snapshot, for re-integrating restored nodes into an MM mesh.\\nI think a recovery.conf option to remove all slots during archive recovery could be handy. But mostly it comes down to tools not copying them.\",\n \"msg_date\": \"Sun, 21 Aug 2016 12:24:50 +0800\",\n \"msg_from\": \"Craig Ringer <craig.ringer@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: replication slots replicated to standbys?\"\n },\n {\n \"msg_contents\": \"On Sun, Aug 21, 2016 at 1:24 PM, Craig Ringer\\n<craig.ringer@2ndquadrant.com> wrote:\\n> On 21 Aug 2016 12:36 AM, \\\"Bruce Momjian\\\" <bruce@momjian.us> wrote:\\n>> Seems like another good idea to use pg_basebackup rather than manually\\n>> doing base backups; Magnus has been saying this for a while.\\n>\\n> The main time that's an issue is when you're rsync'ing to save bandwidth,\\n> using CoW volume snapshots, etc. pg_basebackup becomes totally impractical\\n> on big systems.\\n\\nYes, and that's not fun. Particularly when the backup takes so long\\nthat WAL has already been recycled... Replication slots help here but\\nthe partitions dedicated to pg_xlog have their limit as well.\\n\\n>> I supposed there is no way we could remove this error-prone behavior\\n>> because replication slots must survive server restarts. Is there no way\\n>> to know if we are starting a standby from a fresh base backup vs.\\n>> restarting a standby? In that case we could clear the replication\\n>> slots. Are there any other error-prone things copied from the master?\\n>\\n> We could remove slots when we enter archive recovery. But I've recently\\n> implemented support for logical decoding from standbys, which needs slots.\\n> Physical slot use on standby is also handy. We cannot tell whether a slot\\n> was created on the replica or created on the master and copied in the base\\n> backup and don't want to drop slots created on the replica.\\n>\\n> I also have use cases for slots being retained in restore from snapshot, for\\n> re-integrating restored nodes into an MM mesh.\\n>\\n> I think a recovery.conf option to remove all slots during archive recovery\\n> could be handy. But mostly it comes down to tools not copying them.\\n\\nYes, I'd personally let recovery.conf out of that, as well as the\\nremoval of replication slot data when archive recovery begins to keep\\nthe configuration simple. The decision-making of the data included in\\nany backup will be done by the tool itself anyway..\\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, 22 Aug 2016 11:31:42 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: replication slots replicated to standbys?\"\n },\n {\n \"msg_contents\": \"On Sun, Aug 21, 2016 at 1:35 AM, Bruce Momjian <bruce@momjian.us> wrote:\\n> On Sat, Aug 20, 2016 at 01:43:42PM +0900, Michael Paquier wrote:\\n>> Note as well that pg_basebackup omits its content and creates an empty\\n>> directory.\\n>\\n> Are there any other error-prone things copied from the master?\\n\\nThe contents of pg_snapshots get copied by pg_basebackup. Those are\\nuseless in a backup, but harmless.\\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, 22 Aug 2016 11:49:02 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: replication slots replicated to standbys?\"\n },\n {\n \"msg_contents\": \"On 22 August 2016 at 10:31, Michael Paquier <michael.paquier@gmail.com>\\nwrote:\\n\\n> On Sun, Aug 21, 2016 at 1:24 PM, Craig Ringer\\n> <craig.ringer@2ndquadrant.com> wrote:\\n> > On 21 Aug 2016 12:36 AM, \\\"Bruce Momjian\\\" <bruce@momjian.us> wrote:\\n> >> Seems like another good idea to use pg_basebackup rather than manually\\n> >> doing base backups; Magnus has been saying this for a while.\\n> >\\n> > The main time that's an issue is when you're rsync'ing to save bandwidth,\\n> > using CoW volume snapshots, etc. pg_basebackup becomes totally\\n> impractical\\n> > on big systems.\\n>\\n> Yes, and that's not fun. Particularly when the backup takes so long\\n> that WAL has already been recycled... Replication slots help here but\\n> the partitions dedicated to pg_xlog have their limit as well.\\n>\\n\\nWe can and probably should allow XLogReader to invoke restore_command to\\nfetch WAL, read it, and discard/recycle it again. This would greatly\\nalleviate the pain of indefinite xlog retention.\\n\\nIt's a pain to do so while recovery.conf is its own separate magic though,\\nnot part of postgresql.conf.\\n\\nI have no plans to work on this at this time.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\nOn 22 August 2016 at 10:31, Michael Paquier <michael.paquier@gmail.com> wrote:On Sun, Aug 21, 2016 at 1:24 PM, Craig Ringer\\n<craig.ringer@2ndquadrant.com> wrote:\\n> On 21 Aug 2016 12:36 AM, \\\"Bruce Momjian\\\" <bruce@momjian.us> wrote:\\n>> Seems like another good idea to use pg_basebackup rather than manually\\n>> doing base backups; Magnus has been saying this for a while.\\n>\\n> The main time that's an issue is when you're rsync'ing to save bandwidth,\\n> using CoW volume snapshots, etc. pg_basebackup becomes totally impractical\\n> on big systems.\\n\\nYes, and that's not fun. Particularly when the backup takes so long\\nthat WAL has already been recycled... Replication slots help here but\\nthe partitions dedicated to pg_xlog have their limit as well.We can and probably should allow XLogReader to invoke restore_command to fetch WAL, read it, and discard/recycle it again. This would greatly alleviate the pain of indefinite xlog retention.It's a pain to do so while recovery.conf is its own separate magic though, not part of postgresql.conf. I have no plans to work on this at this time.-- Craig Ringer http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Mon, 22 Aug 2016 11:17:25 +0800\",\n \"msg_from\": \"Craig Ringer <craig.ringer@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: replication slots replicated to standbys?\"\n }\n]","isConversational":false}}},{"rowIdx":42993,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi Yury,\n\n\nOn Fri, Aug 19, 2016 at 9:46 AM, Yury Zhuravlev\n<u.zhuravlev@postgrespro.ru> wrote:\n> Christian Convey wrote:\n>>\n>> I'm interested in helping with your CMake effort. I don't have any\n>> experience contributing to PG, but I do have some free time at the\n>> moment. Please let me know if I can help.\n>\n> I glad to hear it. I suppose you can just try build postgres and send all\n> problems to github tracker.\n> https://github.com/stalkerg/postgres_cmake/issues\n\nThanks, I'll be happy to do that. There's been a lot of discussion on\nthis thread regarding the minimum required CMake version. If you'd\nlike me to test with a particular version (or versions) of CMake,\nplease let me know.\n\n- Christian\n\n\n-- \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, 20 Aug 2016 01:18:29 -0400","msg_from":"Christian Convey <christian.convey@gmail.com>","msg_from_op":true,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Hi Yury,\n\n>> I glad to hear it. I suppose you can just try build postgres and send all\n>> problems to github tracker.\n>> https://github.com/stalkerg/postgres_cmake/issues\n\nFYI, I had success using your \"postgres_cmake\" repo. I tested it up\nthrough \"make check\" and \"make install\".\n\nHere are the details:\n\n* postgres_cmake commit e7e75160d4533cd8caa9f3f0dd7b485dbd4e7bdf\n* compiler = cc (Ubuntu 4.8.4-2ubuntu1~14.04.3) 4.8.4\n* cmake version = 3.5.3\n\nKind regards,\nChristian\n\n\n-- \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, 21 Aug 2016 18:48:39 -0400","msg_from":"Christian Convey <christian.convey@gmail.com>","msg_from_op":true,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Yury,\n\nOn Mon, Aug 22, 2016 at 7:48 AM, Christian Convey\n<christian.convey@gmail.com> wrote:\n>>> I glad to hear it. I suppose you can just try build postgres and send all\n>>> problems to github tracker.\n>>> https://github.com/stalkerg/postgres_cmake/issues\n>\n> FYI, I had success using your \"postgres_cmake\" repo. I tested it up\n> through \"make check\" and \"make install\".\n>\n> Here are the details:\n>\n> * postgres_cmake commit e7e75160d4533cd8caa9f3f0dd7b485dbd4e7bdf\n> * compiler = cc (Ubuntu 4.8.4-2ubuntu1~14.04.3) 4.8.4\n> * cmake version = 3.5.3\n\nCould it be possible to get a refreshed patch on this thread for at\nleast the sake of the archives? I'd really like to see somehting\nhappening here and do some progress for this CF.\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 Sep 2016 21:41:35 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Michael Paquier wrote:\n> Could it be possible to get a refreshed patch on this thread for at\n> least the sake of the archives? I'd really like to see somehting\n> happening here and do some progress for this CF.\n\nSure, I will do it on Friday.\nToday I finished mingw+msys support. (mingw without msys has not passed \nsome tests)\n-- \nYury Zhuravlev\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres 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, 15 Sep 2016 22:38:16 +0300","msg_from":"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Fri, Sep 16, 2016 at 4:38 AM, Yury Zhuravlev\n<u.zhuravlev@postgrespro.ru> wrote:\n> Michael Paquier wrote:\n>>\n>> Could it be possible to get a refreshed patch on this thread for at\n>> least the sake of the archives? I'd really like to see somehting\n>> happening here and do some progress for this CF.\n>\n>\n> Sure, I will do it on Friday.\n> Today I finished mingw+msys support. (mingw without msys has not passed some\n> tests)\n\nOkay. That sounds good to me. I don't recall what your patch is\nexactly doing but could you still keep the vanilla Makefiles around?\nThis will reduce the diff of the patch, and we'd need anyway to keep\nthe former Makefile methods around until the buildfarm scripts are\nupdated, the animals do the switch and then get green. So a period\nwhere both live in parallel is unavoidable.\n\nI heard as well that MySQL is using cmake... It may be interesting to\nsee how they are integrating with it as a large project and avoid\ntheir past mistakes.\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, 16 Sep 2016 09:58:42 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Fri, Sep 16, 2016 at 9:58 AM, Michael Paquier\n<michael.paquier@gmail.com> wrote:\n> Okay. That sounds good to me. I donas well a look around, and cmake really looks like a robust alternative to ./configure. Now I am aware of the fact that your patch recommends to build 't recall what your patch is\n> exactly doing but could you still keep the vanilla Makefiles around?\n> This will reduce the diff of the patch, and we'd need anyway to keep\n> the former Makefile methods around until the buildfarm scripts are\n> updated, the animals do the switch and then get green. So a period\n> where both live in parallel is unavoidable.\n>\n> I heard as well that MySQL is using cmake... It may be interesting to\n> see how they are integrating with it as a large project and avoid\n> their past mistakes.\n\nI could not resist so I just had a look at your patch. I had as well a\nlook around, and cmake really looks like a robust alternative to\n./configure. In short, people doing now that:\n./configure\nmake\nmake install\nWould just do that:\ncmake .\nmake\nmake install\n\nI am no cmake guru yet, but VPATH builds are supported out of the box,\nwhich is cool.\n\nYour patch recommends to build with cmake after creating build/, now I\nwould expect most users to run cmake from the root folder. However\nthis causes all the Makefiles to get overwritten. As supporting all\nplatforms at once with cmake is going to be uncommitable, we are going\nto need both methods able to live together for a while. Well, they can\ncoexist with this patch as long as cmake is not run from the root of\nthe code tree, which is acceptable for me as long as the switch is not\ncompleted. However others may think differently on the matter.\n\nInstead of getting support for all existing platforms, I would\nrecommend as well focusing only on one platform for the time being,\nLinux, and get the work done correctly for that first. Once there is\nsomething committed, we will be able to patch the buildfarm, and get\nmachines to switch to cmake one by one. After those are switched, we\ncould extend that. Another point of contention is support for\nextensions. How long should we keep support for the existing PGXS? How\nexternal extensions would compile with the new thing infrastructure?\n\nWhich brings me to another point, your patch is focused on features,\nmeaning that per-OS checks are all grouped by feature, but it may be a\ngood idea to split checks by OS if necessary, with for example\nper-platform files and scripts in cmake/os/. And we could have just\nsomething for Linux now.\n\nA couple of issues I have noticed with your patch after a first set of tests:\n- root's .gitignore needs to add entries for CMakeFiles and\ncmake_install.cmake. You need as well to ignore CMakeCache\n- A couple of headers are generated, like cubeparse.h (?)\n- Currently a lot of users do things like that:\ncd src/test/regress/ && make check\nBut this patch breaks that, and that's not cool. Recovery tests in\nsrc/test/regress won't run either.\n\nThat's all I have for now. Looking forward to seeing some progress here.\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, 16 Sep 2016 22:33:50 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Michael Paquier wrote:\n> Your patch recommends to build with cmake after creating build/, now I\n> would expect most users to run cmake from the root folder. However\n> this causes all the Makefiles to get overwritten. As supporting all\n> platforms at once with cmake is going to be uncommitable, we are going\n> to need both methods able to live together for a while. Well, they can\n> coexist with this patch as long as cmake is not run from the root of\n> the code tree, which is acceptable for me as long as the switch is not\n> completed. However others may think differently on the matter.\n\nIt's really a good point. Forbidden run cmake from root it is better \ndecision for me (of course for start). \n\n>Instead of getting support for all existing platforms, I would\n>recommend as well focusing only on one platform for the time being,\n>Linux, and get the work done correctly for that first. Once there is\n>something committed, we will be able to patch the buildfarm, and get\n>machines to switch to cmake one by one. After those are switched, we\n>could extend that.\n\nYou mean in first version of patch I can focus on Linux systems?\n\n>Another point of contention is support for\n>extensions. How long should we keep support for the existing PGXS? How\n>external extensions would compile with the new thing infrastructure?\n\nAs long as possible. I hope I can make PGXS Makefiles generator.\n\n>Which brings me to another point, your patch is focused on features,\n>meaning that per-OS checks are all grouped by feature, but it may be a\n>good idea to split checks by OS if necessary, with for example\n>per-platform files and scripts in cmake/os/. And we could have just\n>something for Linux now.\n\nCurrently I do not have a lot OS specific tests. All checks are doing in \nsame manner. \n\n>- root's .gitignore needs to add entries for CMakeFiles and\n>cmake_install.cmake. You need as well to ignore CMakeCache\n\nThanks I done this in last commit. \n\n>- A couple of headers are generated, like cubeparse.h (?)\n\nBecause BISON generate header by default. I suppose author of cube launched \nbison by hand but I made it automatic. \n\n>- Currently a lot of users do things like that:\n>cd src/test/regress/ && make check\n>But this patch breaks that, and that's not cool. Recovery tests in\n>src/test/regress won't run either.\n\nIt seems restriction by design because in CMake you have only one enter \npoint. \n\n>That's all I have for now. Looking forward to seeing some progress here.\n\nI merged master to my branch and I spent time to porting all changes. I \nhope send patch in the weekend without terrible flaws.\n\n-- \nYury Zhuravlev\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres 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, 16 Sep 2016 19:40:48 +0300","msg_from":"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Sat, Sep 17, 2016 at 1:40 AM, Yury Zhuravlev\n<u.zhuravlev@postgrespro.ru> wrote:\n> Michael Paquier wrote:\n> I merged master to my branch and I spent time to porting all changes. I hope\n> send patch in the weekend without terrible flaws.\n\nBy the way, I noticed that you did not register this patch in the current CF..\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":"Sat, 17 Sep 2016 08:07:43 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Hi,\n\nOn 2016-09-16 22:33:50 +0900, Michael Paquier wrote:\n> As supporting all platforms at once with cmake is going to be\n> uncommitable, we are going to need both methods able to live together\n> for a while.\n\nI very strongly disagree with this. It's way too likely that we end up\nreleasing with multiple buildsystems that way. I think we'll have to\nrequire support for the most common OSs, and then fix up the fallout\nafterwards.\n\nAndres\n\n\n-- \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 Sep 2016 16:11:55 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Sat, Sep 17, 2016 at 8:11 AM, Andres Freund <andres@anarazel.de> wrote:\n> Hi,\n>\n> On 2016-09-16 22:33:50 +0900, Michael Paquier wrote:\n>> As supporting all platforms at once with cmake is going to be\n>> uncommitable, we are going to need both methods able to live together\n>> for a while.\n>\n> I very strongly disagree with this. It's way too likely that we end up\n> releasing with multiple buildsystems that way. I think we'll have to\n> require support for the most common OSs, and then fix up the fallout\n> afterwards.\n\nOk, then if we go through the hard method the switch patch had better\nremove as well all the Makefiles in the tree and recommend cmake run\nfrom base root as the default build method. By the way, there will be\na strong need for docs in the patch sooner or later...\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":"Sat, 17 Sep 2016 08:17:02 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Michael Paquier wrote:\n> On Sat, Sep 17, 2016 at 1:40 AM, Yury Zhuravlev\n> <u.zhuravlev@postgrespro.ru> wrote:\n>> Michael Paquier wrote:\n>> I merged master to my branch and I spent time to porting all \n>> changes. I hope\n>> send patch in the weekend without terrible flaws.\n>\n> By the way, I noticed that you did not register this patch in \n> the current CF..\n\nNow, I published the first version of the patch. \nThis patch not ready for commit yet and all current task you can read here:\nhttps://github.com/stalkerg/postgres_cmake/issues\n\nI hope we realy close to end. \n\nIn this patch I forbade in-source build for avoid overwriting current \nMakefiles. \nWe will can remove all Makefiles only after shall see in CMake. You don't \nneed support two system. During commitfests CMake build system will be \nsupported by me. \n\nI need help with buildfarm because my knowledge of Perl is very bad \n(thought about rewrite buildfarm to Python). \n\n\nI hope for your support.\n\n-- \nYury Zhuravlev\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres 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","msg_date":"Sat, 17 Sep 2016 20:21:55 +0300","msg_from":"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Andres Freund wrote:\n> It's way too likely that we end up\n> releasing with multiple buildsystems that way.\n\nIf we fail in the CMake integration we can easily remove all CMake files \nbefore code freeze. \nI hope step by step CMake integration better way for current situation. \nBut I do not insist, just I trying to find the easiest way.\n\n-- \nYury Zhuravlev\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres 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, 17 Sep 2016 20:43:07 +0300","msg_from":"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 9/16/16 9:33 AM, Michael Paquier wrote:\n> I am no cmake guru yet, but VPATH builds are supported out of the box,\n> which is cool.\n> \n> Your patch recommends to build with cmake after creating build/, now I\n> would expect most users to run cmake from the root folder.\n\nMy understanding is that cmake strongly recommends building in a\nseparate directory, which is usually a subdirectory named something like\n\"build\". So the above is likely going to be the standard workflow.\n\n-- \nPeter Eisentraut 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":"Sat, 17 Sep 2016 20:12:54 -0400","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"> On 17 Sep 2016, at 20:21, Yury Zhuravlev <u.zhuravlev@postgrespro.ru> wrote:\n> \n> Michael Paquier wrote:\n>> On Sat, Sep 17, 2016 at 1:40 AM, Yury Zhuravlev\n>> <u.zhuravlev@postgrespro.ru> wrote:\n>>> Michael Paquier wrote:\n>>> I merged master to my branch and I spent time to porting all changes. I hope\n>>> send patch in the weekend without terrible flaws.\n>> \n>> By the way, I noticed that you did not register this patch in the current CF..\n> \n> Now, I published the first version of the patch. This patch not ready for commit yet and all current task you can read here:\n> https://github.com/stalkerg/postgres_cmake/issues\n> \n> I hope we realy close to end. \n> In this patch I forbade in-source build for avoid overwriting current Makefiles. We will can remove all Makefiles only after shall see in CMake. You don't need support two system. During commitfests CMake build system will be supported by me. \n> I need help with buildfarm because my knowledge of Perl is very bad (thought about rewrite buildfarm to Python). \n> \n> I hope for your support.\n\n\nTried to generate Xcode project out of cmake, build fails on genbki.pl: can't locate Catalog.pm (which itself lives in src/backend/catalog/Catalog.pm)\n\n-- \nStas Kelvich\nPostgres Professional: http://www.postgrespro.com\nRussian Postgres Company\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, 3 Oct 2016 13:01:46 +0300","msg_from":"Stas Kelvich <s.kelvich@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Stas Kelvich wrote:\n>> On 17 Sep 2016, at 20:21, Yury Zhuravlev \n>> <u.zhuravlev@postgrespro.ru> wrote:\n>> \n>> Michael Paquier wrote: ...\n>\n>\n> Tried to generate Xcode project out of cmake, build fails on \n> genbki.pl: can't locate Catalog.pm (which itself lives in \n> src/backend/catalog/Catalog.pm)\n>\n\nCan you try again? On my Xcode 7.3 / ElCapitan everything is working \ncorrectly.\n-- \nYury Zhuravlev\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres 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, 07 Oct 2016 01:22:11 +0300","msg_from":"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Stas Kelvich wrote:\n> Tried to generate Xcode project out of cmake, build fails on \n> genbki.pl: can't locate Catalog.pm (which itself lives in \n> src/backend/catalog/Catalog.pm)\n\nThanks again! I have corrected your issue.\n-- \nYury Zhuravlev\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres 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, 14 Oct 2016 12:51:57 +0300","msg_from":"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 9/17/16 1:21 PM, Yury Zhuravlev wrote:\n> Now, I published the first version of the patch. \n\nI tried this out. Because of some file moves in initdb and\npg_basebackup, the build fails:\n\n[ 74%] Linking C executable initdb\n Undefined symbols for architecture x86_64:\n \"_fsync_pgdata\", referenced from:\n _main in initdb.c.o\n ld: symbol(s) not found for architecture x86_64\n collect2: error: ld returned 1 exit status\n make[2]: *** [src/bin/initdb/CMakeFiles/initdb.dir/build.make:177:\nsrc/bin/initdb/initdb] Error 1\n make[1]: *** [CMakeFiles/Makefile2:2893:\nsrc/bin/initdb/CMakeFiles/initdb.dir/all] Error 2\n make: *** [Makefile:128: all] Error 2\n\nPlease submit an updated patch.\n\nI suggest you use git format-patch to produce patches. This is easier\nto apply, especially when there are a lot of new files involved. Also\nuse the git facilities to check for whitespace errors.\n\nPlease supply some documentation, such as\n\n- what are the basic commands\n- how to set include/library paths, choose configure options\n- how to set CFLAGS\n- how to see raw build commands\n- what are the targets for all/world/check/docs etc.\n- explain directory structure\n\nI suggest for now you could put this into a README.cmake file in your\npatch. We don't need to commit it that way, but it would help in the\nmeantime.\n\nWhen I run cmake without options, it seems to do opportunistic feature\nchecking. For example, it turns on OpenSSL or Python support if it can\nfind it, otherwise it turns it off. We need this to be deterministic.\nWithout options, choose the basic feature set, require all other\nfeatures to be turned on explicitly, fail if they can't be found.\nWhatever the Autoconf-based build does now has been fairly deliberately\ntuned, so there should be very little reason to deviate from that.\n\nThe Python check appears to be picking up pieces from two different\nPython installations:\n\n -- Found PythonInterp: /usr/local/bin/python (found version \"2.7.12\")\n -- Found PythonLibs: /usr/lib/libpython2.7.dylib (found version \"2.7.10\")\n\nThe check results otherwise look OK, but I'm a bit confused about the\norder. It checks for some functions before all the header files are\nchecked for. Is that intentional?\n\nThere are a number of changes in .[ch] and .pl files that are unclear\nand not explained. Please explain them. You can also submit separate\npreliminary patches if you need to do some refactoring. Ultimately, I\nwould expect this patch not to require C code changes.\n\n-- \nPeter Eisentraut 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":"Tue, 8 Nov 2016 14:37:17 -0500","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Hello.\n\nPeter Eisentraut wrote:\n> I tried this out. Because of some file moves in initdb and\n> pg_basebackup, the build fails:\nYes, I need rebase patch to latest master. I do it as soon as possible. \nAlso I have some new achievements. \n\n>I suggest you use git format-patch to produce patches. This is easier\n>to apply, especially when there are a lot of new files involved. Also\n>use the git facilities to check for whitespace errors.\nI agree. \n\n>I suggest for now you could put this into a README.cmake file in your\n>patch. We don't need to commit it that way, but it would help in the\n>meantime.\nGood idea, currently I write all documentation on github. I will try to \nmove it to patch. \n\n>When I run cmake without options, it seems to do opportunistic feature\n>checking. For example, it turns on OpenSSL or Python support if it can\n>find it, otherwise it turns it off. We need this to be deterministic.\n>Without options, choose the basic feature set, require all other\n>features to be turned on explicitly, fail if they can't be found.\n>Whatever the Autoconf-based build does now has been fairly deliberately\n>tuned, so there should be very little reason to deviate from that.\nThis approach I see only in Postgres project and not fully understood.\nCan you explain me more what reasons led to this approach?\nNot big deal to make behavior like postgres Autoconf but I think it's \nimportant clear view here. \n\n>The Python check appears to be picking up pieces from two different\n>Python installations:\nOoops you right. For Python detecting I use standard CMake module and\nhis behavior depending on CMake version. We should really careful here.\n\n>The check results otherwise look OK, but I'm a bit confused about the\n>order. It checks for some functions before all the header files are\n>checked for. Is that intentional?\nI have plans to change order checks. \n\n>There are a number of changes in .[ch] and .pl files that are unclear\n>and not explained. Please explain them. You can also submit separate\n>preliminary patches if you need to do some refactoring. Ultimately, I\n>would expect this patch not to require C code changes.\nI suppose separate patches with comments will be best way. I will do it.\nI think we can talks about details after that.\n\nBig thanks for your remarks it's very important for me and this \"small\" \nproject.\nI will try to do all tasks quickly.\n\nBest regards. \n\n-- \nYury Zhuravlev\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres 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, 08 Nov 2016 23:40:57 +0300","msg_from":"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 9 Nov. 2016 06:37, \"Yury Zhuravlev\" <u.zhuravlev@postgrespro.ru> wrote:\n>\n>> When I run cmake without options, it seems to do opportunistic feature\n>> checking. For example, it turns on OpenSSL or Python support if it can\n>> find it, otherwise it turns it off. We need this to be deterministic.\n>> Without options, choose the basic feature set, require all other\n>> features to be turned on explicitly, fail if they can't be found.\n>> Whatever the Autoconf-based build does now has been fairly deliberately\n>> tuned, so there should be very little reason to deviate from that.\n>\n> This approach I see only in Postgres project and not fully understood.\n> Can you explain me more what reasons led to this approach?\n\nIt's predictable. The default has the same result for everyone. I quite\nlike it myself.\n\n\nOn 9 Nov. 2016 06:37, \"Yury Zhuravlev\" <u.zhuravlev@postgrespro.ru> wrote:\n>\n>> When I run cmake without options, it seems to do opportunistic feature\n>> checking. For example, it turns on OpenSSL or Python support if it can\n>> find it, otherwise it turns it off. We need this to be deterministic.\n>> Without options, choose the basic feature set, require all other\n>> features to be turned on explicitly, fail if they can't be found.\n>> Whatever the Autoconf-based build does now has been fairly deliberately\n>> tuned, so there should be very little reason to deviate from that.\n>\n> This approach I see only in Postgres project and not fully understood.\n> Can you explain me more what reasons led to this approach?\nIt's predictable. The default has the same result for everyone. I quite like it myself.","msg_date":"Wed, 9 Nov 2016 06:54:03 +0800","msg_from":"Craig Ringer <craig.ringer@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Wed, Nov 9, 2016 at 7:54 AM, Craig Ringer\n<craig.ringer@2ndquadrant.com> wrote:\n> On 9 Nov. 2016 06:37, \"Yury Zhuravlev\" <u.zhuravlev@postgrespro.ru> wrote:\n>> This approach I see only in Postgres project and not fully understood.\n>> Can you explain me more what reasons led to this approach?\n>\n> It's predictable. The default has the same result for everyone. I quite like\n> it myself.\n\n+1. Let's tell to the system what we want him to do and not let him\nguess what we'd like to be done or it will get harder to test and\ndevelop code for all kind of code paths with #ifdef's. That's one step\naway from Skynet.\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, 9 Nov 2016 11:12:10 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 9 November 2016 at 10:12, Michael Paquier <michael.paquier@gmail.com> wrote:\n> On Wed, Nov 9, 2016 at 7:54 AM, Craig Ringer\n> <craig.ringer@2ndquadrant.com> wrote:\n>> On 9 Nov. 2016 06:37, \"Yury Zhuravlev\" <u.zhuravlev@postgrespro.ru> wrote:\n>>> This approach I see only in Postgres project and not fully understood.\n>>> Can you explain me more what reasons led to this approach?\n>>\n>> It's predictable. The default has the same result for everyone. I quite like\n>> it myself.\n>\n> +1. Let's tell to the system what we want him to do and not let him\n> guess what we'd like to be done or it will get harder to test and\n> develop code for all kind of code paths with #ifdef's. That's one step\n> away from Skynet.\n\nEr... ok then. (Backs away slowly).\n\nMore seriously, I like it for development where a stable and\npredictable default is great.\n\nFor users it slightly sucks, as most users will want us to find\nwhatever is on the system without being manually told to enable each\nfeature. \"Of course I want SSL, I have openssl installed don't I?\"\nIt's not like we require users to specify --enable-largefile\n--enable-atomics --enable-getopt --enable-ipv6 .... we do detect a\nlot automatically.\n\nSo personally I think it'd be fine if a cmake build defaulted to\nfinding and using what it could, but offered a --minimal mode or\nwhatever that gets us just core postgres + whatever we enable\nexplicitly. But our current behaviour is OK too.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL 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, 9 Nov 2016 10:52:48 +0800","msg_from":"Craig Ringer <craig.ringer@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Craig Ringer <craig.ringer@2ndquadrant.com> writes:\n> On 9 Nov. 2016 06:37, \"Yury Zhuravlev\" <u.zhuravlev@postgrespro.ru> wrote:\n>> This approach I see only in Postgres project and not fully understood.\n>> Can you explain me more what reasons led to this approach?\n\n> It's predictable. The default has the same result for everyone. I quite\n> like it myself.\n\nIt's advantageous from a packaging standpoint, precisely because it's\npredictable: either you get the set of features you expect, or you get\na build failure. Years ago we were more on the \"opportunistic\" side\nof things, and we had problems with packages sometimes silently losing\nfeatures. A pretty-recent example is that OpenSSL changed their APIs\nin 1.1.0 so that our configure probe failed. With an opportunistic\napproach, that would have meant builds silently going from \"ssl yes\"\nto \"ssl no\". That's not good.\n\nSo this is really not open for negotiation. As Peter said upthread,\nwhat we are looking for in a CMake reimplementation is that it behaves\nexactly like the Autoconf version does. To the extent that you are unable\nor unwilling to duplicate that behavior, you increase the odds that\nwe'll reject this work.\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, 08 Nov 2016 22:47:34 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Tom Lane wrote:\n> So this is really not open for negotiation. As Peter said upthread,\n> what we are looking for in a CMake reimplementation is that it behaves\n> exactly like the Autoconf version does. To the extent that you are unable\n> or unwilling to duplicate that behavior, you increase the odds that\n> we'll reject this work.\nWho asking about negotiation? I just wanted an explanation for the clear \nunderstanding and nothing more. \nNow I know about reasons. Thanks.\n\nregards\n-- \nYury Zhuravlev\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres 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, 09 Nov 2016 12:52:07 +0300","msg_from":"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Craig Ringer wrote:\n> So personally I think it'd be fine if a cmake build defaulted to\n> finding and using what it could, but offered a --minimal mode or\n> whatever that gets us just core postgres + whatever we enable\n> explicitly. But our current behaviour is OK too.\n\nTo me it's best way. But I'm not sure what Tom Lane will accept this.\n\n\n-- \nYury Zhuravlev\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres 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, 09 Nov 2016 19:58:38 +0300","msg_from":"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Craig Ringer wrote:\n> So personally I think it'd be fine if a cmake build defaulted to\n> finding and using what it could, but offered a --minimal mode or\n> whatever that gets us just core postgres + whatever we enable\n> explicitly. But our current behaviour is OK too.\n\nTo me it's best way. But I'm not sure what Tom Lane will accept this.\n\n\n-- \nYury Zhuravlev\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres 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, 10 Nov 2016 22:15:04 +0300","msg_from":"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 11/11/16 08:15, Yury Zhuravlev wrote:\n\n> Craig Ringer wrote:\n>> So personally I think it'd be fine if a cmake build defaulted to\n>> finding and using what it could, but offered a --minimal mode or\n>> whatever that gets us just core postgres + whatever we enable\n>> explicitly. But our current behaviour is OK too.\n>\n> To me it's best way. But I'm not sure what Tom Lane will accept this.\n>\n>\n\nI just had a play with this patch - nice! (ok so it needs a fix so that \nthe compile completes as mentioned prev).\n\nI would recommend making it behave as Tom suggested. *Then* add an \n--autodetect or similar option that makes\nbehave in the 'finding and using what it could' manner as a 2nd patch.\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":"Fri, 11 Nov 2016 08:32:27 +1300","msg_from":"Mark Kirkwood <mark.kirkwood@catalyst.net.nz>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Mark Kirkwood <mark.kirkwood@catalyst.net.nz> writes:\n> I would recommend making it behave as Tom suggested. *Then* add an \n> --autodetect or similar option that makes\n> behave in the 'finding and using what it could' manner as a 2nd patch.\n\nAn \"--autodetect\" switch would be fine with me. I just don't want it\nto behave that way by default, because then you get into the unexpected\nresults problem.\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, 10 Nov 2016 14:40:08 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"I had a bit a play around to see if I could get this building properly \nagain with v10 (i.e master). I've attached a minimal *incremental* patch \nthat needs to be applied after v1. This gets everything under the src \ntree building (added the new file_utils.c and reordered some link libs \nand removed some duplicates).\n\n\nI haven't made any changes with respect to it trying to detect and build \neverything. One added nit I see is that unless I'm missing something \nthere appears to be no way to stop it trying to build all the \ncontribs...so an option to enable/disable their build is needed.\n\n\nTo make it display what options there are I use:\n\n$ mkdir build; cd build ; cmake .. -LH\n\n\nAnd to do a build that works:\n\n$ cmake .. -DCMAKE_INSTALL_PREFIX=/usr/local/pgsql/10 -DWITH_PERL=OFF \n-DWITH_OPENSSL=OFF -DWITH_PYTHON=OFF\n\n\nWith reference to Tom's comments, I'm thinking that these should all \ndefault to 'OFF' plus an additional variable WITH_CONTRIB (or similar) \nshould default to OFF too.\n\n\nregards\n\n\nMark\n\n\n\nOn 09/11/16 08:37, Peter Eisentraut wrote:\n> On 9/17/16 1:21 PM, Yury Zhuravlev wrote:\n>> Now, I published the first version of the patch.\n> I tried this out. Because of some file moves in initdb and\n> pg_basebackup, the build fails:\n>\n> [ 74%] Linking C executable initdb\n> Undefined symbols for architecture x86_64:\n> \"_fsync_pgdata\", referenced from:\n> _main in initdb.c.o\n> ld: symbol(s) not found for architecture x86_64\n> collect2: error: ld returned 1 exit status\n> make[2]: *** [src/bin/initdb/CMakeFiles/initdb.dir/build.make:177:\n> src/bin/initdb/initdb] Error 1\n> make[1]: *** [CMakeFiles/Makefile2:2893:\n> src/bin/initdb/CMakeFiles/initdb.dir/all] Error 2\n> make: *** [Makefile:128: all] Error 2\n>\n> Please submit an updated patch.\n>\n> I suggest you use git format-patch to produce patches. This is easier\n> to apply, especially when there are a lot of new files involved. Also\n> use the git facilities to check for whitespace errors.\n>\n> Please supply some documentation, such as\n>\n> - what are the basic commands\n> - how to set include/library paths, choose configure options\n> - how to set CFLAGS\n> - how to see raw build commands\n> - what are the targets for all/world/check/docs etc.\n> - explain directory structure\n>\n> I suggest for now you could put this into a README.cmake file in your\n> patch. We don't need to commit it that way, but it would help in the\n> meantime.\n>\n> When I run cmake without options, it seems to do opportunistic feature\n> checking. For example, it turns on OpenSSL or Python support if it can\n> find it, otherwise it turns it off. We need this to be deterministic.\n> Without options, choose the basic feature set, require all other\n> features to be turned on explicitly, fail if they can't be found.\n> Whatever the Autoconf-based build does now has been fairly deliberately\n> tuned, so there should be very little reason to deviate from that.\n>\n> The Python check appears to be picking up pieces from two different\n> Python installations:\n>\n> -- Found PythonInterp: /usr/local/bin/python (found version \"2.7.12\")\n> -- Found PythonLibs: /usr/lib/libpython2.7.dylib (found version \"2.7.10\")\n>\n> The check results otherwise look OK, but I'm a bit confused about the\n> order. It checks for some functions before all the header files are\n> checked for. Is that intentional?\n>\n> There are a number of changes in .[ch] and .pl files that are unclear\n> and not explained. Please explain them. You can also submit separate\n> preliminary patches if you need to do some refactoring. Ultimately, I\n> would expect this patch not to require C code changes.\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","msg_date":"Wed, 16 Nov 2016 14:43:32 +1300","msg_from":"Mark Kirkwood <mark.kirkwood@catalyst.net.nz>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Yeah, there seems to be a lot of these. Looking through them almost all \nconcern the addition of piece of code to wrap putenv. e.g:\n\n--- a/src/bin/psql/command.c\n+++ b/src/bin/psql/command.c\n@@ -1350,7 +1350,7 @@ exec_command(const char *cmd,\n char *newval;\n\n newval = psprintf(\"%s=%s\", envvar, envval);\n- putenv(newval);\n+ pg_putenv_proxy(newval);\n success = true;\n\n /*\n\nWhere pg_putenv_proxy either calls putenv or pgwin32_putenv (the latter \non windows I'd guess). I wonder if this could have been avoided, since \nthe original code handles this sort of thing. There are also some minor \n- and not immediately obvious - changes to a number of macros in various \nincludes...If I'm feeling keen I'll experiment to see how far I can get \nwithout any source changes at all.\n\n\nregards\n\n\nMark\n\n\nOn 09/11/16 08:37, Peter Eisentraut wrote:\n>\n> There are a number of changes in .[ch] and .pl files that are unclear\n> and not explained. Please explain them. You can also submit separate\n> preliminary patches if you need to do some refactoring. Ultimately, I\n> would expect this patch not to require C code changes.\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, 16 Nov 2016 16:00:48 +1300","msg_from":"Mark Kirkwood <mark.kirkwood@catalyst.net.nz>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Actually, it was not that tricky to separate out the cmake only changes, \nand test this on unmodified sources. It appears to work fine for me - \npasses 'make check' (needs the v1_1 incremental patch applied of \ncourse). The Patch is attached. I wonder if the original had some \nchanges for building under latest Windows...(I'm using Ubuntu 16.10, \nwith cmake 3.5).\n\n\nregards\n\n\nMark\n\n\nOn 16/11/16 16:00, Mark Kirkwood wrote:\n> Yeah, there seems to be a lot of these. Looking through them almost \n> all concern the addition of piece of code to wrap putenv. e.g:\n>\n> --- a/src/bin/psql/command.c\n> +++ b/src/bin/psql/command.c\n> @@ -1350,7 +1350,7 @@ exec_command(const char *cmd,\n> char *newval;\n>\n> newval = psprintf(\"%s=%s\", envvar, envval);\n> - putenv(newval);\n> + pg_putenv_proxy(newval);\n> success = true;\n>\n> /*\n>\n> Where pg_putenv_proxy either calls putenv or pgwin32_putenv (the \n> latter on windows I'd guess). I wonder if this could have been \n> avoided, since the original code handles this sort of thing. There are \n> also some minor - and not immediately obvious - changes to a number of \n> macros in various includes...If I'm feeling keen I'll experiment to \n> see how far I can get without any source changes at all.\n>\n>\n> regards\n>\n>\n> Mark\n>\n>\n> On 09/11/16 08:37, Peter Eisentraut wrote:\n>>\n>> There are a number of changes in .[ch] and .pl files that are unclear\n>> and not explained. Please explain them. You can also submit separate\n>> preliminary patches if you need to do some refactoring. Ultimately, I\n>> would expect this patch not to require C code changes.\n>>\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","msg_date":"Wed, 16 Nov 2016 16:45:24 +1300","msg_from":"Mark Kirkwood <mark.kirkwood@catalyst.net.nz>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Mark Kirkwood wrote:\n> Actually, it was not that tricky to separate out the cmake only \n> changes, and test this on unmodified sources. It appears to work \n> fine for me - passes 'make check' (needs the v1_1 incremental \n> patch applied of course). The Patch is attached. I wonder if the \n> original had some changes for building under latest \n> Windows...(I'm using Ubuntu 16.10, with cmake 3.5).\n>\nThanks for all your works! Can you make push request here: \nhttps://github.com/stalkerg/postgres_cmake\n\nI have rebased (merge) to master and make other important fix.\n\n-- \nYury Zhuravlev\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres 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, 16 Nov 2016 11:05:55 +0300","msg_from":"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Mark Kirkwood wrote:\n> Yeah, there seems to be a lot of these. Looking through them \n> almost all concern the addition of piece of code to wrap putenv. \n> e.g:\nI made this small wrapper special for MSVC 2015 without Service Packs \nbecause postgres macross were in conflict with MS internal functions. \nAfter some time and some updates for MSVC Michael Paquier could not \nreproduce my problem but I keep this patch to avoid problems in the \nfuture. \nI can check old behavior again and revert all changes if needed and \nofcourse I have plans to make separate patch for this changes. \n\nThanks!\n-- \nYury Zhuravlev\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres 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, 16 Nov 2016 11:22:43 +0300","msg_from":"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Tue, Nov 8, 2016 at 9:12 PM, Michael Paquier\n<michael.paquier@gmail.com> wrote:\n> On Wed, Nov 9, 2016 at 7:54 AM, Craig Ringer\n> <craig.ringer@2ndquadrant.com> wrote:\n>> On 9 Nov. 2016 06:37, \"Yury Zhuravlev\" <u.zhuravlev@postgrespro.ru> wrote:\n>>> This approach I see only in Postgres project and not fully understood.\n>>> Can you explain me more what reasons led to this approach?\n>>\n>> It's predictable. The default has the same result for everyone. I quite like\n>> it myself.\n>\n> +1. Let's tell to the system what we want him to do and not let him\n> guess what we'd like to be done or it will get harder to test and\n> develop code for all kind of code paths with #ifdef's. That's one step\n> away from Skynet.\n\nExaggerate much?\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, 16 Nov 2016 10:40:35 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"I see there are some patches for the putenv issue with Visual studio \n2013 in progress on this list - it is probably best to work with the \nauthor to see if 2015 has any new issues and keep all changes for that \n*out* of the cmake patches.\n\n\nregards\n\n\nMark\n\n\nOn 16/11/16 21:22, Yury Zhuravlev wrote:\n> I made this small wrapper special for MSVC 2015 without Service Packs \n> because postgres macross were in conflict with MS internal functions. \n> After some time and some updates for MSVC Michael Paquier could not \n> reproduce my problem but I keep this patch to avoid problems in the \n> future. I can check old behavior again and revert all changes if \n> needed and ofcourse I have plans to make separate patch for this changes.\n> Thanks!\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, 17 Nov 2016 11:14:24 +1300","msg_from":"Mark Kirkwood <mark.kirkwood@catalyst.net.nz>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Wed, Nov 16, 2016 at 2:14 PM, Mark Kirkwood\n<mark.kirkwood@catalyst.net.nz> wrote:\n> I see there are some patches for the putenv issue with Visual studio 2013 in\n> progress on this list - it is probably best to work with the author to see\n> if 2015 has any new issues and keep all changes for that *out* of the cmake\n> patches.\n\nI don't recall all the details here, but no wrappers should be needed,\nparticularly on Windows where we already do that:\nsrc/include/port/win32.h:#define putenv(x) pgwin32_putenv(x)\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, 16 Nov 2016 14:22:30 -0800","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Michael Paquier wrote:\n> src/include/port/win32.h:#define putenv(x) pgwin32_putenv(x)\nand my MSVC2015 drop down here because pgwin32_putenv has wrong signature.\nI hope it is not true now. \n\n-- \nYury Zhuravlev\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres 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, 17 Nov 2016 12:35:51 +0300","msg_from":"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Hello hackers.\n\n\n----\n\nI think you notice what I changed email, it is happened because I quit from\nPgPro.\n\nGood news is that I have time to work on CMake. Bad news I have to find a\nnew job. ;)\n\n----\n\nWhat is new in that patchset:\n\n 1.\n\n I separated changes to several patches. Main patch include only new\n CMake files without postgres files changes.\n 2.\n\n I cleaned all the unnecessary postgres changes.\n 3.\n\n I update cmake with latest master, add STRONG_RANDOM option and etc.\n 4.\n\n Add small document about CMake and Postgres. (README.cmake bad idea for\n name but I think now it’s ok)\n 5.\n\n I have stopped use gendef.pl for MSVC build.\n 6.\n\n Fix for AIX 7.1\n\n\nAbout patches:\n\ncmake_v2_1_main_files.patch\n\nOnly CMake files. CMakeFiles.txt is main files with rules. *.cmake is\nspecial modules with macros and functions. *.in is template for generate\nheaders. (like autoconf)\n\ncmake_v2_2_c_define.patch\n\nSmall chages in c.h . At first it is “#pragma fenv_access (off)” it is\nnecessary if we use /fp:strict for MSVC compiler. Without this pragma we\ncan’t calc floats for const variables in compiller time (2 * M_PI for\nexample). Strict mode important if we want to be close with ieee754 float\nformat on MSVC (1.0 / 0.0 = inf for example). Detail info here:\nhttps://msdn.microsoft.com/en-us/library/e7s85ffb.aspx\n\nSecond change is because I find and set HAVE_INT128 directly from CMake.\nPG_INT128_TYPE used only for autoconfig scripts.\n\ncmake_v2_3_rijndael.patch\n\nFirst I added special wraparound because here CMake have circular\ndependency (cmake very smart here). Second I removed rijndael.tbl because\nit generated during build process every time.\n\ncmake_v2_4_uuid.patch\n\nAnother small patch. Right place for uuid.h I find by CMake and not\nnecessary this ifdef hell.\n\ncmake_v2_5_readme.patch\n\nSmall exercise to CMake world for all who want to try.\n\nQuestions for discussion:\n\nIn generated project by CMake we always have only one enter point. Also\nINSTALL macross support only including to “all” targets. It follows that it\nis impossible build contrib modules separately only with “all” target. Here\nwrite about this behavior:\nhttps://cmake.org/cmake/help/v3.7/prop_tgt/EXCLUDE_FROM_ALL.html\n\nInteresting note:\n\nMany folks here love only command line but many novices developers from\nmodern generation love GUI. Also using command line on Windows it's really\nbig pain. With CMake you can work with Postgres under MSVC more\ncomfortable. For example my screenshots of regress tests result runing from\nMSVC: https://twitter.com/stalkerg/status/814423972263657472\n\nFin:\n\nI hope this patchset one step closer to merge. In future I see a lot of\nwork but I suppose it is right and important direction. I'm ready for this\nlong journey.\n\nBig thanks Peter Eisentraut for advices and the right vector of\ndevelopment.\n\nAlso big thanks Mark Kirkwood for help and discussion.\n\nHappy New Year, everyone!\n\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","msg_date":"Fri, 30 Dec 2016 17:10:45 +0300","msg_from":"Yuriy Zhuravlev <stalkerg@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 12/30/16 9:10 AM, Yuriy Zhuravlev wrote:\n> cmake_v2_2_c_define.patch\n> \n> Small chages in c.h . At first it is �#pragma fenv_access (off)� it is\n> necessary if we use /fp:strict for MSVC compiler. Without this pragma we\n> can�t calc floats for const variables in compiller time (2 * M_PI for\n> example). Strict mode important if we want to be close with ieee754\n> float format on MSVC (1.0 / 0.0 = inf for example). Detail info here:\n> https://msdn.microsoft.com/en-us/library/e7s85ffb.aspx\n\nI don't understand what this has to do with cmake. If this is a\nworthwhile improvement for the Windows build, then please explain why,\nwith a \"before\" and \"after\" output and a patch for the existing build\nsystem as well.\n\n> Second change is because I find and set HAVE_INT128 directly from CMake.\n> PG_INT128_TYPE used only for autoconfig scripts.\n\nIt might also be worth refactoring the existing Autoconf code here to\nmake this consistent.\n\n(My assumption is that if we were to move forward with cmake or any\nother build system change, we would have to keep the old one alongside\nat least for a little while. So any changes to the C code would need to\nbe side-ported.)\n\n> cmake_v2_3_rijndael.patch\n> \n> First I added special wraparound because here CMake have circular\n> dependency (cmake very smart here). Second I removed rijndael.tbl\n> because it generated during build process every time.\n\nPlease explain what the circular dependency is. If there is one, we\nshould also side-port this change.\n\n> cmake_v2_4_uuid.patch\n> \n> Another small patch. Right place for uuid.h I find by CMake and not\n> necessary this ifdef hell.\n\nThis patch removes the uuid.h include but doesn't add it anywhere else.\nHow does it work?\n\n> Questions for discussion:\n> \n> In generated project by CMake we always have only one enter point. Also\n> INSTALL macross support only including to �all� targets. It follows that\n> it is impossible build contrib modules separately only with �all�\n> target. Here write about this behavior:\n> https://cmake.org/cmake/help/v3.7/prop_tgt/EXCLUDE_FROM_ALL.html\n\nYeah, I think this is how the MSVC stuff effectively works right now as\nwell.\n\n-- \nPeter Eisentraut 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":"Tue, 3 Jan 2017 09:11:45 -0500","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Tue, Jan 3, 2017 at 11:11 PM, Peter Eisentraut\n<peter.eisentraut@2ndquadrant.com> wrote:\n> On 12/30/16 9:10 AM, Yuriy Zhuravlev wrote:\n>> cmake_v2_2_c_define.patch\n>>\n>> Small chages in c.h . At first it is “#pragma fenv_access (off)” it is\n>> necessary if we use /fp:strict for MSVC compiler. Without this pragma we\n>> can’t calc floats for const variables in compiller time (2 * M_PI for\n>> example). Strict mode important if we want to be close with ieee754\n>> float format on MSVC (1.0 / 0.0 = inf for example). Detail info here:\n>> https://msdn.microsoft.com/en-us/library/e7s85ffb.aspx\n>\n> I don't understand what this has to do with cmake. If this is a\n> worthwhile improvement for the Windows build, then please explain why,\n> with a \"before\" and \"after\" output and a patch for the existing build\n> system as well.\n>\n>> Second change is because I find and set HAVE_INT128 directly from CMake.\n>> PG_INT128_TYPE used only for autoconfig scripts.\n>\n> It might also be worth refactoring the existing Autoconf code here to\n> make this consistent.\n>\n> (My assumption is that if we were to move forward with cmake or any\n> other build system change, we would have to keep the old one alongside\n> at least for a little while. So any changes to the C code would need to\n> be side-ported.)\n>\n>> cmake_v2_3_rijndael.patch\n>>\n>> First I added special wraparound because here CMake have circular\n>> dependency (cmake very smart here). Second I removed rijndael.tbl\n>> because it generated during build process every time.\n>\n> Please explain what the circular dependency is. If there is one, we\n> should also side-port this change.\n>\n>> cmake_v2_4_uuid.patch\n>>\n>> Another small patch. Right place for uuid.h I find by CMake and not\n>> necessary this ifdef hell.\n>\n> This patch removes the uuid.h include but doesn't add it anywhere else.\n> How does it work?\n>\n>> Questions for discussion:\n>>\n>> In generated project by CMake we always have only one enter point. Also\n>> INSTALL macross support only including to “all” targets. It follows that\n>> it is impossible build contrib modules separately only with “all”\n>> target. Here write about this behavior:\n>> https://cmake.org/cmake/help/v3.7/prop_tgt/EXCLUDE_FROM_ALL.html\n>\n> Yeah, I think this is how the MSVC stuff effectively works right now as\n> well.\n\nI am marking this patch as \"returned with feedback\". That's quite a\nheavy change and it looks to be too late in the development cycle of\nPG10 to consider it. Peter's commit bits, who is also the reviewer,\nare beginning to smoke as well after everything that has happened for\nthe logical replication changes.\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":"Tue, 24 Jan 2017 15:50:47 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 2017-01-24 15:50:47 +0900, Michael Paquier wrote:\n> I am marking this patch as \"returned with feedback\". That's quite a\n> heavy change and it looks to be too late in the development cycle of\n> PG10 to consider it. Peter's commit bits, who is also the reviewer,\n> are beginning to smoke as well after everything that has happened for\n> the logical replication changes.\n\nI'm doubtful about this being ready in time too, but it seems a might\nheavyhanded to make that call on your own. Including the judgement about\nPeter's capability to handle more.\n\nGreetings,\n\nAndres Freund\n\n\n-- \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, 23 Jan 2017 22:58:48 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Tue, Jan 24, 2017 at 3:58 PM, Andres Freund <andres@anarazel.de> wrote:\n> On 2017-01-24 15:50:47 +0900, Michael Paquier wrote:\n>> I am marking this patch as \"returned with feedback\". That's quite a\n>> heavy change and it looks to be too late in the development cycle of\n>> PG10 to consider it. Peter's commit bits, who is also the reviewer,\n>> are beginning to smoke as well after everything that has happened for\n>> the logical replication changes.\n>\n> I'm doubtful about this being ready in time too, but it seems a might\n> heavyhanded to make that call on your own. Including the judgement about\n> Peter's capability to handle more.\n\nSure, sorry for that. I am switching back the patch, let's see what\nhappens by the end of the CF.\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":"Tue, 24 Jan 2017 16:00:16 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 24 Jan. 2017 18:00, \"Andres Freund\" <andres@anarazel.de> wrote:\n\nOn 2017-01-24 15:50:47 +0900, Michael Paquier wrote:\n> I am marking this patch as \"returned with feedback\". That's quite a\n> heavy change and it looks to be too late in the development cycle of\n> PG10 to consider it. Peter's commit bits, who is also the reviewer,\n> are beginning to smoke as well after everything that has happened for\n> the logical replication changes.\n\nI'm doubtful about this being ready in time too, but it seems a might\nheavyhanded to make that call on your own. Including the judgement about\nPeter's capability to handle more.\n\n\nPersonally I think we should aim to have this in as a non default build\nmode in pg10 if it can be made ready, and aim to make it default in pg11 at\nleast for Windows.\n\nOn 24 Jan. 2017 18:00, \"Andres Freund\" <andres@anarazel.de> wrote:On 2017-01-24 15:50:47 +0900, Michael Paquier wrote:\n> I am marking this patch as \"returned with feedback\". That's quite a\n> heavy change and it looks to be too late in the development cycle of\n> PG10 to consider it. Peter's commit bits, who is also the reviewer,\n> are beginning to smoke as well after everything that has happened for\n> the logical replication changes.\n\nI'm doubtful about this being ready in time too, but it seems a might\nheavyhanded to make that call on your own. Including the judgement about\nPeter's capability to handle more.\nPersonally I think we should aim to have this in as a non default build mode in pg10 if it can be made ready, and aim to make it default in pg11 at least for Windows.","msg_date":"Tue, 24 Jan 2017 16:27:58 +0800","msg_from":"Craig Ringer <craig.ringer@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Craig Ringer <craig.ringer@2ndquadrant.com> writes:\n> Personally I think we should aim to have this in as a non default build\n> mode in pg10 if it can be made ready, and aim to make it default in pg11 at\n> least for Windows.\n\nAFAIK we haven't committed to accepting this at all, let alone trying\nto do so on a tight schedule. And I believe there was general agreement\nthat we would not accept it as something to maintain in parallel with\nthe existing makefiles. If we have to maintain two build systems, we\nhave that already.\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, 24 Jan 2017 08:37:47 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 1/24/17 8:37 AM, Tom Lane wrote:\n> Craig Ringer <craig.ringer@2ndquadrant.com> writes:\n>> Personally I think we should aim to have this in as a non default build\n>> mode in pg10 if it can be made ready, and aim to make it default in pg11 at\n>> least for Windows.\n> \n> AFAIK we haven't committed to accepting this at all, let alone trying\n> to do so on a tight schedule. And I believe there was general agreement\n> that we would not accept it as something to maintain in parallel with\n> the existing makefiles. If we have to maintain two build systems, we\n> have that already.\n\nMy preferred scenario would be to replace the Windows build system by\nthis first, then refine it, then get rid of Autoconf.\n\nThe ideal timeline would be to have a ready patch to commit early in a\ndevelopment cycle, then get rid of the Windows build system by the end\nof it. Naturally, this would need buy-in from Windows developers.\n\nI don't foresee replacing the Autoconf build system by this immediately.\n\nRight now, however, the patch isn't moving at all, and I don't see it\ngoing into PG10, so I'm fine with returning with feedback.\n\n-- \nPeter Eisentraut 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":"Fri, 27 Jan 2017 09:09:36 -0500","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Peter Eisentraut wrote:\n\n> Right now, however, the patch isn't moving at all, and I don't see it\n> going into PG10, so I'm fine with returning with feedback.\n\nThere are a bunch of side patches that we should apply separately, such\nas the pgcrypto fix. I don't understand why they are part of this patch\nseries, really, given that they seem largely independent prerequisites.\n\n-- \n�lvaro Herrera https://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":"Fri, 27 Jan 2017 11:40:55 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Fri, Jan 27, 2017 at 11:09 PM, Peter Eisentraut\n<peter.eisentraut@2ndquadrant.com> wrote:\n> On 1/24/17 8:37 AM, Tom Lane wrote:\n>> Craig Ringer <craig.ringer@2ndquadrant.com> writes:\n>>> Personally I think we should aim to have this in as a non default build\n>>> mode in pg10 if it can be made ready, and aim to make it default in pg11 at\n>>> least for Windows.\n>>\n>> AFAIK we haven't committed to accepting this at all, let alone trying\n>> to do so on a tight schedule. And I believe there was general agreement\n>> that we would not accept it as something to maintain in parallel with\n>> the existing makefiles. If we have to maintain two build systems, we\n>> have that already.\n>\n> My preferred scenario would be to replace the Windows build system by\n> this first, then refine it, then get rid of Autoconf.\n>\n> The ideal timeline would be to have a ready patch to commit early in a\n> development cycle, then get rid of the Windows build system by the end\n> of it. Naturally, this would need buy-in from Windows developers.\n\nThis looks like a really good plan to me.\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":"Sat, 28 Jan 2017 07:50:31 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 2017-01-27 09:09:36 -0500, Peter Eisentraut wrote:\n> My preferred scenario would be to replace the Windows build system by\n> this first, then refine it, then get rid of Autoconf.\n> \n> The ideal timeline would be to have a ready patch to commit early in a\n> development cycle, then get rid of the Windows build system by the end\n> of it. Naturally, this would need buy-in from Windows developers.\n> \n> I don't foresee replacing the Autoconf build system by this immediately.\n\nI'm very strongly against this path, it seems way too likely that we'll\nend up with yet another fragile thing that nobody from the *nix side\nwill be able to test.\n\n- Andres\n\n\n-- \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, 27 Jan 2017 15:11:20 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 1/27/17 6:11 PM, Andres Freund wrote:\n> On 2017-01-27 09:09:36 -0500, Peter Eisentraut wrote:\n>> My preferred scenario would be to replace the Windows build system by\n>> this first, then refine it, then get rid of Autoconf.\n>>\n>> The ideal timeline would be to have a ready patch to commit early in a\n>> development cycle, then get rid of the Windows build system by the end\n>> of it. Naturally, this would need buy-in from Windows developers.\n>>\n>> I don't foresee replacing the Autoconf build system by this immediately.\n> \n> I'm very strongly against this path, it seems way too likely that we'll\n> end up with yet another fragile thing that nobody from the *nix side\n> will be able to test.\n\nThat's a fair concern, but at least with CMake, someone from the *nix\nside *can* test it, whereas right now it's completely separate.\n\nWhat kind of strategy do you have in mind?\n\n-- \nPeter Eisentraut 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":"Fri, 27 Jan 2017 22:20:41 -0500","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Sat, Jan 28, 2017 at 12:20 PM, Peter Eisentraut\n<peter.eisentraut@2ndquadrant.com> wrote:\n> On 1/27/17 6:11 PM, Andres Freund wrote:\n>> On 2017-01-27 09:09:36 -0500, Peter Eisentraut wrote:\n>>> My preferred scenario would be to replace the Windows build system by\n>>> this first, then refine it, then get rid of Autoconf.\n>>>\n>>> The ideal timeline would be to have a ready patch to commit early in a\n>>> development cycle, then get rid of the Windows build system by the end\n>>> of it. Naturally, this would need buy-in from Windows developers.\n>>>\n>>> I don't foresee replacing the Autoconf build system by this immediately.\n>>\n>> I'm very strongly against this path, it seems way too likely that we'll\n>> end up with yet another fragile thing that nobody from the *nix side\n>> will be able to test.\n>\n> That's a fair concern, but at least with CMake, someone from the *nix\n> side *can* test it, whereas right now it's completely separate.\n\nAnd people complain all the time that the MSVC build scripts are hacky\nand complicated.. So by beginning from there we switch from one build\nto the other, not increasing the number of builds that need to be\nmaintained. Based on that Peter's strategy looks appealing to me. By\nthe way, I am marking the patch as returned with feedback for this CF.\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, 30 Jan 2017 15:22:16 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 2017-01-27 22:20:41 -0500, Peter Eisentraut wrote:\n> On 1/27/17 6:11 PM, Andres Freund wrote:\n> > On 2017-01-27 09:09:36 -0500, Peter Eisentraut wrote:\n> >> My preferred scenario would be to replace the Windows build system by\n> >> this first, then refine it, then get rid of Autoconf.\n> >>\n> >> The ideal timeline would be to have a ready patch to commit early in a\n> >> development cycle, then get rid of the Windows build system by the end\n> >> of it. Naturally, this would need buy-in from Windows developers.\n> >>\n> >> I don't foresee replacing the Autoconf build system by this immediately.\n> > \n> > I'm very strongly against this path, it seems way too likely that we'll\n> > end up with yet another fragile thing that nobody from the *nix side\n> > will be able to test.\n> \n> That's a fair concern, but at least with CMake, someone from the *nix\n> side *can* test it, whereas right now it's completely separate.\n\nGiven that fact, I just don't buy why it's a good idea to not also\nreplace autoconf initially. Either we're able to properly test it -\ni.e. it runs all tests - on *nix or we're not. There's not a a whole of\neffort between those if you also want to do the windows side of things\nproperly.\n\n\n> What kind of strategy do you have in mind?\n\nDo all of it. I'm unconvinced that a windows only version buys us\nmeaningful savings, and I think the dangers of adding more duplication\n(msvc stuff after all gets some information from the makefiles) and\nlong-term coexistence are quite severe.\n\n\nGreetings,\n\nAndres Freund\n\n\n-- \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, 29 Jan 2017 22:28:55 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 1/30/17 1:28 AM, Andres Freund wrote:\n> Given that fact, I just don't buy why it's a good idea to not also\n> replace autoconf initially.\n\nWell, I find it a bit scary. If you do the big switch all at once, then\nyou will have to dedicate the following 3 months to fixing complaints\nfrom developers and build farmers.\n\n> Either we're able to properly test it - i.e. it runs all tests - on *nix or we're not.\n\nThat would work if there were a single entry point into the build\nsystem. But in practice there are many, and every one of them is\nsomeone's favorite. It's unlikely that we will be able to enumerate all\nof them during patch review.\n\n-- \nPeter Eisentraut 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":"Mon, 30 Jan 2017 10:26:18 -0500","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":">\n> I don't understand what this has to do with cmake. If this is a\n> worthwhile improvement for the Windows build, then please explain why,\n> with a \"before\" and \"after\" output and a patch for the existing build\n> system as well.\n\nDuring the porting process, I meet such situations when I should fix\nsomething. It's happening because I build with different way also current\nbuild system is trying to avoid many sharp corners.\nIf talk about this situation - without strict mode many \"floats\" checks\ndon't work correctly. You can read the link above. Besides this option puts\nby build system. I think we can make a new thread for this approach. (with\npatch for current perl system)\n\nIt might also be worth refactoring the existing Autoconf code here to\n> make this consistent.\n\n\nI do it because it's convenient in CMake. I can change this it's not big\ndeal.\n\nPlease explain what the circular dependency is. If there is one, we\n> should also side-port this change.\n\n\nIt's an important part. I have a rule for generate rijndael.tbl by\ngen-rtab who make from rijndael.c (with special define) who include\nrijndael.tbl .\nIf I generate rijndael.tbl it's to force build gen-rtab and generate\nrijndael.tbl again.\nCMake knows about \"includes\" in files but we can make the wraparound macro\nto hide include.\n\nThis patch removes the uuid.h include but doesn't add it anywhere else.\n\nHow does it work?\n\n\nCMake sends to compiler right place for uuid.h (I mean -I/usr/include and\netc for gcc).\n\n\n> Yeah, I think this is how the MSVC stuff effectively works right now as\n> well.\n\n\nI glad to hear it.\n\n2017-01-03 17:11 GMT+03:00 Peter Eisentraut <\npeter.eisentraut@2ndquadrant.com>:\n\n> On 12/30/16 9:10 AM, Yuriy Zhuravlev wrote:\n> > cmake_v2_2_c_define.patch\n> >\n> > Small chages in c.h . At first it is “#pragma fenv_access (off)” it is\n> > necessary if we use /fp:strict for MSVC compiler. Without this pragma we\n> > can’t calc floats for const variables in compiller time (2 * M_PI for\n> > example). Strict mode important if we want to be close with ieee754\n> > float format on MSVC (1.0 / 0.0 = inf for example). Detail info here:\n> > https://msdn.microsoft.com/en-us/library/e7s85ffb.aspx\n>\n> I don't understand what this has to do with cmake. If this is a\n> worthwhile improvement for the Windows build, then please explain why,\n> with a \"before\" and \"after\" output and a patch for the existing build\n> system as well.\n>\n> > Second change is because I find and set HAVE_INT128 directly from CMake.\n> > PG_INT128_TYPE used only for autoconfig scripts.\n>\n> It might also be worth refactoring the existing Autoconf code here to\n> make this consistent.\n>\n> (My assumption is that if we were to move forward with cmake or any\n> other build system change, we would have to keep the old one alongside\n> at least for a little while. So any changes to the C code would need to\n> be side-ported.)\n>\n> > cmake_v2_3_rijndael.patch\n> >\n> > First I added special wraparound because here CMake have circular\n> > dependency (cmake very smart here). Second I removed rijndael.tbl\n> > because it generated during build process every time.\n>\n> Please explain what the circular dependency is. If there is one, we\n> should also side-port this change.\n>\n> > cmake_v2_4_uuid.patch\n> >\n> > Another small patch. Right place for uuid.h I find by CMake and not\n> > necessary this ifdef hell.\n>\n> This patch removes the uuid.h include but doesn't add it anywhere else.\n> How does it work?\n>\n> > Questions for discussion:\n> >\n> > In generated project by CMake we always have only one enter point. Also\n> > INSTALL macross support only including to “all” targets. It follows that\n> > it is impossible build contrib modules separately only with “all”\n> > target. Here write about this behavior:\n> > https://cmake.org/cmake/help/v3.7/prop_tgt/EXCLUDE_FROM_ALL.html\n>\n> Yeah, I think this is how the MSVC stuff effectively works right now as\n> well.\n>\n> --\n> Peter Eisentraut http://www.2ndQuadrant.com/\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>\n\nI don't understand what this has to do with cmake. If this is aworthwhile improvement for the Windows build, then please explain why,with a \"before\" and \"after\" output and a patch for the existing buildsystem as well.During the porting process, I meet such situations when I should fix something. It's happening because I build with different way also current build system is trying to avoid many sharp corners. If talk about this situation - without strict mode many \"floats\" checks don't work correctly. You can read the link above. Besides this option puts by build system. I think we can make a new thread for this approach. (with patch for current perl system) It might also be worth refactoring the existing Autoconf code here tomake this consistent.I do it because it's convenient in CMake. I can change this it's not big deal. Please explain what the circular dependency is. If there is one, weshould also side-port this change.It's an important part. I have a rule for generate rijndael.tbl by gen-rtab who make from rijndael.c (with special define) who include rijndael.tbl . If I generate rijndael.tbl it's to force build gen-rtab and generate rijndael.tbl again. CMake knows about \"includes\" in files but we can make the wraparound macro to hide include. This patch removes the uuid.h include but doesn't add it anywhere else.How does it work? CMake sends to compiler right place for uuid.h (I mean -I/usr/include and etc for gcc). Yeah, I think this is how the MSVC stuff effectively works right now aswell.I glad to hear it. 2017-01-03 17:11 GMT+03:00 Peter Eisentraut <peter.eisentraut@2ndquadrant.com>:On 12/30/16 9:10 AM, Yuriy Zhuravlev wrote:\n> cmake_v2_2_c_define.patch\n>\n> Small chages in c.h . At first it is “#pragma fenv_access (off)” it is\n> necessary if we use /fp:strict for MSVC compiler. Without this pragma we\n> can’t calc floats for const variables in compiller time (2 * M_PI for\n> example). Strict mode important if we want to be close with ieee754\n> float format on MSVC (1.0 / 0.0 = inf for example). Detail info here:\n> https://msdn.microsoft.com/en-us/library/e7s85ffb.aspx\n\nI don't understand what this has to do with cmake. If this is a\nworthwhile improvement for the Windows build, then please explain why,\nwith a \"before\" and \"after\" output and a patch for the existing build\nsystem as well.\n\n> Second change is because I find and set HAVE_INT128 directly from CMake.\n> PG_INT128_TYPE used only for autoconfig scripts.\n\nIt might also be worth refactoring the existing Autoconf code here to\nmake this consistent.\n\n(My assumption is that if we were to move forward with cmake or any\nother build system change, we would have to keep the old one alongside\nat least for a little while. So any changes to the C code would need to\nbe side-ported.)\n\n> cmake_v2_3_rijndael.patch\n>\n> First I added special wraparound because here CMake have circular\n> dependency (cmake very smart here). Second I removed rijndael.tbl\n> because it generated during build process every time.\n\nPlease explain what the circular dependency is. If there is one, we\nshould also side-port this change.\n\n> cmake_v2_4_uuid.patch\n>\n> Another small patch. Right place for uuid.h I find by CMake and not\n> necessary this ifdef hell.\n\nThis patch removes the uuid.h include but doesn't add it anywhere else.\nHow does it work?\n\n> Questions for discussion:\n>\n> In generated project by CMake we always have only one enter point. Also\n> INSTALL macross support only including to “all” targets. It follows that\n> it is impossible build contrib modules separately only with “all”\n> target. Here write about this behavior:\n> https://cmake.org/cmake/help/v3.7/prop_tgt/EXCLUDE_FROM_ALL.html\n\nYeah, I think this is how the MSVC stuff effectively works right now as\nwell.\n\n--\nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Wed, 8 Feb 2017 14:04:39 +0300","msg_from":"Yuriy Zhuravlev <stalkerg@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"2017-01-28 1:50 GMT+03:00 Michael Paquier <michael.paquier@gmail.com>:\n\n> On Fri, Jan 27, 2017 at 11:09 PM, Peter Eisentraut\n> <peter.eisentraut@2ndquadrant.com> wrote:\n> > On 1/24/17 8:37 AM, Tom Lane wrote:\n> >> Craig Ringer <craig.ringer@2ndquadrant.com> writes:\n> >>> Personally I think we should aim to have this in as a non default build\n> >>> mode in pg10 if it can be made ready, and aim to make it default in\n> pg11 at\n> >>> least for Windows.\n> >>\n> >> AFAIK we haven't committed to accepting this at all, let alone trying\n> >> to do so on a tight schedule. And I believe there was general agreement\n> >> that we would not accept it as something to maintain in parallel with\n> >> the existing makefiles. If we have to maintain two build systems, we\n> >> have that already.\n> >\n> > My preferred scenario would be to replace the Windows build system by\n> > this first, then refine it, then get rid of Autoconf.\n> >\n> > The ideal timeline would be to have a ready patch to commit early in a\n> > development cycle, then get rid of the Windows build system by the end\n> > of it. Naturally, this would need buy-in from Windows developers.\n>\n> This looks like a really good plan to me.\n\n\nI think it's best plan because when this patch will be in Postgres guys\nfrom community can test it for Unix systems too.\nSupport two build systems it's not big deal really. I have been working on\nthis past year without any big troubles.\nAlso we have second perl build system...\n\n2017-01-28 1:50 GMT+03:00 Michael Paquier <michael.paquier@gmail.com>:On Fri, Jan 27, 2017 at 11:09 PM, Peter Eisentraut\n<peter.eisentraut@2ndquadrant.com> wrote:\n> On 1/24/17 8:37 AM, Tom Lane wrote:\n>> Craig Ringer <craig.ringer@2ndquadrant.com> writes:\n>>> Personally I think we should aim to have this in as a non default build\n>>> mode in pg10 if it can be made ready, and aim to make it default in pg11 at\n>>> least for Windows.\n>>\n>> AFAIK we haven't committed to accepting this at all, let alone trying\n>> to do so on a tight schedule. And I believe there was general agreement\n>> that we would not accept it as something to maintain in parallel with\n>> the existing makefiles. If we have to maintain two build systems, we\n>> have that already.\n>\n> My preferred scenario would be to replace the Windows build system by\n> this first, then refine it, then get rid of Autoconf.\n>\n> The ideal timeline would be to have a ready patch to commit early in a\n> development cycle, then get rid of the Windows build system by the end\n> of it. Naturally, this would need buy-in from Windows developers.\n\nThis looks like a really good plan to me. I think it's best plan because when this patch will be in Postgres guys from community can test it for Unix systems too. Support two build systems it's not big deal really. I have been working on this past year without any big troubles. Also we have second perl build system...","msg_date":"Wed, 8 Feb 2017 14:21:56 +0300","msg_from":"Yuriy Zhuravlev <stalkerg@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 2/8/17 6:21 AM, Yuriy Zhuravlev wrote:\n> Support two build systems it's not big deal really. I have been working\n> on this past year without any big troubles. \n> Also we have second perl build system...\n\nThe perl/msvc build system pulls in information from the makefiles. So\nwhen you add a file or something basic like that, you don't have to\nupdate it. So it's really more like 1.5 build systems.\n\n-- \nPeter Eisentraut 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, 8 Feb 2017 16:36:42 -0500","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Hi,\n\nOn 2017-01-30 10:26:18 -0500, Peter Eisentraut wrote:\n> On 1/30/17 1:28 AM, Andres Freund wrote:\n> > Given that fact, I just don't buy why it's a good idea to not also\n> > replace autoconf initially.\n> \n> Well, I find it a bit scary. If you do the big switch all at once, then\n> you will have to dedicate the following 3 months to fixing complaints\n> from developers and build farmers.\n\nThat'll be the case just as well if we spread it out over two cycles,\nexcept that we'll have it in multiple phases, and we run into the danger\nof a half-done conversion.\n\nI'd rather not change systems at all than run into the danger of that.\n\n\n> > Either we're able to properly test it - i.e. it runs all tests - on *nix or we're not.\n> \n> That would work if there were a single entry point into the build\n> system. But in practice there are many, and every one of them is\n> someone's favorite. It's unlikely that we will be able to enumerate all\n> of them during patch review.\n\nNot sure what you mean with \"entry point\"?\n\n\n- Andres\n\n\n-- \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, 8 Feb 2017 13:44:50 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n> On 2/8/17 6:21 AM, Yuriy Zhuravlev wrote:\n>> Support two build systems it's not big deal really. I have been working\n>> on this past year without any big troubles. \n>> Also we have second perl build system...\n\n> The perl/msvc build system pulls in information from the makefiles. So\n> when you add a file or something basic like that, you don't have to\n> update it. So it's really more like 1.5 build systems.\n\nReally it's more like 1.1 build systems, in that the MSVC scripts do that\njust well enough that you *usually* don't have to think about them. But\nthen when they fail, and you have to figure out why, it can be a pain.\n\nFor my own purposes, the only thing that I find seriously annoying about\nthe status quo is the amount of time required to run \"configure\". For\nme, that step is usually comparable to or even more than the time to\ndo the build proper, because (a) ccache and (b) multiple CPUs.\nconfigure isn't parallelizable, and there's probably nothing that\ncan be done about that. If CMake offers a substantial improvement\nin configuration time then that would be attractive. Otherwise I'd\njust as soon see us put the effort into making the MSVC scripts more\nrobust and able to pull more data from the makefiles.\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, 08 Feb 2017 16:52:19 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 2/8/17 3:52 PM, Tom Lane wrote:\n> For my own purposes, the only thing that I find seriously annoying about\n> the status quo is the amount of time required to run \"configure\". For\n> me, that step is usually comparable to or even more than the time to\n> do the build proper, because (a) ccache and (b) multiple CPUs.\n> configure isn't parallelizable, and there's probably nothing that\n> can be done about that. If CMake offers a substantial improvement\n> in configuration time then that would be attractive. Otherwise I'd\n> just as soon see us put the effort into making the MSVC scripts more\n> robust and able to pull more data from the makefiles.\n\nFWIW, I've had good luck adding -C to configure to cache the output. I'd \nguess it's at least 10x faster on my laptop.\n\nObviously doesn't help if you're doing where you're testing something \nthat alters config output. In those cases I'll either edit config.cache \nand delete the relevant lines or just temporarily move it out of the way \n(or just nuke it...).\n-- \nJim Nasby, Data Architect, Blue Treble Consulting, Austin TX\nExperts in Analytics, Data Architecture and PostgreSQL\nData in Trouble? Get it in Treble! http://BlueTreble.com\n855-TREBLE2 (855-873-2532)\n\n\n-- \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, 8 Feb 2017 15:55:34 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"Hi,\n\nOn 2017-02-08 16:52:19 -0500, Tom Lane wrote:\n> For my own purposes, the only thing that I find seriously annoying about\n> the status quo is the amount of time required to run \"configure\". For\n> me, that step is usually comparable to or even more than the time to\n> do the build proper, because (a) ccache and (b) multiple CPUs.\n> configure isn't parallelizable, and there's probably nothing that\n> can be done about that.\n\nI use autoconf caching feature to make that a bit less painful (plus\nsome scripting about when to scrap the cache file...). I find that\nseriously annoying too.\n\n\n> If CMake offers a substantial improvement\n> in configuration time then that would be attractive. Otherwise I'd\n> just as soon see us put the effort into making the MSVC scripts more\n> robust and able to pull more data from the makefiles.\n\nSome of the build-tooling in cmake is quite nice, I have to admit. I've\ne.g. grown to like using ninja instead of make to build the resulting\nfiles. Primarily in projects that take longer than pg to compile - a\nclean build in llvm with ninja takes like 0.1 seconds.\n\nBeing more able to rely on things working on windows when doing them on\nlinux does seem like an advantage to me - fiddlin with Mkvcbuild.pm is\nquite annoying.\n\n- Andres\n\n\n-- \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, 8 Feb 2017 13:59:11 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Mon, Jan 30, 2017 at 10:26 AM, Peter Eisentraut\n<peter.eisentraut@2ndquadrant.com> wrote:\n> On 1/30/17 1:28 AM, Andres Freund wrote:\n>> Given that fact, I just don't buy why it's a good idea to not also\n>> replace autoconf initially.\n>\n> Well, I find it a bit scary. If you do the big switch all at once, then\n> you will have to dedicate the following 3 months to fixing complaints\n> from developers and build farmers.\n\nI agree with that. I think replacing the Windows build system first\nand then the non-Windows build system later is a better plan than\nreplacing both at the same time.\n\nBut also, I'm not really sure whether this conversion makes sense. I\nmean, any build system is going to require some work, and accordingly\nour present build systems require some work. cmake will require\ndifferent work, but not necessarily less. The current system has a\nlong history; we pretty much know it works. Switching will inevitably\nbreak some things. Maybe we'll end up better off in the long term,\nbut maybe we won't. Things are pretty good now, so it seems like it\nwould be easy for them to get worse rather than better. If nothing\nelse, everybody who has to learn the new system either to use it for\ndevelopment or because they are doing packaging will have to do some\namount of extra work as a result of any switch.\n\nI do agree that - in theory - one build system is better than two.\nBut two well-tested, reliable build systems could easily be better\nthan one system with a bunch of problems. And the points downthread\nabout our two existing systems being not entirely separate are on\npoint, too.\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, 10 Feb 2017 12:07:15 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 2017-02-10 12:07:15 -0500, Robert Haas wrote:\n> On Mon, Jan 30, 2017 at 10:26 AM, Peter Eisentraut\n> <peter.eisentraut@2ndquadrant.com> wrote:\n> > On 1/30/17 1:28 AM, Andres Freund wrote:\n> >> Given that fact, I just don't buy why it's a good idea to not also\n> >> replace autoconf initially.\n> >\n> > Well, I find it a bit scary. If you do the big switch all at once, then\n> > you will have to dedicate the following 3 months to fixing complaints\n> > from developers and build farmers.\n> \n> I agree with that. I think replacing the Windows build system first\n> and then the non-Windows build system later is a better plan than\n> replacing both at the same time.\n\nObviously I disagree ;)\n\n\nBut I'm more replying because of:\n\n> But also, I'm not really sure whether this conversion makes sense. I\n> mean, any build system is going to require some work, and accordingly\n> our present build systems require some work. cmake will require\n> different work, but not necessarily less. The current system has a\n> long history; we pretty much know it works. Switching will inevitably\n> break some things. Maybe we'll end up better off in the long term,\n> but maybe we won't. Things are pretty good now, so it seems like it\n> would be easy for them to get worse rather than better.\n\nI do think it's kinda ok for people who've dealt with this for some\ntime. But for the first few times having to write autoconf macros is\nquite intimidating. A somewhat less obscure way to deal with that is\nworth something. As is better file dependency management, across\ndifferent environments. As is the IDE integration cmake is more and\nmore getting. As is properly builtin working caching of configure tests\n(llvm first cmake run, 7.7s, second 2.54s). As is the fact that a lot of\nbig projects (llvm, kde, qt, and a lot of others) migrated to it.\n\nFor me personally those benefits aren't worth that much. I've learned\nautoconf stuff. I've scripting around autoconf caching. But for newer\npeople that's not true.\n\nGreetings,\n\nAndres Freund\n\n\n-- \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, 10 Feb 2017 10:22:34 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Fri, Feb 10, 2017 at 6:07 PM, Robert Haas <robertmhaas@gmail.com> wrote:\n\n> On Mon, Jan 30, 2017 at 10:26 AM, Peter Eisentraut\n> <peter.eisentraut@2ndquadrant.com> wrote:\n> > On 1/30/17 1:28 AM, Andres Freund wrote:\n> >> Given that fact, I just don't buy why it's a good idea to not also\n> >> replace autoconf initially.\n> >\n> > Well, I find it a bit scary. If you do the big switch all at once, then\n> > you will have to dedicate the following 3 months to fixing complaints\n> > from developers and build farmers.\n>\n> I agree with that. I think replacing the Windows build system first\n> and then the non-Windows build system later is a better plan than\n> replacing both at the same time.\n>\n> But also, I'm not really sure whether this conversion makes sense. I\n> mean, any build system is going to require some work, and accordingly\n> our present build systems require some work. cmake will require\n> different work, but not necessarily less. The current system has a\n> long history; we pretty much know it works. Switching will inevitably\n> break some things. Maybe we'll end up better off in the long term,\n> but maybe we won't. Things are pretty good now, so it seems like it\n> would be easy for them to get worse rather than better. If nothing\n> else, everybody who has to learn the new system either to use it for\n> development or because they are doing packaging will have to do some\n> amount of extra work as a result of any switch.\n>\n>\nFor me a killer feature would be if/when we can get to a point where we can\nhave something pgxs-style on cmake that also works on windows.\n\nOur homemade Windows build system works OK for postgres, and while ugly it\nis as you say well tested by now. But it doesn't do *anything* to help\npeople build extensions on Windows.\n\n-- \n Magnus Hagander\n Me: http://www.hagander.net/\n Work: http://www.redpill-linpro.com/\n\nOn Fri, Feb 10, 2017 at 6:07 PM, Robert Haas <robertmhaas@gmail.com> wrote:On Mon, Jan 30, 2017 at 10:26 AM, Peter Eisentraut\n<peter.eisentraut@2ndquadrant.com> wrote:\n> On 1/30/17 1:28 AM, Andres Freund wrote:\n>> Given that fact, I just don't buy why it's a good idea to not also\n>> replace autoconf initially.\n>\n> Well, I find it a bit scary. If you do the big switch all at once, then\n> you will have to dedicate the following 3 months to fixing complaints\n> from developers and build farmers.\n\nI agree with that. I think replacing the Windows build system first\nand then the non-Windows build system later is a better plan than\nreplacing both at the same time.\n\nBut also, I'm not really sure whether this conversion makes sense. I\nmean, any build system is going to require some work, and accordingly\nour present build systems require some work. cmake will require\ndifferent work, but not necessarily less. The current system has a\nlong history; we pretty much know it works. Switching will inevitably\nbreak some things. Maybe we'll end up better off in the long term,\nbut maybe we won't. Things are pretty good now, so it seems like it\nwould be easy for them to get worse rather than better. If nothing\nelse, everybody who has to learn the new system either to use it for\ndevelopment or because they are doing packaging will have to do some\namount of extra work as a result of any switch.For me a killer feature would be if/when we can get to a point where we can have something pgxs-style on cmake that also works on windows.Our homemade Windows build system works OK for postgres, and while ugly it is as you say well tested by now. But it doesn't do *anything* to help people build extensions on Windows. -- Magnus Hagander Me: http://www.hagander.net/ Work: http://www.redpill-linpro.com/","msg_date":"Fri, 10 Feb 2017 19:27:55 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 02/10/2017 08:27 PM, Magnus Hagander wrote:\n> For me a killer feature would be if/when we can get to a point where we can\n> have something pgxs-style on cmake that also works on windows.\n>\n> Our homemade Windows build system works OK for postgres, and while ugly it\n> is as you say well tested by now. But it doesn't do *anything* to help\n> people build extensions on Windows.\n\nDo we need to build PostgreSQL itself using cmake, to achieve that? \nCould we write something like pgxs for cmake, only for extensions?\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":"Fri, 10 Feb 2017 20:31:12 +0200","msg_from":"Heikki Linnakangas <hlinnaka@iki.fi>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 2017-02-10 19:27:55 +0100, Magnus Hagander wrote:\n> For me a killer feature would be if/when we can get to a point where we can\n> have something pgxs-style on cmake that also works on windows.\n\nThat should be quite possible. The ugliest part will be to determine\nwhere to include a cmake file from (since that'll be copied in every\nsuch project), but after that it should be nearly trivial.\n\n- Andres\n\n\n-- \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, 10 Feb 2017 10:32:45 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Fri, Feb 10, 2017 at 7:31 PM, Heikki Linnakangas <hlinnaka@iki.fi> wrote:\n\n> On 02/10/2017 08:27 PM, Magnus Hagander wrote:\n>\n>> For me a killer feature would be if/when we can get to a point where we\n>> can\n>> have something pgxs-style on cmake that also works on windows.\n>>\n>> Our homemade Windows build system works OK for postgres, and while ugly it\n>> is as you say well tested by now. But it doesn't do *anything* to help\n>> people build extensions on Windows.\n>>\n>\n> Do we need to build PostgreSQL itself using cmake, to achieve that? Could\n> we write something like pgxs for cmake, only for extensions?\n>\n>\nI guess we wouldn't, but we'd still need the \"replacement for autoconf\"\npart. So then we're back to maintaining multiple buildsystems.\n\n\n-- \n Magnus Hagander\n Me: http://www.hagander.net/\n Work: http://www.redpill-linpro.com/\n\nOn Fri, Feb 10, 2017 at 7:31 PM, Heikki Linnakangas <hlinnaka@iki.fi> wrote:On 02/10/2017 08:27 PM, Magnus Hagander wrote:\n\nFor me a killer feature would be if/when we can get to a point where we can\nhave something pgxs-style on cmake that also works on windows.\n\nOur homemade Windows build system works OK for postgres, and while ugly it\nis as you say well tested by now. But it doesn't do *anything* to help\npeople build extensions on Windows.\n\n\nDo we need to build PostgreSQL itself using cmake, to achieve that? Could we write something like pgxs for cmake, only for extensions?I guess we wouldn't, but we'd still need the \"replacement for autoconf\" part. So then we're back to maintaining multiple buildsystems.-- Magnus Hagander Me: http://www.hagander.net/ Work: http://www.redpill-linpro.com/","msg_date":"Fri, 10 Feb 2017 19:33:18 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 2017-02-10 20:31:12 +0200, Heikki Linnakangas wrote:\n> On 02/10/2017 08:27 PM, Magnus Hagander wrote:\n> > For me a killer feature would be if/when we can get to a point where we can\n> > have something pgxs-style on cmake that also works on windows.\n> > \n> > Our homemade Windows build system works OK for postgres, and while ugly it\n> > is as you say well tested by now. But it doesn't do *anything* to help\n> > people build extensions on Windows.\n> \n> Do we need to build PostgreSQL itself using cmake, to achieve that? Could we\n> write something like pgxs for cmake, only for extensions?\n\nI don't see why it'd need to be done together. The minimal version\nwould be a simple cmake file that just sets a bunch of variables from\npg_config, provides a few rules for specifying the current pgxs stuff,\nand an example stanza how to include that file. We'd have to duplicate\nsome of the pgxs specific logic, but that's probably not too bad.\n\n- Andres\n\n\n-- \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, 10 Feb 2017 10:39:04 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 2017-02-10 19:33:18 +0100, Magnus Hagander wrote:\n> I guess we wouldn't, but we'd still need the \"replacement for autoconf\"\n> part. So then we're back to maintaining multiple buildsystems.\n\nHm? Do we really need that? Most of the things in an extension you do\n*not* want to determine separately from the backend. It's not like pgxs\natm really allows to differ wildly from autoconf's results. And most of\nthe relevant determinations made by autoconf are available in headers\nand/or we can generate a cmake include file with the results of\nautoconf.\n\n- Andres\n\n\n-- \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, 10 Feb 2017 10:41:27 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Feb 10, 2017 19:41, \"Andres Freund\" <andres@anarazel.de> wrote:\n\nOn 2017-02-10 19:33:18 +0100, Magnus Hagander wrote:\n> I guess we wouldn't, but we'd still need the \"replacement for autoconf\"\n> part. So then we're back to maintaining multiple buildsystems.\n\nHm? Do we really need that? Most of the things in an extension you do\n*not* want to determine separately from the backend. It's not like pgxs\natm really allows to differ wildly from autoconf's results. And most of\nthe relevant determinations made by autoconf are available in headers\nand/or we can generate a cmake include file with the results of\nautoconf.\n\n\nYeah, you're right. You need the output from the process, it mot the\nprocess itself.\n\n/Magnus\n\nOn Feb 10, 2017 19:41, \"Andres Freund\" <andres@anarazel.de> wrote:On 2017-02-10 19:33:18 +0100, Magnus Hagander wrote:\n> I guess we wouldn't, but we'd still need the \"replacement for autoconf\"\n> part. So then we're back to maintaining multiple buildsystems.\n\nHm? Do we really need that? Most of the things in an extension you do\n*not* want to determine separately from the backend. It's not like pgxs\natm really allows to differ wildly from autoconf's results. And most of\nthe relevant determinations made by autoconf are available in headers\nand/or we can generate a cmake include file with the results of\nautoconf.Yeah, you're right. You need the output from the process, it mot the process itself. /Magnus","msg_date":"Fri, 10 Feb 2017 19:55:30 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"\n\nOn 02/10/2017 01:27 PM, Magnus Hagander wrote:\n> On Fri, Feb 10, 2017 at 6:07 PM, Robert Haas <robertmhaas@gmail.com\n> <mailto:robertmhaas@gmail.com>> wrote:\n>\n> On Mon, Jan 30, 2017 at 10:26 AM, Peter Eisentraut\n> <peter.eisentraut@2ndquadrant.com\n> <mailto:peter.eisentraut@2ndquadrant.com>> wrote:\n> > On 1/30/17 1:28 AM, Andres Freund wrote:\n> >> Given that fact, I just don't buy why it's a good idea to not also\n> >> replace autoconf initially.\n> >\n> > Well, I find it a bit scary. If you do the big switch all at\n> once, then\n> > you will have to dedicate the following 3 months to fixing\n> complaints\n> > from developers and build farmers.\n>\n> I agree with that. I think replacing the Windows build system first\n> and then the non-Windows build system later is a better plan than\n> replacing both at the same time.\n>\n> But also, I'm not really sure whether this conversion makes sense. I\n> mean, any build system is going to require some work, and accordingly\n> our present build systems require some work. cmake will require\n> different work, but not necessarily less. The current system has a\n> long history; we pretty much know it works. Switching will inevitably\n> break some things. Maybe we'll end up better off in the long term,\n> but maybe we won't. Things are pretty good now, so it seems like it\n> would be easy for them to get worse rather than better. If nothing\n> else, everybody who has to learn the new system either to use it for\n> development or because they are doing packaging will have to do some\n> amount of extra work as a result of any switch.\n>\n>\n> For me a killer feature would be if/when we can get to a point where\n> we can have something pgxs-style on cmake that also works on windows.\n>\n> Our homemade Windows build system works OK for postgres, and while\n> ugly it is as you say well tested by now. But it doesn't do *anything*\n> to help people build extensions on Windows. \n>\n>\n\n\nWatch this space. There is work being done on building extensions out of\ntree using CMake on Windows. It's pretty neat, and I'm hoping to demo it\npublicly soon. But it's not reliant on CMake in core.\n\ncheers\n\nandrew\n\n-- \nAndrew Dunstan https://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":"Fri, 10 Feb 2017 14:42:16 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 11 Feb. 2017 08:42, \"Andrew Dunstan\" <andrew.dunstan@2ndquadrant.com>\nwrote:\n\n\n\nOn 02/10/2017 01:27 PM, Magnus Hagander wrote:\n> On Fri, Feb 10, 2017 at 6:07 PM, Robert Haas <robertmhaas@gmail.com\n> <mailto:robertmhaas@gmail.com>> wrote:\n>\n> On Mon, Jan 30, 2017 at 10:26 AM, Peter Eisentraut\n> <peter.eisentraut@2ndquadrant.com\n> <mailto:peter.eisentraut@2ndquadrant.com>> wrote:\n> > On 1/30/17 1:28 AM, Andres Freund wrote:\n> >> Given that fact, I just don't buy why it's a good idea to not also\n> >> replace autoconf initially.\n> >\n> > Well, I find it a bit scary. If you do the big switch all at\n> once, then\n> > you will have to dedicate the following 3 months to fixing\n> complaints\n> > from developers and build farmers.\n>\n> I agree with that. I think replacing the Windows build system first\n> and then the non-Windows build system later is a better plan than\n> replacing both at the same time.\n>\n> But also, I'm not really sure whether this conversion makes sense. I\n> mean, any build system is going to require some work, and accordingly\n> our present build systems require some work. cmake will require\n> different work, but not necessarily less. The current system has a\n> long history; we pretty much know it works. Switching will inevitably\n> break some things. Maybe we'll end up better off in the long term,\n> but maybe we won't. Things are pretty good now, so it seems like it\n> would be easy for them to get worse rather than better. If nothing\n> else, everybody who has to learn the new system either to use it for\n> development or because they are doing packaging will have to do some\n> amount of extra work as a result of any switch.\n>\n>\n> For me a killer feature would be if/when we can get to a point where\n> we can have something pgxs-style on cmake that also works on windows.\n>\n> Our homemade Windows build system works OK for postgres, and while\n> ugly it is as you say well tested by now. But it doesn't do *anything*\n> to help people build extensions on Windows.\n>\n>\n\n\nWatch this space. There is work being done on building extensions out of\ntree using CMake on Windows. It's pretty neat, and I'm hoping to demo it\npublicly soon. But it's not reliant on CMake in core.\n\n\nYeah. In fact it's completely independent of core's build system it and\nworks with any supported Pg, using pg_config for discovery.\n\nWe don't need cmake in core for pgxs-like functionality on Windows.\n\nI'd like to use it instead of our Perl windows stuff msbuild stuff in core\nthough. I wonder if it's practical to generate the cmake file lists etc\nfrom our Makefiles initialy like we do for the current system, and just\nreplace the MSVC/msbuild project generators with cmake initially.\n\nOn 11 Feb. 2017 08:42, \"Andrew Dunstan\" <andrew.dunstan@2ndquadrant.com> wrote:\n\nOn 02/10/2017 01:27 PM, Magnus Hagander wrote:\n> On Fri, Feb 10, 2017 at 6:07 PM, Robert Haas <robertmhaas@gmail.com\n> <mailto:robertmhaas@gmail.com>> wrote:\n>\n> On Mon, Jan 30, 2017 at 10:26 AM, Peter Eisentraut\n> <peter.eisentraut@2ndquadrant.com\n> <mailto:peter.eisentraut@2ndquadrant.com>> wrote:\n> > On 1/30/17 1:28 AM, Andres Freund wrote:\n> >> Given that fact, I just don't buy why it's a good idea to not also\n> >> replace autoconf initially.\n> >\n> > Well, I find it a bit scary. If you do the big switch all at\n> once, then\n> > you will have to dedicate the following 3 months to fixing\n> complaints\n> > from developers and build farmers.\n>\n> I agree with that. I think replacing the Windows build system first\n> and then the non-Windows build system later is a better plan than\n> replacing both at the same time.\n>\n> But also, I'm not really sure whether this conversion makes sense. I\n> mean, any build system is going to require some work, and accordingly\n> our present build systems require some work. cmake will require\n> different work, but not necessarily less. The current system has a\n> long history; we pretty much know it works. Switching will inevitably\n> break some things. Maybe we'll end up better off in the long term,\n> but maybe we won't. Things are pretty good now, so it seems like it\n> would be easy for them to get worse rather than better. If nothing\n> else, everybody who has to learn the new system either to use it for\n> development or because they are doing packaging will have to do some\n> amount of extra work as a result of any switch.\n>\n>\n> For me a killer feature would be if/when we can get to a point where\n> we can have something pgxs-style on cmake that also works on windows.\n>\n> Our homemade Windows build system works OK for postgres, and while\n> ugly it is as you say well tested by now. But it doesn't do *anything*\n> to help people build extensions on Windows.\n>\n>\n\n\nWatch this space. There is work being done on building extensions out of\ntree using CMake on Windows. It's pretty neat, and I'm hoping to demo it\npublicly soon. But it's not reliant on CMake in core.Yeah. In fact it's completely independent of core's build system it and works with any supported Pg, using pg_config for discovery.We don't need cmake in core for pgxs-like functionality on Windows.I'd like to use it instead of our Perl windows stuff msbuild stuff in core though. I wonder if it's practical to generate the cmake file lists etc from our Makefiles initialy like we do for the current system, and just replace the MSVC/msbuild project generators with cmake initially.","msg_date":"Sat, 11 Feb 2017 07:06:37 +0800","msg_from":"Craig Ringer <craig.ringer@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 2017-02-10 10:22:34 -0800, Andres Freund wrote:\n> On 2017-02-10 12:07:15 -0500, Robert Haas wrote:\n> > But also, I'm not really sure whether this conversion makes sense. I\n> > mean, any build system is going to require some work, and accordingly\n> > our present build systems require some work. cmake will require\n> > different work, but not necessarily less. The current system has a\n> > long history; we pretty much know it works. Switching will inevitably\n> > break some things. Maybe we'll end up better off in the long term,\n> > but maybe we won't. Things are pretty good now, so it seems like it\n> > would be easy for them to get worse rather than better.\n> \n> I do think it's kinda ok for people who've dealt with this for some\n> time. But for the first few times having to write autoconf macros is\n> quite intimidating. A somewhat less obscure way to deal with that is\n> worth something. As is better file dependency management, across\n> different environments. As is the IDE integration cmake is more and\n> more getting. As is properly builtin working caching of configure tests\n> (llvm first cmake run, 7.7s, second 2.54s). As is the fact that a lot of\n> big projects (llvm, kde, qt, and a lot of others) migrated to it.\n> \n> For me personally those benefits aren't worth that much. I've learned\n> autoconf stuff. I've scripting around autoconf caching. But for newer\n> people that's not true.\n\nCraig's email just now reminded me of another advantage: Using cmake\nacross the board, would mean we'd use the same ./configure alike logic\non both windows and linux. To me that seems quite and advantage over\nmanaging pg_config.h.win32/config_default.pl manually/separately - we\nobviously have screwed up quite badly there in the past\n(cf376a4adc0805b0).\n\n\n-- \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, 10 Feb 2017 15:15:36 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Fri, Feb 10, 2017 at 5:07 PM, Robert Haas <robertmhaas@gmail.com> wrote:\n\n> But also, I'm not really sure whether this conversion makes sense. I\n> mean, any build system is going to require some work, and accordingly\n> our present build systems require some work. cmake will require\n> different work, but not necessarily less. The current system has a\n> long history; we pretty much know it works. Switching will inevitably\n> break some things. Maybe we'll end up better off in the long term,\n> but maybe we won't. Things are pretty good now, so it seems like it\n> would be easy for them to get worse rather than better. If nothing\n> else, everybody who has to learn the new system either to use it for\n> development or because they are doing packaging will have to do some\n> amount of extra work as a result of any switch.\n>\n\nFor what it's worth (even well-maintained) cmake is a usability regression\nfrom well-maintained autotools from syseng/packager perspective.\n\nTwo anecdotes:\n\n- ./configure is much nicer from user perspective. Compare:\n./configure --prefix=/bla --enable-foo --disable-bar --with-ssl=/opt/myssl\ncmake -DCMAKE_INSTALL_PREFIX:PATH=/usr -DCMAKE_BLAH -DCMAKE_FOO\n-DCMAKE_WHO_KNOWS_WHAT:PATH=/opt/myssl\n\n- Things like set(_PYTHON3_VERSIONS 3.7 3.6 3.5 3.4 3.3 3.2 3.1 3.0)\n<https://github.com/Kitware/CMake/blob/master/Modules/FindPythonInterp.cmake#L43>\nGuess what, on older cmake versions this list did not include anything\nolder that 3.3, so it was failing with in-comprehensive generic \"not found\"\nerror even though 3.4 was installed.\nWith this \"fix\" somebody somewhere will be banging their head against the\nwall again once 3.8 is out.\nOverall, when things go wrong debugging cmake requires cmake knowledge,\nwhile autotools mostly require shell knowledge which is much more common\n(again, for sysadmins/packagers).\n\nSo while cmake is better for developers it is usually worse off for\npackagers and advanced users. I'm not saying migration is not worth it, I'm\npointing out costs of such migration.\n\nPS: I personally like Google's internal version of https://bazel.build/ a\nlot. I've never used open-source version but I presume it's similar. While\nit has many problems (Java, lack of popular IDE support, lack of popularity\nand, again, Java) good parts are rules are both machine- and human-\nreadable and writable and generally easy to debug. I'm not bold enough to\npropose PostgreSQL to use it, but I'd be happy to see ideas from it to be\nused elsewhere.\n\n-- \nVladimir Rusinov\nStorage SRE, Google Ireland\n\nGoogle Ireland Ltd.,Gordon House, Barrow Street, Dublin 4, Ireland\nRegistered in Dublin, Ireland\nRegistration Number: 368047","msg_date":"Sun, 12 Feb 2017 17:55:56 +0000","msg_from":"Vladimir Rusinov <vrusinov@google.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"2017-02-12 20:55 GMT+03:00 Vladimir Rusinov <vrusinov@google.com>:\n\n> Overall, when things go wrong debugging cmake requires cmake knowledge,\n> while autotools mostly require shell knowledge which is much more common\n> (again, for sysadmins/packagers).\n\n\nIt's not really true because of CMake scripts much easier than tons of crap\nbash (configure) and m4 scripts in Autotools, also please don't forget\nWindows MSVC, Xcode and etc usage.\n\nPS: I personally like Google's internal version of https://bazel.build/ a\n> lot. I've never used open-source version but I presume it's similar. While\n> it has many problems (Java, lack of popular IDE support, lack of popularity\n> and, again, Java) good parts are rules are both machine- and human-\n> readable and writable and generally easy to debug. I'm not bold enough to\n> propose PostgreSQL to use it, but I'd be happy to see ideas from it to be\n> used elsewhere.\n\n\nWe have many build systems, for example, another one http://mesonbuild.com/\nbut CMake the best today as meta build system.\n\n2017-02-12 20:55 GMT+03:00 Vladimir Rusinov <vrusinov@google.com>:Overall, when things go wrong debugging cmake requires cmake knowledge, while autotools mostly require shell knowledge which is much more common (again, for sysadmins/packagers).It's not really true because of CMake scripts much easier than tons of crap bash (configure) and m4 scripts in Autotools, also please don't forget Windows MSVC, Xcode and etc usage. PS: I personally like Google's internal version of https://bazel.build/ a lot. I've never used open-source version but I presume it's similar. While it has many problems (Java, lack of popular IDE support, lack of popularity and, again, Java) good parts are rules are both machine- and human- readable and writable and generally easy to debug. I'm not bold enough to propose PostgreSQL to use it, but I'd be happy to see ideas from it to be used elsewhere.We have many build systems, for example, another one http://mesonbuild.com/ but CMake the best today as meta build system.","msg_date":"Mon, 13 Feb 2017 11:53:22 +0300","msg_from":"Yuriy Zhuravlev <stalkerg@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On 2/8/17 4:44 PM, Andres Freund wrote:\n>> Well, I find it a bit scary. If you do the big switch all at once, then\n>> you will have to dedicate the following 3 months to fixing complaints\n>> from developers and build farmers.\n> \n> That'll be the case just as well if we spread it out over two cycles,\n> except that we'll have it in multiple phases, and we run into the danger\n> of a half-done conversion.\n\nThey key term is \"dedicated\". I don't think anyone wants to spend 3\nmonths doing nothing but fixing the build system. If we spread it over\nseveral releases and bring more people up to speed along the way, that\nlooks more manageable.\n\n>> That would work if there were a single entry point into the build\n>> system. But in practice there are many, and every one of them is\n>> someone's favorite. It's unlikely that we will be able to enumerate all\n>> of them during patch review.\n> \n> Not sure what you mean with \"entry point\"?\n\nPeople have all kinds of expectations on how the build system behaves.\nIt's not just ./configure; make; make install. All kinds of niche and\nedge cases have evolved over the years. Variables you can set,\noverrides, targets in some subdirectories, building in subdirectories\nand having it build another subdirectory first, testing this way and\ntesting that way, testing with a custom locale or a custom database\nname, building before testing or not, testing without reinstalling, and\nso on and so on. You'd have to make sure at least some of that\ncontinues to work or be able to explain it away. And most of it is not\nreally documented.\n\n-- \nPeter Eisentraut 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":"Mon, 13 Feb 2017 10:52:41 -0500","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Mon, Feb 13, 2017 at 10:52 AM, Peter Eisentraut\n<peter.eisentraut@2ndquadrant.com> wrote:\n> People have all kinds of expectations on how the build system behaves.\n> It's not just ./configure; make; make install. All kinds of niche and\n> edge cases have evolved over the years. Variables you can set,\n> overrides, targets in some subdirectories, building in subdirectories\n> and having it build another subdirectory first, testing this way and\n> testing that way, testing with a custom locale or a custom database\n> name, building before testing or not, testing without reinstalling, and\n> so on and so on. You'd have to make sure at least some of that\n> continues to work or be able to explain it away. And most of it is not\n> really documented.\n\n...which is another argument for just not changing anything. I mean,\nagain, the current Windows build system is unbeautiful and\noccasionally takes some work, but if we switch to cmake, that has to\nget enough better to make up for all of the things that get worse.\nAnd it's far from obvious than this will be so, especially when you\nconsider the fact that many people have detailed knowledge of how to\ntweak the current system to do what they want. As you point out here,\na lot of that stuff may stop working if we replace the system\nwholesale.\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, 13 Feb 2017 13:48:08 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"},{"msg_contents":"On Wed, Feb 8, 2017 at 04:52:19PM -0500, Tom Lane wrote:\n> Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n> > On 2/8/17 6:21 AM, Yuriy Zhuravlev wrote:\n> >> Support two build systems it's not big deal really. I have been working\n> >> on this past year without any big troubles. \n> >> Also we have second perl build system...\n> \n> > The perl/msvc build system pulls in information from the makefiles. So\n> > when you add a file or something basic like that, you don't have to\n> > update it. So it's really more like 1.5 build systems.\n> \n> Really it's more like 1.1 build systems, in that the MSVC scripts do that\n> just well enough that you *usually* don't have to think about them. But\n> then when they fail, and you have to figure out why, it can be a pain.\n\nIf cmake isn't going to be able to query the Makefiles and adjust to\nchanges we make there, changing our Windows build system from MSVC to\ncmake takes us from maintaining 1.1 build systems to two build systems,\nand I don't think anyone wants that.\n\nIf we go to cmake, I think we need to agree we will eventually _only_\nuse cmake. I don't think having two build systems we have to maintain\nis better than 1.1 build systems where we can mostly ignore 0.1 of that.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n\n-- \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 Feb 2017 16:11:21 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: WIP: About CMake v2"}],"string":"[\n {\n \"msg_contents\": \"Hi Yury,\\n\\n\\nOn Fri, Aug 19, 2016 at 9:46 AM, Yury Zhuravlev\\n<u.zhuravlev@postgrespro.ru> wrote:\\n> Christian Convey wrote:\\n>>\\n>> I'm interested in helping with your CMake effort. I don't have any\\n>> experience contributing to PG, but I do have some free time at the\\n>> moment. Please let me know if I can help.\\n>\\n> I glad to hear it. I suppose you can just try build postgres and send all\\n> problems to github tracker.\\n> https://github.com/stalkerg/postgres_cmake/issues\\n\\nThanks, I'll be happy to do that. There's been a lot of discussion on\\nthis thread regarding the minimum required CMake version. If you'd\\nlike me to test with a particular version (or versions) of CMake,\\nplease let me know.\\n\\n- Christian\\n\\n\\n-- \\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, 20 Aug 2016 01:18:29 -0400\",\n \"msg_from\": \"Christian Convey <christian.convey@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Hi Yury,\\n\\n>> I glad to hear it. I suppose you can just try build postgres and send all\\n>> problems to github tracker.\\n>> https://github.com/stalkerg/postgres_cmake/issues\\n\\nFYI, I had success using your \\\"postgres_cmake\\\" repo. I tested it up\\nthrough \\\"make check\\\" and \\\"make install\\\".\\n\\nHere are the details:\\n\\n* postgres_cmake commit e7e75160d4533cd8caa9f3f0dd7b485dbd4e7bdf\\n* compiler = cc (Ubuntu 4.8.4-2ubuntu1~14.04.3) 4.8.4\\n* cmake version = 3.5.3\\n\\nKind regards,\\nChristian\\n\\n\\n-- \\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, 21 Aug 2016 18:48:39 -0400\",\n \"msg_from\": \"Christian Convey <christian.convey@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Yury,\\n\\nOn Mon, Aug 22, 2016 at 7:48 AM, Christian Convey\\n<christian.convey@gmail.com> wrote:\\n>>> I glad to hear it. I suppose you can just try build postgres and send all\\n>>> problems to github tracker.\\n>>> https://github.com/stalkerg/postgres_cmake/issues\\n>\\n> FYI, I had success using your \\\"postgres_cmake\\\" repo. I tested it up\\n> through \\\"make check\\\" and \\\"make install\\\".\\n>\\n> Here are the details:\\n>\\n> * postgres_cmake commit e7e75160d4533cd8caa9f3f0dd7b485dbd4e7bdf\\n> * compiler = cc (Ubuntu 4.8.4-2ubuntu1~14.04.3) 4.8.4\\n> * cmake version = 3.5.3\\n\\nCould it be possible to get a refreshed patch on this thread for at\\nleast the sake of the archives? I'd really like to see somehting\\nhappening here and do some progress for this CF.\\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 Sep 2016 21:41:35 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Michael Paquier wrote:\\n> Could it be possible to get a refreshed patch on this thread for at\\n> least the sake of the archives? I'd really like to see somehting\\n> happening here and do some progress for this CF.\\n\\nSure, I will do it on Friday.\\nToday I finished mingw+msys support. (mingw without msys has not passed \\nsome tests)\\n-- \\nYury Zhuravlev\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres 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, 15 Sep 2016 22:38:16 +0300\",\n \"msg_from\": \"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Fri, Sep 16, 2016 at 4:38 AM, Yury Zhuravlev\\n<u.zhuravlev@postgrespro.ru> wrote:\\n> Michael Paquier wrote:\\n>>\\n>> Could it be possible to get a refreshed patch on this thread for at\\n>> least the sake of the archives? I'd really like to see somehting\\n>> happening here and do some progress for this CF.\\n>\\n>\\n> Sure, I will do it on Friday.\\n> Today I finished mingw+msys support. (mingw without msys has not passed some\\n> tests)\\n\\nOkay. That sounds good to me. I don't recall what your patch is\\nexactly doing but could you still keep the vanilla Makefiles around?\\nThis will reduce the diff of the patch, and we'd need anyway to keep\\nthe former Makefile methods around until the buildfarm scripts are\\nupdated, the animals do the switch and then get green. So a period\\nwhere both live in parallel is unavoidable.\\n\\nI heard as well that MySQL is using cmake... It may be interesting to\\nsee how they are integrating with it as a large project and avoid\\ntheir past mistakes.\\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, 16 Sep 2016 09:58:42 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Fri, Sep 16, 2016 at 9:58 AM, Michael Paquier\\n<michael.paquier@gmail.com> wrote:\\n> Okay. That sounds good to me. I donas well a look around, and cmake really looks like a robust alternative to ./configure. Now I am aware of the fact that your patch recommends to build 't recall what your patch is\\n> exactly doing but could you still keep the vanilla Makefiles around?\\n> This will reduce the diff of the patch, and we'd need anyway to keep\\n> the former Makefile methods around until the buildfarm scripts are\\n> updated, the animals do the switch and then get green. So a period\\n> where both live in parallel is unavoidable.\\n>\\n> I heard as well that MySQL is using cmake... It may be interesting to\\n> see how they are integrating with it as a large project and avoid\\n> their past mistakes.\\n\\nI could not resist so I just had a look at your patch. I had as well a\\nlook around, and cmake really looks like a robust alternative to\\n./configure. In short, people doing now that:\\n./configure\\nmake\\nmake install\\nWould just do that:\\ncmake .\\nmake\\nmake install\\n\\nI am no cmake guru yet, but VPATH builds are supported out of the box,\\nwhich is cool.\\n\\nYour patch recommends to build with cmake after creating build/, now I\\nwould expect most users to run cmake from the root folder. However\\nthis causes all the Makefiles to get overwritten. As supporting all\\nplatforms at once with cmake is going to be uncommitable, we are going\\nto need both methods able to live together for a while. Well, they can\\ncoexist with this patch as long as cmake is not run from the root of\\nthe code tree, which is acceptable for me as long as the switch is not\\ncompleted. However others may think differently on the matter.\\n\\nInstead of getting support for all existing platforms, I would\\nrecommend as well focusing only on one platform for the time being,\\nLinux, and get the work done correctly for that first. Once there is\\nsomething committed, we will be able to patch the buildfarm, and get\\nmachines to switch to cmake one by one. After those are switched, we\\ncould extend that. Another point of contention is support for\\nextensions. How long should we keep support for the existing PGXS? How\\nexternal extensions would compile with the new thing infrastructure?\\n\\nWhich brings me to another point, your patch is focused on features,\\nmeaning that per-OS checks are all grouped by feature, but it may be a\\ngood idea to split checks by OS if necessary, with for example\\nper-platform files and scripts in cmake/os/. And we could have just\\nsomething for Linux now.\\n\\nA couple of issues I have noticed with your patch after a first set of tests:\\n- root's .gitignore needs to add entries for CMakeFiles and\\ncmake_install.cmake. You need as well to ignore CMakeCache\\n- A couple of headers are generated, like cubeparse.h (?)\\n- Currently a lot of users do things like that:\\ncd src/test/regress/ && make check\\nBut this patch breaks that, and that's not cool. Recovery tests in\\nsrc/test/regress won't run either.\\n\\nThat's all I have for now. Looking forward to seeing some progress here.\\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, 16 Sep 2016 22:33:50 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Michael Paquier wrote:\\n> Your patch recommends to build with cmake after creating build/, now I\\n> would expect most users to run cmake from the root folder. However\\n> this causes all the Makefiles to get overwritten. As supporting all\\n> platforms at once with cmake is going to be uncommitable, we are going\\n> to need both methods able to live together for a while. Well, they can\\n> coexist with this patch as long as cmake is not run from the root of\\n> the code tree, which is acceptable for me as long as the switch is not\\n> completed. However others may think differently on the matter.\\n\\nIt's really a good point. Forbidden run cmake from root it is better \\ndecision for me (of course for start). \\n\\n>Instead of getting support for all existing platforms, I would\\n>recommend as well focusing only on one platform for the time being,\\n>Linux, and get the work done correctly for that first. Once there is\\n>something committed, we will be able to patch the buildfarm, and get\\n>machines to switch to cmake one by one. After those are switched, we\\n>could extend that.\\n\\nYou mean in first version of patch I can focus on Linux systems?\\n\\n>Another point of contention is support for\\n>extensions. How long should we keep support for the existing PGXS? How\\n>external extensions would compile with the new thing infrastructure?\\n\\nAs long as possible. I hope I can make PGXS Makefiles generator.\\n\\n>Which brings me to another point, your patch is focused on features,\\n>meaning that per-OS checks are all grouped by feature, but it may be a\\n>good idea to split checks by OS if necessary, with for example\\n>per-platform files and scripts in cmake/os/. And we could have just\\n>something for Linux now.\\n\\nCurrently I do not have a lot OS specific tests. All checks are doing in \\nsame manner. \\n\\n>- root's .gitignore needs to add entries for CMakeFiles and\\n>cmake_install.cmake. You need as well to ignore CMakeCache\\n\\nThanks I done this in last commit. \\n\\n>- A couple of headers are generated, like cubeparse.h (?)\\n\\nBecause BISON generate header by default. I suppose author of cube launched \\nbison by hand but I made it automatic. \\n\\n>- Currently a lot of users do things like that:\\n>cd src/test/regress/ && make check\\n>But this patch breaks that, and that's not cool. Recovery tests in\\n>src/test/regress won't run either.\\n\\nIt seems restriction by design because in CMake you have only one enter \\npoint. \\n\\n>That's all I have for now. Looking forward to seeing some progress here.\\n\\nI merged master to my branch and I spent time to porting all changes. I \\nhope send patch in the weekend without terrible flaws.\\n\\n-- \\nYury Zhuravlev\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres 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, 16 Sep 2016 19:40:48 +0300\",\n \"msg_from\": \"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Sat, Sep 17, 2016 at 1:40 AM, Yury Zhuravlev\\n<u.zhuravlev@postgrespro.ru> wrote:\\n> Michael Paquier wrote:\\n> I merged master to my branch and I spent time to porting all changes. I hope\\n> send patch in the weekend without terrible flaws.\\n\\nBy the way, I noticed that you did not register this patch in the current CF..\\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\": \"Sat, 17 Sep 2016 08:07:43 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2016-09-16 22:33:50 +0900, Michael Paquier wrote:\\n> As supporting all platforms at once with cmake is going to be\\n> uncommitable, we are going to need both methods able to live together\\n> for a while.\\n\\nI very strongly disagree with this. It's way too likely that we end up\\nreleasing with multiple buildsystems that way. I think we'll have to\\nrequire support for the most common OSs, and then fix up the fallout\\nafterwards.\\n\\nAndres\\n\\n\\n-- \\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 Sep 2016 16:11:55 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Sat, Sep 17, 2016 at 8:11 AM, Andres Freund <andres@anarazel.de> wrote:\\n> Hi,\\n>\\n> On 2016-09-16 22:33:50 +0900, Michael Paquier wrote:\\n>> As supporting all platforms at once with cmake is going to be\\n>> uncommitable, we are going to need both methods able to live together\\n>> for a while.\\n>\\n> I very strongly disagree with this. It's way too likely that we end up\\n> releasing with multiple buildsystems that way. I think we'll have to\\n> require support for the most common OSs, and then fix up the fallout\\n> afterwards.\\n\\nOk, then if we go through the hard method the switch patch had better\\nremove as well all the Makefiles in the tree and recommend cmake run\\nfrom base root as the default build method. By the way, there will be\\na strong need for docs in the patch sooner or later...\\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\": \"Sat, 17 Sep 2016 08:17:02 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Michael Paquier wrote:\\n> On Sat, Sep 17, 2016 at 1:40 AM, Yury Zhuravlev\\n> <u.zhuravlev@postgrespro.ru> wrote:\\n>> Michael Paquier wrote:\\n>> I merged master to my branch and I spent time to porting all \\n>> changes. I hope\\n>> send patch in the weekend without terrible flaws.\\n>\\n> By the way, I noticed that you did not register this patch in \\n> the current CF..\\n\\nNow, I published the first version of the patch. \\nThis patch not ready for commit yet and all current task you can read here:\\nhttps://github.com/stalkerg/postgres_cmake/issues\\n\\nI hope we realy close to end. \\n\\nIn this patch I forbade in-source build for avoid overwriting current \\nMakefiles. \\nWe will can remove all Makefiles only after shall see in CMake. You don't \\nneed support two system. During commitfests CMake build system will be \\nsupported by me. \\n\\nI need help with buildfarm because my knowledge of Perl is very bad \\n(thought about rewrite buildfarm to Python). \\n\\n\\nI hope for your support.\\n\\n-- \\nYury Zhuravlev\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres 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, 17 Sep 2016 20:21:55 +0300\",\n \"msg_from\": \"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Andres Freund wrote:\\n> It's way too likely that we end up\\n> releasing with multiple buildsystems that way.\\n\\nIf we fail in the CMake integration we can easily remove all CMake files \\nbefore code freeze. \\nI hope step by step CMake integration better way for current situation. \\nBut I do not insist, just I trying to find the easiest way.\\n\\n-- \\nYury Zhuravlev\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres 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, 17 Sep 2016 20:43:07 +0300\",\n \"msg_from\": \"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 9/16/16 9:33 AM, Michael Paquier wrote:\\n> I am no cmake guru yet, but VPATH builds are supported out of the box,\\n> which is cool.\\n> \\n> Your patch recommends to build with cmake after creating build/, now I\\n> would expect most users to run cmake from the root folder.\\n\\nMy understanding is that cmake strongly recommends building in a\\nseparate directory, which is usually a subdirectory named something like\\n\\\"build\\\". So the above is likely going to be the standard workflow.\\n\\n-- \\nPeter Eisentraut 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\": \"Sat, 17 Sep 2016 20:12:54 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"> On 17 Sep 2016, at 20:21, Yury Zhuravlev <u.zhuravlev@postgrespro.ru> wrote:\\n> \\n> Michael Paquier wrote:\\n>> On Sat, Sep 17, 2016 at 1:40 AM, Yury Zhuravlev\\n>> <u.zhuravlev@postgrespro.ru> wrote:\\n>>> Michael Paquier wrote:\\n>>> I merged master to my branch and I spent time to porting all changes. I hope\\n>>> send patch in the weekend without terrible flaws.\\n>> \\n>> By the way, I noticed that you did not register this patch in the current CF..\\n> \\n> Now, I published the first version of the patch. This patch not ready for commit yet and all current task you can read here:\\n> https://github.com/stalkerg/postgres_cmake/issues\\n> \\n> I hope we realy close to end. \\n> In this patch I forbade in-source build for avoid overwriting current Makefiles. We will can remove all Makefiles only after shall see in CMake. You don't need support two system. During commitfests CMake build system will be supported by me. \\n> I need help with buildfarm because my knowledge of Perl is very bad (thought about rewrite buildfarm to Python). \\n> \\n> I hope for your support.\\n\\n\\nTried to generate Xcode project out of cmake, build fails on genbki.pl: can't locate Catalog.pm (which itself lives in src/backend/catalog/Catalog.pm)\\n\\n-- \\nStas Kelvich\\nPostgres Professional: http://www.postgrespro.com\\nRussian Postgres Company\\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, 3 Oct 2016 13:01:46 +0300\",\n \"msg_from\": \"Stas Kelvich <s.kelvich@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Stas Kelvich wrote:\\n>> On 17 Sep 2016, at 20:21, Yury Zhuravlev \\n>> <u.zhuravlev@postgrespro.ru> wrote:\\n>> \\n>> Michael Paquier wrote: ...\\n>\\n>\\n> Tried to generate Xcode project out of cmake, build fails on \\n> genbki.pl: can't locate Catalog.pm (which itself lives in \\n> src/backend/catalog/Catalog.pm)\\n>\\n\\nCan you try again? On my Xcode 7.3 / ElCapitan everything is working \\ncorrectly.\\n-- \\nYury Zhuravlev\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres 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, 07 Oct 2016 01:22:11 +0300\",\n \"msg_from\": \"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Stas Kelvich wrote:\\n> Tried to generate Xcode project out of cmake, build fails on \\n> genbki.pl: can't locate Catalog.pm (which itself lives in \\n> src/backend/catalog/Catalog.pm)\\n\\nThanks again! I have corrected your issue.\\n-- \\nYury Zhuravlev\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres 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, 14 Oct 2016 12:51:57 +0300\",\n \"msg_from\": \"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 9/17/16 1:21 PM, Yury Zhuravlev wrote:\\n> Now, I published the first version of the patch. \\n\\nI tried this out. Because of some file moves in initdb and\\npg_basebackup, the build fails:\\n\\n[ 74%] Linking C executable initdb\\n Undefined symbols for architecture x86_64:\\n \\\"_fsync_pgdata\\\", referenced from:\\n _main in initdb.c.o\\n ld: symbol(s) not found for architecture x86_64\\n collect2: error: ld returned 1 exit status\\n make[2]: *** [src/bin/initdb/CMakeFiles/initdb.dir/build.make:177:\\nsrc/bin/initdb/initdb] Error 1\\n make[1]: *** [CMakeFiles/Makefile2:2893:\\nsrc/bin/initdb/CMakeFiles/initdb.dir/all] Error 2\\n make: *** [Makefile:128: all] Error 2\\n\\nPlease submit an updated patch.\\n\\nI suggest you use git format-patch to produce patches. This is easier\\nto apply, especially when there are a lot of new files involved. Also\\nuse the git facilities to check for whitespace errors.\\n\\nPlease supply some documentation, such as\\n\\n- what are the basic commands\\n- how to set include/library paths, choose configure options\\n- how to set CFLAGS\\n- how to see raw build commands\\n- what are the targets for all/world/check/docs etc.\\n- explain directory structure\\n\\nI suggest for now you could put this into a README.cmake file in your\\npatch. We don't need to commit it that way, but it would help in the\\nmeantime.\\n\\nWhen I run cmake without options, it seems to do opportunistic feature\\nchecking. For example, it turns on OpenSSL or Python support if it can\\nfind it, otherwise it turns it off. We need this to be deterministic.\\nWithout options, choose the basic feature set, require all other\\nfeatures to be turned on explicitly, fail if they can't be found.\\nWhatever the Autoconf-based build does now has been fairly deliberately\\ntuned, so there should be very little reason to deviate from that.\\n\\nThe Python check appears to be picking up pieces from two different\\nPython installations:\\n\\n -- Found PythonInterp: /usr/local/bin/python (found version \\\"2.7.12\\\")\\n -- Found PythonLibs: /usr/lib/libpython2.7.dylib (found version \\\"2.7.10\\\")\\n\\nThe check results otherwise look OK, but I'm a bit confused about the\\norder. It checks for some functions before all the header files are\\nchecked for. Is that intentional?\\n\\nThere are a number of changes in .[ch] and .pl files that are unclear\\nand not explained. Please explain them. You can also submit separate\\npreliminary patches if you need to do some refactoring. Ultimately, I\\nwould expect this patch not to require C code changes.\\n\\n-- \\nPeter Eisentraut 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\": \"Tue, 8 Nov 2016 14:37:17 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Hello.\\n\\nPeter Eisentraut wrote:\\n> I tried this out. Because of some file moves in initdb and\\n> pg_basebackup, the build fails:\\nYes, I need rebase patch to latest master. I do it as soon as possible. \\nAlso I have some new achievements. \\n\\n>I suggest you use git format-patch to produce patches. This is easier\\n>to apply, especially when there are a lot of new files involved. Also\\n>use the git facilities to check for whitespace errors.\\nI agree. \\n\\n>I suggest for now you could put this into a README.cmake file in your\\n>patch. We don't need to commit it that way, but it would help in the\\n>meantime.\\nGood idea, currently I write all documentation on github. I will try to \\nmove it to patch. \\n\\n>When I run cmake without options, it seems to do opportunistic feature\\n>checking. For example, it turns on OpenSSL or Python support if it can\\n>find it, otherwise it turns it off. We need this to be deterministic.\\n>Without options, choose the basic feature set, require all other\\n>features to be turned on explicitly, fail if they can't be found.\\n>Whatever the Autoconf-based build does now has been fairly deliberately\\n>tuned, so there should be very little reason to deviate from that.\\nThis approach I see only in Postgres project and not fully understood.\\nCan you explain me more what reasons led to this approach?\\nNot big deal to make behavior like postgres Autoconf but I think it's \\nimportant clear view here. \\n\\n>The Python check appears to be picking up pieces from two different\\n>Python installations:\\nOoops you right. For Python detecting I use standard CMake module and\\nhis behavior depending on CMake version. We should really careful here.\\n\\n>The check results otherwise look OK, but I'm a bit confused about the\\n>order. It checks for some functions before all the header files are\\n>checked for. Is that intentional?\\nI have plans to change order checks. \\n\\n>There are a number of changes in .[ch] and .pl files that are unclear\\n>and not explained. Please explain them. You can also submit separate\\n>preliminary patches if you need to do some refactoring. Ultimately, I\\n>would expect this patch not to require C code changes.\\nI suppose separate patches with comments will be best way. I will do it.\\nI think we can talks about details after that.\\n\\nBig thanks for your remarks it's very important for me and this \\\"small\\\" \\nproject.\\nI will try to do all tasks quickly.\\n\\nBest regards. \\n\\n-- \\nYury Zhuravlev\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres 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, 08 Nov 2016 23:40:57 +0300\",\n \"msg_from\": \"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 9 Nov. 2016 06:37, \\\"Yury Zhuravlev\\\" <u.zhuravlev@postgrespro.ru> wrote:\\n>\\n>> When I run cmake without options, it seems to do opportunistic feature\\n>> checking. For example, it turns on OpenSSL or Python support if it can\\n>> find it, otherwise it turns it off. We need this to be deterministic.\\n>> Without options, choose the basic feature set, require all other\\n>> features to be turned on explicitly, fail if they can't be found.\\n>> Whatever the Autoconf-based build does now has been fairly deliberately\\n>> tuned, so there should be very little reason to deviate from that.\\n>\\n> This approach I see only in Postgres project and not fully understood.\\n> Can you explain me more what reasons led to this approach?\\n\\nIt's predictable. The default has the same result for everyone. I quite\\nlike it myself.\\n\\n\\nOn 9 Nov. 2016 06:37, \\\"Yury Zhuravlev\\\" <u.zhuravlev@postgrespro.ru> wrote:\\n>\\n>> When I run cmake without options, it seems to do opportunistic feature\\n>> checking. For example, it turns on OpenSSL or Python support if it can\\n>> find it, otherwise it turns it off. We need this to be deterministic.\\n>> Without options, choose the basic feature set, require all other\\n>> features to be turned on explicitly, fail if they can't be found.\\n>> Whatever the Autoconf-based build does now has been fairly deliberately\\n>> tuned, so there should be very little reason to deviate from that.\\n>\\n> This approach I see only in Postgres project and not fully understood.\\n> Can you explain me more what reasons led to this approach?\\nIt's predictable. The default has the same result for everyone. I quite like it myself.\",\n \"msg_date\": \"Wed, 9 Nov 2016 06:54:03 +0800\",\n \"msg_from\": \"Craig Ringer <craig.ringer@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 9, 2016 at 7:54 AM, Craig Ringer\\n<craig.ringer@2ndquadrant.com> wrote:\\n> On 9 Nov. 2016 06:37, \\\"Yury Zhuravlev\\\" <u.zhuravlev@postgrespro.ru> wrote:\\n>> This approach I see only in Postgres project and not fully understood.\\n>> Can you explain me more what reasons led to this approach?\\n>\\n> It's predictable. The default has the same result for everyone. I quite like\\n> it myself.\\n\\n+1. Let's tell to the system what we want him to do and not let him\\nguess what we'd like to be done or it will get harder to test and\\ndevelop code for all kind of code paths with #ifdef's. That's one step\\naway from Skynet.\\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, 9 Nov 2016 11:12:10 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 9 November 2016 at 10:12, Michael Paquier <michael.paquier@gmail.com> wrote:\\n> On Wed, Nov 9, 2016 at 7:54 AM, Craig Ringer\\n> <craig.ringer@2ndquadrant.com> wrote:\\n>> On 9 Nov. 2016 06:37, \\\"Yury Zhuravlev\\\" <u.zhuravlev@postgrespro.ru> wrote:\\n>>> This approach I see only in Postgres project and not fully understood.\\n>>> Can you explain me more what reasons led to this approach?\\n>>\\n>> It's predictable. The default has the same result for everyone. I quite like\\n>> it myself.\\n>\\n> +1. Let's tell to the system what we want him to do and not let him\\n> guess what we'd like to be done or it will get harder to test and\\n> develop code for all kind of code paths with #ifdef's. That's one step\\n> away from Skynet.\\n\\nEr... ok then. (Backs away slowly).\\n\\nMore seriously, I like it for development where a stable and\\npredictable default is great.\\n\\nFor users it slightly sucks, as most users will want us to find\\nwhatever is on the system without being manually told to enable each\\nfeature. \\\"Of course I want SSL, I have openssl installed don't I?\\\"\\nIt's not like we require users to specify --enable-largefile\\n--enable-atomics --enable-getopt --enable-ipv6 .... we do detect a\\nlot automatically.\\n\\nSo personally I think it'd be fine if a cmake build defaulted to\\nfinding and using what it could, but offered a --minimal mode or\\nwhatever that gets us just core postgres + whatever we enable\\nexplicitly. But our current behaviour is OK too.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL 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, 9 Nov 2016 10:52:48 +0800\",\n \"msg_from\": \"Craig Ringer <craig.ringer@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Craig Ringer <craig.ringer@2ndquadrant.com> writes:\\n> On 9 Nov. 2016 06:37, \\\"Yury Zhuravlev\\\" <u.zhuravlev@postgrespro.ru> wrote:\\n>> This approach I see only in Postgres project and not fully understood.\\n>> Can you explain me more what reasons led to this approach?\\n\\n> It's predictable. The default has the same result for everyone. I quite\\n> like it myself.\\n\\nIt's advantageous from a packaging standpoint, precisely because it's\\npredictable: either you get the set of features you expect, or you get\\na build failure. Years ago we were more on the \\\"opportunistic\\\" side\\nof things, and we had problems with packages sometimes silently losing\\nfeatures. A pretty-recent example is that OpenSSL changed their APIs\\nin 1.1.0 so that our configure probe failed. With an opportunistic\\napproach, that would have meant builds silently going from \\\"ssl yes\\\"\\nto \\\"ssl no\\\". That's not good.\\n\\nSo this is really not open for negotiation. As Peter said upthread,\\nwhat we are looking for in a CMake reimplementation is that it behaves\\nexactly like the Autoconf version does. To the extent that you are unable\\nor unwilling to duplicate that behavior, you increase the odds that\\nwe'll reject this work.\\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, 08 Nov 2016 22:47:34 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Tom Lane wrote:\\n> So this is really not open for negotiation. As Peter said upthread,\\n> what we are looking for in a CMake reimplementation is that it behaves\\n> exactly like the Autoconf version does. To the extent that you are unable\\n> or unwilling to duplicate that behavior, you increase the odds that\\n> we'll reject this work.\\nWho asking about negotiation? I just wanted an explanation for the clear \\nunderstanding and nothing more. \\nNow I know about reasons. Thanks.\\n\\nregards\\n-- \\nYury Zhuravlev\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres 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, 09 Nov 2016 12:52:07 +0300\",\n \"msg_from\": \"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Craig Ringer wrote:\\n> So personally I think it'd be fine if a cmake build defaulted to\\n> finding and using what it could, but offered a --minimal mode or\\n> whatever that gets us just core postgres + whatever we enable\\n> explicitly. But our current behaviour is OK too.\\n\\nTo me it's best way. But I'm not sure what Tom Lane will accept this.\\n\\n\\n-- \\nYury Zhuravlev\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres 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, 09 Nov 2016 19:58:38 +0300\",\n \"msg_from\": \"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Craig Ringer wrote:\\n> So personally I think it'd be fine if a cmake build defaulted to\\n> finding and using what it could, but offered a --minimal mode or\\n> whatever that gets us just core postgres + whatever we enable\\n> explicitly. But our current behaviour is OK too.\\n\\nTo me it's best way. But I'm not sure what Tom Lane will accept this.\\n\\n\\n-- \\nYury Zhuravlev\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres 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, 10 Nov 2016 22:15:04 +0300\",\n \"msg_from\": \"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 11/11/16 08:15, Yury Zhuravlev wrote:\\n\\n> Craig Ringer wrote:\\n>> So personally I think it'd be fine if a cmake build defaulted to\\n>> finding and using what it could, but offered a --minimal mode or\\n>> whatever that gets us just core postgres + whatever we enable\\n>> explicitly. But our current behaviour is OK too.\\n>\\n> To me it's best way. But I'm not sure what Tom Lane will accept this.\\n>\\n>\\n\\nI just had a play with this patch - nice! (ok so it needs a fix so that \\nthe compile completes as mentioned prev).\\n\\nI would recommend making it behave as Tom suggested. *Then* add an \\n--autodetect or similar option that makes\\nbehave in the 'finding and using what it could' manner as a 2nd patch.\\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\": \"Fri, 11 Nov 2016 08:32:27 +1300\",\n \"msg_from\": \"Mark Kirkwood <mark.kirkwood@catalyst.net.nz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Mark Kirkwood <mark.kirkwood@catalyst.net.nz> writes:\\n> I would recommend making it behave as Tom suggested. *Then* add an \\n> --autodetect or similar option that makes\\n> behave in the 'finding and using what it could' manner as a 2nd patch.\\n\\nAn \\\"--autodetect\\\" switch would be fine with me. I just don't want it\\nto behave that way by default, because then you get into the unexpected\\nresults problem.\\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, 10 Nov 2016 14:40:08 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"I had a bit a play around to see if I could get this building properly \\nagain with v10 (i.e master). I've attached a minimal *incremental* patch \\nthat needs to be applied after v1. This gets everything under the src \\ntree building (added the new file_utils.c and reordered some link libs \\nand removed some duplicates).\\n\\n\\nI haven't made any changes with respect to it trying to detect and build \\neverything. One added nit I see is that unless I'm missing something \\nthere appears to be no way to stop it trying to build all the \\ncontribs...so an option to enable/disable their build is needed.\\n\\n\\nTo make it display what options there are I use:\\n\\n$ mkdir build; cd build ; cmake .. -LH\\n\\n\\nAnd to do a build that works:\\n\\n$ cmake .. -DCMAKE_INSTALL_PREFIX=/usr/local/pgsql/10 -DWITH_PERL=OFF \\n-DWITH_OPENSSL=OFF -DWITH_PYTHON=OFF\\n\\n\\nWith reference to Tom's comments, I'm thinking that these should all \\ndefault to 'OFF' plus an additional variable WITH_CONTRIB (or similar) \\nshould default to OFF too.\\n\\n\\nregards\\n\\n\\nMark\\n\\n\\n\\nOn 09/11/16 08:37, Peter Eisentraut wrote:\\n> On 9/17/16 1:21 PM, Yury Zhuravlev wrote:\\n>> Now, I published the first version of the patch.\\n> I tried this out. Because of some file moves in initdb and\\n> pg_basebackup, the build fails:\\n>\\n> [ 74%] Linking C executable initdb\\n> Undefined symbols for architecture x86_64:\\n> \\\"_fsync_pgdata\\\", referenced from:\\n> _main in initdb.c.o\\n> ld: symbol(s) not found for architecture x86_64\\n> collect2: error: ld returned 1 exit status\\n> make[2]: *** [src/bin/initdb/CMakeFiles/initdb.dir/build.make:177:\\n> src/bin/initdb/initdb] Error 1\\n> make[1]: *** [CMakeFiles/Makefile2:2893:\\n> src/bin/initdb/CMakeFiles/initdb.dir/all] Error 2\\n> make: *** [Makefile:128: all] Error 2\\n>\\n> Please submit an updated patch.\\n>\\n> I suggest you use git format-patch to produce patches. This is easier\\n> to apply, especially when there are a lot of new files involved. Also\\n> use the git facilities to check for whitespace errors.\\n>\\n> Please supply some documentation, such as\\n>\\n> - what are the basic commands\\n> - how to set include/library paths, choose configure options\\n> - how to set CFLAGS\\n> - how to see raw build commands\\n> - what are the targets for all/world/check/docs etc.\\n> - explain directory structure\\n>\\n> I suggest for now you could put this into a README.cmake file in your\\n> patch. We don't need to commit it that way, but it would help in the\\n> meantime.\\n>\\n> When I run cmake without options, it seems to do opportunistic feature\\n> checking. For example, it turns on OpenSSL or Python support if it can\\n> find it, otherwise it turns it off. We need this to be deterministic.\\n> Without options, choose the basic feature set, require all other\\n> features to be turned on explicitly, fail if they can't be found.\\n> Whatever the Autoconf-based build does now has been fairly deliberately\\n> tuned, so there should be very little reason to deviate from that.\\n>\\n> The Python check appears to be picking up pieces from two different\\n> Python installations:\\n>\\n> -- Found PythonInterp: /usr/local/bin/python (found version \\\"2.7.12\\\")\\n> -- Found PythonLibs: /usr/lib/libpython2.7.dylib (found version \\\"2.7.10\\\")\\n>\\n> The check results otherwise look OK, but I'm a bit confused about the\\n> order. It checks for some functions before all the header files are\\n> checked for. Is that intentional?\\n>\\n> There are a number of changes in .[ch] and .pl files that are unclear\\n> and not explained. Please explain them. You can also submit separate\\n> preliminary patches if you need to do some refactoring. Ultimately, I\\n> would expect this patch not to require C code changes.\\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\": \"Wed, 16 Nov 2016 14:43:32 +1300\",\n \"msg_from\": \"Mark Kirkwood <mark.kirkwood@catalyst.net.nz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Yeah, there seems to be a lot of these. Looking through them almost all \\nconcern the addition of piece of code to wrap putenv. e.g:\\n\\n--- a/src/bin/psql/command.c\\n+++ b/src/bin/psql/command.c\\n@@ -1350,7 +1350,7 @@ exec_command(const char *cmd,\\n char *newval;\\n\\n newval = psprintf(\\\"%s=%s\\\", envvar, envval);\\n- putenv(newval);\\n+ pg_putenv_proxy(newval);\\n success = true;\\n\\n /*\\n\\nWhere pg_putenv_proxy either calls putenv or pgwin32_putenv (the latter \\non windows I'd guess). I wonder if this could have been avoided, since \\nthe original code handles this sort of thing. There are also some minor \\n- and not immediately obvious - changes to a number of macros in various \\nincludes...If I'm feeling keen I'll experiment to see how far I can get \\nwithout any source changes at all.\\n\\n\\nregards\\n\\n\\nMark\\n\\n\\nOn 09/11/16 08:37, Peter Eisentraut wrote:\\n>\\n> There are a number of changes in .[ch] and .pl files that are unclear\\n> and not explained. Please explain them. You can also submit separate\\n> preliminary patches if you need to do some refactoring. Ultimately, I\\n> would expect this patch not to require C code changes.\\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\": \"Wed, 16 Nov 2016 16:00:48 +1300\",\n \"msg_from\": \"Mark Kirkwood <mark.kirkwood@catalyst.net.nz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Actually, it was not that tricky to separate out the cmake only changes, \\nand test this on unmodified sources. It appears to work fine for me - \\npasses 'make check' (needs the v1_1 incremental patch applied of \\ncourse). The Patch is attached. I wonder if the original had some \\nchanges for building under latest Windows...(I'm using Ubuntu 16.10, \\nwith cmake 3.5).\\n\\n\\nregards\\n\\n\\nMark\\n\\n\\nOn 16/11/16 16:00, Mark Kirkwood wrote:\\n> Yeah, there seems to be a lot of these. Looking through them almost \\n> all concern the addition of piece of code to wrap putenv. e.g:\\n>\\n> --- a/src/bin/psql/command.c\\n> +++ b/src/bin/psql/command.c\\n> @@ -1350,7 +1350,7 @@ exec_command(const char *cmd,\\n> char *newval;\\n>\\n> newval = psprintf(\\\"%s=%s\\\", envvar, envval);\\n> - putenv(newval);\\n> + pg_putenv_proxy(newval);\\n> success = true;\\n>\\n> /*\\n>\\n> Where pg_putenv_proxy either calls putenv or pgwin32_putenv (the \\n> latter on windows I'd guess). I wonder if this could have been \\n> avoided, since the original code handles this sort of thing. There are \\n> also some minor - and not immediately obvious - changes to a number of \\n> macros in various includes...If I'm feeling keen I'll experiment to \\n> see how far I can get without any source changes at all.\\n>\\n>\\n> regards\\n>\\n>\\n> Mark\\n>\\n>\\n> On 09/11/16 08:37, Peter Eisentraut wrote:\\n>>\\n>> There are a number of changes in .[ch] and .pl files that are unclear\\n>> and not explained. Please explain them. You can also submit separate\\n>> preliminary patches if you need to do some refactoring. Ultimately, I\\n>> would expect this patch not to require C code changes.\\n>>\\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\": \"Wed, 16 Nov 2016 16:45:24 +1300\",\n \"msg_from\": \"Mark Kirkwood <mark.kirkwood@catalyst.net.nz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Mark Kirkwood wrote:\\n> Actually, it was not that tricky to separate out the cmake only \\n> changes, and test this on unmodified sources. It appears to work \\n> fine for me - passes 'make check' (needs the v1_1 incremental \\n> patch applied of course). The Patch is attached. I wonder if the \\n> original had some changes for building under latest \\n> Windows...(I'm using Ubuntu 16.10, with cmake 3.5).\\n>\\nThanks for all your works! Can you make push request here: \\nhttps://github.com/stalkerg/postgres_cmake\\n\\nI have rebased (merge) to master and make other important fix.\\n\\n-- \\nYury Zhuravlev\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres 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, 16 Nov 2016 11:05:55 +0300\",\n \"msg_from\": \"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Mark Kirkwood wrote:\\n> Yeah, there seems to be a lot of these. Looking through them \\n> almost all concern the addition of piece of code to wrap putenv. \\n> e.g:\\nI made this small wrapper special for MSVC 2015 without Service Packs \\nbecause postgres macross were in conflict with MS internal functions. \\nAfter some time and some updates for MSVC Michael Paquier could not \\nreproduce my problem but I keep this patch to avoid problems in the \\nfuture. \\nI can check old behavior again and revert all changes if needed and \\nofcourse I have plans to make separate patch for this changes. \\n\\nThanks!\\n-- \\nYury Zhuravlev\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres 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, 16 Nov 2016 11:22:43 +0300\",\n \"msg_from\": \"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 8, 2016 at 9:12 PM, Michael Paquier\\n<michael.paquier@gmail.com> wrote:\\n> On Wed, Nov 9, 2016 at 7:54 AM, Craig Ringer\\n> <craig.ringer@2ndquadrant.com> wrote:\\n>> On 9 Nov. 2016 06:37, \\\"Yury Zhuravlev\\\" <u.zhuravlev@postgrespro.ru> wrote:\\n>>> This approach I see only in Postgres project and not fully understood.\\n>>> Can you explain me more what reasons led to this approach?\\n>>\\n>> It's predictable. The default has the same result for everyone. I quite like\\n>> it myself.\\n>\\n> +1. Let's tell to the system what we want him to do and not let him\\n> guess what we'd like to be done or it will get harder to test and\\n> develop code for all kind of code paths with #ifdef's. That's one step\\n> away from Skynet.\\n\\nExaggerate much?\\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, 16 Nov 2016 10:40:35 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"I see there are some patches for the putenv issue with Visual studio \\n2013 in progress on this list - it is probably best to work with the \\nauthor to see if 2015 has any new issues and keep all changes for that \\n*out* of the cmake patches.\\n\\n\\nregards\\n\\n\\nMark\\n\\n\\nOn 16/11/16 21:22, Yury Zhuravlev wrote:\\n> I made this small wrapper special for MSVC 2015 without Service Packs \\n> because postgres macross were in conflict with MS internal functions. \\n> After some time and some updates for MSVC Michael Paquier could not \\n> reproduce my problem but I keep this patch to avoid problems in the \\n> future. I can check old behavior again and revert all changes if \\n> needed and ofcourse I have plans to make separate patch for this changes.\\n> Thanks!\\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, 17 Nov 2016 11:14:24 +1300\",\n \"msg_from\": \"Mark Kirkwood <mark.kirkwood@catalyst.net.nz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 16, 2016 at 2:14 PM, Mark Kirkwood\\n<mark.kirkwood@catalyst.net.nz> wrote:\\n> I see there are some patches for the putenv issue with Visual studio 2013 in\\n> progress on this list - it is probably best to work with the author to see\\n> if 2015 has any new issues and keep all changes for that *out* of the cmake\\n> patches.\\n\\nI don't recall all the details here, but no wrappers should be needed,\\nparticularly on Windows where we already do that:\\nsrc/include/port/win32.h:#define putenv(x) pgwin32_putenv(x)\\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, 16 Nov 2016 14:22:30 -0800\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Michael Paquier wrote:\\n> src/include/port/win32.h:#define putenv(x) pgwin32_putenv(x)\\nand my MSVC2015 drop down here because pgwin32_putenv has wrong signature.\\nI hope it is not true now. \\n\\n-- \\nYury Zhuravlev\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres 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, 17 Nov 2016 12:35:51 +0300\",\n \"msg_from\": \"Yury Zhuravlev <u.zhuravlev@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Hello hackers.\\n\\n\\n----\\n\\nI think you notice what I changed email, it is happened because I quit from\\nPgPro.\\n\\nGood news is that I have time to work on CMake. Bad news I have to find a\\nnew job. ;)\\n\\n----\\n\\nWhat is new in that patchset:\\n\\n 1.\\n\\n I separated changes to several patches. Main patch include only new\\n CMake files without postgres files changes.\\n 2.\\n\\n I cleaned all the unnecessary postgres changes.\\n 3.\\n\\n I update cmake with latest master, add STRONG_RANDOM option and etc.\\n 4.\\n\\n Add small document about CMake and Postgres. (README.cmake bad idea for\\n name but I think now it’s ok)\\n 5.\\n\\n I have stopped use gendef.pl for MSVC build.\\n 6.\\n\\n Fix for AIX 7.1\\n\\n\\nAbout patches:\\n\\ncmake_v2_1_main_files.patch\\n\\nOnly CMake files. CMakeFiles.txt is main files with rules. *.cmake is\\nspecial modules with macros and functions. *.in is template for generate\\nheaders. (like autoconf)\\n\\ncmake_v2_2_c_define.patch\\n\\nSmall chages in c.h . At first it is “#pragma fenv_access (off)” it is\\nnecessary if we use /fp:strict for MSVC compiler. Without this pragma we\\ncan’t calc floats for const variables in compiller time (2 * M_PI for\\nexample). Strict mode important if we want to be close with ieee754 float\\nformat on MSVC (1.0 / 0.0 = inf for example). Detail info here:\\nhttps://msdn.microsoft.com/en-us/library/e7s85ffb.aspx\\n\\nSecond change is because I find and set HAVE_INT128 directly from CMake.\\nPG_INT128_TYPE used only for autoconfig scripts.\\n\\ncmake_v2_3_rijndael.patch\\n\\nFirst I added special wraparound because here CMake have circular\\ndependency (cmake very smart here). Second I removed rijndael.tbl because\\nit generated during build process every time.\\n\\ncmake_v2_4_uuid.patch\\n\\nAnother small patch. Right place for uuid.h I find by CMake and not\\nnecessary this ifdef hell.\\n\\ncmake_v2_5_readme.patch\\n\\nSmall exercise to CMake world for all who want to try.\\n\\nQuestions for discussion:\\n\\nIn generated project by CMake we always have only one enter point. Also\\nINSTALL macross support only including to “all” targets. It follows that it\\nis impossible build contrib modules separately only with “all” target. Here\\nwrite about this behavior:\\nhttps://cmake.org/cmake/help/v3.7/prop_tgt/EXCLUDE_FROM_ALL.html\\n\\nInteresting note:\\n\\nMany folks here love only command line but many novices developers from\\nmodern generation love GUI. Also using command line on Windows it's really\\nbig pain. With CMake you can work with Postgres under MSVC more\\ncomfortable. For example my screenshots of regress tests result runing from\\nMSVC: https://twitter.com/stalkerg/status/814423972263657472\\n\\nFin:\\n\\nI hope this patchset one step closer to merge. In future I see a lot of\\nwork but I suppose it is right and important direction. I'm ready for this\\nlong journey.\\n\\nBig thanks Peter Eisentraut for advices and the right vector of\\ndevelopment.\\n\\nAlso big thanks Mark Kirkwood for help and discussion.\\n\\nHappy New Year, everyone!\\n\\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\": \"Fri, 30 Dec 2016 17:10:45 +0300\",\n \"msg_from\": \"Yuriy Zhuravlev <stalkerg@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 12/30/16 9:10 AM, Yuriy Zhuravlev wrote:\\n> cmake_v2_2_c_define.patch\\n> \\n> Small chages in c.h . At first it is �#pragma fenv_access (off)� it is\\n> necessary if we use /fp:strict for MSVC compiler. Without this pragma we\\n> can�t calc floats for const variables in compiller time (2 * M_PI for\\n> example). Strict mode important if we want to be close with ieee754\\n> float format on MSVC (1.0 / 0.0 = inf for example). Detail info here:\\n> https://msdn.microsoft.com/en-us/library/e7s85ffb.aspx\\n\\nI don't understand what this has to do with cmake. If this is a\\nworthwhile improvement for the Windows build, then please explain why,\\nwith a \\\"before\\\" and \\\"after\\\" output and a patch for the existing build\\nsystem as well.\\n\\n> Second change is because I find and set HAVE_INT128 directly from CMake.\\n> PG_INT128_TYPE used only for autoconfig scripts.\\n\\nIt might also be worth refactoring the existing Autoconf code here to\\nmake this consistent.\\n\\n(My assumption is that if we were to move forward with cmake or any\\nother build system change, we would have to keep the old one alongside\\nat least for a little while. So any changes to the C code would need to\\nbe side-ported.)\\n\\n> cmake_v2_3_rijndael.patch\\n> \\n> First I added special wraparound because here CMake have circular\\n> dependency (cmake very smart here). Second I removed rijndael.tbl\\n> because it generated during build process every time.\\n\\nPlease explain what the circular dependency is. If there is one, we\\nshould also side-port this change.\\n\\n> cmake_v2_4_uuid.patch\\n> \\n> Another small patch. Right place for uuid.h I find by CMake and not\\n> necessary this ifdef hell.\\n\\nThis patch removes the uuid.h include but doesn't add it anywhere else.\\nHow does it work?\\n\\n> Questions for discussion:\\n> \\n> In generated project by CMake we always have only one enter point. Also\\n> INSTALL macross support only including to �all� targets. It follows that\\n> it is impossible build contrib modules separately only with �all�\\n> target. Here write about this behavior:\\n> https://cmake.org/cmake/help/v3.7/prop_tgt/EXCLUDE_FROM_ALL.html\\n\\nYeah, I think this is how the MSVC stuff effectively works right now as\\nwell.\\n\\n-- \\nPeter Eisentraut 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\": \"Tue, 3 Jan 2017 09:11:45 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 3, 2017 at 11:11 PM, Peter Eisentraut\\n<peter.eisentraut@2ndquadrant.com> wrote:\\n> On 12/30/16 9:10 AM, Yuriy Zhuravlev wrote:\\n>> cmake_v2_2_c_define.patch\\n>>\\n>> Small chages in c.h . At first it is “#pragma fenv_access (off)” it is\\n>> necessary if we use /fp:strict for MSVC compiler. Without this pragma we\\n>> can’t calc floats for const variables in compiller time (2 * M_PI for\\n>> example). Strict mode important if we want to be close with ieee754\\n>> float format on MSVC (1.0 / 0.0 = inf for example). Detail info here:\\n>> https://msdn.microsoft.com/en-us/library/e7s85ffb.aspx\\n>\\n> I don't understand what this has to do with cmake. If this is a\\n> worthwhile improvement for the Windows build, then please explain why,\\n> with a \\\"before\\\" and \\\"after\\\" output and a patch for the existing build\\n> system as well.\\n>\\n>> Second change is because I find and set HAVE_INT128 directly from CMake.\\n>> PG_INT128_TYPE used only for autoconfig scripts.\\n>\\n> It might also be worth refactoring the existing Autoconf code here to\\n> make this consistent.\\n>\\n> (My assumption is that if we were to move forward with cmake or any\\n> other build system change, we would have to keep the old one alongside\\n> at least for a little while. So any changes to the C code would need to\\n> be side-ported.)\\n>\\n>> cmake_v2_3_rijndael.patch\\n>>\\n>> First I added special wraparound because here CMake have circular\\n>> dependency (cmake very smart here). Second I removed rijndael.tbl\\n>> because it generated during build process every time.\\n>\\n> Please explain what the circular dependency is. If there is one, we\\n> should also side-port this change.\\n>\\n>> cmake_v2_4_uuid.patch\\n>>\\n>> Another small patch. Right place for uuid.h I find by CMake and not\\n>> necessary this ifdef hell.\\n>\\n> This patch removes the uuid.h include but doesn't add it anywhere else.\\n> How does it work?\\n>\\n>> Questions for discussion:\\n>>\\n>> In generated project by CMake we always have only one enter point. Also\\n>> INSTALL macross support only including to “all” targets. It follows that\\n>> it is impossible build contrib modules separately only with “all”\\n>> target. Here write about this behavior:\\n>> https://cmake.org/cmake/help/v3.7/prop_tgt/EXCLUDE_FROM_ALL.html\\n>\\n> Yeah, I think this is how the MSVC stuff effectively works right now as\\n> well.\\n\\nI am marking this patch as \\\"returned with feedback\\\". That's quite a\\nheavy change and it looks to be too late in the development cycle of\\nPG10 to consider it. Peter's commit bits, who is also the reviewer,\\nare beginning to smoke as well after everything that has happened for\\nthe logical replication changes.\\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\": \"Tue, 24 Jan 2017 15:50:47 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 2017-01-24 15:50:47 +0900, Michael Paquier wrote:\\n> I am marking this patch as \\\"returned with feedback\\\". That's quite a\\n> heavy change and it looks to be too late in the development cycle of\\n> PG10 to consider it. Peter's commit bits, who is also the reviewer,\\n> are beginning to smoke as well after everything that has happened for\\n> the logical replication changes.\\n\\nI'm doubtful about this being ready in time too, but it seems a might\\nheavyhanded to make that call on your own. Including the judgement about\\nPeter's capability to handle more.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n-- \\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, 23 Jan 2017 22:58:48 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 24, 2017 at 3:58 PM, Andres Freund <andres@anarazel.de> wrote:\\n> On 2017-01-24 15:50:47 +0900, Michael Paquier wrote:\\n>> I am marking this patch as \\\"returned with feedback\\\". That's quite a\\n>> heavy change and it looks to be too late in the development cycle of\\n>> PG10 to consider it. Peter's commit bits, who is also the reviewer,\\n>> are beginning to smoke as well after everything that has happened for\\n>> the logical replication changes.\\n>\\n> I'm doubtful about this being ready in time too, but it seems a might\\n> heavyhanded to make that call on your own. Including the judgement about\\n> Peter's capability to handle more.\\n\\nSure, sorry for that. I am switching back the patch, let's see what\\nhappens by the end of the CF.\\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\": \"Tue, 24 Jan 2017 16:00:16 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 24 Jan. 2017 18:00, \\\"Andres Freund\\\" <andres@anarazel.de> wrote:\\n\\nOn 2017-01-24 15:50:47 +0900, Michael Paquier wrote:\\n> I am marking this patch as \\\"returned with feedback\\\". That's quite a\\n> heavy change and it looks to be too late in the development cycle of\\n> PG10 to consider it. Peter's commit bits, who is also the reviewer,\\n> are beginning to smoke as well after everything that has happened for\\n> the logical replication changes.\\n\\nI'm doubtful about this being ready in time too, but it seems a might\\nheavyhanded to make that call on your own. Including the judgement about\\nPeter's capability to handle more.\\n\\n\\nPersonally I think we should aim to have this in as a non default build\\nmode in pg10 if it can be made ready, and aim to make it default in pg11 at\\nleast for Windows.\\n\\nOn 24 Jan. 2017 18:00, \\\"Andres Freund\\\" <andres@anarazel.de> wrote:On 2017-01-24 15:50:47 +0900, Michael Paquier wrote:\\n> I am marking this patch as \\\"returned with feedback\\\". That's quite a\\n> heavy change and it looks to be too late in the development cycle of\\n> PG10 to consider it. Peter's commit bits, who is also the reviewer,\\n> are beginning to smoke as well after everything that has happened for\\n> the logical replication changes.\\n\\nI'm doubtful about this being ready in time too, but it seems a might\\nheavyhanded to make that call on your own. Including the judgement about\\nPeter's capability to handle more.\\nPersonally I think we should aim to have this in as a non default build mode in pg10 if it can be made ready, and aim to make it default in pg11 at least for Windows.\",\n \"msg_date\": \"Tue, 24 Jan 2017 16:27:58 +0800\",\n \"msg_from\": \"Craig Ringer <craig.ringer@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Craig Ringer <craig.ringer@2ndquadrant.com> writes:\\n> Personally I think we should aim to have this in as a non default build\\n> mode in pg10 if it can be made ready, and aim to make it default in pg11 at\\n> least for Windows.\\n\\nAFAIK we haven't committed to accepting this at all, let alone trying\\nto do so on a tight schedule. And I believe there was general agreement\\nthat we would not accept it as something to maintain in parallel with\\nthe existing makefiles. If we have to maintain two build systems, we\\nhave that already.\\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, 24 Jan 2017 08:37:47 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 1/24/17 8:37 AM, Tom Lane wrote:\\n> Craig Ringer <craig.ringer@2ndquadrant.com> writes:\\n>> Personally I think we should aim to have this in as a non default build\\n>> mode in pg10 if it can be made ready, and aim to make it default in pg11 at\\n>> least for Windows.\\n> \\n> AFAIK we haven't committed to accepting this at all, let alone trying\\n> to do so on a tight schedule. And I believe there was general agreement\\n> that we would not accept it as something to maintain in parallel with\\n> the existing makefiles. If we have to maintain two build systems, we\\n> have that already.\\n\\nMy preferred scenario would be to replace the Windows build system by\\nthis first, then refine it, then get rid of Autoconf.\\n\\nThe ideal timeline would be to have a ready patch to commit early in a\\ndevelopment cycle, then get rid of the Windows build system by the end\\nof it. Naturally, this would need buy-in from Windows developers.\\n\\nI don't foresee replacing the Autoconf build system by this immediately.\\n\\nRight now, however, the patch isn't moving at all, and I don't see it\\ngoing into PG10, so I'm fine with returning with feedback.\\n\\n-- \\nPeter Eisentraut 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\": \"Fri, 27 Jan 2017 09:09:36 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut wrote:\\n\\n> Right now, however, the patch isn't moving at all, and I don't see it\\n> going into PG10, so I'm fine with returning with feedback.\\n\\nThere are a bunch of side patches that we should apply separately, such\\nas the pgcrypto fix. I don't understand why they are part of this patch\\nseries, really, given that they seem largely independent prerequisites.\\n\\n-- \\n�lvaro Herrera https://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\": \"Fri, 27 Jan 2017 11:40:55 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 27, 2017 at 11:09 PM, Peter Eisentraut\\n<peter.eisentraut@2ndquadrant.com> wrote:\\n> On 1/24/17 8:37 AM, Tom Lane wrote:\\n>> Craig Ringer <craig.ringer@2ndquadrant.com> writes:\\n>>> Personally I think we should aim to have this in as a non default build\\n>>> mode in pg10 if it can be made ready, and aim to make it default in pg11 at\\n>>> least for Windows.\\n>>\\n>> AFAIK we haven't committed to accepting this at all, let alone trying\\n>> to do so on a tight schedule. And I believe there was general agreement\\n>> that we would not accept it as something to maintain in parallel with\\n>> the existing makefiles. If we have to maintain two build systems, we\\n>> have that already.\\n>\\n> My preferred scenario would be to replace the Windows build system by\\n> this first, then refine it, then get rid of Autoconf.\\n>\\n> The ideal timeline would be to have a ready patch to commit early in a\\n> development cycle, then get rid of the Windows build system by the end\\n> of it. Naturally, this would need buy-in from Windows developers.\\n\\nThis looks like a really good plan to me.\\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\": \"Sat, 28 Jan 2017 07:50:31 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 2017-01-27 09:09:36 -0500, Peter Eisentraut wrote:\\n> My preferred scenario would be to replace the Windows build system by\\n> this first, then refine it, then get rid of Autoconf.\\n> \\n> The ideal timeline would be to have a ready patch to commit early in a\\n> development cycle, then get rid of the Windows build system by the end\\n> of it. Naturally, this would need buy-in from Windows developers.\\n> \\n> I don't foresee replacing the Autoconf build system by this immediately.\\n\\nI'm very strongly against this path, it seems way too likely that we'll\\nend up with yet another fragile thing that nobody from the *nix side\\nwill be able to test.\\n\\n- Andres\\n\\n\\n-- \\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, 27 Jan 2017 15:11:20 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 1/27/17 6:11 PM, Andres Freund wrote:\\n> On 2017-01-27 09:09:36 -0500, Peter Eisentraut wrote:\\n>> My preferred scenario would be to replace the Windows build system by\\n>> this first, then refine it, then get rid of Autoconf.\\n>>\\n>> The ideal timeline would be to have a ready patch to commit early in a\\n>> development cycle, then get rid of the Windows build system by the end\\n>> of it. Naturally, this would need buy-in from Windows developers.\\n>>\\n>> I don't foresee replacing the Autoconf build system by this immediately.\\n> \\n> I'm very strongly against this path, it seems way too likely that we'll\\n> end up with yet another fragile thing that nobody from the *nix side\\n> will be able to test.\\n\\nThat's a fair concern, but at least with CMake, someone from the *nix\\nside *can* test it, whereas right now it's completely separate.\\n\\nWhat kind of strategy do you have in mind?\\n\\n-- \\nPeter Eisentraut 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\": \"Fri, 27 Jan 2017 22:20:41 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Sat, Jan 28, 2017 at 12:20 PM, Peter Eisentraut\\n<peter.eisentraut@2ndquadrant.com> wrote:\\n> On 1/27/17 6:11 PM, Andres Freund wrote:\\n>> On 2017-01-27 09:09:36 -0500, Peter Eisentraut wrote:\\n>>> My preferred scenario would be to replace the Windows build system by\\n>>> this first, then refine it, then get rid of Autoconf.\\n>>>\\n>>> The ideal timeline would be to have a ready patch to commit early in a\\n>>> development cycle, then get rid of the Windows build system by the end\\n>>> of it. Naturally, this would need buy-in from Windows developers.\\n>>>\\n>>> I don't foresee replacing the Autoconf build system by this immediately.\\n>>\\n>> I'm very strongly against this path, it seems way too likely that we'll\\n>> end up with yet another fragile thing that nobody from the *nix side\\n>> will be able to test.\\n>\\n> That's a fair concern, but at least with CMake, someone from the *nix\\n> side *can* test it, whereas right now it's completely separate.\\n\\nAnd people complain all the time that the MSVC build scripts are hacky\\nand complicated.. So by beginning from there we switch from one build\\nto the other, not increasing the number of builds that need to be\\nmaintained. Based on that Peter's strategy looks appealing to me. By\\nthe way, I am marking the patch as returned with feedback for this CF.\\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, 30 Jan 2017 15:22:16 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 2017-01-27 22:20:41 -0500, Peter Eisentraut wrote:\\n> On 1/27/17 6:11 PM, Andres Freund wrote:\\n> > On 2017-01-27 09:09:36 -0500, Peter Eisentraut wrote:\\n> >> My preferred scenario would be to replace the Windows build system by\\n> >> this first, then refine it, then get rid of Autoconf.\\n> >>\\n> >> The ideal timeline would be to have a ready patch to commit early in a\\n> >> development cycle, then get rid of the Windows build system by the end\\n> >> of it. Naturally, this would need buy-in from Windows developers.\\n> >>\\n> >> I don't foresee replacing the Autoconf build system by this immediately.\\n> > \\n> > I'm very strongly against this path, it seems way too likely that we'll\\n> > end up with yet another fragile thing that nobody from the *nix side\\n> > will be able to test.\\n> \\n> That's a fair concern, but at least with CMake, someone from the *nix\\n> side *can* test it, whereas right now it's completely separate.\\n\\nGiven that fact, I just don't buy why it's a good idea to not also\\nreplace autoconf initially. Either we're able to properly test it -\\ni.e. it runs all tests - on *nix or we're not. There's not a a whole of\\neffort between those if you also want to do the windows side of things\\nproperly.\\n\\n\\n> What kind of strategy do you have in mind?\\n\\nDo all of it. I'm unconvinced that a windows only version buys us\\nmeaningful savings, and I think the dangers of adding more duplication\\n(msvc stuff after all gets some information from the makefiles) and\\nlong-term coexistence are quite severe.\\n\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n-- \\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, 29 Jan 2017 22:28:55 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 1/30/17 1:28 AM, Andres Freund wrote:\\n> Given that fact, I just don't buy why it's a good idea to not also\\n> replace autoconf initially.\\n\\nWell, I find it a bit scary. If you do the big switch all at once, then\\nyou will have to dedicate the following 3 months to fixing complaints\\nfrom developers and build farmers.\\n\\n> Either we're able to properly test it - i.e. it runs all tests - on *nix or we're not.\\n\\nThat would work if there were a single entry point into the build\\nsystem. But in practice there are many, and every one of them is\\nsomeone's favorite. It's unlikely that we will be able to enumerate all\\nof them during patch review.\\n\\n-- \\nPeter Eisentraut 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\": \"Mon, 30 Jan 2017 10:26:18 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \">\\n> I don't understand what this has to do with cmake. If this is a\\n> worthwhile improvement for the Windows build, then please explain why,\\n> with a \\\"before\\\" and \\\"after\\\" output and a patch for the existing build\\n> system as well.\\n\\nDuring the porting process, I meet such situations when I should fix\\nsomething. It's happening because I build with different way also current\\nbuild system is trying to avoid many sharp corners.\\nIf talk about this situation - without strict mode many \\\"floats\\\" checks\\ndon't work correctly. You can read the link above. Besides this option puts\\nby build system. I think we can make a new thread for this approach. (with\\npatch for current perl system)\\n\\nIt might also be worth refactoring the existing Autoconf code here to\\n> make this consistent.\\n\\n\\nI do it because it's convenient in CMake. I can change this it's not big\\ndeal.\\n\\nPlease explain what the circular dependency is. If there is one, we\\n> should also side-port this change.\\n\\n\\nIt's an important part. I have a rule for generate rijndael.tbl by\\ngen-rtab who make from rijndael.c (with special define) who include\\nrijndael.tbl .\\nIf I generate rijndael.tbl it's to force build gen-rtab and generate\\nrijndael.tbl again.\\nCMake knows about \\\"includes\\\" in files but we can make the wraparound macro\\nto hide include.\\n\\nThis patch removes the uuid.h include but doesn't add it anywhere else.\\n\\nHow does it work?\\n\\n\\nCMake sends to compiler right place for uuid.h (I mean -I/usr/include and\\netc for gcc).\\n\\n\\n> Yeah, I think this is how the MSVC stuff effectively works right now as\\n> well.\\n\\n\\nI glad to hear it.\\n\\n2017-01-03 17:11 GMT+03:00 Peter Eisentraut <\\npeter.eisentraut@2ndquadrant.com>:\\n\\n> On 12/30/16 9:10 AM, Yuriy Zhuravlev wrote:\\n> > cmake_v2_2_c_define.patch\\n> >\\n> > Small chages in c.h . At first it is “#pragma fenv_access (off)” it is\\n> > necessary if we use /fp:strict for MSVC compiler. Without this pragma we\\n> > can’t calc floats for const variables in compiller time (2 * M_PI for\\n> > example). Strict mode important if we want to be close with ieee754\\n> > float format on MSVC (1.0 / 0.0 = inf for example). Detail info here:\\n> > https://msdn.microsoft.com/en-us/library/e7s85ffb.aspx\\n>\\n> I don't understand what this has to do with cmake. If this is a\\n> worthwhile improvement for the Windows build, then please explain why,\\n> with a \\\"before\\\" and \\\"after\\\" output and a patch for the existing build\\n> system as well.\\n>\\n> > Second change is because I find and set HAVE_INT128 directly from CMake.\\n> > PG_INT128_TYPE used only for autoconfig scripts.\\n>\\n> It might also be worth refactoring the existing Autoconf code here to\\n> make this consistent.\\n>\\n> (My assumption is that if we were to move forward with cmake or any\\n> other build system change, we would have to keep the old one alongside\\n> at least for a little while. So any changes to the C code would need to\\n> be side-ported.)\\n>\\n> > cmake_v2_3_rijndael.patch\\n> >\\n> > First I added special wraparound because here CMake have circular\\n> > dependency (cmake very smart here). Second I removed rijndael.tbl\\n> > because it generated during build process every time.\\n>\\n> Please explain what the circular dependency is. If there is one, we\\n> should also side-port this change.\\n>\\n> > cmake_v2_4_uuid.patch\\n> >\\n> > Another small patch. Right place for uuid.h I find by CMake and not\\n> > necessary this ifdef hell.\\n>\\n> This patch removes the uuid.h include but doesn't add it anywhere else.\\n> How does it work?\\n>\\n> > Questions for discussion:\\n> >\\n> > In generated project by CMake we always have only one enter point. Also\\n> > INSTALL macross support only including to “all” targets. It follows that\\n> > it is impossible build contrib modules separately only with “all”\\n> > target. Here write about this behavior:\\n> > https://cmake.org/cmake/help/v3.7/prop_tgt/EXCLUDE_FROM_ALL.html\\n>\\n> Yeah, I think this is how the MSVC stuff effectively works right now as\\n> well.\\n>\\n> --\\n> Peter Eisentraut http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\nI don't understand what this has to do with cmake. If this is aworthwhile improvement for the Windows build, then please explain why,with a \\\"before\\\" and \\\"after\\\" output and a patch for the existing buildsystem as well.During the porting process, I meet such situations when I should fix something. It's happening because I build with different way also current build system is trying to avoid many sharp corners. If talk about this situation - without strict mode many \\\"floats\\\" checks don't work correctly. You can read the link above. Besides this option puts by build system. I think we can make a new thread for this approach. (with patch for current perl system) It might also be worth refactoring the existing Autoconf code here tomake this consistent.I do it because it's convenient in CMake. I can change this it's not big deal. Please explain what the circular dependency is. If there is one, weshould also side-port this change.It's an important part. I have a rule for generate rijndael.tbl by gen-rtab who make from rijndael.c (with special define) who include rijndael.tbl . If I generate rijndael.tbl it's to force build gen-rtab and generate rijndael.tbl again. CMake knows about \\\"includes\\\" in files but we can make the wraparound macro to hide include. This patch removes the uuid.h include but doesn't add it anywhere else.How does it work? CMake sends to compiler right place for uuid.h (I mean -I/usr/include and etc for gcc). Yeah, I think this is how the MSVC stuff effectively works right now aswell.I glad to hear it. 2017-01-03 17:11 GMT+03:00 Peter Eisentraut <peter.eisentraut@2ndquadrant.com>:On 12/30/16 9:10 AM, Yuriy Zhuravlev wrote:\\n> cmake_v2_2_c_define.patch\\n>\\n> Small chages in c.h . At first it is “#pragma fenv_access (off)” it is\\n> necessary if we use /fp:strict for MSVC compiler. Without this pragma we\\n> can’t calc floats for const variables in compiller time (2 * M_PI for\\n> example). Strict mode important if we want to be close with ieee754\\n> float format on MSVC (1.0 / 0.0 = inf for example). Detail info here:\\n> https://msdn.microsoft.com/en-us/library/e7s85ffb.aspx\\n\\nI don't understand what this has to do with cmake. If this is a\\nworthwhile improvement for the Windows build, then please explain why,\\nwith a \\\"before\\\" and \\\"after\\\" output and a patch for the existing build\\nsystem as well.\\n\\n> Second change is because I find and set HAVE_INT128 directly from CMake.\\n> PG_INT128_TYPE used only for autoconfig scripts.\\n\\nIt might also be worth refactoring the existing Autoconf code here to\\nmake this consistent.\\n\\n(My assumption is that if we were to move forward with cmake or any\\nother build system change, we would have to keep the old one alongside\\nat least for a little while. So any changes to the C code would need to\\nbe side-ported.)\\n\\n> cmake_v2_3_rijndael.patch\\n>\\n> First I added special wraparound because here CMake have circular\\n> dependency (cmake very smart here). Second I removed rijndael.tbl\\n> because it generated during build process every time.\\n\\nPlease explain what the circular dependency is. If there is one, we\\nshould also side-port this change.\\n\\n> cmake_v2_4_uuid.patch\\n>\\n> Another small patch. Right place for uuid.h I find by CMake and not\\n> necessary this ifdef hell.\\n\\nThis patch removes the uuid.h include but doesn't add it anywhere else.\\nHow does it work?\\n\\n> Questions for discussion:\\n>\\n> In generated project by CMake we always have only one enter point. Also\\n> INSTALL macross support only including to “all” targets. It follows that\\n> it is impossible build contrib modules separately only with “all”\\n> target. Here write about this behavior:\\n> https://cmake.org/cmake/help/v3.7/prop_tgt/EXCLUDE_FROM_ALL.html\\n\\nYeah, I think this is how the MSVC stuff effectively works right now as\\nwell.\\n\\n--\\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Wed, 8 Feb 2017 14:04:39 +0300\",\n \"msg_from\": \"Yuriy Zhuravlev <stalkerg@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"2017-01-28 1:50 GMT+03:00 Michael Paquier <michael.paquier@gmail.com>:\\n\\n> On Fri, Jan 27, 2017 at 11:09 PM, Peter Eisentraut\\n> <peter.eisentraut@2ndquadrant.com> wrote:\\n> > On 1/24/17 8:37 AM, Tom Lane wrote:\\n> >> Craig Ringer <craig.ringer@2ndquadrant.com> writes:\\n> >>> Personally I think we should aim to have this in as a non default build\\n> >>> mode in pg10 if it can be made ready, and aim to make it default in\\n> pg11 at\\n> >>> least for Windows.\\n> >>\\n> >> AFAIK we haven't committed to accepting this at all, let alone trying\\n> >> to do so on a tight schedule. And I believe there was general agreement\\n> >> that we would not accept it as something to maintain in parallel with\\n> >> the existing makefiles. If we have to maintain two build systems, we\\n> >> have that already.\\n> >\\n> > My preferred scenario would be to replace the Windows build system by\\n> > this first, then refine it, then get rid of Autoconf.\\n> >\\n> > The ideal timeline would be to have a ready patch to commit early in a\\n> > development cycle, then get rid of the Windows build system by the end\\n> > of it. Naturally, this would need buy-in from Windows developers.\\n>\\n> This looks like a really good plan to me.\\n\\n\\nI think it's best plan because when this patch will be in Postgres guys\\nfrom community can test it for Unix systems too.\\nSupport two build systems it's not big deal really. I have been working on\\nthis past year without any big troubles.\\nAlso we have second perl build system...\\n\\n2017-01-28 1:50 GMT+03:00 Michael Paquier <michael.paquier@gmail.com>:On Fri, Jan 27, 2017 at 11:09 PM, Peter Eisentraut\\n<peter.eisentraut@2ndquadrant.com> wrote:\\n> On 1/24/17 8:37 AM, Tom Lane wrote:\\n>> Craig Ringer <craig.ringer@2ndquadrant.com> writes:\\n>>> Personally I think we should aim to have this in as a non default build\\n>>> mode in pg10 if it can be made ready, and aim to make it default in pg11 at\\n>>> least for Windows.\\n>>\\n>> AFAIK we haven't committed to accepting this at all, let alone trying\\n>> to do so on a tight schedule. And I believe there was general agreement\\n>> that we would not accept it as something to maintain in parallel with\\n>> the existing makefiles. If we have to maintain two build systems, we\\n>> have that already.\\n>\\n> My preferred scenario would be to replace the Windows build system by\\n> this first, then refine it, then get rid of Autoconf.\\n>\\n> The ideal timeline would be to have a ready patch to commit early in a\\n> development cycle, then get rid of the Windows build system by the end\\n> of it. Naturally, this would need buy-in from Windows developers.\\n\\nThis looks like a really good plan to me. I think it's best plan because when this patch will be in Postgres guys from community can test it for Unix systems too. Support two build systems it's not big deal really. I have been working on this past year without any big troubles. Also we have second perl build system...\",\n \"msg_date\": \"Wed, 8 Feb 2017 14:21:56 +0300\",\n \"msg_from\": \"Yuriy Zhuravlev <stalkerg@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 2/8/17 6:21 AM, Yuriy Zhuravlev wrote:\\n> Support two build systems it's not big deal really. I have been working\\n> on this past year without any big troubles. \\n> Also we have second perl build system...\\n\\nThe perl/msvc build system pulls in information from the makefiles. So\\nwhen you add a file or something basic like that, you don't have to\\nupdate it. So it's really more like 1.5 build systems.\\n\\n-- \\nPeter Eisentraut 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, 8 Feb 2017 16:36:42 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2017-01-30 10:26:18 -0500, Peter Eisentraut wrote:\\n> On 1/30/17 1:28 AM, Andres Freund wrote:\\n> > Given that fact, I just don't buy why it's a good idea to not also\\n> > replace autoconf initially.\\n> \\n> Well, I find it a bit scary. If you do the big switch all at once, then\\n> you will have to dedicate the following 3 months to fixing complaints\\n> from developers and build farmers.\\n\\nThat'll be the case just as well if we spread it out over two cycles,\\nexcept that we'll have it in multiple phases, and we run into the danger\\nof a half-done conversion.\\n\\nI'd rather not change systems at all than run into the danger of that.\\n\\n\\n> > Either we're able to properly test it - i.e. it runs all tests - on *nix or we're not.\\n> \\n> That would work if there were a single entry point into the build\\n> system. But in practice there are many, and every one of them is\\n> someone's favorite. It's unlikely that we will be able to enumerate all\\n> of them during patch review.\\n\\nNot sure what you mean with \\\"entry point\\\"?\\n\\n\\n- Andres\\n\\n\\n-- \\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, 8 Feb 2017 13:44:50 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n> On 2/8/17 6:21 AM, Yuriy Zhuravlev wrote:\\n>> Support two build systems it's not big deal really. I have been working\\n>> on this past year without any big troubles. \\n>> Also we have second perl build system...\\n\\n> The perl/msvc build system pulls in information from the makefiles. So\\n> when you add a file or something basic like that, you don't have to\\n> update it. So it's really more like 1.5 build systems.\\n\\nReally it's more like 1.1 build systems, in that the MSVC scripts do that\\njust well enough that you *usually* don't have to think about them. But\\nthen when they fail, and you have to figure out why, it can be a pain.\\n\\nFor my own purposes, the only thing that I find seriously annoying about\\nthe status quo is the amount of time required to run \\\"configure\\\". For\\nme, that step is usually comparable to or even more than the time to\\ndo the build proper, because (a) ccache and (b) multiple CPUs.\\nconfigure isn't parallelizable, and there's probably nothing that\\ncan be done about that. If CMake offers a substantial improvement\\nin configuration time then that would be attractive. Otherwise I'd\\njust as soon see us put the effort into making the MSVC scripts more\\nrobust and able to pull more data from the makefiles.\\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, 08 Feb 2017 16:52:19 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 2/8/17 3:52 PM, Tom Lane wrote:\\n> For my own purposes, the only thing that I find seriously annoying about\\n> the status quo is the amount of time required to run \\\"configure\\\". For\\n> me, that step is usually comparable to or even more than the time to\\n> do the build proper, because (a) ccache and (b) multiple CPUs.\\n> configure isn't parallelizable, and there's probably nothing that\\n> can be done about that. If CMake offers a substantial improvement\\n> in configuration time then that would be attractive. Otherwise I'd\\n> just as soon see us put the effort into making the MSVC scripts more\\n> robust and able to pull more data from the makefiles.\\n\\nFWIW, I've had good luck adding -C to configure to cache the output. I'd \\nguess it's at least 10x faster on my laptop.\\n\\nObviously doesn't help if you're doing where you're testing something \\nthat alters config output. In those cases I'll either edit config.cache \\nand delete the relevant lines or just temporarily move it out of the way \\n(or just nuke it...).\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting, Austin TX\\nExperts in Analytics, Data Architecture and PostgreSQL\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n855-TREBLE2 (855-873-2532)\\n\\n\\n-- \\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, 8 Feb 2017 15:55:34 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2017-02-08 16:52:19 -0500, Tom Lane wrote:\\n> For my own purposes, the only thing that I find seriously annoying about\\n> the status quo is the amount of time required to run \\\"configure\\\". For\\n> me, that step is usually comparable to or even more than the time to\\n> do the build proper, because (a) ccache and (b) multiple CPUs.\\n> configure isn't parallelizable, and there's probably nothing that\\n> can be done about that.\\n\\nI use autoconf caching feature to make that a bit less painful (plus\\nsome scripting about when to scrap the cache file...). I find that\\nseriously annoying too.\\n\\n\\n> If CMake offers a substantial improvement\\n> in configuration time then that would be attractive. Otherwise I'd\\n> just as soon see us put the effort into making the MSVC scripts more\\n> robust and able to pull more data from the makefiles.\\n\\nSome of the build-tooling in cmake is quite nice, I have to admit. I've\\ne.g. grown to like using ninja instead of make to build the resulting\\nfiles. Primarily in projects that take longer than pg to compile - a\\nclean build in llvm with ninja takes like 0.1 seconds.\\n\\nBeing more able to rely on things working on windows when doing them on\\nlinux does seem like an advantage to me - fiddlin with Mkvcbuild.pm is\\nquite annoying.\\n\\n- Andres\\n\\n\\n-- \\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, 8 Feb 2017 13:59:11 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 30, 2017 at 10:26 AM, Peter Eisentraut\\n<peter.eisentraut@2ndquadrant.com> wrote:\\n> On 1/30/17 1:28 AM, Andres Freund wrote:\\n>> Given that fact, I just don't buy why it's a good idea to not also\\n>> replace autoconf initially.\\n>\\n> Well, I find it a bit scary. If you do the big switch all at once, then\\n> you will have to dedicate the following 3 months to fixing complaints\\n> from developers and build farmers.\\n\\nI agree with that. I think replacing the Windows build system first\\nand then the non-Windows build system later is a better plan than\\nreplacing both at the same time.\\n\\nBut also, I'm not really sure whether this conversion makes sense. I\\nmean, any build system is going to require some work, and accordingly\\nour present build systems require some work. cmake will require\\ndifferent work, but not necessarily less. The current system has a\\nlong history; we pretty much know it works. Switching will inevitably\\nbreak some things. Maybe we'll end up better off in the long term,\\nbut maybe we won't. Things are pretty good now, so it seems like it\\nwould be easy for them to get worse rather than better. If nothing\\nelse, everybody who has to learn the new system either to use it for\\ndevelopment or because they are doing packaging will have to do some\\namount of extra work as a result of any switch.\\n\\nI do agree that - in theory - one build system is better than two.\\nBut two well-tested, reliable build systems could easily be better\\nthan one system with a bunch of problems. And the points downthread\\nabout our two existing systems being not entirely separate are on\\npoint, too.\\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, 10 Feb 2017 12:07:15 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 2017-02-10 12:07:15 -0500, Robert Haas wrote:\\n> On Mon, Jan 30, 2017 at 10:26 AM, Peter Eisentraut\\n> <peter.eisentraut@2ndquadrant.com> wrote:\\n> > On 1/30/17 1:28 AM, Andres Freund wrote:\\n> >> Given that fact, I just don't buy why it's a good idea to not also\\n> >> replace autoconf initially.\\n> >\\n> > Well, I find it a bit scary. If you do the big switch all at once, then\\n> > you will have to dedicate the following 3 months to fixing complaints\\n> > from developers and build farmers.\\n> \\n> I agree with that. I think replacing the Windows build system first\\n> and then the non-Windows build system later is a better plan than\\n> replacing both at the same time.\\n\\nObviously I disagree ;)\\n\\n\\nBut I'm more replying because of:\\n\\n> But also, I'm not really sure whether this conversion makes sense. I\\n> mean, any build system is going to require some work, and accordingly\\n> our present build systems require some work. cmake will require\\n> different work, but not necessarily less. The current system has a\\n> long history; we pretty much know it works. Switching will inevitably\\n> break some things. Maybe we'll end up better off in the long term,\\n> but maybe we won't. Things are pretty good now, so it seems like it\\n> would be easy for them to get worse rather than better.\\n\\nI do think it's kinda ok for people who've dealt with this for some\\ntime. But for the first few times having to write autoconf macros is\\nquite intimidating. A somewhat less obscure way to deal with that is\\nworth something. As is better file dependency management, across\\ndifferent environments. As is the IDE integration cmake is more and\\nmore getting. As is properly builtin working caching of configure tests\\n(llvm first cmake run, 7.7s, second 2.54s). As is the fact that a lot of\\nbig projects (llvm, kde, qt, and a lot of others) migrated to it.\\n\\nFor me personally those benefits aren't worth that much. I've learned\\nautoconf stuff. I've scripting around autoconf caching. But for newer\\npeople that's not true.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n-- \\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, 10 Feb 2017 10:22:34 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 10, 2017 at 6:07 PM, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On Mon, Jan 30, 2017 at 10:26 AM, Peter Eisentraut\\n> <peter.eisentraut@2ndquadrant.com> wrote:\\n> > On 1/30/17 1:28 AM, Andres Freund wrote:\\n> >> Given that fact, I just don't buy why it's a good idea to not also\\n> >> replace autoconf initially.\\n> >\\n> > Well, I find it a bit scary. If you do the big switch all at once, then\\n> > you will have to dedicate the following 3 months to fixing complaints\\n> > from developers and build farmers.\\n>\\n> I agree with that. I think replacing the Windows build system first\\n> and then the non-Windows build system later is a better plan than\\n> replacing both at the same time.\\n>\\n> But also, I'm not really sure whether this conversion makes sense. I\\n> mean, any build system is going to require some work, and accordingly\\n> our present build systems require some work. cmake will require\\n> different work, but not necessarily less. The current system has a\\n> long history; we pretty much know it works. Switching will inevitably\\n> break some things. Maybe we'll end up better off in the long term,\\n> but maybe we won't. Things are pretty good now, so it seems like it\\n> would be easy for them to get worse rather than better. If nothing\\n> else, everybody who has to learn the new system either to use it for\\n> development or because they are doing packaging will have to do some\\n> amount of extra work as a result of any switch.\\n>\\n>\\nFor me a killer feature would be if/when we can get to a point where we can\\nhave something pgxs-style on cmake that also works on windows.\\n\\nOur homemade Windows build system works OK for postgres, and while ugly it\\nis as you say well tested by now. But it doesn't do *anything* to help\\npeople build extensions on Windows.\\n\\n-- \\n Magnus Hagander\\n Me: http://www.hagander.net/\\n Work: http://www.redpill-linpro.com/\\n\\nOn Fri, Feb 10, 2017 at 6:07 PM, Robert Haas <robertmhaas@gmail.com> wrote:On Mon, Jan 30, 2017 at 10:26 AM, Peter Eisentraut\\n<peter.eisentraut@2ndquadrant.com> wrote:\\n> On 1/30/17 1:28 AM, Andres Freund wrote:\\n>> Given that fact, I just don't buy why it's a good idea to not also\\n>> replace autoconf initially.\\n>\\n> Well, I find it a bit scary. If you do the big switch all at once, then\\n> you will have to dedicate the following 3 months to fixing complaints\\n> from developers and build farmers.\\n\\nI agree with that. I think replacing the Windows build system first\\nand then the non-Windows build system later is a better plan than\\nreplacing both at the same time.\\n\\nBut also, I'm not really sure whether this conversion makes sense. I\\nmean, any build system is going to require some work, and accordingly\\nour present build systems require some work. cmake will require\\ndifferent work, but not necessarily less. The current system has a\\nlong history; we pretty much know it works. Switching will inevitably\\nbreak some things. Maybe we'll end up better off in the long term,\\nbut maybe we won't. Things are pretty good now, so it seems like it\\nwould be easy for them to get worse rather than better. If nothing\\nelse, everybody who has to learn the new system either to use it for\\ndevelopment or because they are doing packaging will have to do some\\namount of extra work as a result of any switch.For me a killer feature would be if/when we can get to a point where we can have something pgxs-style on cmake that also works on windows.Our homemade Windows build system works OK for postgres, and while ugly it is as you say well tested by now. But it doesn't do *anything* to help people build extensions on Windows. -- Magnus Hagander Me: http://www.hagander.net/ Work: http://www.redpill-linpro.com/\",\n \"msg_date\": \"Fri, 10 Feb 2017 19:27:55 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 02/10/2017 08:27 PM, Magnus Hagander wrote:\\n> For me a killer feature would be if/when we can get to a point where we can\\n> have something pgxs-style on cmake that also works on windows.\\n>\\n> Our homemade Windows build system works OK for postgres, and while ugly it\\n> is as you say well tested by now. But it doesn't do *anything* to help\\n> people build extensions on Windows.\\n\\nDo we need to build PostgreSQL itself using cmake, to achieve that? \\nCould we write something like pgxs for cmake, only for extensions?\\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\": \"Fri, 10 Feb 2017 20:31:12 +0200\",\n \"msg_from\": \"Heikki Linnakangas <hlinnaka@iki.fi>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 2017-02-10 19:27:55 +0100, Magnus Hagander wrote:\\n> For me a killer feature would be if/when we can get to a point where we can\\n> have something pgxs-style on cmake that also works on windows.\\n\\nThat should be quite possible. The ugliest part will be to determine\\nwhere to include a cmake file from (since that'll be copied in every\\nsuch project), but after that it should be nearly trivial.\\n\\n- Andres\\n\\n\\n-- \\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, 10 Feb 2017 10:32:45 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 10, 2017 at 7:31 PM, Heikki Linnakangas <hlinnaka@iki.fi> wrote:\\n\\n> On 02/10/2017 08:27 PM, Magnus Hagander wrote:\\n>\\n>> For me a killer feature would be if/when we can get to a point where we\\n>> can\\n>> have something pgxs-style on cmake that also works on windows.\\n>>\\n>> Our homemade Windows build system works OK for postgres, and while ugly it\\n>> is as you say well tested by now. But it doesn't do *anything* to help\\n>> people build extensions on Windows.\\n>>\\n>\\n> Do we need to build PostgreSQL itself using cmake, to achieve that? Could\\n> we write something like pgxs for cmake, only for extensions?\\n>\\n>\\nI guess we wouldn't, but we'd still need the \\\"replacement for autoconf\\\"\\npart. So then we're back to maintaining multiple buildsystems.\\n\\n\\n-- \\n Magnus Hagander\\n Me: http://www.hagander.net/\\n Work: http://www.redpill-linpro.com/\\n\\nOn Fri, Feb 10, 2017 at 7:31 PM, Heikki Linnakangas <hlinnaka@iki.fi> wrote:On 02/10/2017 08:27 PM, Magnus Hagander wrote:\\n\\nFor me a killer feature would be if/when we can get to a point where we can\\nhave something pgxs-style on cmake that also works on windows.\\n\\nOur homemade Windows build system works OK for postgres, and while ugly it\\nis as you say well tested by now. But it doesn't do *anything* to help\\npeople build extensions on Windows.\\n\\n\\nDo we need to build PostgreSQL itself using cmake, to achieve that? Could we write something like pgxs for cmake, only for extensions?I guess we wouldn't, but we'd still need the \\\"replacement for autoconf\\\" part. So then we're back to maintaining multiple buildsystems.-- Magnus Hagander Me: http://www.hagander.net/ Work: http://www.redpill-linpro.com/\",\n \"msg_date\": \"Fri, 10 Feb 2017 19:33:18 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 2017-02-10 20:31:12 +0200, Heikki Linnakangas wrote:\\n> On 02/10/2017 08:27 PM, Magnus Hagander wrote:\\n> > For me a killer feature would be if/when we can get to a point where we can\\n> > have something pgxs-style on cmake that also works on windows.\\n> > \\n> > Our homemade Windows build system works OK for postgres, and while ugly it\\n> > is as you say well tested by now. But it doesn't do *anything* to help\\n> > people build extensions on Windows.\\n> \\n> Do we need to build PostgreSQL itself using cmake, to achieve that? Could we\\n> write something like pgxs for cmake, only for extensions?\\n\\nI don't see why it'd need to be done together. The minimal version\\nwould be a simple cmake file that just sets a bunch of variables from\\npg_config, provides a few rules for specifying the current pgxs stuff,\\nand an example stanza how to include that file. We'd have to duplicate\\nsome of the pgxs specific logic, but that's probably not too bad.\\n\\n- Andres\\n\\n\\n-- \\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, 10 Feb 2017 10:39:04 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 2017-02-10 19:33:18 +0100, Magnus Hagander wrote:\\n> I guess we wouldn't, but we'd still need the \\\"replacement for autoconf\\\"\\n> part. So then we're back to maintaining multiple buildsystems.\\n\\nHm? Do we really need that? Most of the things in an extension you do\\n*not* want to determine separately from the backend. It's not like pgxs\\natm really allows to differ wildly from autoconf's results. And most of\\nthe relevant determinations made by autoconf are available in headers\\nand/or we can generate a cmake include file with the results of\\nautoconf.\\n\\n- Andres\\n\\n\\n-- \\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, 10 Feb 2017 10:41:27 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Feb 10, 2017 19:41, \\\"Andres Freund\\\" <andres@anarazel.de> wrote:\\n\\nOn 2017-02-10 19:33:18 +0100, Magnus Hagander wrote:\\n> I guess we wouldn't, but we'd still need the \\\"replacement for autoconf\\\"\\n> part. So then we're back to maintaining multiple buildsystems.\\n\\nHm? Do we really need that? Most of the things in an extension you do\\n*not* want to determine separately from the backend. It's not like pgxs\\natm really allows to differ wildly from autoconf's results. And most of\\nthe relevant determinations made by autoconf are available in headers\\nand/or we can generate a cmake include file with the results of\\nautoconf.\\n\\n\\nYeah, you're right. You need the output from the process, it mot the\\nprocess itself.\\n\\n/Magnus\\n\\nOn Feb 10, 2017 19:41, \\\"Andres Freund\\\" <andres@anarazel.de> wrote:On 2017-02-10 19:33:18 +0100, Magnus Hagander wrote:\\n> I guess we wouldn't, but we'd still need the \\\"replacement for autoconf\\\"\\n> part. So then we're back to maintaining multiple buildsystems.\\n\\nHm? Do we really need that? Most of the things in an extension you do\\n*not* want to determine separately from the backend. It's not like pgxs\\natm really allows to differ wildly from autoconf's results. And most of\\nthe relevant determinations made by autoconf are available in headers\\nand/or we can generate a cmake include file with the results of\\nautoconf.Yeah, you're right. You need the output from the process, it mot the process itself. /Magnus\",\n \"msg_date\": \"Fri, 10 Feb 2017 19:55:30 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 02/10/2017 01:27 PM, Magnus Hagander wrote:\\n> On Fri, Feb 10, 2017 at 6:07 PM, Robert Haas <robertmhaas@gmail.com\\n> <mailto:robertmhaas@gmail.com>> wrote:\\n>\\n> On Mon, Jan 30, 2017 at 10:26 AM, Peter Eisentraut\\n> <peter.eisentraut@2ndquadrant.com\\n> <mailto:peter.eisentraut@2ndquadrant.com>> wrote:\\n> > On 1/30/17 1:28 AM, Andres Freund wrote:\\n> >> Given that fact, I just don't buy why it's a good idea to not also\\n> >> replace autoconf initially.\\n> >\\n> > Well, I find it a bit scary. If you do the big switch all at\\n> once, then\\n> > you will have to dedicate the following 3 months to fixing\\n> complaints\\n> > from developers and build farmers.\\n>\\n> I agree with that. I think replacing the Windows build system first\\n> and then the non-Windows build system later is a better plan than\\n> replacing both at the same time.\\n>\\n> But also, I'm not really sure whether this conversion makes sense. I\\n> mean, any build system is going to require some work, and accordingly\\n> our present build systems require some work. cmake will require\\n> different work, but not necessarily less. The current system has a\\n> long history; we pretty much know it works. Switching will inevitably\\n> break some things. Maybe we'll end up better off in the long term,\\n> but maybe we won't. Things are pretty good now, so it seems like it\\n> would be easy for them to get worse rather than better. If nothing\\n> else, everybody who has to learn the new system either to use it for\\n> development or because they are doing packaging will have to do some\\n> amount of extra work as a result of any switch.\\n>\\n>\\n> For me a killer feature would be if/when we can get to a point where\\n> we can have something pgxs-style on cmake that also works on windows.\\n>\\n> Our homemade Windows build system works OK for postgres, and while\\n> ugly it is as you say well tested by now. But it doesn't do *anything*\\n> to help people build extensions on Windows. \\n>\\n>\\n\\n\\nWatch this space. There is work being done on building extensions out of\\ntree using CMake on Windows. It's pretty neat, and I'm hoping to demo it\\npublicly soon. But it's not reliant on CMake in core.\\n\\ncheers\\n\\nandrew\\n\\n-- \\nAndrew Dunstan https://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\",\n \"msg_date\": \"Fri, 10 Feb 2017 14:42:16 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 11 Feb. 2017 08:42, \\\"Andrew Dunstan\\\" <andrew.dunstan@2ndquadrant.com>\\nwrote:\\n\\n\\n\\nOn 02/10/2017 01:27 PM, Magnus Hagander wrote:\\n> On Fri, Feb 10, 2017 at 6:07 PM, Robert Haas <robertmhaas@gmail.com\\n> <mailto:robertmhaas@gmail.com>> wrote:\\n>\\n> On Mon, Jan 30, 2017 at 10:26 AM, Peter Eisentraut\\n> <peter.eisentraut@2ndquadrant.com\\n> <mailto:peter.eisentraut@2ndquadrant.com>> wrote:\\n> > On 1/30/17 1:28 AM, Andres Freund wrote:\\n> >> Given that fact, I just don't buy why it's a good idea to not also\\n> >> replace autoconf initially.\\n> >\\n> > Well, I find it a bit scary. If you do the big switch all at\\n> once, then\\n> > you will have to dedicate the following 3 months to fixing\\n> complaints\\n> > from developers and build farmers.\\n>\\n> I agree with that. I think replacing the Windows build system first\\n> and then the non-Windows build system later is a better plan than\\n> replacing both at the same time.\\n>\\n> But also, I'm not really sure whether this conversion makes sense. I\\n> mean, any build system is going to require some work, and accordingly\\n> our present build systems require some work. cmake will require\\n> different work, but not necessarily less. The current system has a\\n> long history; we pretty much know it works. Switching will inevitably\\n> break some things. Maybe we'll end up better off in the long term,\\n> but maybe we won't. Things are pretty good now, so it seems like it\\n> would be easy for them to get worse rather than better. If nothing\\n> else, everybody who has to learn the new system either to use it for\\n> development or because they are doing packaging will have to do some\\n> amount of extra work as a result of any switch.\\n>\\n>\\n> For me a killer feature would be if/when we can get to a point where\\n> we can have something pgxs-style on cmake that also works on windows.\\n>\\n> Our homemade Windows build system works OK for postgres, and while\\n> ugly it is as you say well tested by now. But it doesn't do *anything*\\n> to help people build extensions on Windows.\\n>\\n>\\n\\n\\nWatch this space. There is work being done on building extensions out of\\ntree using CMake on Windows. It's pretty neat, and I'm hoping to demo it\\npublicly soon. But it's not reliant on CMake in core.\\n\\n\\nYeah. In fact it's completely independent of core's build system it and\\nworks with any supported Pg, using pg_config for discovery.\\n\\nWe don't need cmake in core for pgxs-like functionality on Windows.\\n\\nI'd like to use it instead of our Perl windows stuff msbuild stuff in core\\nthough. I wonder if it's practical to generate the cmake file lists etc\\nfrom our Makefiles initialy like we do for the current system, and just\\nreplace the MSVC/msbuild project generators with cmake initially.\\n\\nOn 11 Feb. 2017 08:42, \\\"Andrew Dunstan\\\" <andrew.dunstan@2ndquadrant.com> wrote:\\n\\nOn 02/10/2017 01:27 PM, Magnus Hagander wrote:\\n> On Fri, Feb 10, 2017 at 6:07 PM, Robert Haas <robertmhaas@gmail.com\\n> <mailto:robertmhaas@gmail.com>> wrote:\\n>\\n> On Mon, Jan 30, 2017 at 10:26 AM, Peter Eisentraut\\n> <peter.eisentraut@2ndquadrant.com\\n> <mailto:peter.eisentraut@2ndquadrant.com>> wrote:\\n> > On 1/30/17 1:28 AM, Andres Freund wrote:\\n> >> Given that fact, I just don't buy why it's a good idea to not also\\n> >> replace autoconf initially.\\n> >\\n> > Well, I find it a bit scary. If you do the big switch all at\\n> once, then\\n> > you will have to dedicate the following 3 months to fixing\\n> complaints\\n> > from developers and build farmers.\\n>\\n> I agree with that. I think replacing the Windows build system first\\n> and then the non-Windows build system later is a better plan than\\n> replacing both at the same time.\\n>\\n> But also, I'm not really sure whether this conversion makes sense. I\\n> mean, any build system is going to require some work, and accordingly\\n> our present build systems require some work. cmake will require\\n> different work, but not necessarily less. The current system has a\\n> long history; we pretty much know it works. Switching will inevitably\\n> break some things. Maybe we'll end up better off in the long term,\\n> but maybe we won't. Things are pretty good now, so it seems like it\\n> would be easy for them to get worse rather than better. If nothing\\n> else, everybody who has to learn the new system either to use it for\\n> development or because they are doing packaging will have to do some\\n> amount of extra work as a result of any switch.\\n>\\n>\\n> For me a killer feature would be if/when we can get to a point where\\n> we can have something pgxs-style on cmake that also works on windows.\\n>\\n> Our homemade Windows build system works OK for postgres, and while\\n> ugly it is as you say well tested by now. But it doesn't do *anything*\\n> to help people build extensions on Windows.\\n>\\n>\\n\\n\\nWatch this space. There is work being done on building extensions out of\\ntree using CMake on Windows. It's pretty neat, and I'm hoping to demo it\\npublicly soon. But it's not reliant on CMake in core.Yeah. In fact it's completely independent of core's build system it and works with any supported Pg, using pg_config for discovery.We don't need cmake in core for pgxs-like functionality on Windows.I'd like to use it instead of our Perl windows stuff msbuild stuff in core though. I wonder if it's practical to generate the cmake file lists etc from our Makefiles initialy like we do for the current system, and just replace the MSVC/msbuild project generators with cmake initially.\",\n \"msg_date\": \"Sat, 11 Feb 2017 07:06:37 +0800\",\n \"msg_from\": \"Craig Ringer <craig.ringer@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 2017-02-10 10:22:34 -0800, Andres Freund wrote:\\n> On 2017-02-10 12:07:15 -0500, Robert Haas wrote:\\n> > But also, I'm not really sure whether this conversion makes sense. I\\n> > mean, any build system is going to require some work, and accordingly\\n> > our present build systems require some work. cmake will require\\n> > different work, but not necessarily less. The current system has a\\n> > long history; we pretty much know it works. Switching will inevitably\\n> > break some things. Maybe we'll end up better off in the long term,\\n> > but maybe we won't. Things are pretty good now, so it seems like it\\n> > would be easy for them to get worse rather than better.\\n> \\n> I do think it's kinda ok for people who've dealt with this for some\\n> time. But for the first few times having to write autoconf macros is\\n> quite intimidating. A somewhat less obscure way to deal with that is\\n> worth something. As is better file dependency management, across\\n> different environments. As is the IDE integration cmake is more and\\n> more getting. As is properly builtin working caching of configure tests\\n> (llvm first cmake run, 7.7s, second 2.54s). As is the fact that a lot of\\n> big projects (llvm, kde, qt, and a lot of others) migrated to it.\\n> \\n> For me personally those benefits aren't worth that much. I've learned\\n> autoconf stuff. I've scripting around autoconf caching. But for newer\\n> people that's not true.\\n\\nCraig's email just now reminded me of another advantage: Using cmake\\nacross the board, would mean we'd use the same ./configure alike logic\\non both windows and linux. To me that seems quite and advantage over\\nmanaging pg_config.h.win32/config_default.pl manually/separately - we\\nobviously have screwed up quite badly there in the past\\n(cf376a4adc0805b0).\\n\\n\\n-- \\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, 10 Feb 2017 15:15:36 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 10, 2017 at 5:07 PM, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> But also, I'm not really sure whether this conversion makes sense. I\\n> mean, any build system is going to require some work, and accordingly\\n> our present build systems require some work. cmake will require\\n> different work, but not necessarily less. The current system has a\\n> long history; we pretty much know it works. Switching will inevitably\\n> break some things. Maybe we'll end up better off in the long term,\\n> but maybe we won't. Things are pretty good now, so it seems like it\\n> would be easy for them to get worse rather than better. If nothing\\n> else, everybody who has to learn the new system either to use it for\\n> development or because they are doing packaging will have to do some\\n> amount of extra work as a result of any switch.\\n>\\n\\nFor what it's worth (even well-maintained) cmake is a usability regression\\nfrom well-maintained autotools from syseng/packager perspective.\\n\\nTwo anecdotes:\\n\\n- ./configure is much nicer from user perspective. Compare:\\n./configure --prefix=/bla --enable-foo --disable-bar --with-ssl=/opt/myssl\\ncmake -DCMAKE_INSTALL_PREFIX:PATH=/usr -DCMAKE_BLAH -DCMAKE_FOO\\n-DCMAKE_WHO_KNOWS_WHAT:PATH=/opt/myssl\\n\\n- Things like set(_PYTHON3_VERSIONS 3.7 3.6 3.5 3.4 3.3 3.2 3.1 3.0)\\n<https://github.com/Kitware/CMake/blob/master/Modules/FindPythonInterp.cmake#L43>\\nGuess what, on older cmake versions this list did not include anything\\nolder that 3.3, so it was failing with in-comprehensive generic \\\"not found\\\"\\nerror even though 3.4 was installed.\\nWith this \\\"fix\\\" somebody somewhere will be banging their head against the\\nwall again once 3.8 is out.\\nOverall, when things go wrong debugging cmake requires cmake knowledge,\\nwhile autotools mostly require shell knowledge which is much more common\\n(again, for sysadmins/packagers).\\n\\nSo while cmake is better for developers it is usually worse off for\\npackagers and advanced users. I'm not saying migration is not worth it, I'm\\npointing out costs of such migration.\\n\\nPS: I personally like Google's internal version of https://bazel.build/ a\\nlot. I've never used open-source version but I presume it's similar. While\\nit has many problems (Java, lack of popular IDE support, lack of popularity\\nand, again, Java) good parts are rules are both machine- and human-\\nreadable and writable and generally easy to debug. I'm not bold enough to\\npropose PostgreSQL to use it, but I'd be happy to see ideas from it to be\\nused elsewhere.\\n\\n-- \\nVladimir Rusinov\\nStorage SRE, Google Ireland\\n\\nGoogle Ireland Ltd.,Gordon House, Barrow Street, Dublin 4, Ireland\\nRegistered in Dublin, Ireland\\nRegistration Number: 368047\",\n \"msg_date\": \"Sun, 12 Feb 2017 17:55:56 +0000\",\n \"msg_from\": \"Vladimir Rusinov <vrusinov@google.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"2017-02-12 20:55 GMT+03:00 Vladimir Rusinov <vrusinov@google.com>:\\n\\n> Overall, when things go wrong debugging cmake requires cmake knowledge,\\n> while autotools mostly require shell knowledge which is much more common\\n> (again, for sysadmins/packagers).\\n\\n\\nIt's not really true because of CMake scripts much easier than tons of crap\\nbash (configure) and m4 scripts in Autotools, also please don't forget\\nWindows MSVC, Xcode and etc usage.\\n\\nPS: I personally like Google's internal version of https://bazel.build/ a\\n> lot. I've never used open-source version but I presume it's similar. While\\n> it has many problems (Java, lack of popular IDE support, lack of popularity\\n> and, again, Java) good parts are rules are both machine- and human-\\n> readable and writable and generally easy to debug. I'm not bold enough to\\n> propose PostgreSQL to use it, but I'd be happy to see ideas from it to be\\n> used elsewhere.\\n\\n\\nWe have many build systems, for example, another one http://mesonbuild.com/\\nbut CMake the best today as meta build system.\\n\\n2017-02-12 20:55 GMT+03:00 Vladimir Rusinov <vrusinov@google.com>:Overall, when things go wrong debugging cmake requires cmake knowledge, while autotools mostly require shell knowledge which is much more common (again, for sysadmins/packagers).It's not really true because of CMake scripts much easier than tons of crap bash (configure) and m4 scripts in Autotools, also please don't forget Windows MSVC, Xcode and etc usage. PS: I personally like Google's internal version of https://bazel.build/ a lot. I've never used open-source version but I presume it's similar. While it has many problems (Java, lack of popular IDE support, lack of popularity and, again, Java) good parts are rules are both machine- and human- readable and writable and generally easy to debug. I'm not bold enough to propose PostgreSQL to use it, but I'd be happy to see ideas from it to be used elsewhere.We have many build systems, for example, another one http://mesonbuild.com/ but CMake the best today as meta build system.\",\n \"msg_date\": \"Mon, 13 Feb 2017 11:53:22 +0300\",\n \"msg_from\": \"Yuriy Zhuravlev <stalkerg@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On 2/8/17 4:44 PM, Andres Freund wrote:\\n>> Well, I find it a bit scary. If you do the big switch all at once, then\\n>> you will have to dedicate the following 3 months to fixing complaints\\n>> from developers and build farmers.\\n> \\n> That'll be the case just as well if we spread it out over two cycles,\\n> except that we'll have it in multiple phases, and we run into the danger\\n> of a half-done conversion.\\n\\nThey key term is \\\"dedicated\\\". I don't think anyone wants to spend 3\\nmonths doing nothing but fixing the build system. If we spread it over\\nseveral releases and bring more people up to speed along the way, that\\nlooks more manageable.\\n\\n>> That would work if there were a single entry point into the build\\n>> system. But in practice there are many, and every one of them is\\n>> someone's favorite. It's unlikely that we will be able to enumerate all\\n>> of them during patch review.\\n> \\n> Not sure what you mean with \\\"entry point\\\"?\\n\\nPeople have all kinds of expectations on how the build system behaves.\\nIt's not just ./configure; make; make install. All kinds of niche and\\nedge cases have evolved over the years. Variables you can set,\\noverrides, targets in some subdirectories, building in subdirectories\\nand having it build another subdirectory first, testing this way and\\ntesting that way, testing with a custom locale or a custom database\\nname, building before testing or not, testing without reinstalling, and\\nso on and so on. You'd have to make sure at least some of that\\ncontinues to work or be able to explain it away. And most of it is not\\nreally documented.\\n\\n-- \\nPeter Eisentraut 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\": \"Mon, 13 Feb 2017 10:52:41 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 13, 2017 at 10:52 AM, Peter Eisentraut\\n<peter.eisentraut@2ndquadrant.com> wrote:\\n> People have all kinds of expectations on how the build system behaves.\\n> It's not just ./configure; make; make install. All kinds of niche and\\n> edge cases have evolved over the years. Variables you can set,\\n> overrides, targets in some subdirectories, building in subdirectories\\n> and having it build another subdirectory first, testing this way and\\n> testing that way, testing with a custom locale or a custom database\\n> name, building before testing or not, testing without reinstalling, and\\n> so on and so on. You'd have to make sure at least some of that\\n> continues to work or be able to explain it away. And most of it is not\\n> really documented.\\n\\n...which is another argument for just not changing anything. I mean,\\nagain, the current Windows build system is unbeautiful and\\noccasionally takes some work, but if we switch to cmake, that has to\\nget enough better to make up for all of the things that get worse.\\nAnd it's far from obvious than this will be so, especially when you\\nconsider the fact that many people have detailed knowledge of how to\\ntweak the current system to do what they want. As you point out here,\\na lot of that stuff may stop working if we replace the system\\nwholesale.\\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, 13 Feb 2017 13:48:08 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n },\n {\n \"msg_contents\": \"On Wed, Feb 8, 2017 at 04:52:19PM -0500, Tom Lane wrote:\\n> Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n> > On 2/8/17 6:21 AM, Yuriy Zhuravlev wrote:\\n> >> Support two build systems it's not big deal really. I have been working\\n> >> on this past year without any big troubles. \\n> >> Also we have second perl build system...\\n> \\n> > The perl/msvc build system pulls in information from the makefiles. So\\n> > when you add a file or something basic like that, you don't have to\\n> > update it. So it's really more like 1.5 build systems.\\n> \\n> Really it's more like 1.1 build systems, in that the MSVC scripts do that\\n> just well enough that you *usually* don't have to think about them. But\\n> then when they fail, and you have to figure out why, it can be a pain.\\n\\nIf cmake isn't going to be able to query the Makefiles and adjust to\\nchanges we make there, changing our Windows build system from MSVC to\\ncmake takes us from maintaining 1.1 build systems to two build systems,\\nand I don't think anyone wants that.\\n\\nIf we go to cmake, I think we need to agree we will eventually _only_\\nuse cmake. I don't think having two build systems we have to maintain\\nis better than 1.1 build systems where we can mostly ignore 0.1 of that.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\\n-- \\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 Feb 2017 16:11:21 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: WIP: About CMake v2\"\n }\n]","isConversational":false}}},{"rowIdx":42994,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"This is a new statistics view that is used to provide the number of\nSQL operations that are\nhappened on a particular interval of time. This view is useful for the\nsystem to find out the\npattern of the operations that are happening in the instance during\nparticular interval of\ntime.\n\nFollowing is the more or less columns and their details of the pg_stat_sql view.\n\npostgres=# \\d pg_stat_sql\n View \"pg_catalog.pg_stat_sql\"\n Column | Type | Modifiers\n-------------+--------------------------+-----------\n selects | bigint |\n inserts | bigint |\n deletes | bigint |\n updates | bigint |\n declares | bigint |\n fetches | bigint |\n copies | bigint |\n reindexes | bigint |\n truncates | bigint |\n stats_reset | timestamp with time zone |\n\nThe SQL counters gets updated only for the external SQL queries, not for\nthe SQL queries that are generated internally. The counters gets updated\nat exec_simple_query and exec_execute_message functions.\n\nUser can reset the SQL counter statistics using an exposed function.\nThe Stats collection can be turned on/off using a GUC also.\n\nAny comments/objections in providing a patch for the same?\n\nRegards,\nHari Babu\nFujitsu Australia\n\n\n-- \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, 20 Aug 2016 22:21:36 +1000","msg_from":"Haribabu Kommi <kommi.haribabu@gmail.com>","msg_from_op":true,"msg_subject":"New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"Haribabu Kommi <kommi.haribabu@gmail.com> writes:\n> This is a new statistics view that is used to provide the number of\n> SQL operations that are\n> happened on a particular interval of time. This view is useful for the\n> system to find out the\n> pattern of the operations that are happening in the instance during\n> particular interval of\n> time.\n\n> Following is the more or less columns and their details of the pg_stat_sql view.\n\n> postgres=# \\d pg_stat_sql\n> View \"pg_catalog.pg_stat_sql\"\n> Column | Type | Modifiers\n> -------------+--------------------------+-----------\n> selects | bigint |\n> inserts | bigint |\n> deletes | bigint |\n> updates | bigint |\n> declares | bigint |\n> fetches | bigint |\n> copies | bigint |\n> reindexes | bigint |\n> truncates | bigint |\n> stats_reset | timestamp with time zone |\n\n\n1. This set of counters seems remarkably random. Why, for instance,\ncount reindexes but not original index creations?\n\n2. What will you do with cases such as SELECTs containing modifying CTEs?\nOr EXPLAIN ANALYZE? Does CLUSTER count as a REINDEX? Does REFRESH\nMATERIALIZED VIEW count as a SELECT? (Or maybe it's an INSERT?)\n\nIf you're going to be selective about what you count, you're forever\ngoing to be fielding complaints from users who are unhappy that you\ndidn't count some statement type they care about, or feel that you\nmisclassified some complex case.\n\nI'm inclined to suggest you forget this approach and propose a single\ncounter for \"SQL commands executed\", which avoids all of the above\ndefinitional problems. People who need more detail than that are\nprobably best advised to look to contrib/pg_stat_statements, anyway.\n\nAlso, if you do that, there hardly seems a need for a whole new view.\nYou could just add a column to pg_stat_database. The incremental\nmaintenance effort doesn't seem enough to justify its own GUC, either.\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, 20 Aug 2016 11:17:01 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Sun, Aug 21, 2016 at 1:17 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Haribabu Kommi <kommi.haribabu@gmail.com> writes:\n>> This is a new statistics view that is used to provide the number of\n>> SQL operations that are\n>> happened on a particular interval of time. This view is useful for the\n>> system to find out the\n>> pattern of the operations that are happening in the instance during\n>> particular interval of\n>> time.\n>\n> 1. This set of counters seems remarkably random. Why, for instance,\n> count reindexes but not original index creations?\n\nI thought of adding the columns to the view that are mostly used(excluding\nall DDL commands), otherwise the view columns fill with all the commands\nthat are supported by PostgreSQL.\n\n> 2. What will you do with cases such as SELECTs containing modifying CTEs?\n> Or EXPLAIN ANALYZE? Does CLUSTER count as a REINDEX? Does REFRESH\n> MATERIALIZED VIEW count as a SELECT? (Or maybe it's an INSERT?)\n\nCluster commands will be calculated as clusters, explain needs some\nlogic to identify what SQL operation exactly. This is because of using\nnodeTag data from the parseTree structure to identify what type of SQL\nstatement it is.\n\nThe counters will not updated for any SQL type that is not as part of the view.\n\n> If you're going to be selective about what you count, you're forever\n> going to be fielding complaints from users who are unhappy that you\n> didn't count some statement type they care about, or feel that you\n> misclassified some complex case.\n\nYes, I agree that users may complain regarding with limited columns that are not\nsufficient for their use. But with the mostly used columns, this view\nmay be useful for\nsome of the users to find out the pattern of the SQL operations that\nare happening\non the instance.\n\n> I'm inclined to suggest you forget this approach and propose a single\n> counter for \"SQL commands executed\", which avoids all of the above\n> definitional problems. People who need more detail than that are\n> probably best advised to look to contrib/pg_stat_statements, anyway.\n\nI will add a new column to the pg_stat_database, to track the number of SQL\noperations that are happened on this particular database.\n\nRegards,\nHari Babu\nFujitsu Australia\n\n\n-- \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 Aug 2016 11:05:44 +1000","msg_from":"Haribabu Kommi <kommi.haribabu@gmail.com>","msg_from_op":true,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> I'm inclined to suggest you forget this approach and propose a single\n> counter for \"SQL commands executed\", which avoids all of the above\n> definitional problems. People who need more detail than that are\n> probably best advised to look to contrib/pg_stat_statements, anyway.\n\nI disagree. I think SQL commands executed, lumping absolutely\neverything together, really isn't much use. Haribabu's categorization\nscheme seems to need some work, but the idea of categorizing\nstatements by type and counting executions per type seems very\nreasonable.\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, 22 Aug 2016 13:54:43 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On 2016-08-22 13:54:43 -0400, Robert Haas wrote:\n> On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> > I'm inclined to suggest you forget this approach and propose a single\n> > counter for \"SQL commands executed\", which avoids all of the above\n> > definitional problems. People who need more detail than that are\n> > probably best advised to look to contrib/pg_stat_statements, anyway.\n> \n> I disagree. I think SQL commands executed, lumping absolutely\n> everything together, really isn't much use.\n\nI'm inclined to agree. I think that's a quite useful stat when looking\nat an installation one previously didn't have a lot of interaction with.\n\n\n> Haribabu's categorization\n> scheme seems to need some work, but the idea of categorizing\n> statements by type and counting executions per type seems very\n> reasonable.\n\nI'd consider instead using something like COALESCE(commandType,\nnodeTag(Query->utilityStmt)) as the categories. Not sure if I'd even\npivot that.\n\nAndres\n\n\n-- \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 Aug 2016 10:59:12 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> On 2016-08-22 13:54:43 -0400, Robert Haas wrote:\n>> On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>>> I'm inclined to suggest you forget this approach and propose a single\n>>> counter for \"SQL commands executed\", which avoids all of the above\n>>> definitional problems. People who need more detail than that are\n>>> probably best advised to look to contrib/pg_stat_statements, anyway.\n\n>> I disagree. I think SQL commands executed, lumping absolutely\n>> everything together, really isn't much use.\n\n> I'm inclined to agree. I think that's a quite useful stat when looking\n> at an installation one previously didn't have a lot of interaction with.\n\nWell, let's at least have an \"other\" category so you can add up the\ncounters and get a meaningful total.\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, 22 Aug 2016 16:27:06 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On 23/08/16 08:27, Tom Lane wrote:\n> Andres Freund <andres@anarazel.de> writes:\n>> On 2016-08-22 13:54:43 -0400, Robert Haas wrote:\n>>> On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>>>> I'm inclined to suggest you forget this approach and propose a single\n>>>> counter for \"SQL commands executed\", which avoids all of the above\n>>>> definitional problems. People who need more detail than that are\n>>>> probably best advised to look to contrib/pg_stat_statements, anyway.\n>>> I disagree. I think SQL commands executed, lumping absolutely\n>>> everything together, really isn't much use.\n>> I'm inclined to agree. I think that's a quite useful stat when looking\n>> at an installation one previously didn't have a lot of interaction with.\n> Well, let's at least have an \"other\" category so you can add up the\n> counters and get a meaningful total.\n>\n> \t\t\tregards, tom lane\n>\n>\nInitially I thought of something I thought was factious, but then \nrealized it might actually be both practicable & useful...\n\nHow about 2 extra categories (if appropriate!!!):\n\n 1. Things that actually might be sort of as fitting in 2 or more of the\n existing categories, or there is an ambiguity as to which category\n is appropriate.\n\n 2. Things that don't fit into any existing category\n\nThis was inspired by a real use case, in a totally unrelated area - but \nwhere I attempted to ensure counts were in the most useful categories I \nwas able to provide. The user had listed categories, but I found that \nthe data didn't always fit neatly into them as specified.\n\n\nCheers,\n\nGavin\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, 23 Aug 2016 09:31:36 +1200","msg_from":"Gavin Flower <GavinFlower@archidevsys.co.nz>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":">Andres Freund <andres@anarazel.de> writes:\n>> On 2016-08-22 13:54:43 -0400, Robert Haas wrote:\n>> On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>>>> I'm inclined to suggest you forget this approach and propose a single\n>>>> counter for \"SQL commands executed\", which avoids all of the above\n>>>> definitional problems. People who need more detail than that are\n>>>> probably best advised to look to contrib/pg_stat_statements, anyway.\n\n>>> I disagree. I think SQL commands executed, lumping absolutely\n>>> everything together, really isn't much use.\n\n>> I'm inclined to agree. I think that's a quite useful stat when looking\n>> at an installation one previously didn't have a lot of interaction with.\n\n>Well, let's at least have an \"other\" category so you can add up the\n>counters and get a meaningful total.\n\nHow would that meaningful total might help a user. What can a user might\nanalyse with the counter in 'other' category.\n\n\nNeha\n\n>Andres Freund <andres@anarazel.de> writes:>> On 2016-08-22 13:54:43 -0400, Robert Haas wrote:>> On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:>>>> I'm inclined to suggest you forget this approach and propose a single>>>> counter for \"SQL commands executed\", which avoids all of the above>>>> definitional problems. People who need more detail than that are>>>> probably best advised to look to contrib/pg_stat_statements, anyway.>>> I disagree. I think SQL commands executed, lumping absolutely>>> everything together, really isn't much use.>> I'm inclined to agree. I think that's a quite useful stat when looking>> at an installation one previously didn't have a lot of interaction with.>Well, let's at least have an \"other\" category so you can add up the>counters and get a meaningful total.How would that meaningful total might help a user. What can a user might analyse with the counter in 'other' category.Neha","msg_date":"Wed, 24 Aug 2016 10:02:26 +1000","msg_from":"neha khatri <nehakhatri5@gmail.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On 24/08/16 12:02, neha khatri wrote:\n> >Andres Freund <andres@anarazel.de <mailto:andres@anarazel.de>> writes:\n> >> On 2016-08-22 13:54:43 -0400, Robert Haas wrote:\n> >> On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us \n> <mailto:tgl@sss.pgh.pa.us>> wrote:\n> >>>> I'm inclined to suggest you forget this approach and propose a single\n> >>>> counter for \"SQL commands executed\", which avoids all of the above\n> >>>> definitional problems. People who need more detail than that are\n> >>>> probably best advised to look to contrib/pg_stat_statements, anyway.\n>\n> >>> I disagree. I think SQL commands executed, lumping absolutely\n> >>> everything together, really isn't much use.\n>\n> >> I'm inclined to agree. I think that's a quite useful stat when looking\n> >> at an installation one previously didn't have a lot of interaction \n> with.\n>\n> >Well, let's at least have an \"other\" category so you can add up the\n> >counters and get a meaningful total.\n>\n> How would that meaningful total might help a user. What can a user \n> might analyse with the counter in 'other' category.\n>\n>\n> Neha\n>\nThe user could then judge if there were a significant number of examples \nnot covered in the other categories - this may, or may not, be a \nproblem; depending on the use case.\n\n\nCheers,\nGavin\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 Aug 2016 12:07:34 +1200","msg_from":"Gavin Flower <GavinFlower@archidevsys.co.nz>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Wed, Aug 24, 2016 at 10:07 AM, Gavin Flower <\nGavinFlower@archidevsys.co.nz> wrote:\n\n> On 24/08/16 12:02, neha khatri wrote:\n>\n>> >Andres Freund <andres@anarazel.de <mailto:andres@anarazel.de>> writes:\n>> >> On 2016-08-22 13:54:43 -0400, Robert Haas wrote:\n>> >> On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us <mailto:\n>> tgl@sss.pgh.pa.us>> wrote:\n>> >>>> I'm inclined to suggest you forget this approach and propose a single\n>> >>>> counter for \"SQL commands executed\", which avoids all of the above\n>> >>>> definitional problems. People who need more detail than that are\n>> >>>> probably best advised to look to contrib/pg_stat_statements, anyway.\n>>\n>> >>> I disagree. I think SQL commands executed, lumping absolutely\n>> >>> everything together, really isn't much use.\n>>\n>> >> I'm inclined to agree. I think that's a quite useful stat when looking\n>> >> at an installation one previously didn't have a lot of interaction\n>> with.\n>>\n>> >Well, let's at least have an \"other\" category so you can add up the\n>> >counters and get a meaningful total.\n>>\n>> How would that meaningful total might help a user. What can a user might\n>> analyse with the counter in 'other' category.\n>>\n>>\n>>\n>> The user could then judge if there were a significant number of examples\n> not covered in the other categories - this may, or may not, be a problem;\n> depending on the use case.\n>\n>\n> Cheers,\n> Gavin\n>\n> For the user to be able to judge that whether the number in the 'other'\ncategory is a problem or not, the user is also required to know what all\nmight fall under the 'other' category. It may not be good to say that\n_anything_ that is not part of the already defined category is part of\n'other'. Probably, 'other' should also be a set of predefined operations.\n\nNeha\n\nOn Wed, Aug 24, 2016 at 10:07 AM, Gavin Flower <GavinFlower@archidevsys.co.nz> wrote:On 24/08/16 12:02, neha khatri wrote:\n\n>Andres Freund <andres@anarazel.de <mailto:andres@anarazel.de>> writes:\n>> On 2016-08-22 13:54:43 -0400, Robert Haas wrote:\n>> On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us <mailto:tgl@sss.pgh.pa.us>> wrote:\n>>>> I'm inclined to suggest you forget this approach and propose a single\n>>>> counter for \"SQL commands executed\", which avoids all of the above\n>>>> definitional problems. People who need more detail than that are\n>>>> probably best advised to look to contrib/pg_stat_statements, anyway.\n\n>>> I disagree. I think SQL commands executed, lumping absolutely\n>>> everything together, really isn't much use.\n\n>> I'm inclined to agree. I think that's a quite useful stat when looking\n>> at an installation one previously didn't have a lot of interaction with.\n\n>Well, let's at least have an \"other\" category so you can add up the\n>counters and get a meaningful total.\n\nHow would that meaningful total might help a user. What can a user might analyse with the counter in 'other' category.\n\n\n\nThe user could then judge if there were a significant number of examples not covered in the other categories - this may, or may not, be a problem; depending on the use case.\n\n\nCheers,\nGavin\n\nFor the user to be able to judge that whether the number in the 'other' category is a problem or not, the user is also required to know what all might fall under the 'other' category. It may not be good to say that _anything_ that is not part of the already defined category is part of 'other'. Probably, 'other' should also be a set of predefined operations.Neha","msg_date":"Wed, 24 Aug 2016 11:40:31 +1000","msg_from":"neha khatri <nehakhatri5@gmail.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Tue, Aug 23, 2016 at 3:59 AM, Andres Freund <andres@anarazel.de> wrote:\n>> Haribabu's categorization\n>> scheme seems to need some work, but the idea of categorizing\n>> statements by type and counting executions per type seems very\n>> reasonable.\n>\n> I'd consider instead using something like COALESCE(commandType,\n> nodeTag(Query->utilityStmt)) as the categories. Not sure if I'd even\n> pivot that.\n\nCommandType means to use select, insert, update, delete and utility\nas the categorization to display the counters? I think I am not getting\nyour point correctly.\n\nApart from the above, here are the following list of command tags that\nare generated in the code, I took only the first word of the command tag\njust to see how many categories present. The number indicates the\nsubset of operations or number of types it is used. Like create table,\ncreate function and etc.\n\ninsert\ndelete\nupdate\nselect (6)\ntransaction (10)\ndeclare\nclose (2)\nmove\nfetch\ncreate (37)\ndrop (36)\nAlter (56)\nimport\ntruncate\ncomment\nsecurity\ncopy\ngrant\nrevoke\nnotify\nlisten\nunlisten\nload\ncluster\nvacuum\nanalyze\nexplain\nrefresh\nset (2)\nreset\nshow\ndiscard (4)\nreassign\nlock\ncheckpoint\nreindex\nprepare\nexecute\ndeallocate\n\nwe can try to pick something from the above list also for main\ncategories and rest will fall under\nother.\n\n\nRegards,\nHari Babu\nFujitsu Australia\n\n\n-- \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 Aug 2016 18:07:54 +1000","msg_from":"Haribabu Kommi <kommi.haribabu@gmail.com>","msg_from_op":true,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"Haribabu Kommi wrote:\n\n> Apart from the above, here are the following list of command tags that\n> are generated in the code, I took only the first word of the command tag\n> just to see how many categories present. The number indicates the\n> subset of operations or number of types it is used. Like create table,\n> create function and etc.\n\nSounds about right. I suppose all those cases that you aggregated here\nwould expand to full tags in the actual code. I furthermore suppose\nthat some of these could be ignored, such as the transaction ones and\nthings like load, lock, move, fetch, discard, deallocate (maybe lump\nthem all together under \"other\", or some other rough categorization, as\nTom suggests). Also, for many of these commands it's probably relevant\nwhether they are acting on a temporary object or not; we should either\ncount these separately, or not count the temp ones at all.\n\n> insert\n> delete\n> update\n> select (6)\n> transaction (10)\n> declare\n> close (2)\n> move\n> fetch\n> create (37)\n> drop (36)\n> Alter (56)\n> import\n> truncate\n> comment\n> security\n> copy\n> grant\n> revoke\n> notify\n> listen\n> unlisten\n> load\n> cluster\n> vacuum\n> analyze\n> explain\n> refresh\n> set (2)\n> reset\n> show\n> discard (4)\n> reassign\n> lock\n> checkpoint\n> reindex\n> prepare\n> execute\n> deallocate\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":"Tue, 30 Aug 2016 14:19:24 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Wed, Aug 31, 2016 at 3:19 AM, Alvaro Herrera <alvherre@2ndquadrant.com>\nwrote:\n> Haribabu Kommi wrote:\n>\n>> Apart from the above, here are the following list of command tags that\n>> are generated in the code, I took only the first word of the command tag\n>> just to see how many categories present. The number indicates the\n>> subset of operations or number of types it is used. Like create table,\n>> create function and etc.\n>\n> Sounds about right. I suppose all those cases that you aggregated here\n> would expand to full tags in the actual code. I furthermore suppose\n> that some of these could be ignored, such as the transaction ones and\n> things like load, lock, move, fetch, discard, deallocate (maybe lump\n> them all together under \"other\", or some other rough categorization, as\n> Tom suggests).\n\nFollowing is the pg_stat_sql view with the SQL categories that I considered\nthat are important. Rest of the them will be shown under others category.\n\npostgres=# \\d pg_stat_sql\n View \"pg_catalog.pg_stat_sql\"\n Column | Type | Modifiers\n-----------------+--------------------------+-----------\n inserts | bigint |\n deletes | bigint |\n updates | bigint |\n selects | bigint |\n declare_cursors | bigint |\n closes | bigint |\n creates | bigint |\n drops | bigint |\n alters | bigint |\n imports | bigint |\n truncates | bigint |\n copies | bigint |\n grants | bigint |\n revokes | bigint |\n clusters | bigint |\n vacuums | bigint |\n analyzes | bigint |\n refreshs | bigint |\n locks | bigint |\n checkpoints | bigint |\n reindexes | bigint |\n deallocates | bigint |\n others | bigint |\n stats_reset | timestamp with time zone |\n\n\nIf any additions/deletions, I can accommodate them.\n\nThe stats data gets updated in exec_simple_query and exec_execute_message\nfunctions and the collected stats will be sent to stats collector similar\nlike function usage stats in pgstat_report_stat function.\n\nThese SQL statistics data is stored in the stats file similar like global\nstatistics. The STAT file format is updated to accommodate the new stats.\n\nA new GUC \"track_sql\" is added to track the SQL statement\nstatistics, by default this is off. Only superuser can change this\nparameter.\n\nAttached a patch for the same.\n\n> Also, for many of these commands it's probably relevant\n> whether they are acting on a temporary object or not; we should either\n> count these separately, or not count the temp ones at all.\n\nCurrently the SQL stats are not checking any object level that is a temp\none or not? The temp objects are specific to a backend only. But what\nI feel, any way that is an SQL query that was executed on a temp object,\nso we need to count that operation.\n\nI feel this SQL stats should not worry about the object type. May be I am\nwrong.\n\nRegards,\nHari Babu\nFujitsu Australia\n\n\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, 2 Sep 2016 16:33:53 +1000","msg_from":"Haribabu Kommi <kommi.haribabu@gmail.com>","msg_from_op":true,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Fri, Sep 2, 2016 at 2:33 AM, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:\n> On Wed, Aug 31, 2016 at 3:19 AM, Alvaro Herrera <alvherre@2ndquadrant.com>\n> wrote:\n>> Haribabu Kommi wrote:\n>>\n>>> Apart from the above, here are the following list of command tags that\n>>> are generated in the code, I took only the first word of the command tag\n>>> just to see how many categories present. The number indicates the\n>>> subset of operations or number of types it is used. Like create table,\n>>> create function and etc.\n>>\n>> Sounds about right. I suppose all those cases that you aggregated here\n>> would expand to full tags in the actual code. I furthermore suppose\n>> that some of these could be ignored, such as the transaction ones and\n>> things like load, lock, move, fetch, discard, deallocate (maybe lump\n>> them all together under \"other\", or some other rough categorization, as\n>> Tom suggests).\n>\n> Following is the pg_stat_sql view with the SQL categories that I considered\n> that are important. Rest of the them will be shown under others category.\n>\n> postgres=# \\d pg_stat_sql\n> View \"pg_catalog.pg_stat_sql\"\n> Column | Type | Modifiers\n> -----------------+--------------------------+-----------\n> inserts | bigint |\n> deletes | bigint |\n> updates | bigint |\n> selects | bigint |\n> declare_cursors | bigint |\n> closes | bigint |\n> creates | bigint |\n> drops | bigint |\n> alters | bigint |\n> imports | bigint |\n> truncates | bigint |\n> copies | bigint |\n> grants | bigint |\n> revokes | bigint |\n> clusters | bigint |\n> vacuums | bigint |\n> analyzes | bigint |\n> refreshs | bigint |\n> locks | bigint |\n> checkpoints | bigint |\n> reindexes | bigint |\n> deallocates | bigint |\n> others | bigint |\n> stats_reset | timestamp with time zone |\n>\n>\n> If any additions/deletions, I can accommodate them.\n\nI think it is not a good idea to make the command names used here the\nplural forms of the command tags. Instead of \"inserts\", \"updates\",\n\"imports\", etc. just use \"INSERT\", \"UPDATE\", \"IMPORT\". That's simpler\nand less error prone - e.g. you won't end up with things like\n\"refreshs\", which is not a word.\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 Sep 2016 16:01:25 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Thu, Sep 15, 2016 at 6:01 AM, Robert Haas <robertmhaas@gmail.com> wrote:\n\n> On Fri, Sep 2, 2016 at 2:33 AM, Haribabu Kommi <kommi.haribabu@gmail.com>\n> wrote:\n> > On Wed, Aug 31, 2016 at 3:19 AM, Alvaro Herrera <\n> alvherre@2ndquadrant.com>\n> > wrote:\n> >> Haribabu Kommi wrote:\n> >>\n> >>> Apart from the above, here are the following list of command tags that\n> >>> are generated in the code, I took only the first word of the command\n> tag\n> >>> just to see how many categories present. The number indicates the\n> >>> subset of operations or number of types it is used. Like create table,\n> >>> create function and etc.\n> >>\n> >> Sounds about right. I suppose all those cases that you aggregated here\n> >> would expand to full tags in the actual code. I furthermore suppose\n> >> that some of these could be ignored, such as the transaction ones and\n> >> things like load, lock, move, fetch, discard, deallocate (maybe lump\n> >> them all together under \"other\", or some other rough categorization, as\n> >> Tom suggests).\n> >\n> > Following is the pg_stat_sql view with the SQL categories that I\n> considered\n> > that are important. Rest of the them will be shown under others category.\n> >\n> > postgres=# \\d pg_stat_sql\n> > View \"pg_catalog.pg_stat_sql\"\n> > Column | Type | Modifiers\n> > -----------------+--------------------------+-----------\n> > inserts | bigint |\n> > deletes | bigint |\n> > updates | bigint |\n> > selects | bigint |\n> > declare_cursors | bigint |\n> > closes | bigint |\n> > creates | bigint |\n> > drops | bigint |\n> > alters | bigint |\n> > imports | bigint |\n> > truncates | bigint |\n> > copies | bigint |\n> > grants | bigint |\n> > revokes | bigint |\n> > clusters | bigint |\n> > vacuums | bigint |\n> > analyzes | bigint |\n> > refreshs | bigint |\n> > locks | bigint |\n> > checkpoints | bigint |\n> > reindexes | bigint |\n> > deallocates | bigint |\n> > others | bigint |\n> > stats_reset | timestamp with time zone |\n> >\n> >\n> > If any additions/deletions, I can accommodate them.\n>\n> I think it is not a good idea to make the command names used here the\n> plural forms of the command tags. Instead of \"inserts\", \"updates\",\n> \"imports\", etc. just use \"INSERT\", \"UPDATE\", \"IMPORT\". That's simpler\n> and less error prone - e.g. you won't end up with things like\n> \"refreshs\", which is not a word.\n\n\nThanks for your suggestion. I also thought of changing the name while\nwriting\n\"refreshs\" as a column name of the view originally. A small restriction\nwith the\nchange of names to command names is that, user needs to execute the query\nas follows.\n\nselect pg_stat_sql.select from pg_stat_sql;\n\nUpdated patch is attached with the corrections. Apart from name change,\nadded documentation and tests.\n\nRegards,\nHari Babu\nFujitsu Australia\n\n\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, 21 Sep 2016 12:23:02 +1000","msg_from":"Haribabu Kommi <kommi.haribabu@gmail.com>","msg_from_op":true,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On 9/14/16 4:01 PM, Robert Haas wrote:\n> I think it is not a good idea to make the command names used here the\n> plural forms of the command tags. Instead of \"inserts\", \"updates\",\n> \"imports\", etc. just use \"INSERT\", \"UPDATE\", \"IMPORT\". That's simpler\n> and less error prone - e.g. you won't end up with things like\n> \"refreshs\", which is not a word.\n\nHow about having the tag not be a column name but a row entry. So you'd\ndo something like\n\nSELECT * FROM pg_stat_sql WHERE tag = 'ALTER VIEW';\n\nThat way, we don't have to keep updating (and re-debating) this when new\ncommand types or subtypes are added. And queries written for future\nversions will not fail when run against old servers.\n\n-- \nPeter Eisentraut 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, 21 Sep 2016 11:25:14 -0400","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Wed, Sep 21, 2016 at 11:25:14AM -0400, Peter Eisentraut wrote:\n> On 9/14/16 4:01 PM, Robert Haas wrote:\n> > I think it is not a good idea to make the command names used here the\n> > plural forms of the command tags. Instead of \"inserts\", \"updates\",\n> > \"imports\", etc. just use \"INSERT\", \"UPDATE\", \"IMPORT\". That's simpler\n> > and less error prone - e.g. you won't end up with things like\n> > \"refreshs\", which is not a word.\n> \n> How about having the tag not be a column name but a row entry. So you'd\n> do something like\n> \n> SELECT * FROM pg_stat_sql WHERE tag = 'ALTER VIEW';\n\n+1 for this. It's MUCH easier to deal with changes in row counts than\nchanges in row type.\n\nBest,\nDavid.\n-- \nDavid Fetter <david(at)fetter(dot)org> http://fetter.org/\nPhone: +1 415 235 3778 AIM: dfetter666 Yahoo!: dfetter\nSkype: davidfetter XMPP: david(dot)fetter(at)gmail(dot)com\n\nRemember to vote!\nConsider donating to Postgres: http://www.postgresql.org/about/donate\n\n\n-- \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, 21 Sep 2016 09:22:59 -0700","msg_from":"David Fetter <david@fetter.org>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"Peter Eisentraut wrote:\n\n> How about having the tag not be a column name but a row entry. So you'd\n> do something like\n> \n> SELECT * FROM pg_stat_sql WHERE tag = 'ALTER VIEW';\n> \n> That way, we don't have to keep updating (and re-debating) this when new\n> command types or subtypes are added. And queries written for future\n> versions will not fail when run against old servers.\n\nYeah, good idea.\n\nLet's also discuss the interface from the stats collector. Currently we\nhave some 20 new SQL functions, all alike, each loading the whole data\nand returning a single counter, and then the view invokes each function\nseparately. That doesn't seem great to me. How about having a single C\nfunction that returns the whole thing as a SRF instead, and the view is\njust a single function invocation -- something like pg_lock_status\nfilling pg_locks in one go.\n\nAnother consideration is that the present patch lumps together all ALTER\ncases in a single counter. This isn't great, but at the same time we\ndon't want to bloat the stat files by having hundreds of counters per\ndatabase, do we?\n\n-- \n�lvaro Herrera https://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, 21 Sep 2016 14:05:24 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Wed, Sep 21, 2016 at 02:05:24PM -0300, Alvaro Herrera wrote:\n> Another consideration is that the present patch lumps together all\n> ALTER cases in a single counter. This isn't great, but at the same\n> time we don't want to bloat the stat files by having hundreds of\n> counters per database, do we?\n\nI count 37 documented versions of ALTER as of git master. Is there\nsome multiplier I'm missing?\n\nBest,\nDavid.\n-- \nDavid Fetter <david(at)fetter(dot)org> http://fetter.org/\nPhone: +1 415 235 3778 AIM: dfetter666 Yahoo!: dfetter\nSkype: davidfetter XMPP: david(dot)fetter(at)gmail(dot)com\n\nRemember to vote!\nConsider donating to Postgres: http://www.postgresql.org/about/donate\n\n\n-- \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, 21 Sep 2016 11:46:24 -0700","msg_from":"David Fetter <david@fetter.org>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Thu, Sep 22, 2016 at 3:05 AM, Alvaro Herrera <alvherre@2ndquadrant.com>\nwrote:\n\n> Peter Eisentraut wrote:\n>\n> > How about having the tag not be a column name but a row entry. So you'd\n> > do something like\n> >\n> > SELECT * FROM pg_stat_sql WHERE tag = 'ALTER VIEW';\n> >\n> > That way, we don't have to keep updating (and re-debating) this when new\n> > command types or subtypes are added. And queries written for future\n> > versions will not fail when run against old servers.\n>\n> Yeah, good idea.\n>\n\nYes, Having it as a row entry is good.\n\n\n\n> Let's also discuss the interface from the stats collector. Currently we\n> have some 20 new SQL functions, all alike, each loading the whole data\n> and returning a single counter, and then the view invokes each function\n> separately. That doesn't seem great to me. How about having a single C\n> function that returns the whole thing as a SRF instead, and the view is\n> just a single function invocation -- something like pg_lock_status\n> filling pg_locks in one go.\n>\n> Another consideration is that the present patch lumps together all ALTER\n> cases in a single counter. This isn't great, but at the same time we\n> don't want to bloat the stat files by having hundreds of counters per\n> database, do we?\n\n\nCurrently, The SQL stats is a fixed size counter to track the all the ALTER\ncases as single counter. So while sending the stats from the backend to\nstats collector at the end of the transaction, the cost is same, because of\nit's fixed size. This approach adds overhead to send and read the stats\nis minimal.\n\nWith the following approach, I feel it is possible to support the counter at\ncommand tag level.\n\nAdd a Global and local Hash to keep track of the counters by using the\ncommand tag as the key, this hash table increases dynamically whenever\na new type of SQL command gets executed. The Local Hash data is passed\nto stats collector whenever the transaction gets committed.\n\nThe problem I am thinking is that, Sending data from Hash and populating\nthe Hash from stats file for all the command tags adds some overhead.\n\n\nRegards,\nHari Babu\nFujitsu Australia\n\nOn Thu, Sep 22, 2016 at 3:05 AM, Alvaro Herrera <alvherre@2ndquadrant.com> wrote:Peter Eisentraut wrote:\n\n> How about having the tag not be a column name but a row entry. So you'd\n> do something like\n>\n> SELECT * FROM pg_stat_sql WHERE tag = 'ALTER VIEW';\n>\n> That way, we don't have to keep updating (and re-debating) this when new\n> command types or subtypes are added. And queries written for future\n> versions will not fail when run against old servers.\n\nYeah, good idea.Yes, Having it as a row entry is good. \nLet's also discuss the interface from the stats collector. Currently we\nhave some 20 new SQL functions, all alike, each loading the whole data\nand returning a single counter, and then the view invokes each function\nseparately. That doesn't seem great to me. How about having a single C\nfunction that returns the whole thing as a SRF instead, and the view is\njust a single function invocation -- something like pg_lock_status\nfilling pg_locks in one go.\n\nAnother consideration is that the present patch lumps together all ALTER\ncases in a single counter. This isn't great, but at the same time we\ndon't want to bloat the stat files by having hundreds of counters per\ndatabase, do we?Currently, The SQL stats is a fixed size counter to track the all the ALTER cases as single counter. So while sending the stats from the backend tostats collector at the end of the transaction, the cost is same, because ofit's fixed size. This approach adds overhead to send and read the statsis minimal.With the following approach, I feel it is possible to support the counter atcommand tag level.Add a Global and local Hash to keep track of the counters by using thecommand tag as the key, this hash table increases dynamically whenevera new type of SQL command gets executed. The Local Hash data is passedto stats collector whenever the transaction gets committed.The problem I am thinking is that, Sending data from Hash and populatingthe Hash from stats file for all the command tags adds some overhead.Regards,Hari BabuFujitsu Australia","msg_date":"Thu, 29 Sep 2016 15:45:35 +1000","msg_from":"Haribabu Kommi <kommi.haribabu@gmail.com>","msg_from_op":true,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Thu, Sep 29, 2016 at 3:45 PM, Haribabu Kommi <kommi.haribabu@gmail.com>\nwrote:\n\n>\n>\n> On Thu, Sep 22, 2016 at 3:05 AM, Alvaro Herrera <alvherre@2ndquadrant.com>\n> wrote:\n>\n>> Peter Eisentraut wrote:\n>>\n>> > How about having the tag not be a column name but a row entry. So you'd\n>> > do something like\n>> >\n>> > SELECT * FROM pg_stat_sql WHERE tag = 'ALTER VIEW';\n>> >\n>> > That way, we don't have to keep updating (and re-debating) this when new\n>> > command types or subtypes are added. And queries written for future\n>> > versions will not fail when run against old servers.\n>>\n>> Yeah, good idea.\n>>\n>\n> Yes, Having it as a row entry is good.\n>\n>\n>\n>> Let's also discuss the interface from the stats collector. Currently we\n>> have some 20 new SQL functions, all alike, each loading the whole data\n>> and returning a single counter, and then the view invokes each function\n>> separately. That doesn't seem great to me. How about having a single C\n>> function that returns the whole thing as a SRF instead, and the view is\n>> just a single function invocation -- something like pg_lock_status\n>> filling pg_locks in one go.\n>>\n>> Another consideration is that the present patch lumps together all ALTER\n>> cases in a single counter. This isn't great, but at the same time we\n>> don't want to bloat the stat files by having hundreds of counters per\n>> database, do we?\n>\n>\n> Currently, The SQL stats is a fixed size counter to track the all the\n> ALTER\n> cases as single counter. So while sending the stats from the backend to\n> stats collector at the end of the transaction, the cost is same, because of\n> it's fixed size. This approach adds overhead to send and read the stats\n> is minimal.\n>\n> With the following approach, I feel it is possible to support the counter\n> at\n> command tag level.\n>\n> Add a Global and local Hash to keep track of the counters by using the\n> command tag as the key, this hash table increases dynamically whenever\n> a new type of SQL command gets executed. The Local Hash data is passed\n> to stats collector whenever the transaction gets committed.\n>\n> The problem I am thinking is that, Sending data from Hash and populating\n> the Hash from stats file for all the command tags adds some overhead.\n>\n\n\nI tried changing the pg_stat_sql into row mode and ran the regress suite to\nadd different type of SQL commands to the view and ran the pgbench test\non my laptop to find out any performance impact with this patch.\n\n HEAD PATCH\npgbench - select 828 816\n\nHere I attached the pg_stat_sql patch to keep track of all SQL commands\nbased on the commandTag and their counts. I attached the result of this\nview that is run on the database after \"make installcheck\" just for\nreference.\n\nRegards,\nHari Babu\nFujitsu Australia\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, 12 Oct 2016 14:21:25 +1100","msg_from":"Haribabu Kommi <kommi.haribabu@gmail.com>","msg_from_op":true,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On 2016/10/12 12:21, Haribabu Kommi wrote:\n>\n>\n> On Thu, Sep 29, 2016 at 3:45 PM, Haribabu Kommi \n> <kommi.haribabu@gmail.com <mailto:kommi.haribabu@gmail.com>> wrote:\n>\n>\n>\n> On Thu, Sep 22, 2016 at 3:05 AM, Alvaro Herrera\n> <alvherre@2ndquadrant.com <mailto:alvherre@2ndquadrant.com>> wrote:\n>\n> Peter Eisentraut wrote:\n>\n> > How about having the tag not be a column name but a row\n> entry. So you'd\n> > do something like\n> >\n> > SELECT * FROM pg_stat_sql WHERE tag = 'ALTER VIEW';\n> >\n> > That way, we don't have to keep updating (and re-debating)\n> this when new\n> > command types or subtypes are added. And queries written\n> for future\n> > versions will not fail when run against old servers.\n>\n> Yeah, good idea.\n>\n>\n> Yes, Having it as a row entry is good.\n>\n> Let's also discuss the interface from the stats collector. \n> Currently we\n> have some 20 new SQL functions, all alike, each loading the\n> whole data\n> and returning a single counter, and then the view invokes each\n> function\n> separately. That doesn't seem great to me. How about having\n> a single C\n> function that returns the whole thing as a SRF instead, and\n> the view is\n> just a single function invocation -- something like pg_lock_status\n> filling pg_locks in one go.\n>\n> Another consideration is that the present patch lumps together\n> all ALTER\n> cases in a single counter. This isn't great, but at the same\n> time we\n> don't want to bloat the stat files by having hundreds of\n> counters per\n> database, do we?\n>\n>\n> Currently, The SQL stats is a fixed size counter to track the all\n> the ALTER\n> cases as single counter. So while sending the stats from the\n> backend to\n> stats collector at the end of the transaction, the cost is same,\n> because of\n> it's fixed size. This approach adds overhead to send and read the\n> stats\n> is minimal.\n>\n> With the following approach, I feel it is possible to support the\n> counter at\n> command tag level.\n>\n> Add a Global and local Hash to keep track of the counters by using the\n> command tag as the key, this hash table increases dynamically whenever\n> a new type of SQL command gets executed. The Local Hash data is passed\n> to stats collector whenever the transaction gets committed.\n>\n> The problem I am thinking is that, Sending data from Hash and\n> populating\n> the Hash from stats file for all the command tags adds some overhead.\n>\n>\n>\n> I tried changing the pg_stat_sql into row mode and ran the regress \n> suite to\n> add different type of SQL commands to the view and ran the pgbench test\n> on my laptop to find out any performance impact with this patch.\n>\n> HEAD PATCH\n> pgbench - select 828 816\n>\n> Here I attached the pg_stat_sql patch to keep track of all SQL commands\n> based on the commandTag and their counts. I attached the result of this\n> view that is run on the database after \"make installcheck\" just for \n> reference.\n>\nThank you for the patch.\n\nTest: Commands with uppercase and lowercase\n====\nIf the tag='select' then it returns the 0 rows but count is actually \nincrement by 1.\n\ntag='select' vs tag='SELECT'\n\npostgres=# SET track_sql TO ON;\nSET\npostgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\n tag | count\n--------+-------\n SELECT | 12\n(1 row)\n\npostgres=# SELECT * FROM pg_stat_sql where tag=*'SELECT'*;\n tag | count\n--------+-------\n* SELECT | 13*\n(1 row)\n\npostgres=# SELECT * FROM pg_stat_sql where tag=*'select'*;\n tag | count\n-----+-------\n*(0 rows)*\n\npostgres=# SELECT * FROM pg_stat_sql where tag=*'SELECT'*;\n tag | count\n--------+-------\n* SELECT | 15*\n(1 row)\n\nI think all command works same as above.\n\nRegards,\nVinayak Pokale\nNTT Open Source Software Center\n\n\n\n\n\n\n\n\n\nOn 2016/10/12 12:21, Haribabu Kommi\n wrote:\n\n\n\n\nOn Thu, Sep 29, 2016 at 3:45 PM,\n Haribabu Kommi <kommi.haribabu@gmail.com>\n wrote:\n\n\n\nOn Thu, Sep\n 22, 2016 at 3:05 AM, Alvaro Herrera <alvherre@2ndquadrant.com>\n wrote:\nPeter Eisentraut\n wrote:\n\n > How about having the tag not be a column\n name but a row entry.� So you'd\n > do something like\n >\n > SELECT * FROM pg_stat_sql WHERE tag =\n 'ALTER VIEW';\n >\n > That way, we don't have to keep updating\n (and re-debating) this when new\n > command types or subtypes are added.� And\n queries written for future\n > versions will not fail when run against\n old servers.\n\nYeah, good idea.\n\n\n\n\nYes, Having it as a row entry is good.\n\n�\n�\n\n Let's also discuss the interface from the stats\n collector.� Currently we\n have some 20 new SQL functions, all alike, each\n loading the whole data\n and returning a single counter, and then the\n view invokes each function\n separately.� That doesn't seem great to me.� How\n about having a single C\n function that returns the whole thing as a SRF\n instead, and the view is\n just a single function invocation -- something\n like pg_lock_status\n filling pg_locks in one go.\n\n Another consideration is that the present patch\n lumps together all ALTER\n cases in a single counter.� This isn't great,\n but at the same time we\n don't want to bloat the stat files by having\n hundreds of counters per\n database, do we?\n\n\nCurrently, The SQL stats is a fixed size counter\n to track the all the ALTER�\ncases as single counter. So while sending the\n stats from the backend to\nstats collector at the end of the transaction,\n the cost is same, because of\nit's fixed size. This approach adds overhead to\n send and read the stats\nis minimal.\n\n\nWith the following approach, I feel it is\n possible to support the counter at\ncommand tag level.\n\n\nAdd a Global and local Hash to keep track of the\n counters by using the\ncommand tag as the key, this hash table increases\n dynamically whenever\na new type of SQL command gets executed. The\n Local Hash data is passed\nto stats collector whenever the transaction gets\n committed.\n\n\nThe problem I am thinking is\n that, Sending data from Hash and populating\nthe Hash from stats file for\n all the command tags adds some overhead.\n\n\n\n\n\n\nI tried changing the pg_stat_sql into row mode and ran\n the regress suite to�\nadd different type of SQL commands to the view and ran\n the pgbench test\non my laptop to find out any performance impact with this\n patch.\n\n\n� � � � � � � � � � � � � � � � HEAD � � �PATCH\npgbench - select � � �828 � � � � �816\n\n\nHere I attached the pg_stat_sql patch\n to keep track of all SQL commands\nbased on the commandTag and their\n counts. I attached the result of this\nview that is run on the database\n after \"make installcheck\" just for reference.\n\n\n\n\n Thank you for the patch.\n\n Test: Commands with uppercase and lowercase \n ====\n If the tag='select' then it returns the 0 rows but count is actually\n increment by 1.\n\n tag='select' vs tag='SELECT'\n\n postgres=# SET track_sql TO ON;\n SET\n postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\n � tag�� | count\n --------+-------\n �SELECT |��� 12\n (1 row)\n\n postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\n � tag�� | count\n --------+-------\n�SELECT |��� 13\n (1 row)\n\n postgres=# SELECT * FROM pg_stat_sql where tag='select';\n �tag | count\n -----+-------\n(0 rows)\n\n postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\n � tag�� | count\n --------+-------\n�SELECT |��� 15\n (1 row)\n\n I think all command works same as above.\n\n Regards,\n Vinayak Pokale\n NTT Open Source Software Center","msg_date":"Wed, 12 Oct 2016 14:06:34 +0900","msg_from":"vinayak <Pokale_Vinayak_q3@lab.ntt.co.jp>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On 2016/10/12 12:21, Haribabu Kommi wrote:\n>\n>\n> On Thu, Sep 29, 2016 at 3:45 PM, Haribabu Kommi \n> <kommi.haribabu@gmail.com <mailto:kommi.haribabu@gmail.com>> wrote:\n>\n>\n>\n> On Thu, Sep 22, 2016 at 3:05 AM, Alvaro Herrera\n> <alvherre@2ndquadrant.com <mailto:alvherre@2ndquadrant.com>> wrote:\n>\n> Peter Eisentraut wrote:\n>\n> > How about having the tag not be a column name but a row\n> entry. So you'd\n> > do something like\n> >\n> > SELECT * FROM pg_stat_sql WHERE tag = 'ALTER VIEW';\n> >\n> > That way, we don't have to keep updating (and re-debating)\n> this when new\n> > command types or subtypes are added. And queries written\n> for future\n> > versions will not fail when run against old servers.\n>\n> Yeah, good idea.\n>\n>\n> Yes, Having it as a row entry is good.\n>\n> Let's also discuss the interface from the stats collector. \n> Currently we\n> have some 20 new SQL functions, all alike, each loading the\n> whole data\n> and returning a single counter, and then the view invokes each\n> function\n> separately. That doesn't seem great to me. How about having\n> a single C\n> function that returns the whole thing as a SRF instead, and\n> the view is\n> just a single function invocation -- something like pg_lock_status\n> filling pg_locks in one go.\n>\n> Another consideration is that the present patch lumps together\n> all ALTER\n> cases in a single counter. This isn't great, but at the same\n> time we\n> don't want to bloat the stat files by having hundreds of\n> counters per\n> database, do we?\n>\n>\n> Currently, The SQL stats is a fixed size counter to track the all\n> the ALTER\n> cases as single counter. So while sending the stats from the\n> backend to\n> stats collector at the end of the transaction, the cost is same,\n> because of\n> it's fixed size. This approach adds overhead to send and read the\n> stats\n> is minimal.\n>\n> With the following approach, I feel it is possible to support the\n> counter at\n> command tag level.\n>\n> Add a Global and local Hash to keep track of the counters by using the\n> command tag as the key, this hash table increases dynamically whenever\n> a new type of SQL command gets executed. The Local Hash data is passed\n> to stats collector whenever the transaction gets committed.\n>\n> The problem I am thinking is that, Sending data from Hash and\n> populating\n> the Hash from stats file for all the command tags adds some overhead.\n>\n>\n>\n> I tried changing the pg_stat_sql into row mode and ran the regress \n> suite to\n> add different type of SQL commands to the view and ran the pgbench test\n> on my laptop to find out any performance impact with this patch.\n>\n> HEAD PATCH\n> pgbench - select 828 816\n>\n> Here I attached the pg_stat_sql patch to keep track of all SQL commands\n> based on the commandTag and their counts. I attached the result of this\n> view that is run on the database after \"make installcheck\" just for \n> reference.\nSome comments:\nI think we can use pgstat_* instead of pgstat* for code consistency.\n\n+static HTAB *pgStatBackendSql = NULL;\nHow about *pgstat_backend_sql\n\n+HTAB *pgStatSql = NULL;\nHow about *pgstat_sql\n\n+static void pgstat_recv_sqlstmt(PgStat_MsgSqlstmt * msg, int len);\nHow about PgStat_MsgSqlstmt *msg instead of PgStat_MsgSqlstmt * msg\n\n+typedef struct PgStatSqlstmtEntry\nHow about PgStat_SqlstmtEntry\n\n+typedef struct PgStatMsgSqlstmt\nHow about PgStat_MsgSqlstmt\n\nI have observed below behavior.\nI have SET track_sql to ON and then execute the SELECT command and it \nreturn 0 rows.\nIt start counting from the second command.\npostgres=# SET track_sql TO ON;\nSET\npostgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\n tag | count\n-----+-------\n(0 rows)\n\npostgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\n tag | count\n--------+-------\n SELECT | 1\n(1 row)\nIs this a correct behavior?\n\nRegards,\nVinayak Pokale\nNTT Open Source Software Center\n\n\n\n\n\n\n\n\n\nOn 2016/10/12 12:21, Haribabu Kommi\n wrote:\n\n\n\n\nOn Thu, Sep 29, 2016 at 3:45 PM,\n Haribabu Kommi <kommi.haribabu@gmail.com>\n wrote:\n\n\n\nOn Thu, Sep\n 22, 2016 at 3:05 AM, Alvaro Herrera <alvherre@2ndquadrant.com>\n wrote:\nPeter Eisentraut\n wrote:\n\n > How about having the tag not be a column\n name but a row entry.� So you'd\n > do something like\n >\n > SELECT * FROM pg_stat_sql WHERE tag =\n 'ALTER VIEW';\n >\n > That way, we don't have to keep updating\n (and re-debating) this when new\n > command types or subtypes are added.� And\n queries written for future\n > versions will not fail when run against\n old servers.\n\nYeah, good idea.\n\n\n\n\nYes, Having it as a row entry is good.\n\n�\n�\n\n Let's also discuss the interface from the stats\n collector.� Currently we\n have some 20 new SQL functions, all alike, each\n loading the whole data\n and returning a single counter, and then the\n view invokes each function\n separately.� That doesn't seem great to me.� How\n about having a single C\n function that returns the whole thing as a SRF\n instead, and the view is\n just a single function invocation -- something\n like pg_lock_status\n filling pg_locks in one go.\n\n Another consideration is that the present patch\n lumps together all ALTER\n cases in a single counter.� This isn't great,\n but at the same time we\n don't want to bloat the stat files by having\n hundreds of counters per\n database, do we?\n\n\nCurrently, The SQL stats is a fixed size counter\n to track the all the ALTER�\ncases as single counter. So while sending the\n stats from the backend to\nstats collector at the end of the transaction,\n the cost is same, because of\nit's fixed size. This approach adds overhead to\n send and read the stats\nis minimal.\n\n\nWith the following approach, I feel it is\n possible to support the counter at\ncommand tag level.\n\n\nAdd a Global and local Hash to keep track of the\n counters by using the\ncommand tag as the key, this hash table increases\n dynamically whenever\na new type of SQL command gets executed. The\n Local Hash data is passed\nto stats collector whenever the transaction gets\n committed.\n\n\nThe problem I am thinking is\n that, Sending data from Hash and populating\nthe Hash from stats file for\n all the command tags adds some overhead.\n\n\n\n\n\n\nI tried changing the pg_stat_sql into row mode and ran\n the regress suite to�\nadd different type of SQL commands to the view and ran\n the pgbench test\non my laptop to find out any performance impact with this\n patch.\n\n\n� � � � � � � � � � � � � � � � HEAD � � �PATCH\npgbench - select � � �828 � � � � �816\n\n\nHere I attached the pg_stat_sql patch\n to keep track of all SQL commands\nbased on the commandTag and their\n counts. I attached the result of this\nview that is run on the database\n after \"make installcheck\" just for reference.\n\n\n\n Some comments:\n I think we can use pgstat_* instead of pgstat* for code consistency.\n �\n +static HTAB *pgStatBackendSql = NULL;\n How about *pgstat_backend_sql\n\n +HTAB������ *pgStatSql = NULL;\n How about *pgstat_sql\n\n +static void pgstat_recv_sqlstmt(PgStat_MsgSqlstmt * msg, int len);\n How about PgStat_MsgSqlstmt *msg instead of PgStat_MsgSqlstmt * msg\n\n +typedef struct PgStatSqlstmtEntry\n How about PgStat_SqlstmtEntry\n\n +typedef struct PgStatMsgSqlstmt\n How about PgStat_MsgSqlstmt\n\n I have observed below behavior.\n I have SET track_sql to ON and then execute the SELECT command and\n it return 0 rows.\n It start counting from the second command. \n postgres=# SET track_sql TO ON;\n SET\n postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\n �tag | count\n -----+-------\n (0 rows)\n\n postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\n � tag�� | count\n --------+-------\n �SELECT |���� 1\n (1 row)\n Is this a correct behavior?\n\n Regards,\n Vinayak Pokale\n NTT Open Source Software Center","msg_date":"Fri, 14 Oct 2016 17:48:27 +0900","msg_from":"vinayak <Pokale_Vinayak_q3@lab.ntt.co.jp>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Wed, Oct 12, 2016 at 4:06 PM, vinayak <Pokale_Vinayak_q3@lab.ntt.co.jp>\nwrote:\n>\n>\n> Thank you for the patch.\n>\n> Test: Commands with uppercase and lowercase\n> ====\n> If the tag='select' then it returns the 0 rows but count is actually\n> increment by 1.\n>\n> tag='select' vs tag='SELECT'\n>\n> postgres=# SET track_sql TO ON;\n> SET\n> postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\n> tag | count\n> --------+-------\n> SELECT | 12\n> (1 row)\n>\n> postgres=# SELECT * FROM pg_stat_sql where tag=*'SELECT'*;\n> tag | count\n> --------+-------\n> * SELECT | 13*\n> (1 row)\n>\n> postgres=# SELECT * FROM pg_stat_sql where tag=*'select'*;\n> tag | count\n> -----+-------\n> *(0 rows)*\n>\n> postgres=# SELECT * FROM pg_stat_sql where tag=*'SELECT'*;\n> tag | count\n> --------+-------\n> * SELECT | 15*\n> (1 row)\n>\n> I think all command works same as above.\n>\n\nThanks for checking the patch.\n\nYes, that's correct. Currently the CAPS letters are used as tag names\nthat are getting displayed whenever any SQL command is executed.\nSo I used the same names as entries to store the details of stats of\nSQL statements.\n\nIf anyone feels the other way is better, I am fine for it.\n\n\nRegards,\nHari Babu\nFujitsu Australia\n\nOn Wed, Oct 12, 2016 at 4:06 PM, vinayak <Pokale_Vinayak_q3@lab.ntt.co.jp> wrote:\n\n\n Thank you for the patch.\n\n Test: Commands with uppercase and lowercase \n ====\n If the tag='select' then it returns the 0 rows but count is actually\n increment by 1.\n\n tag='select' vs tag='SELECT'\n\n postgres=# SET track_sql TO ON;\n SET\n postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\n tag | count\n --------+-------\n SELECT | 12\n (1 row)\n\n postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\n tag | count\n --------+-------\n SELECT | 13\n (1 row)\n\n postgres=# SELECT * FROM pg_stat_sql where tag='select';\n tag | count\n -----+-------\n(0 rows)\n\n postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\n tag | count\n --------+-------\n SELECT | 15\n (1 row)\n\n I think all command works same as above.Thanks for checking the patch.Yes, that's correct. Currently the CAPS letters are used as tag namesthat are getting displayed whenever any SQL command is executed.So I used the same names as entries to store the details of stats ofSQL statements.If anyone feels the other way is better, I am fine for it.Regards,Hari BabuFujitsu Australia","msg_date":"Mon, 17 Oct 2016 12:22:20 +1100","msg_from":"Haribabu Kommi <kommi.haribabu@gmail.com>","msg_from_op":true,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Fri, Oct 14, 2016 at 7:48 PM, vinayak <Pokale_Vinayak_q3@lab.ntt.co.jp>\nwrote:\n\n>\n> On 2016/10/12 12:21, Haribabu Kommi wrote:\n>\n> I tried changing the pg_stat_sql into row mode and ran the regress suite\n> to\n> add different type of SQL commands to the view and ran the pgbench test\n> on my laptop to find out any performance impact with this patch.\n>\n> HEAD PATCH\n> pgbench - select 828 816\n>\n> Here I attached the pg_stat_sql patch to keep track of all SQL commands\n> based on the commandTag and their counts. I attached the result of this\n> view that is run on the database after \"make installcheck\" just for\n> reference.\n>\n> Some comments:\n> I think we can use pgstat_* instead of pgstat* for code consistency.\n>\n> +static HTAB *pgStatBackendSql = NULL;\n> How about *pgstat_backend_sql\n>\n> +HTAB *pgStatSql = NULL;\n> How about *pgstat_sql\n>\n\nChanged.\n\n\n> +static void pgstat_recv_sqlstmt(PgStat_MsgSqlstmt * msg, int len);\n> How about PgStat_MsgSqlstmt *msg instead of PgStat_MsgSqlstmt * msg\n>\n\nAdded the typdef into typdef.list file so this problem never occurs.\n\n\n> +typedef struct PgStatSqlstmtEntry\n> How about PgStat_SqlstmtEntry\n>\n> +typedef struct PgStatMsgSqlstmt\n> How about PgStat_MsgSqlstmt\n>\n\n\n changed.\n\n\n> I have observed below behavior.\n> I have SET track_sql to ON and then execute the SELECT command and it\n> return 0 rows.\n> It start counting from the second command.\n> postgres=# SET track_sql TO ON;\n> SET\n> postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\n> tag | count\n> -----+-------\n> (0 rows)\n>\n> postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\n> tag | count\n> --------+-------\n> SELECT | 1\n> (1 row)\n> Is this a correct behavior?\n>\n\n\nYes. This is because the stats of the SQL statements that are collected are\nsent to\nthe stats collector at the end of the SQL statement execution. The current\nSQL statement\nis not counted in the result.\n\nUpdated patch is attached.\n\nRegards,\nHari Babu\nFujitsu Australia\n\n\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, 17 Oct 2016 12:22:51 +1100","msg_from":"Haribabu Kommi <kommi.haribabu@gmail.com>","msg_from_op":true,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On 2016/10/17 10:22, Haribabu Kommi wrote:\n>\n>\n> On Fri, Oct 14, 2016 at 7:48 PM, vinayak \n> <Pokale_Vinayak_q3@lab.ntt.co.jp \n> <mailto:Pokale_Vinayak_q3@lab.ntt.co.jp>> wrote:\n>\n>\n> On 2016/10/12 12:21, Haribabu Kommi wrote:\n>> I tried changing the pg_stat_sql into row mode and ran the\n>> regress suite to\n>> add different type of SQL commands to the view and ran the\n>> pgbench test\n>> on my laptop to find out any performance impact with this patch.\n>>\n>> HEAD PATCH\n>> pgbench - select 828 816\n>>\n>> Here I attached the pg_stat_sql patch to keep track of all SQL\n>> commands\n>> based on the commandTag and their counts. I attached the result\n>> of this\n>> view that is run on the database after \"make installcheck\" just\n>> for reference.\n> Some comments:\n> I think we can use pgstat_* instead of pgstat* for code consistency.\n>\n> +static HTAB *pgStatBackendSql = NULL;\n> How about *pgstat_backend_sql\n>\n> +HTAB *pgStatSql = NULL;\n> How about *pgstat_sql\n>\n>\n> Changed.\n>\n> +static void pgstat_recv_sqlstmt(PgStat_MsgSqlstmt * msg, int len);\n> How about PgStat_MsgSqlstmt *msg instead of PgStat_MsgSqlstmt * msg\n>\n>\n> Added the typdef into typdef.list file so this problem never occurs.\n>\n> +typedef struct PgStatSqlstmtEntry\n> How about PgStat_SqlstmtEntry\n>\n> +typedef struct PgStatMsgSqlstmt\n> How about PgStat_MsgSqlstmt\n>\n>\n>\n> changed.\n>\n> Updated patch is attached.\nThe updated patch gives following error.\nError:\n../../../../src/include/pgstat.h:557: error: ‘PgStatSqlstmtEntry’ \nundeclared here (not in a function)\n\n- ((PGSTAT_MSG_PAYLOAD - sizeof(int)) / sizeof(PgStatSqlstmtEntry))\n+ ((PGSTAT_MSG_PAYLOAD - sizeof(int)) / sizeof(PgStat_SqlstmtEntry))\nThe attached patch fixed the error.\n\nRegards,\nVinayak Pokale\n\n\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, 17 Oct 2016 14:35:39 +0900","msg_from":"vinayak <Pokale_Vinayak_q3@lab.ntt.co.jp>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Thu, Sep 29, 2016 at 1:45 AM, Haribabu Kommi\n<kommi.haribabu@gmail.com> wrote:\n> Currently, The SQL stats is a fixed size counter to track the all the ALTER\n> cases as single counter. So while sending the stats from the backend to\n> stats collector at the end of the transaction, the cost is same, because of\n> it's fixed size. This approach adds overhead to send and read the stats\n> is minimal.\n>\n> With the following approach, I feel it is possible to support the counter at\n> command tag level.\n>\n> Add a Global and local Hash to keep track of the counters by using the\n> command tag as the key, this hash table increases dynamically whenever\n> a new type of SQL command gets executed. The Local Hash data is passed\n> to stats collector whenever the transaction gets committed.\n>\n> The problem I am thinking is that, Sending data from Hash and populating\n> the Hash from stats file for all the command tags adds some overhead.\n\nYeah, I'm not very excited about that overhead. This seems useful as\nfar as it goes, but I don't really want to incur measurable overhead\nwhen it's in use. Having a hash table rather than a fixed array of\nslots means that you have to pass this through the stats collector\nrather than updating shared memory directly, which is fairly heavy\nweight. If each backend could have its own copy of the slot array and\njust update that, and readers added up the values across the whole\narray, this could be done without any locking at all, and it would\ngenerally be much lighter-weight than this approach.\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, 18 Oct 2016 14:11:57 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Wed, Oct 19, 2016 at 5:11 AM, Robert Haas <robertmhaas@gmail.com> wrote:\n\n> On Thu, Sep 29, 2016 at 1:45 AM, Haribabu Kommi\n> <kommi.haribabu@gmail.com> wrote:\n> > Currently, The SQL stats is a fixed size counter to track the all the\n> ALTER\n> > cases as single counter. So while sending the stats from the backend to\n> > stats collector at the end of the transaction, the cost is same, because\n> of\n> > it's fixed size. This approach adds overhead to send and read the stats\n> > is minimal.\n> >\n> > With the following approach, I feel it is possible to support the\n> counter at\n> > command tag level.\n> >\n> > Add a Global and local Hash to keep track of the counters by using the\n> > command tag as the key, this hash table increases dynamically whenever\n> > a new type of SQL command gets executed. The Local Hash data is passed\n> > to stats collector whenever the transaction gets committed.\n> >\n> > The problem I am thinking is that, Sending data from Hash and populating\n> > the Hash from stats file for all the command tags adds some overhead.\n>\n> Yeah, I'm not very excited about that overhead. This seems useful as\n> far as it goes, but I don't really want to incur measurable overhead\n> when it's in use. Having a hash table rather than a fixed array of\n> slots means that you have to pass this through the stats collector\n> rather than updating shared memory directly, which is fairly heavy\n> weight. If each backend could have its own copy of the slot array and\n> just update that, and readers added up the values across the whole\n> array, this could be done without any locking at all, and it would\n> generally be much lighter-weight than this approach.\n\n\nUsing limited information like combining all ALTER XXX into Alter counter,\nthe fixed array logic works fine without having any problems. But people\nare suggesting to provide more details like ALTER VIEW and etc, so I\nchecked the following ways,\n\n1. Using of nodetag as an array of index to update the counter. But this\napproach doesn't work for some cases like T_DropStmt where the tag\nvaries based on the removeType of the DropStmt. So I decide to drop\nthis approach.\n\n2. Using of tag name for searching the fixed size array that is sorted with\nall command tags that are possible in PostgreSQL. Using a binary search,\nfind out the location and update the counter. In this approach, first the\narray\nneeds to be filed with TAG information in the sorted order.\n\n3. Using of Hash table to store the counters information based on the TAG\nname as key.\n\nI choose the approach-3 as it can scale for any additional commands.\n\nAny better alternatives with that directly it can point to the array index?\n\nif we are going with ALTER XXX into single Alter counter is approach, then\nI will change the code with a fixed array approach.\n\nRegards,\nHari Babu\nFujitsu Australia\n\nOn Wed, Oct 19, 2016 at 5:11 AM, Robert Haas <robertmhaas@gmail.com> wrote:On Thu, Sep 29, 2016 at 1:45 AM, Haribabu Kommi\n<kommi.haribabu@gmail.com> wrote:\n> Currently, The SQL stats is a fixed size counter to track the all the ALTER\n> cases as single counter. So while sending the stats from the backend to\n> stats collector at the end of the transaction, the cost is same, because of\n> it's fixed size. This approach adds overhead to send and read the stats\n> is minimal.\n>\n> With the following approach, I feel it is possible to support the counter at\n> command tag level.\n>\n> Add a Global and local Hash to keep track of the counters by using the\n> command tag as the key, this hash table increases dynamically whenever\n> a new type of SQL command gets executed. The Local Hash data is passed\n> to stats collector whenever the transaction gets committed.\n>\n> The problem I am thinking is that, Sending data from Hash and populating\n> the Hash from stats file for all the command tags adds some overhead.\n\nYeah, I'm not very excited about that overhead. This seems useful as\nfar as it goes, but I don't really want to incur measurable overhead\nwhen it's in use. Having a hash table rather than a fixed array of\nslots means that you have to pass this through the stats collector\nrather than updating shared memory directly, which is fairly heavy\nweight. If each backend could have its own copy of the slot array and\njust update that, and readers added up the values across the whole\narray, this could be done without any locking at all, and it would\ngenerally be much lighter-weight than this approach.Using limited information like combining all ALTER XXX into Alter counter,the fixed array logic works fine without having any problems. But peopleare suggesting to provide more details like ALTER VIEW and etc, so Ichecked the following ways,1. Using of nodetag as an array of index to update the counter. But thisapproach doesn't work for some cases like T_DropStmt where the tagvaries based on the removeType of the DropStmt. So I decide to dropthis approach.2. Using of tag name for searching the fixed size array that is sorted withall command tags that are possible in PostgreSQL. Using a binary search,find out the location and update the counter. In this approach, first the arrayneeds to be filed with TAG information in the sorted order.3. Using of Hash table to store the counters information based on the TAGname as key.I choose the approach-3 as it can scale for any additional commands.Any better alternatives with that directly it can point to the array index?if we are going with ALTER XXX into single Alter counter is approach, thenI will change the code with a fixed array approach.Regards,Hari BabuFujitsu Australia","msg_date":"Wed, 19 Oct 2016 11:36:11 +1100","msg_from":"Haribabu Kommi <kommi.haribabu@gmail.com>","msg_from_op":true,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"Hi Vinayak,\n\nThis is a gentle reminder.\n\nyou assigned as reviewer to the current patch in the 11-2016 commitfest.\nIf you have any more review comments / performance results regarding the\napproach of the patch, please share it. Otherwise, you can mark the patch\nas \"Ready for committer\" for committer feedback. This will help us in\nsmoother\noperation of the commitfest.\n\n\nRegards,\nHari Babu\nFujitsu Australia\n\nHi Vinayak,This is a gentle reminder.you assigned as reviewer to the current patch in the 11-2016 commitfest.If you have any more review comments / performance results regarding theapproach of the patch, please share it. Otherwise, you can mark the patchas \"Ready for committer\" for committer feedback. This will help us in smootheroperation of the commitfest. Regards,Hari BabuFujitsu Australia","msg_date":"Tue, 22 Nov 2016 17:57:55 +1100","msg_from":"Haribabu Kommi <kommi.haribabu@gmail.com>","msg_from_op":true,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Tue, Nov 22, 2016 at 5:57 PM, Haribabu Kommi <kommi.haribabu@gmail.com>\nwrote:\n\n> Hi Vinayak,\n>\n> This is a gentle reminder.\n>\n> you assigned as reviewer to the current patch in the 11-2016 commitfest.\n> If you have any more review comments / performance results regarding the\n> approach of the patch, please share it. Otherwise, you can mark the patch\n> as \"Ready for committer\" for committer feedback. This will help us in\n> smoother\n> operation of the commitfest.\n>\n\nMoved to next CF with \"needs review\" status.\n\n\nRegards,\nHari Babu\nFujitsu Australia\n\nOn Tue, Nov 22, 2016 at 5:57 PM, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:Hi Vinayak,This is a gentle reminder.you assigned as reviewer to the current patch in the 11-2016 commitfest.If you have any more review comments / performance results regarding theapproach of the patch, please share it. Otherwise, you can mark the patchas \"Ready for committer\" for committer feedback. This will help us in smootheroperation of the commitfest.\n\nMoved to next CF with \"needs review\" status.Regards,Hari BabuFujitsu Australia","msg_date":"Mon, 5 Dec 2016 13:31:20 +1100","msg_from":"Haribabu Kommi <kommi.haribabu@gmail.com>","msg_from_op":true,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Mon, Dec 5, 2016 at 8:01 AM, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:\n> Moved to next CF with \"needs review\" status.\n\nI have started reviewing the patch, Some initial comments.\n\n$ git apply ../pg_stat_sql_row_mode_2.patch\nerror: patch failed: src/include/catalog/pg_proc.h:2893\nerror: src/include/catalog/pg_proc.h: patch does not apply\n\nPatch is not applying on the head, I think it needs to be rebased.\n\n+void\n+pgstat_count_sqlstmt(const char *commandTag)\n+{\n+ PgStat_SqlstmtEntry *htabent;\n+ bool found;\n+\n+ if (!pgstat_track_sql)\n+ return\n\nCallers of pgstat_count_sqlstmt are always ensuring that\npgstat_track_sql is true, seems it's redundant here.\n\n-- \nRegards,\nDilip Kumar\nEnterpriseDB: http://www.enterprisedb.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, 11 Jan 2017 19:47:35 +0530","msg_from":"Dilip Kumar <dilipbalaut@gmail.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Wed, Jan 11, 2017 at 9:17 AM, Dilip Kumar <dilipbalaut@gmail.com> wrote:\n> On Mon, Dec 5, 2016 at 8:01 AM, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:\n>> Moved to next CF with \"needs review\" status.\n>\n> I have started reviewing the patch, Some initial comments.\n>\n> $ git apply ../pg_stat_sql_row_mode_2.patch\n> error: patch failed: src/include/catalog/pg_proc.h:2893\n> error: src/include/catalog/pg_proc.h: patch does not apply\n\nI suggest using 'patch' directly. 'git apply' is far too picky.\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, 11 Jan 2017 15:06:26 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Wed, Jan 11, 2017 at 7:47 PM, Dilip Kumar <dilipbalaut@gmail.com> wrote:\n> +void\n> +pgstat_count_sqlstmt(const char *commandTag)\n> +{\n> + PgStat_SqlstmtEntry *htabent;\n> + bool found;\n> +\n> + if (!pgstat_track_sql)\n> + return\n>\n> Callers of pgstat_count_sqlstmt are always ensuring that\n> pgstat_track_sql is true, seems it's redundant here.\n\nI have done testing of the feature, it's mostly working as per the expectation.\n\nI have few comments/questions.\n\n--------\nIf you execute the query with EXECUTE then commandTag will be EXECUTE\nthat way we will not show the actual query type, I mean all the\nstatements will get the common tag \"EXECUTE\".\n\nYou can try this.\nPREPARE fooplan (int) AS INSERT INTO t VALUES($1);\nEXECUTE fooplan(1);\n\n------\n\n+ /* Destroy the SQL statement counter stats HashTable */\n+ hash_destroy(pgstat_sql);\n+\n+ /* Create SQL statement counter Stats HashTable */\n\nI think in the patch we have used mixed of \"Stats HashTable\" and\n\"stats HashTable\", I think better\nto be consistent throughout the patch. Check other similar instances.\n\n----------------\n\n@@ -1102,6 +1102,12 @@ exec_simple_query(const char *query_string)\n\n PortalDrop(portal, false);\n\n+ /*\n+ * Count SQL statement for pg_stat_sql view\n+ */\n+ if (pgstat_track_sql)\n+ pgstat_count_sqlstmt(commandTag);\n\nWe are only counting the successful SQL statement, Is that intentional?\n\n------\nI have noticed that pgstat_count_sqlstmt is called from\nexec_simple_query and exec_execute_message. Don't we want to track the\nstatement executed from functions?\n-------\n\n-- \nRegards,\nDilip Kumar\nEnterpriseDB: http://www.enterprisedb.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, 13 Jan 2017 21:28:43 +0530","msg_from":"Dilip Kumar <dilipbalaut@gmail.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Sat, Jan 14, 2017 at 12:58 AM, Dilip Kumar <dilipbalaut@gmail.com> wrote:\n> On Wed, Jan 11, 2017 at 7:47 PM, Dilip Kumar <dilipbalaut@gmail.com> wrote:\n>> +void\n>> +pgstat_count_sqlstmt(const char *commandTag)\n>> +{\n>> + PgStat_SqlstmtEntry *htabent;\n>> + bool found;\n>> +\n>> + if (!pgstat_track_sql)\n>> + return\n>>\n>> Callers of pgstat_count_sqlstmt are always ensuring that\n>> pgstat_track_sql is true, seems it's redundant here.\n>\n> I have done testing of the feature, it's mostly working as per the expectation.\n>\n> I have few comments/questions.\n>\n> --------\n> If you execute the query with EXECUTE then commandTag will be EXECUTE\n> that way we will not show the actual query type, I mean all the\n> statements will get the common tag \"EXECUTE\".\n>\n> You can try this.\n> PREPARE fooplan (int) AS INSERT INTO t VALUES($1);\n> EXECUTE fooplan(1);\n>\n> ------\n>\n> + /* Destroy the SQL statement counter stats HashTable */\n> + hash_destroy(pgstat_sql);\n> +\n> + /* Create SQL statement counter Stats HashTable */\n>\n> I think in the patch we have used mixed of \"Stats HashTable\" and\n> \"stats HashTable\", I think better\n> to be consistent throughout the patch. Check other similar instances.\n>\n> ----------------\n>\n> @@ -1102,6 +1102,12 @@ exec_simple_query(const char *query_string)\n>\n> PortalDrop(portal, false);\n>\n> + /*\n> + * Count SQL statement for pg_stat_sql view\n> + */\n> + if (pgstat_track_sql)\n> + pgstat_count_sqlstmt(commandTag);\n>\n> We are only counting the successful SQL statement, Is that intentional?\n>\n> ------\n> I have noticed that pgstat_count_sqlstmt is called from\n> exec_simple_query and exec_execute_message. Don't we want to track the\n> statement executed from functions?\n> -------\n\nThe latest patch available has rotten, could you rebase please?\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":"Tue, 17 Jan 2017 16:38:48 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Fri, Jan 13, 2017 at 10:58 AM, Dilip Kumar <dilipbalaut@gmail.com> wrote:\n> If you execute the query with EXECUTE then commandTag will be EXECUTE\n> that way we will not show the actual query type, I mean all the\n> statements will get the common tag \"EXECUTE\".\n\nSomebody might say that's a feature, not a bug. But then again,\nsomeone else might say the opposite.\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, 17 Jan 2017 16:14:19 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Sat, Jan 14, 2017 at 2:58 AM, Dilip Kumar <dilipbalaut@gmail.com> wrote:\n\n> On Wed, Jan 11, 2017 at 7:47 PM, Dilip Kumar <dilipbalaut@gmail.com>\n> wrote:\n> > +void\n> > +pgstat_count_sqlstmt(const char *commandTag)\n> > +{\n> > + PgStat_SqlstmtEntry *htabent;\n> > + bool found;\n> > +\n> > + if (!pgstat_track_sql)\n> > + return\n> >\n> > Callers of pgstat_count_sqlstmt are always ensuring that\n> > pgstat_track_sql is true, seems it's redundant here.\n>\n> I have done testing of the feature, it's mostly working as per the\n> expectation.\n>\n\nThanks for the review and test.\n\nThe use case for this view is to find out the number of different SQL\nstatements\nthat are getting executed successfully on an instance by the\napplication/user.\n\nI have few comments/questions.\n>\n> --------\n> If you execute the query with EXECUTE then commandTag will be EXECUTE\n> that way we will not show the actual query type, I mean all the\n> statements will get the common tag \"EXECUTE\".\n>\n> You can try this.\n> PREPARE fooplan (int) AS INSERT INTO t VALUES($1);\n> EXECUTE fooplan(1);\n>\n> ------\n>\n\nYes, that's correct. Currently the count is incremented based on the base\ncommand,\nbecause of this reason, the EXECUTE is increased, instead of the actual\noperation.\n\n\n\n> + /* Destroy the SQL statement counter stats HashTable */\n> + hash_destroy(pgstat_sql);\n> +\n> + /* Create SQL statement counter Stats HashTable */\n>\n> I think in the patch we have used mixed of \"Stats HashTable\" and\n> \"stats HashTable\", I think better\n> to be consistent throughout the patch. Check other similar instances.\n>\n> ----------------\n>\n\nCorrected.\n\n\n>\n> @@ -1102,6 +1102,12 @@ exec_simple_query(const char *query_string)\n>\n> PortalDrop(portal, false);\n>\n> + /*\n> + * Count SQL statement for pg_stat_sql view\n> + */\n> + if (pgstat_track_sql)\n> + pgstat_count_sqlstmt(commandTag);\n>\n> We are only counting the successful SQL statement, Is that intentional?\n>\n\nYes, I thought of counting the successful SQL operations that changed the\ndatabase over a period of time. I am not sure whether there will be many\nfailure operations that can occur on an instance to be counted.\n\n\n> ------\n> I have noticed that pgstat_count_sqlstmt is called from\n> exec_simple_query and exec_execute_message. Don't we want to track the\n> statement executed from functions?\n> -------\n\n\nThe view is currently designed to count user/application initiated SQL\nstatements,\nbut not the internal SQL queries that are getting executed from functions\nand etc.\n\n>>+void\n>>+pgstat_count_sqlstmt(const char *commandTag)\n>>+{\n>>+ PgStat_SqlstmtEntry *htabent;\n>>+ bool found;\n>>+\n>>+ if (!pgstat_track_sql)\n>>+ return\n>\n>Callers of pgstat_count_sqlstmt are always ensuring that\n>pgstat_track_sql is true, seems it's redundant here.\n\nRemoved the check in pgstat_count_sqlstmt statement and add it\nexec_execute_message\nfunction where the check is missed.\n\nUpdated patch attached.\n\nRegards,\nHari Babu\nFujitsu Australia\n\n\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, 18 Jan 2017 16:15:30 +1100","msg_from":"Haribabu Kommi <kommi.haribabu@gmail.com>","msg_from_op":true,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On 2017/01/18 14:15, Haribabu Kommi wrote:\n>\n>\n> On Sat, Jan 14, 2017 at 2:58 AM, Dilip Kumar <dilipbalaut@gmail.com \n> <mailto:dilipbalaut@gmail.com>> wrote:\n>\n> On Wed, Jan 11, 2017 at 7:47 PM, Dilip Kumar\n> <dilipbalaut@gmail.com <mailto:dilipbalaut@gmail.com>> wrote:\n> > +void\n> > +pgstat_count_sqlstmt(const char *commandTag)\n> > +{\n> > + PgStat_SqlstmtEntry *htabent;\n> > + bool found;\n> > +\n> > + if (!pgstat_track_sql)\n> > + return\n> >\n> > Callers of pgstat_count_sqlstmt are always ensuring that\n> > pgstat_track_sql is true, seems it's redundant here.\n>\n> I have done testing of the feature, it's mostly working as per the\n> expectation.\n>\n>\n> Thanks for the review and test.\n>\n> The use case for this view is to find out the number of different SQL \n> statements\n> that are getting executed successfully on an instance by the \n> application/user.\n>\n> I have few comments/questions.\n>\n> --------\n> If you execute the query with EXECUTE then commandTag will be EXECUTE\n> that way we will not show the actual query type, I mean all the\n> statements will get the common tag \"EXECUTE\".\n>\n> You can try this.\n> PREPARE fooplan (int) AS INSERT INTO t VALUES($1);\n> EXECUTE fooplan(1);\n>\n> ------\n>\n>\n> Yes, that's correct. Currently the count is incremented based on the \n> base command,\n> because of this reason, the EXECUTE is increased, instead of the \n> actual operation.\n>\n> + /* Destroy the SQL statement counter stats HashTable */\n> + hash_destroy(pgstat_sql);\n> +\n> + /* Create SQL statement counter Stats HashTable */\n>\n> I think in the patch we have used mixed of \"Stats HashTable\" and\n> \"stats HashTable\", I think better\n> to be consistent throughout the patch. Check other similar instances.\n>\n> ----------------\n>\n>\n> Corrected.\n>\n>\n> @@ -1102,6 +1102,12 @@ exec_simple_query(const char *query_string)\n>\n> PortalDrop(portal, false);\n>\n> + /*\n> + * Count SQL statement for pg_stat_sql view\n> + */\n> + if (pgstat_track_sql)\n> + pgstat_count_sqlstmt(commandTag);\n>\n> We are only counting the successful SQL statement, Is that\n> intentional?\n>\n>\n> Yes, I thought of counting the successful SQL operations that changed the\n> database over a period of time. I am not sure whether there will be many\n> failure operations that can occur on an instance to be counted.\n>\n> ------\n> I have noticed that pgstat_count_sqlstmt is called from\n> exec_simple_query and exec_execute_message. Don't we want to track the\n> statement executed from functions?\n> -------\n>\n> The view is currently designed to count user/application initiated SQL \n> statements,\n> but not the internal SQL queries that are getting executed from \n> functions and etc.\n>\n> >>+void\n> >>+pgstat_count_sqlstmt(const char *commandTag)\n> >>+{\n> >>+ PgStat_SqlstmtEntry *htabent;\n> >>+ bool found;\n> >>+\n> >>+ if (!pgstat_track_sql)\n> >>+ return\n> >\n> >Callers of pgstat_count_sqlstmt are always ensuring that\n> >pgstat_track_sql is true, seems it's redundant here.\n>\n> Removed the check in pgstat_count_sqlstmt statement and add it \n> exec_execute_message\n> function where the check is missed.\n>\n> Updated patch attached.\nI have reviewed the patch. All the test cases works fine.\n\nRegards,\nVinayak Pokale\nNTT Open Source Software Center\n\n\n\n\n\n\n\nOn 2017/01/18 14:15, Haribabu Kommi\n wrote:\n\n\n\n\nOn Sat, Jan 14, 2017 at 2:58 AM,\n Dilip Kumar <dilipbalaut@gmail.com>\n wrote:\nOn Wed, Jan 11,\n 2017 at 7:47 PM, Dilip Kumar <dilipbalaut@gmail.com>\n wrote:\n > +void\n > +pgstat_count_sqlstmt(const char *commandTag)\n > +{\n > + PgStat_SqlstmtEntry *htabent;\n > + bool found;\n > +\n > + if (!pgstat_track_sql)\n > + return\n >\n > Callers of pgstat_count_sqlstmt are always ensuring\n that\n > pgstat_track_sql is true, seems it's redundant\n here.\n\nI have done testing of the feature, it's mostly\n working as per the expectation.\n\n\n\nThanks for the review and test.\n\n\nThe use case for this view is to find out the number of\n different SQL statements\nthat are getting executed successfully on an instance\n by the application/user. \n\n\n\n I have few comments/questions.\n\n --------\n If you execute the query with EXECUTE then commandTag will\n be EXECUTE\n that way we will not show the actual query type, I mean\n all the\n statements will get the common tag \"EXECUTE\".\n\n You can try this.\n PREPARE fooplan (int) AS INSERT INTO t VALUES($1);\n EXECUTE fooplan(1);\n\n ------\n\n\n\nYes, that's correct. Currently the count is incremented\n based on the base command,\nbecause of this reason, the EXECUTE is increased,\n instead of the actual operation.\n\n\n \n\n + /* Destroy the SQL statement counter stats HashTable */\n + hash_destroy(pgstat_sql);\n +\n + /* Create SQL statement counter Stats HashTable */\n\n I think in the patch we have used mixed of \"Stats\n HashTable\" and\n \"stats HashTable\", I think better\n to be consistent throughout the patch. Check other similar\n instances.\n\n ----------------\n\n\n\nCorrected.\n \n\n\n @@ -1102,6 +1102,12 @@ exec_simple_query(const char\n *query_string)\n\n PortalDrop(portal, false);\n\n + /*\n + * Count SQL statement for pg_stat_sql view\n + */\n + if (pgstat_track_sql)\n + pgstat_count_sqlstmt(commandTag);\n\n We are only counting the successful SQL statement, Is that\n intentional?\n\n\n\nYes, I thought of counting the successful SQL\n operations that changed the\ndatabase over a period of time. I am not sure whether\n there will be many\nfailure operations that can occur on an instance to be\n counted.\n \n\n ------\n I have noticed that pgstat_count_sqlstmt is called from\n exec_simple_query and exec_execute_message. Don't we want\n to track the\n statement executed from functions?\n -------\n \n\n The view is currently designed to count user/application\n initiated SQL statements,\nbut not the internal SQL queries that\n are getting executed from functions and etc.\n\n\n>>+void\n>>+pgstat_count_sqlstmt(const\n char *commandTag)\n>>+{\n>>+ PgStat_SqlstmtEntry\n *htabent;\n>>+ bool found;\n>>+\n>>+ if\n (!pgstat_track_sql)\n>>+ return\n >\n>Callers of\n pgstat_count_sqlstmt are always ensuring that\n>pgstat_track_sql is true,\n seems it's redundant here.\n\n\n\n Removed the check in pgstat_count_sqlstmt\n statement and add it exec_execute_message\nfunction\n where the check is missed.\n\n\nUpdated\n patch attached.\n\n\n I have reviewed the patch. All the test cases works fine.\n\n Regards,\n Vinayak Pokale\n NTT Open Source Software Center","msg_date":"Fri, 20 Jan 2017 13:31:43 +0900","msg_from":"vinayak <Pokale_Vinayak_q3@lab.ntt.co.jp>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Wed, Jan 18, 2017 at 10:45 AM, Haribabu Kommi\n<kommi.haribabu@gmail.com> wrote:\n> The use case for this view is to find out the number of different SQL\n> statements\n> that are getting executed successfully on an instance by the\n> application/user.\n>\n>> I have few comments/questions.\n>>\n>> --------\n>> If you execute the query with EXECUTE then commandTag will be EXECUTE\n>> that way we will not show the actual query type, I mean all the\n>> statements will get the common tag \"EXECUTE\".\n>>\n>> You can try this.\n>> PREPARE fooplan (int) AS INSERT INTO t VALUES($1);\n>> EXECUTE fooplan(1);\n>>\n>> ------\n>\n>\n> Yes, that's correct. Currently the count is incremented based on the base\n> command,\n> because of this reason, the EXECUTE is increased, instead of the actual\n> operation.\n\nOkay, As per your explaination, and Robert also pointed out that it\ncan be a feature for someone and bug for others. I would have\npreferred it shows actual SQL statement. But I have no objection to\nwhat you are doing.\n>\n>\n>>\n>> + /* Destroy the SQL statement counter stats HashTable */\n>> + hash_destroy(pgstat_sql);\n>> +\n>> + /* Create SQL statement counter Stats HashTable */\n>>\n>> I think in the patch we have used mixed of \"Stats HashTable\" and\n>> \"stats HashTable\", I think better\n>> to be consistent throughout the patch. Check other similar instances.\n>>\n>> ----------------\n>\n>\n> Corrected.\n>\n>>\n>>\n>> @@ -1102,6 +1102,12 @@ exec_simple_query(const char *query_string)\n>>\n>> PortalDrop(portal, false);\n>>\n>> + /*\n>> + * Count SQL statement for pg_stat_sql view\n>> + */\n>> + if (pgstat_track_sql)\n>> + pgstat_count_sqlstmt(commandTag);\n>>\n>> We are only counting the successful SQL statement, Is that intentional?\n>\n>\n> Yes, I thought of counting the successful SQL operations that changed the\n> database over a period of time. I am not sure whether there will be many\n> failure operations that can occur on an instance to be counted.\n\nOkay..\n>\n>>\n>> ------\n>> I have noticed that pgstat_count_sqlstmt is called from\n>> exec_simple_query and exec_execute_message. Don't we want to track the\n>> statement executed from functions?\n>> -------\n> The view is currently designed to count user/application initiated SQL\n> statements,\n> but not the internal SQL queries that are getting executed from functions\n> and etc.\n\nOkay..\n>\n>>>+void\n>>>+pgstat_count_sqlstmt(const char *commandTag)\n>>>+{\n>>>+ PgStat_SqlstmtEntry *htabent;\n>>>+ bool found;\n>>>+\n>>>+ if (!pgstat_track_sql)\n>>>+ return\n>>\n>>Callers of pgstat_count_sqlstmt are always ensuring that\n>>pgstat_track_sql is true, seems it's redundant here.\n>\n> Removed the check in pgstat_count_sqlstmt statement and add it\n> exec_execute_message\n> function where the check is missed.\n>\n> Updated patch attached.\n\nOverall patch looks fine to me and marking it \"ready for committer\".\n\nThere is two design decision, which I leave it for committer's decision.\n\n1. EXECUTE statement should show only as EXECUTE count, not the\nactual SQL query.\n2. should we count statement execute from Functions or only statement\nwhat is executed directly by the user as it's doing now?\n\n\n-- \nRegards,\nDilip Kumar\nEnterpriseDB: http://www.enterprisedb.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, 24 Jan 2017 10:29:32 +0530","msg_from":"Dilip Kumar <dilipbalaut@gmail.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Tue, Jan 24, 2017 at 3:59 PM, Dilip Kumar <dilipbalaut@gmail.com> wrote:\n\n>\n> Overall patch looks fine to me and marking it \"ready for committer\".\n>\n> There is two design decision, which I leave it for committer's decision.\n>\n> 1. EXECUTE statement should show only as EXECUTE count, not the\n> actual SQL query.\n> 2. should we count statement execute from Functions or only statement\n> what is executed directly by the user as it's doing now?\n\n\nThanks for the review.\nLet's wait for the committer's opinion.\n\nRegards,\nHari Babu\nFujitsu Australia\n\nOn Tue, Jan 24, 2017 at 3:59 PM, Dilip Kumar <dilipbalaut@gmail.com> wrote:\nOverall patch looks fine to me and marking it \"ready for committer\".\n\nThere is two design decision, which I leave it for committer's decision.\n\n1. EXECUTE statement should show only as EXECUTE count, not the\nactual SQL query.\n2. should we count statement execute from Functions or only statement\nwhat is executed directly by the user as it's doing now?Thanks for the review.Let's wait for the committer's opinion.Regards,Hari BabuFujitsu Australia","msg_date":"Fri, 27 Jan 2017 12:26:01 +1100","msg_from":"Haribabu Kommi <kommi.haribabu@gmail.com>","msg_from_op":true,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Fri, Jan 27, 2017 at 10:26 AM, Haribabu Kommi\n<kommi.haribabu@gmail.com> wrote:\n> Thanks for the review.\n> Let's wait for the committer's opinion.\n\nI have moved this patch to CF 2017-03 to wait for this to happen.\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, 1 Feb 2017 13:13:11 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Wed, Feb 1, 2017 at 3:13 PM, Michael Paquier <michael.paquier@gmail.com>\nwrote:\n\n> On Fri, Jan 27, 2017 at 10:26 AM, Haribabu Kommi\n> <kommi.haribabu@gmail.com> wrote:\n> > Thanks for the review.\n> > Let's wait for the committer's opinion.\n>\n> I have moved this patch to CF 2017-03 to wait for this to happen.\n>\n\nAttached a rebased patch to resolve the OID conflict.\n\nRegards,\nHari Babu\nFujitsu Australia\n\n\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, 8 Mar 2017 14:39:23 +1100","msg_from":"Haribabu Kommi <kommi.haribabu@gmail.com>","msg_from_op":true,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"Hi,\n\n\nI'm somewhat inclined to think that this really would be better done in\nan extension like pg_stat_statements.\n\n\nOn 2017-03-08 14:39:23 +1100, Haribabu Kommi wrote:\n> On Wed, Feb 1, 2017 at 3:13 PM, Michael Paquier <michael.paquier@gmail.com>\n\n> + <varlistentry id=\"guc-track-sql\" xreflabel=\"track_sql\">\n> + <term><varname>track_sql</varname> (<type>boolean</type>)\n> + <indexterm>\n> + <primary><varname>track_sql</> configuration parameter</primary>\n> + </indexterm>\n> + </term>\n> + <listitem>\n> + <para>\n> + Enables collection of different SQL statement statistics that are\n> + executed on the instance. This parameter is off by default. Only\n> + superusers can change this setting.\n> + </para>\n> + </listitem>\n> + </varlistentry>\n> +\n\nI don't like this name much, seems a bit too generic to\nme. 'track_activities', 'track_io_timings' are at least a bit clearer.\nHow about track_statement_statistics + corresponding view/function renaming?\n\n\n> + <table id=\"pg-stat-sql-view\" xreflabel=\"pg_stat_sql\">\n> + <title><structname>pg_stat_sql</structname> View</title>\n> + <tgroup cols=\"3\">\n> + <thead>\n> + <row>\n> + <entry>Column</entry>\n> + <entry>Type</entry>\n> + <entry>Description</entry>\n> + </row>\n> + </thead>\n> +\n> + <tbody>\n> + <row>\n> + <entry><structfield>tag</></entry>\n> + <entry><type>text</></entry>\n> + <entry>Name of the SQL statement</entry>\n> + </row>\n\nIt's not really the \"name\" of a statement. Internally and IIRC elsewhere\nin the docs we describe something like this as tag?\n\n\n> +/* ----------\n> + * pgstat_send_sqlstmt(void)\n> + *\n> + *\t\tSend SQL statement statistics to the collector\n> + * ----------\n> + */\n> +static void\n> +pgstat_send_sqlstmt(void)\n> +{\n> +\tPgStat_MsgSqlstmt msg;\n> +\tPgStat_SqlstmtEntry *entry;\n> +\tHASH_SEQ_STATUS hstat;\n> +\n> +\tif (pgstat_backend_sql == NULL)\n> +\t\treturn;\n> +\n> +\tpgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_SQLSTMT);\n> +\tmsg.m_nentries = 0;\n> +\n> +\thash_seq_init(&hstat, pgstat_backend_sql);\n> +\twhile ((entry = (PgStat_SqlstmtEntry *) hash_seq_search(&hstat)) != NULL)\n> +\t{\n> +\t\tPgStat_SqlstmtEntry *m_ent;\n> +\n> +\t\t/* Skip it if no counts accumulated since last time */\n> +\t\tif (entry->count == 0)\n> +\t\t\tcontinue;\n> +\n> +\t\t/* need to convert format of time accumulators */\n> +\t\tm_ent = &msg.m_entry[msg.m_nentries];\n> +\t\tmemcpy(m_ent, entry, sizeof(PgStat_SqlstmtEntry));\n> +\n> +\t\tif (++msg.m_nentries >= PGSTAT_NUM_SQLSTMTS)\n> +\t\t{\n> +\t\t\tpgstat_send(&msg, offsetof(PgStat_MsgSqlstmt, m_entry[0]) +\n> +\t\t\t\t\t\tmsg.m_nentries * sizeof(PgStat_SqlstmtEntry));\n> +\t\t\tmsg.m_nentries = 0;\n> +\t\t}\n> +\n> +\t\t/* reset the entry's count */\n> +\t\tentry->count = 0;\n> +\t}\n> +\n> +\tif (msg.m_nentries > 0)\n> +\t\tpgstat_send(&msg, offsetof(PgStat_MsgSqlstmt, m_entry[0]) +\n> +\t\t\t\t\tmsg.m_nentries * sizeof(PgStat_SqlstmtEntry));\n\nHm. So pgstat_backend_sql is never deleted, which'll mean that if a\nbackend lives long we'll continually iterate over a lot of 0 entries. I\nthink performance evaluation of this feature should take that into\naccount.\n\n\n> +}\n> +\n> +/*\n> + * Count SQL statement for pg_stat_sql view\n> + */\n> +void\n> +pgstat_count_sqlstmt(const char *commandTag)\n> +{\n> +\tPgStat_SqlstmtEntry *htabent;\n> +\tbool\t\tfound;\n> +\n> +\tif (!pgstat_backend_sql)\n> +\t{\n> +\t\t/* First time through - initialize SQL statement stat table */\n> +\t\tHASHCTL\t\thash_ctl;\n> +\n> +\t\tmemset(&hash_ctl, 0, sizeof(hash_ctl));\n> +\t\thash_ctl.keysize = NAMEDATALEN;\n> +\t\thash_ctl.entrysize = sizeof(PgStat_SqlstmtEntry);\n> +\t\tpgstat_backend_sql = hash_create(\"SQL statement stat entries\",\n> +\t\t\t\t\t\t\t\t\t\t PGSTAT_SQLSTMT_HASH_SIZE,\n> +\t\t\t\t\t\t\t\t\t\t &hash_ctl,\n> +\t\t\t\t\t\t\t\t\t\t HASH_ELEM | HASH_BLOBS);\n> +\t}\n> +\n> +\t/* Get the stats entry for this SQL statement, create if necessary */\n> +\thtabent = hash_search(pgstat_backend_sql, commandTag,\n> +\t\t\t\t\t\t HASH_ENTER, &found);\n> +\tif (!found)\n> +\t\thtabent->count = 1;\n> +\telse\n> +\t\thtabent->count++;\n> +}\n\n\nThat's not really something in this patch, but a lot of this would be\nbetter if we didn't internally have command tags as strings, but as an\nenum. We should really only convert to a string with needed. That\nwe're doing string comparisons on Portal->commandTag is just plain bad.\n\nIf so, we could also make this whole patch a lot cheaper - have a fixed\nsize array that has an entry for every possible tag (possibly even in\nshared memory, and then use atomics there).\n\n> +++ b/src/backend/tcop/postgres.c\n> @@ -1109,6 +1109,12 @@ exec_simple_query(const char *query_string)\n> \n> \t\tPortalDrop(portal, false);\n> \n> +\t\t/*\n> +\t\t * Count SQL statement for pg_stat_sql view\n> +\t\t */\n> +\t\tif (pgstat_track_sql)\n> +\t\t\tpgstat_count_sqlstmt(commandTag);\n> +\n\nIsn't doing this at the message level quite wrong? What about\nstatements executed from functions and such? Shouldn't this integrate\nat the executor level instead?\n\n\n> \t\tif (IsA(parsetree->stmt, TransactionStmt))\n> \t\t{\n> \t\t\t/*\n> @@ -1991,6 +1997,29 @@ exec_execute_message(const char *portal_name, long max_rows)\n> \n> \t(*receiver->rDestroy) (receiver);\n> \n> +\t/*\n> +\t * Count SQL Statement for pgx_stat_sql\n> +\t */\n> +\tif (pgstat_track_sql)\n> +\t{\n> +\t\tCachedPlanSource *psrc = NULL;\n> +\n> +\t\tif (portal->prepStmtName)\n> +\t\t{\n> +\t\t\tPreparedStatement *pstmt;\n> +\n> +\t\t\tpstmt = FetchPreparedStatement(portal->prepStmtName, false);\n> +\t\t\tif (pstmt)\n> +\t\t\t\tpsrc = pstmt->plansource;\n> +\t\t}\n> +\t\telse\n> +\t\t\tpsrc = unnamed_stmt_psrc;\n> +\n> +\t\t/* psrc should not be NULL here */\n> +\t\tif (psrc && psrc->commandTag && !execute_is_fetch && pgstat_track_sql)\n> +\t\t\tpgstat_count_sqlstmt(psrc->commandTag);\n\nWait, we're re-fetching the statement here? That doesn't sound\nalright.\n\n\n> +Datum\n> +pg_stat_sql(PG_FUNCTION_ARGS)\n> +{\n> +\tTupleDesc\ttupdesc;\n> +\tDatum\t\tvalues[2];\n> +\tbool\t\tnulls[2];\n> +\tReturnSetInfo *rsi;\n> +\tMemoryContext old_cxt;\n> +\tTuplestorestate *tuple_store;\n> +\tPgStat_SqlstmtEntry *sqlstmtentry;\n> +\tHASH_SEQ_STATUS hstat;\n> +\n> +\t/* Function call to let the backend read the stats file */\n> +\tpgstat_fetch_global();\n> +\n> +\t/* Initialize values and nulls arrays */\n> +\tMemSet(values, 0, sizeof(values));\n> +\tMemSet(nulls, 0, sizeof(nulls));\n\nwhy not instead declare values[2] = {0}? Obviously not important.\n\n\n> +\ttuple_store =\n> +\t\ttuplestore_begin_heap(rsi->allowedModes & SFRM_Materialize_Random,\n> +\t\t\t\t\t\t\t false, work_mem);\n\nHuh, why do we need random?\n\n\n\nGreetings,\n\nAndres Freund\n\n\n-- \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 Apr 2017 12:17:57 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Thu, Apr 6, 2017 at 5:17 AM, Andres Freund <andres@anarazel.de> wrote:\n\n> Hi,\n>\n>\n> I'm somewhat inclined to think that this really would be better done in\n> an extension like pg_stat_statements.\n>\n\nThanks for the review.\n\n\n\n>\n> > +}\n> > +\n> > +/*\n> > + * Count SQL statement for pg_stat_sql view\n> > + */\n> > +void\n> > +pgstat_count_sqlstmt(const char *commandTag)\n> > +{\n> > + PgStat_SqlstmtEntry *htabent;\n> > + bool found;\n> > +\n> > + if (!pgstat_backend_sql)\n> > + {\n> > + /* First time through - initialize SQL statement stat\n> table */\n> > + HASHCTL hash_ctl;\n> > +\n> > + memset(&hash_ctl, 0, sizeof(hash_ctl));\n> > + hash_ctl.keysize = NAMEDATALEN;\n> > + hash_ctl.entrysize = sizeof(PgStat_SqlstmtEntry);\n> > + pgstat_backend_sql = hash_create(\"SQL statement stat\n> entries\",\n> > +\n> PGSTAT_SQLSTMT_HASH_SIZE,\n> > +\n> &hash_ctl,\n> > +\n> HASH_ELEM | HASH_BLOBS);\n> > + }\n> > +\n> > + /* Get the stats entry for this SQL statement, create if necessary\n> */\n> > + htabent = hash_search(pgstat_backend_sql, commandTag,\n> > + HASH_ENTER, &found);\n> > + if (!found)\n> > + htabent->count = 1;\n> > + else\n> > + htabent->count++;\n> > +}\n>\n>\n> That's not really something in this patch, but a lot of this would be\n> better if we didn't internally have command tags as strings, but as an\n> enum. We should really only convert to a string with needed. That\n> we're doing string comparisons on Portal->commandTag is just plain bad.\n>\n> If so, we could also make this whole patch a lot cheaper - have a fixed\n> size array that has an entry for every possible tag (possibly even in\n> shared memory, and then use atomics there).\n>\n\nDuring the development, I thought of using an array of all command tags\nand update the counters using the tag name, but not like the enum values.\nNow I have to create a mapping array with enums and tag names for easier\ncounter updates. The problem in this approach is, whenever any addition\nof new commands, this mapping array needs to be updated with both\nnew enum and new tag name, whereas with hash table approach, it works\nfuture command additions also, but there is some performance overhead\ncompared to the array approach.\n\nI will modify the patch to use the array approach and provide it to the next\ncommitfest.\n\n\n\n> > +++ b/src/backend/tcop/postgres.c\n> > @@ -1109,6 +1109,12 @@ exec_simple_query(const char *query_string)\n> >\n> > PortalDrop(portal, false);\n> >\n> > + /*\n> > + * Count SQL statement for pg_stat_sql view\n> > + */\n> > + if (pgstat_track_sql)\n> > + pgstat_count_sqlstmt(commandTag);\n> > +\n>\n> Isn't doing this at the message level quite wrong? What about\n> statements executed from functions and such? Shouldn't this integrate\n> at the executor level instead?\n>\n\nCurrently the patch is designed to find out only the user executed\nstatements\nthat are successfully finished, (no internal statements that are executed\nfrom\nfunctions and etc).\n\nThe same question was asked by dilip also in earlier mails, may be now it is\nthe time that we can decide the approach of statement counts.\n\n\nRegards,\nHari Babu\nFujitsu Australia\n\nOn Thu, Apr 6, 2017 at 5:17 AM, Andres Freund <andres@anarazel.de> wrote:Hi,\n\n\nI'm somewhat inclined to think that this really would be better done in\nan extension like pg_stat_statements.Thanks for the review. \n> +}\n> +\n> +/*\n> + * Count SQL statement for pg_stat_sql view\n> + */\n> +void\n> +pgstat_count_sqlstmt(const char *commandTag)\n> +{\n> + PgStat_SqlstmtEntry *htabent;\n> + bool found;\n> +\n> + if (!pgstat_backend_sql)\n> + {\n> + /* First time through - initialize SQL statement stat table */\n> + HASHCTL hash_ctl;\n> +\n> + memset(&hash_ctl, 0, sizeof(hash_ctl));\n> + hash_ctl.keysize = NAMEDATALEN;\n> + hash_ctl.entrysize = sizeof(PgStat_SqlstmtEntry);\n> + pgstat_backend_sql = hash_create(\"SQL statement stat entries\",\n> + PGSTAT_SQLSTMT_HASH_SIZE,\n> + &hash_ctl,\n> + HASH_ELEM | HASH_BLOBS);\n> + }\n> +\n> + /* Get the stats entry for this SQL statement, create if necessary */\n> + htabent = hash_search(pgstat_backend_sql, commandTag,\n> + HASH_ENTER, &found);\n> + if (!found)\n> + htabent->count = 1;\n> + else\n> + htabent->count++;\n> +}\n\n\nThat's not really something in this patch, but a lot of this would be\nbetter if we didn't internally have command tags as strings, but as an\nenum. We should really only convert to a string with needed. That\nwe're doing string comparisons on Portal->commandTag is just plain bad.\n\nIf so, we could also make this whole patch a lot cheaper - have a fixed\nsize array that has an entry for every possible tag (possibly even in\nshared memory, and then use atomics there).During the development, I thought of using an array of all command tagsand update the counters using the tag name, but not like the enum values.Now I have to create a mapping array with enums and tag names for easiercounter updates. The problem in this approach is, whenever any additionof new commands, this mapping array needs to be updated with bothnew enum and new tag name, whereas with hash table approach, it worksfuture command additions also, but there is some performance overheadcompared to the array approach.I will modify the patch to use the array approach and provide it to the nextcommitfest. \n> +++ b/src/backend/tcop/postgres.c\n> @@ -1109,6 +1109,12 @@ exec_simple_query(const char *query_string)\n>\n> PortalDrop(portal, false);\n>\n> + /*\n> + * Count SQL statement for pg_stat_sql view\n> + */\n> + if (pgstat_track_sql)\n> + pgstat_count_sqlstmt(commandTag);\n> +\n\nIsn't doing this at the message level quite wrong? What about\nstatements executed from functions and such? Shouldn't this integrate\nat the executor level instead?Currently the patch is designed to find out only the user executed statementsthat are successfully finished, (no internal statements that are executed fromfunctions and etc).The same question was asked by dilip also in earlier mails, may be now it isthe time that we can decide the approach of statement counts.Regards,Hari BabuFujitsu Australia","msg_date":"Fri, 7 Apr 2017 15:19:28 +1000","msg_from":"Haribabu Kommi <kommi.haribabu@gmail.com>","msg_from_op":true,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"Dear Hackers,\n\nWe have resurrected this 2 year old \"SQL statements statistics counter\" proposal from Hari.\n\nThe attached patch addresses the previous review issues.\n\nThe commit-fest entry has been moved forward to the next commit-fest\nhttps://commitfest.postgresql.org/25/790/\n\nPlease review again, and consider if this is OK for \"Ready for Committer\" status.\n\nKind Regards\n--\nPeter Smith\nFujitsu Australia\n\n\n\n-----Original Message-----\nFrom: pgsql-hackers-owner@postgresql.org <pgsql-hackers-owner@postgresql.org> On Behalf Of Andres Freund\nSent: Thursday, 6 April 2017 5:18 AM\nTo: Haribabu Kommi <kommi.haribabu@gmail.com>\nCc: Michael Paquier <michael.paquier@gmail.com>; Dilip Kumar <dilipbalaut@gmail.com>; vinayak <Pokale_Vinayak_q3@lab.ntt.co.jp>; pgsql-hackers@postgresql.org\nSubject: Re: New SQL counter statistics view (pg_stat_sql)\n\nHi,\n\n\nI'm somewhat inclined to think that this really would be better done in an extension like pg_stat_statements.\n\n\nOn 2017-03-08 14:39:23 +1100, Haribabu Kommi wrote:\n> On Wed, Feb 1, 2017 at 3:13 PM, Michael Paquier \n> <michael.paquier@gmail.com>\n\n> + <varlistentry id=\"guc-track-sql\" xreflabel=\"track_sql\">\n> + <term><varname>track_sql</varname> (<type>boolean</type>)\n> + <indexterm>\n> + <primary><varname>track_sql</> configuration parameter</primary>\n> + </indexterm>\n> + </term>\n> + <listitem>\n> + <para>\n> + Enables collection of different SQL statement statistics that are\n> + executed on the instance. This parameter is off by default. Only\n> + superusers can change this setting.\n> + </para>\n> + </listitem>\n> + </varlistentry>\n> +\n\nI don't like this name much, seems a bit too generic to me. 'track_activities', 'track_io_timings' are at least a bit clearer.\nHow about track_statement_statistics + corresponding view/function renaming?\n\n\n> + <table id=\"pg-stat-sql-view\" xreflabel=\"pg_stat_sql\">\n> + <title><structname>pg_stat_sql</structname> View</title>\n> + <tgroup cols=\"3\">\n> + <thead>\n> + <row>\n> + <entry>Column</entry>\n> + <entry>Type</entry>\n> + <entry>Description</entry>\n> + </row>\n> + </thead>\n> +\n> + <tbody>\n> + <row>\n> + <entry><structfield>tag</></entry>\n> + <entry><type>text</></entry>\n> + <entry>Name of the SQL statement</entry>\n> + </row>\n\nIt's not really the \"name\" of a statement. Internally and IIRC elsewhere in the docs we describe something like this as tag?\n\n\n> +/* ----------\n> + * pgstat_send_sqlstmt(void)\n> + *\n> + *\t\tSend SQL statement statistics to the collector\n> + * ----------\n> + */\n> +static void\n> +pgstat_send_sqlstmt(void)\n> +{\n> +\tPgStat_MsgSqlstmt msg;\n> +\tPgStat_SqlstmtEntry *entry;\n> +\tHASH_SEQ_STATUS hstat;\n> +\n> +\tif (pgstat_backend_sql == NULL)\n> +\t\treturn;\n> +\n> +\tpgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_SQLSTMT);\n> +\tmsg.m_nentries = 0;\n> +\n> +\thash_seq_init(&hstat, pgstat_backend_sql);\n> +\twhile ((entry = (PgStat_SqlstmtEntry *) hash_seq_search(&hstat)) != NULL)\n> +\t{\n> +\t\tPgStat_SqlstmtEntry *m_ent;\n> +\n> +\t\t/* Skip it if no counts accumulated since last time */\n> +\t\tif (entry->count == 0)\n> +\t\t\tcontinue;\n> +\n> +\t\t/* need to convert format of time accumulators */\n> +\t\tm_ent = &msg.m_entry[msg.m_nentries];\n> +\t\tmemcpy(m_ent, entry, sizeof(PgStat_SqlstmtEntry));\n> +\n> +\t\tif (++msg.m_nentries >= PGSTAT_NUM_SQLSTMTS)\n> +\t\t{\n> +\t\t\tpgstat_send(&msg, offsetof(PgStat_MsgSqlstmt, m_entry[0]) +\n> +\t\t\t\t\t\tmsg.m_nentries * sizeof(PgStat_SqlstmtEntry));\n> +\t\t\tmsg.m_nentries = 0;\n> +\t\t}\n> +\n> +\t\t/* reset the entry's count */\n> +\t\tentry->count = 0;\n> +\t}\n> +\n> +\tif (msg.m_nentries > 0)\n> +\t\tpgstat_send(&msg, offsetof(PgStat_MsgSqlstmt, m_entry[0]) +\n> +\t\t\t\t\tmsg.m_nentries * sizeof(PgStat_SqlstmtEntry));\n\nHm. So pgstat_backend_sql is never deleted, which'll mean that if a backend lives long we'll continually iterate over a lot of 0 entries. I think performance evaluation of this feature should take that into account.\n\n\n> +}\n> +\n> +/*\n> + * Count SQL statement for pg_stat_sql view */ void \n> +pgstat_count_sqlstmt(const char *commandTag) {\n> +\tPgStat_SqlstmtEntry *htabent;\n> +\tbool\t\tfound;\n> +\n> +\tif (!pgstat_backend_sql)\n> +\t{\n> +\t\t/* First time through - initialize SQL statement stat table */\n> +\t\tHASHCTL\t\thash_ctl;\n> +\n> +\t\tmemset(&hash_ctl, 0, sizeof(hash_ctl));\n> +\t\thash_ctl.keysize = NAMEDATALEN;\n> +\t\thash_ctl.entrysize = sizeof(PgStat_SqlstmtEntry);\n> +\t\tpgstat_backend_sql = hash_create(\"SQL statement stat entries\",\n> +\t\t\t\t\t\t\t\t\t\t PGSTAT_SQLSTMT_HASH_SIZE,\n> +\t\t\t\t\t\t\t\t\t\t &hash_ctl,\n> +\t\t\t\t\t\t\t\t\t\t HASH_ELEM | HASH_BLOBS);\n> +\t}\n> +\n> +\t/* Get the stats entry for this SQL statement, create if necessary */\n> +\thtabent = hash_search(pgstat_backend_sql, commandTag,\n> +\t\t\t\t\t\t HASH_ENTER, &found);\n> +\tif (!found)\n> +\t\thtabent->count = 1;\n> +\telse\n> +\t\thtabent->count++;\n> +}\n\n\nThat's not really something in this patch, but a lot of this would be better if we didn't internally have command tags as strings, but as an enum. We should really only convert to a string with needed. That we're doing string comparisons on Portal->commandTag is just plain bad.\n\nIf so, we could also make this whole patch a lot cheaper - have a fixed size array that has an entry for every possible tag (possibly even in shared memory, and then use atomics there).\n\n> +++ b/src/backend/tcop/postgres.c\n> @@ -1109,6 +1109,12 @@ exec_simple_query(const char *query_string)\n> \n> \t\tPortalDrop(portal, false);\n> \n> +\t\t/*\n> +\t\t * Count SQL statement for pg_stat_sql view\n> +\t\t */\n> +\t\tif (pgstat_track_sql)\n> +\t\t\tpgstat_count_sqlstmt(commandTag);\n> +\n\nIsn't doing this at the message level quite wrong? What about statements executed from functions and such? Shouldn't this integrate at the executor level instead?\n\n\n> \t\tif (IsA(parsetree->stmt, TransactionStmt))\n> \t\t{\n> \t\t\t/*\n> @@ -1991,6 +1997,29 @@ exec_execute_message(const char *portal_name, \n> long max_rows)\n> \n> \t(*receiver->rDestroy) (receiver);\n> \n> +\t/*\n> +\t * Count SQL Statement for pgx_stat_sql\n> +\t */\n> +\tif (pgstat_track_sql)\n> +\t{\n> +\t\tCachedPlanSource *psrc = NULL;\n> +\n> +\t\tif (portal->prepStmtName)\n> +\t\t{\n> +\t\t\tPreparedStatement *pstmt;\n> +\n> +\t\t\tpstmt = FetchPreparedStatement(portal->prepStmtName, false);\n> +\t\t\tif (pstmt)\n> +\t\t\t\tpsrc = pstmt->plansource;\n> +\t\t}\n> +\t\telse\n> +\t\t\tpsrc = unnamed_stmt_psrc;\n> +\n> +\t\t/* psrc should not be NULL here */\n> +\t\tif (psrc && psrc->commandTag && !execute_is_fetch && pgstat_track_sql)\n> +\t\t\tpgstat_count_sqlstmt(psrc->commandTag);\n\nWait, we're re-fetching the statement here? That doesn't sound alright.\n\n\n> +Datum\n> +pg_stat_sql(PG_FUNCTION_ARGS)\n> +{\n> +\tTupleDesc\ttupdesc;\n> +\tDatum\t\tvalues[2];\n> +\tbool\t\tnulls[2];\n> +\tReturnSetInfo *rsi;\n> +\tMemoryContext old_cxt;\n> +\tTuplestorestate *tuple_store;\n> +\tPgStat_SqlstmtEntry *sqlstmtentry;\n> +\tHASH_SEQ_STATUS hstat;\n> +\n> +\t/* Function call to let the backend read the stats file */\n> +\tpgstat_fetch_global();\n> +\n> +\t/* Initialize values and nulls arrays */\n> +\tMemSet(values, 0, sizeof(values));\n> +\tMemSet(nulls, 0, sizeof(nulls));\n\nwhy not instead declare values[2] = {0}? Obviously not important.\n\n\n> +\ttuple_store =\n> +\t\ttuplestore_begin_heap(rsi->allowedModes & SFRM_Materialize_Random,\n> +\t\t\t\t\t\t\t false, work_mem);\n\nHuh, why do we need random?\n\n\n\nGreetings,\n\nAndres Freund\n\n\n--\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Thu, 17 Oct 2019 02:22:32 +0000","msg_from":"\"Smith, Peter\" <peters@fast.au.fujitsu.com>","msg_from_op":false,"msg_subject":"RE: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On Thu, Oct 17, 2019 at 3:22 PM Smith, Peter <peters@fast.au.fujitsu.com> wrote:\n> We have resurrected this 2 year old \"SQL statements statistics counter\" proposal from Hari.\n>\n> The attached patch addresses the previous review issues.\n\nHi,\n\nNo comment on the patch but I noticed that the documentation changes\ndon't build. Please make sure you can \"make docs\" successfully,\nhaving installed the documentation tools[1].\n\n[1] https://www.postgresql.org/docs/devel/docguide-toolsets.html\n\n\n","msg_date":"Mon, 4 Nov 2019 15:42:39 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"From: Thomas Munro <thomas.munro@gmail.com> Sent: Monday, 4 November 2019 1:43 PM\r\n\r\n> No comment on the patch but I noticed that the documentation changes don't build. Please make sure you can \"make docs\" successfully, having installed the documentation tools[1].\r\n\r\nThanks for the feedback. An updated patch which fixes the docs issue is attached.\r\n\r\nKind Regards.\r\nPeter Smith\r\n---\r\nFujitsu Australia","msg_date":"Wed, 13 Nov 2019 03:51:05 +0000","msg_from":"\"Smith, Peter\" <peters@fast.au.fujitsu.com>","msg_from_op":false,"msg_subject":"RE: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On 2019-Nov-13, Smith, Peter wrote:\n\n> From: Thomas Munro <thomas.munro@gmail.com> Sent: Monday, 4 November 2019 1:43 PM\n> \n> > No comment on the patch but I noticed that the documentation changes don't build. Please make sure you can \"make docs\" successfully, having installed the documentation tools[1].\n> \n> Thanks for the feedback. An updated patch which fixes the docs issue is attached.\n\nSo, who is updating this patch for the new cmdtaglist.h stuff?\n\n-- \n�lvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Tue, 3 Mar 2020 23:50:04 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"\n\n> On Mar 3, 2020, at 6:50 PM, Alvaro Herrera <alvherre@2ndquadrant.com> wrote:\n> \n> On 2019-Nov-13, Smith, Peter wrote:\n> \n>> From: Thomas Munro <thomas.munro@gmail.com> Sent: Monday, 4 November 2019 1:43 PM\n>> \n>>> No comment on the patch but I noticed that the documentation changes don't build. Please make sure you can \"make docs\" successfully, having installed the documentation tools[1].\n>> \n>> Thanks for the feedback. An updated patch which fixes the docs issue is attached.\n> \n> So, who is updating this patch for the new cmdtaglist.h stuff?\n\nI have a much altered (competing?) patch inspired by pg_stat_sql. I am working to integrate the changes you made to my commandtag patch into my commandstats patch before posting it.\n\nIt might be good if somebody else tackled the pg_stat_sql patch in case my version is rejected.\n\nI don't want to highjack this thread to talk in detail about the other patch, so I'll just give an overview that might be useful for anybody thinking about putting effort into committing this patch first. \n\nThe motivation for the commandtag patch that you've committed (thanks again!) was to make the commandstats patch possible. I didn't like the way pg_stat_sql stored the tags as strings, but I also still don't like the way it needs to take a lock every time a backend increments the counter for a command. That is bad, I think, because the lock overhead for an otherwise fast command (like moving a cursor) may be high enough to discourage users from turning this feature on in production.\n\nI solve that by trading off speed and memory. For most commandtags, there is a single uint64 counter, and incrementing the counter requires taking a lock. This is very similar to how the other patch works. For the few commandtags that we deem worthy of the special treatment, there is an additional uint32 counter *per backend* reserved in shared memory. Backends update this counter using atomics, but do not require locks. When the counter is in danger of overflow, locks are taken to roll the count into the uint64 counters and zero out the backend-specific uint32 counter. I have roughly 20 command tags receiving this special treatment, but the design of the patch doesn't make it too hard to change the exact list of tags worthy of it, so I don't plan to stress too much about the exact set of tags. I'm sure you'll all tell me which ones you think do or do not deserve the treatment.\n\nIf it isn't obvious, the logic of having this treatment for only a subset of tags is that 192 tags * 4 bytes per counter * MaxBackends gets expensive for folks running a high number of backends. Taking that from 192 tags down to one or two dozen seems worthwhile.\n\nI expect to post on a separate thread before the day is over.\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Wed, 4 Mar 2020 13:52:56 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"Hackers,\n\nas mentioned in [1], I have created an implementation of command counter statistics very similar in purpose to the one already pending in the commitfest going by the name \"pg_stat_sql\". I don't really care if this implementation is seen as building on that one or as separate, but I was worried about hijacking that thread, so I'm starting this thead. There are elements of this patch that borrowed from that one, so if this is accepted, authorship should reflect that.\n\nSee the work by Haribabu Kommi (haribabu) at https://commitfest.postgresql.org/27/790/\n\nThe two main differences are that\n\n(1) This implementation is based on commandtags as enums, not strings and\n(2) This implementation uses techniques to reduce lock contention\n\nI think (2) is the more important part.\n\n[1] https://www.postgresql.org/message-id/28F46159-3764-421B-904B-004DEA339310%40enterprisedb.com\n\n\n\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Wed, 4 Mar 2020 19:43:25 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":false,"msg_subject":"Command statistics system (cmdstats)"},{"msg_contents":"> On Mar 4, 2020, at 7:43 PM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n> \n> Hackers,\n> \n> as mentioned in [1], I have created an implementation of command counter statistics very similar in purpose to the one already pending in the commitfest going by the name \"pg_stat_sql\". I don't really care if this implementation is seen as building on that one or as separate, but I was worried about hijacking that thread, so I'm starting this thead. There are elements of this patch that borrowed from that one, so if this is accepted, authorship should reflect that.\n> \n> See the work by Haribabu Kommi (haribabu) at https://commitfest.postgresql.org/27/790/\n> \n> The two main differences are that\n> \n> (1) This implementation is based on commandtags as enums, not strings and\n> (2) This implementation uses techniques to reduce lock contention\n> \n> I think (2) is the more important part.\n> \n> [1] https://www.postgresql.org/message-id/28F46159-3764-421B-904B-004DEA339310%40enterprisedb.com\n> \n> <v1-0001-Implementing-the-cmdstats-subsystem.patch>\n> —\n\nRebased patch against master caa3c4242cf86322e2ed0c86199e6462a2c41565.\n\n\n\n\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Tue, 2 Jun 2020 09:58:25 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Command statistics system (cmdstats)"},{"msg_contents":"> On Jun 2, 2020, at 9:58 AM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n> \n>> \n>> On Mar 4, 2020, at 7:43 PM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n>> \n>> Hackers,\n>> \n>> as mentioned in [1], I have created an implementation of command counter statistics very similar in purpose to the one already pending in the commitfest going by the name \"pg_stat_sql\". I don't really care if this implementation is seen as building on that one or as separate, but I was worried about hijacking that thread, so I'm starting this thead. There are elements of this patch that borrowed from that one, so if this is accepted, authorship should reflect that.\n>> \n>> See the work by Haribabu Kommi (haribabu) at https://commitfest.postgresql.org/27/790/\n>> \n>> The two main differences are that\n>> \n>> (1) This implementation is based on commandtags as enums, not strings and\n>> (2) This implementation uses techniques to reduce lock contention\n>> \n>> I think (2) is the more important part.\n>> \n>> [1] https://www.postgresql.org/message-id/28F46159-3764-421B-904B-004DEA339310%40enterprisedb.com\n>> \n>> <v1-0001-Implementing-the-cmdstats-subsystem.patch>\n>> —\n> \n> Rebased patch against master caa3c4242cf86322e2ed0c86199e6462a2c41565.\n\n\n\n\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Tue, 2 Jun 2020 10:02:24 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Command statistics system (cmdstats)"},{"msg_contents":"On Wed, Mar 4, 2020 at 10:43 PM Mark Dilger\n<mark.dilger@enterprisedb.com> wrote:\n> The two main differences are that\n>\n> (1) This implementation is based on commandtags as enums, not strings and\n> (2) This implementation uses techniques to reduce lock contention\n>\n> I think (2) is the more important part.\n\nMy spidey sense is tingling here, telling me that we need some actual\nbenchmarking. Like, suppose we test the two patches under normal cases\nand under cases that are constructed to be as bad as possible for each\nof them. Or suppose we test this patch with the lock mitigation\nstrategies and then remove the mitigations for some inexpensive\ncommand (e.g. SHOW) and then use pgbench to spam that command. Like\nyou, I suspect that the locking mitigations are important in some\nworkloads, but it would be good to have some figures to back that out,\nas well as to find out whether there's still too much overhead.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Wed, 10 Jun 2020 13:45:02 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Command statistics system (cmdstats)"},{"msg_contents":"On Wed, Jun 10, 2020 at 01:45:02PM -0400, Robert Haas wrote:\n> My spidey sense is tingling here, telling me that we need some actual\n> benchmarking. Like, suppose we test the two patches under normal cases\n> and under cases that are constructed to be as bad as possible for each\n> of them. Or suppose we test this patch with the lock mitigation\n> strategies and then remove the mitigations for some inexpensive\n> command (e.g. SHOW) and then use pgbench to spam that command. Like\n> you, I suspect that the locking mitigations are important in some\n> workloads, but it would be good to have some figures to back that out,\n> as well as to find out whether there's still too much overhead.\n\nThis patch has not received any replies after this comment for three\nmonths, so I am marking it as returned with feedback. I agree that\nthis should be benchmarked carefully.\n--\nMichael","msg_date":"Thu, 17 Sep 2020 13:51:23 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Command statistics system (cmdstats)"},{"msg_contents":"On Tue, Mar 03, 2020 at 11:50:04PM -0300, Alvaro Herrera wrote:\n> So, who is updating this patch for the new cmdtaglist.h stuff?\n\nThis patch had no activity for months, so I am marking it as returned\nwith feedback.\n--\nMichael","msg_date":"Thu, 17 Sep 2020 14:01:19 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: New SQL counter statistics view (pg_stat_sql)"},{"msg_contents":"On 2020-Sep-17, Michael Paquier wrote:\n\n> On Wed, Jun 10, 2020 at 01:45:02PM -0400, Robert Haas wrote:\n> > My spidey sense is tingling here, telling me that we need some actual\n> > benchmarking.\n> \n> This patch has not received any replies after this comment for three\n> months, so I am marking it as returned with feedback. I agree that\n> this should be benchmarked carefully.\n\nIt seems fine to mark the patch as RwF at the end of commitfest, but\nthat's two weeks away. I don't understand what is accomplished by doing\nit ahead of time, other than alienating the patch authors.\n\n-- \n�lvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Thu, 17 Sep 2020 09:40:20 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Command statistics system (cmdstats)"},{"msg_contents":"\n\n> On Sep 17, 2020, at 5:40 AM, Alvaro Herrera <alvherre@2ndquadrant.com> wrote:\n> \n> On 2020-Sep-17, Michael Paquier wrote:\n> \n>> On Wed, Jun 10, 2020 at 01:45:02PM -0400, Robert Haas wrote:\n>>> My spidey sense is tingling here, telling me that we need some actual\n>>> benchmarking.\n>> \n>> This patch has not received any replies after this comment for three\n>> months, so I am marking it as returned with feedback. I agree that\n>> this should be benchmarked carefully.\n\nYes, I have prioritized a couple other patches over this one, with benchmarking this patch lower down my priority list. Thank you, Michael, for the reminder!\n\n> \n> It seems fine to mark the patch as RwF at the end of commitfest, but\n> that's two weeks away. I don't understand what is accomplished by doing\n> it ahead of time, other than alienating the patch authors.\n\nThanks, Álvaro! I do not feel alienated, and I hope Haribabu Kommi does not either.\n\nHaribabu, do you have an opinion on which of the two patches should proceed? I don't recall seeing a response from you about whether you liked what I did with your patch. Perhaps you were waiting on the benchmarking results?\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Thu, 17 Sep 2020 08:03:55 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Command statistics system (cmdstats)"},{"msg_contents":"On Thu, Sep 17, 2020 at 08:03:55AM -0700, Mark Dilger wrote:\n> Yes, I have prioritized a couple other patches over this one, with\n> benchmarking this patch lower down my priority list. Thank you,\n> Michael, for the reminder!\n\nBased on the low level of activity and the fact that the patch was\nmarked as waiting on author for a couple of weeks, it looks like\nlittle could be achieved by the end of the CF, and the attention was\nelsewhere, so it looked better (and it still does IMO) to give more\nattention to the remaining 170~ patches that are still lying on the CF\napp.\n--\nMichael","msg_date":"Fri, 18 Sep 2020 14:55:43 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Command statistics system (cmdstats)"},{"msg_contents":"On 2020-Sep-18, Michael Paquier wrote:\n\n> Based on the low level of activity and the fact that the patch was\n> marked as waiting on author for a couple of weeks, it looks like\n> little could be achieved by the end of the CF, and the attention was\n> elsewhere, so it looked better (and it still does IMO) to give more\n> attention to the remaining 170~ patches that are still lying on the CF\n> app.\n\n\"A couple of weeks\" of inactivity is not sufficient, in my view, to boot\na patch out of the commitfest process. Whenever the patch is\nresurrected, it will be a new entry which won't have the history that it\nhad accumulated in the long time since it was created -- which biases\nit against other new patches.\n\nI'm not really arguing about this one patch only, but more generally\nabout the process you're following. The problem is that you as CFM are\nimposing your personal priorities on the whole process by punting\npatches ahead of time -- you are saying that nobody else should be\nlooking at updating this patch because it is now closed. It seems fair\nto do so at the end of the commitfest, but it does not seem fair to do\nit when it's still mid-cycle.\n\nPutting also in perspective the history that the patch had prior to your\nunfairly closing it, there was a lot of effort in getting it reviewed to\na good state and updated by the author -- yet a single unsubstantiated\ncomment, without itself a lot of effort on the reviewer's part, that\n\"maybe it has a perf drawback\" is sufficient to get it booted? It\ndoesn't seem appropriate.\n\n-- \n�lvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Mon, 21 Sep 2020 10:41:46 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Command statistics system (cmdstats)"},{"msg_contents":"On Mon, Sep 21, 2020 at 9:41 AM Alvaro Herrera <alvherre@2ndquadrant.com> wrote:\n> \"A couple of weeks\" of inactivity is not sufficient, in my view, to boot\n> a patch out of the commitfest process. Whenever the patch is\n> resurrected, it will be a new entry which won't have the history that it\n> had accumulated in the long time since it was created -- which biases\n> it against other new patches.\n\nAs Michael said, it had been inactive for three *months*. I don't\nthink there's a big problem here. I think that the redesign of the\nCommitFest application encourages carrying things along from CF to CF\nforever, instead of getting rid of things that aren't wanted or aren't\nmaking any progress. That's work we don't need. There ought to be a\nway to mark a patch RwF when nothing's happen that lets it be revived\nlater if and when someone gets around to resubmitting, but I think\nthat's not possible now.\n\nThen, too, since we have email integration now, maybe we ought to do\nthat automatically if the thread isn't getting updated and the patch\nis setting there waiting on author. It's a real waste of CFM time to\nchase down and kick out obviously-inactive patches, and if the CFM is\ngoing to get flack for it then that's even worse. Like, do we have 170\npatches now because we have more activity than a few years ago, or\njust because we've become more reluctant to boot things? I don't know,\nbut to the extent it's from unwillingness to boot things, I think\nthat's bad. Discouraging contributors is not good, but wasting CFM and\ncommiter time is also bad. You've got to draw a line someplace or\nyou'll go nuts.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Mon, 21 Sep 2020 17:19:07 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Command statistics system (cmdstats)"},{"msg_contents":"On Mon, Sep 21, 2020 at 10:41:46AM -0300, Alvaro Herrera wrote:\n> \"A couple of weeks\" of inactivity is not sufficient, in my view, to boot\n> a patch out of the commitfest process. Whenever the patch is\n> resurrected, it will be a new entry which won't have the history that it\n> had accumulated in the long time since it was created -- which biases\n> it against other new patches.\n\nAny of the patches that have been marked as RwF last week have been\nwaiting on author since at least the end of the commit fest of July,\nwith zero actions taken on them. In my experience, this points\nstrongly to the fact that these are unlikely going to be worked on\nactively in a short time frame.\n\n> I'm not really arguing about this one patch only, but more generally\n> about the process you're following. The problem is that you as CFM are\n> imposing your personal priorities on the whole process by punting\n> patches ahead of time -- you are saying that nobody else should be\n> looking at updating this patch because it is now closed. It seems fair\n> to do so at the end of the commitfest, but it does not seem fair to do\n> it when it's still mid-cycle.\n\nI think that this remark is a bit unfair. When it comes to any patch\nin the commit fest, and I have managed a couple of them over the\nyears, I have tried to keep a neutral view for all of them, meaning\nthat I only let the numbers speak by themselves, mostly:\n- When has a patch lastly been updated.\n- What's the current state in the CF app If the patch had no reviews\nand still in \"Needs review\", simply switch it to \"waiting on author\".\nIf the patch has been waiting on author for more than two weeks, it\nwould get marked as RwF at the end of the CF. There are also cases\nwhere the status of the patch is incorrect, mostly that a patch\nwaiting on author because of a review was not updated as such in the\nCF app. In this case, and in the middle of a CF, this would get get\nmarked as Rwf, but the author would get a reminder of that.\n\nAny of the patches that have been discarded by me last week were\nwaiting in the stack for much, much, longer than that, and I think\nthat it does not help reviewers and committers in keeping this stuff\nlonger than necessary because this mainly bloats the CF app in a\nnon-necessary way, hiding patches that should deserve more attention.\nSome of those patches have been left idle for half a year.\n\nAnother point that I'd like to make here is that my intention was to\nlift the curve here (a term quite fashionable this year particularly,\nisn't it?), due to the high number of patches to handle in a single\nCF, and the shortage of hands actually doing this kind of work.\n\nThen, let's be clear about one last thing. If at *any* moment people\nhere feel that at least the state of *one* patch of this CF has been\nchanged based on some of my \"personal\" priority views, I will\nimmediately stop any action as CFM and resign from it. So, if that's\nthe case, please feel free to speak up here or on any of the threads\nrelated to those patches.\n\n> Putting also in perspective the history that the patch had prior to your\n> unfairly closing it, there was a lot of effort in getting it reviewed to\n> a good state and updated by the author -- yet a single unsubstantiated\n> comment, without itself a lot of effort on the reviewer's part, that\n> \"maybe it has a perf drawback\" is sufficient to get it booted? It\n> doesn't seem appropriate.\n\nI agree on the fact that not being able to recreate a new CF entry\nthat keeps the history of an old one, with all the past threads or\nannotations, could be helpful in some cases. Now, I have seen as well \ncases where a patch got around for so long with a very long thread\nmade things confusing, and that people also like starting a new thread\nwith a summary of the past state as a first message to attempt\nclarifying things. So both have advantages.\n--\nMichael","msg_date":"Wed, 23 Sep 2020 10:41:10 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Command statistics system (cmdstats)"},{"msg_contents":"On Wed, Sep 23, 2020 at 10:41:10AM +0900, Michael Paquier wrote:\n> I think that this remark is a bit unfair. When it comes to any patch\n> in the commit fest, and I have managed a couple of them over the\n> years, I have tried to keep a neutral view for all of them, meaning\n> that I only let the numbers speak by themselves, mostly:\n> - When has a patch lastly been updated.\n> - What's the current state in the CF app If the patch had no reviews\n> and still in \"Needs review\", simply switch it to \"waiting on author\".\n\nSorry, forgot to mention here that this is in the case where a patch\ndoes not apply anymore or that the CF bot complains, requiring a check\nof the patch from the author.\n\n> If the patch has been waiting on author for more than two weeks, it\n> would get marked as RwF at the end of the CF. There are also cases\n> where the status of the patch is incorrect, mostly that a patch\n> waiting on author because of a review was not updated as such in the\n> CF app. In this case, and in the middle of a CF, this would get get\n> marked as Rwf, but the author would get a reminder of that.\n\nLast sentence here is incorrect as well: s/get get/not get/, and the\nauthor would be reminded to update his/her patch.\n--\nMichael","msg_date":"Wed, 23 Sep 2020 11:12:14 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Command statistics system (cmdstats)"},{"msg_contents":"> >> On Mar 4, 2020, at 7:43 PM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n\n> >> as mentioned in [1], I have created an implementation of command\n> >> counter statistics very similar in purpose to the one already\n> >> pending in the commitfest going by the name \"pg_stat_sql\". I don't\n> >> really care if this implementation is seen as building on that one\n> >> or as separate, but I was worried about hijacking that thread, so\n> >> I'm starting this thead. There are elements of this patch that\n> >> borrowed from that one, so if this is accepted, authorship should\n> >> reflect that.\n\nHello, was this abandoned or is there another thread somewhere? The\npatch in the email I'm replying to doesn't apply, but the conflicts\ndon't look very bad.\n\n-- \n�lvaro Herrera Valdivia, Chile\n\n\n","msg_date":"Fri, 28 May 2021 09:42:24 -0400","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Command statistics system (cmdstats)"},{"msg_contents":"\n\n> On May 28, 2021, at 6:42 AM, Alvaro Herrera <alvherre@2ndquadrant.com> wrote:\n> \n>>>> On Mar 4, 2020, at 7:43 PM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n> \n>>>> as mentioned in [1], I have created an implementation of command\n>>>> counter statistics very similar in purpose to the one already\n>>>> pending in the commitfest going by the name \"pg_stat_sql\". I don't\n>>>> really care if this implementation is seen as building on that one\n>>>> or as separate, but I was worried about hijacking that thread, so\n>>>> I'm starting this thead. There are elements of this patch that\n>>>> borrowed from that one, so if this is accepted, authorship should\n>>>> reflect that.\n> \n> Hello, was this abandoned or is there another thread somewhere? The\n> patch in the email I'm replying to doesn't apply, but the conflicts\n> don't look very bad.\n\nI abandoned it for v14, choosing instead to work on amcheck features. Would you like me to take this up again for v15?\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Fri, 28 May 2021 08:30:49 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Command statistics system (cmdstats)"}],"string":"[\n {\n \"msg_contents\": \"This is a new statistics view that is used to provide the number of\\nSQL operations that are\\nhappened on a particular interval of time. This view is useful for the\\nsystem to find out the\\npattern of the operations that are happening in the instance during\\nparticular interval of\\ntime.\\n\\nFollowing is the more or less columns and their details of the pg_stat_sql view.\\n\\npostgres=# \\\\d pg_stat_sql\\n View \\\"pg_catalog.pg_stat_sql\\\"\\n Column | Type | Modifiers\\n-------------+--------------------------+-----------\\n selects | bigint |\\n inserts | bigint |\\n deletes | bigint |\\n updates | bigint |\\n declares | bigint |\\n fetches | bigint |\\n copies | bigint |\\n reindexes | bigint |\\n truncates | bigint |\\n stats_reset | timestamp with time zone |\\n\\nThe SQL counters gets updated only for the external SQL queries, not for\\nthe SQL queries that are generated internally. The counters gets updated\\nat exec_simple_query and exec_execute_message functions.\\n\\nUser can reset the SQL counter statistics using an exposed function.\\nThe Stats collection can be turned on/off using a GUC also.\\n\\nAny comments/objections in providing a patch for the same?\\n\\nRegards,\\nHari Babu\\nFujitsu Australia\\n\\n\\n-- \\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, 20 Aug 2016 22:21:36 +1000\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"Haribabu Kommi <kommi.haribabu@gmail.com> writes:\\n> This is a new statistics view that is used to provide the number of\\n> SQL operations that are\\n> happened on a particular interval of time. This view is useful for the\\n> system to find out the\\n> pattern of the operations that are happening in the instance during\\n> particular interval of\\n> time.\\n\\n> Following is the more or less columns and their details of the pg_stat_sql view.\\n\\n> postgres=# \\\\d pg_stat_sql\\n> View \\\"pg_catalog.pg_stat_sql\\\"\\n> Column | Type | Modifiers\\n> -------------+--------------------------+-----------\\n> selects | bigint |\\n> inserts | bigint |\\n> deletes | bigint |\\n> updates | bigint |\\n> declares | bigint |\\n> fetches | bigint |\\n> copies | bigint |\\n> reindexes | bigint |\\n> truncates | bigint |\\n> stats_reset | timestamp with time zone |\\n\\n\\n1. This set of counters seems remarkably random. Why, for instance,\\ncount reindexes but not original index creations?\\n\\n2. What will you do with cases such as SELECTs containing modifying CTEs?\\nOr EXPLAIN ANALYZE? Does CLUSTER count as a REINDEX? Does REFRESH\\nMATERIALIZED VIEW count as a SELECT? (Or maybe it's an INSERT?)\\n\\nIf you're going to be selective about what you count, you're forever\\ngoing to be fielding complaints from users who are unhappy that you\\ndidn't count some statement type they care about, or feel that you\\nmisclassified some complex case.\\n\\nI'm inclined to suggest you forget this approach and propose a single\\ncounter for \\\"SQL commands executed\\\", which avoids all of the above\\ndefinitional problems. People who need more detail than that are\\nprobably best advised to look to contrib/pg_stat_statements, anyway.\\n\\nAlso, if you do that, there hardly seems a need for a whole new view.\\nYou could just add a column to pg_stat_database. The incremental\\nmaintenance effort doesn't seem enough to justify its own GUC, either.\\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, 20 Aug 2016 11:17:01 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Sun, Aug 21, 2016 at 1:17 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Haribabu Kommi <kommi.haribabu@gmail.com> writes:\\n>> This is a new statistics view that is used to provide the number of\\n>> SQL operations that are\\n>> happened on a particular interval of time. This view is useful for the\\n>> system to find out the\\n>> pattern of the operations that are happening in the instance during\\n>> particular interval of\\n>> time.\\n>\\n> 1. This set of counters seems remarkably random. Why, for instance,\\n> count reindexes but not original index creations?\\n\\nI thought of adding the columns to the view that are mostly used(excluding\\nall DDL commands), otherwise the view columns fill with all the commands\\nthat are supported by PostgreSQL.\\n\\n> 2. What will you do with cases such as SELECTs containing modifying CTEs?\\n> Or EXPLAIN ANALYZE? Does CLUSTER count as a REINDEX? Does REFRESH\\n> MATERIALIZED VIEW count as a SELECT? (Or maybe it's an INSERT?)\\n\\nCluster commands will be calculated as clusters, explain needs some\\nlogic to identify what SQL operation exactly. This is because of using\\nnodeTag data from the parseTree structure to identify what type of SQL\\nstatement it is.\\n\\nThe counters will not updated for any SQL type that is not as part of the view.\\n\\n> If you're going to be selective about what you count, you're forever\\n> going to be fielding complaints from users who are unhappy that you\\n> didn't count some statement type they care about, or feel that you\\n> misclassified some complex case.\\n\\nYes, I agree that users may complain regarding with limited columns that are not\\nsufficient for their use. But with the mostly used columns, this view\\nmay be useful for\\nsome of the users to find out the pattern of the SQL operations that\\nare happening\\non the instance.\\n\\n> I'm inclined to suggest you forget this approach and propose a single\\n> counter for \\\"SQL commands executed\\\", which avoids all of the above\\n> definitional problems. People who need more detail than that are\\n> probably best advised to look to contrib/pg_stat_statements, anyway.\\n\\nI will add a new column to the pg_stat_database, to track the number of SQL\\noperations that are happened on this particular database.\\n\\nRegards,\\nHari Babu\\nFujitsu Australia\\n\\n\\n-- \\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 Aug 2016 11:05:44 +1000\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> I'm inclined to suggest you forget this approach and propose a single\\n> counter for \\\"SQL commands executed\\\", which avoids all of the above\\n> definitional problems. People who need more detail than that are\\n> probably best advised to look to contrib/pg_stat_statements, anyway.\\n\\nI disagree. I think SQL commands executed, lumping absolutely\\neverything together, really isn't much use. Haribabu's categorization\\nscheme seems to need some work, but the idea of categorizing\\nstatements by type and counting executions per type seems very\\nreasonable.\\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, 22 Aug 2016 13:54:43 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On 2016-08-22 13:54:43 -0400, Robert Haas wrote:\\n> On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> > I'm inclined to suggest you forget this approach and propose a single\\n> > counter for \\\"SQL commands executed\\\", which avoids all of the above\\n> > definitional problems. People who need more detail than that are\\n> > probably best advised to look to contrib/pg_stat_statements, anyway.\\n> \\n> I disagree. I think SQL commands executed, lumping absolutely\\n> everything together, really isn't much use.\\n\\nI'm inclined to agree. I think that's a quite useful stat when looking\\nat an installation one previously didn't have a lot of interaction with.\\n\\n\\n> Haribabu's categorization\\n> scheme seems to need some work, but the idea of categorizing\\n> statements by type and counting executions per type seems very\\n> reasonable.\\n\\nI'd consider instead using something like COALESCE(commandType,\\nnodeTag(Query->utilityStmt)) as the categories. Not sure if I'd even\\npivot that.\\n\\nAndres\\n\\n\\n-- \\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 Aug 2016 10:59:12 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> On 2016-08-22 13:54:43 -0400, Robert Haas wrote:\\n>> On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>>> I'm inclined to suggest you forget this approach and propose a single\\n>>> counter for \\\"SQL commands executed\\\", which avoids all of the above\\n>>> definitional problems. People who need more detail than that are\\n>>> probably best advised to look to contrib/pg_stat_statements, anyway.\\n\\n>> I disagree. I think SQL commands executed, lumping absolutely\\n>> everything together, really isn't much use.\\n\\n> I'm inclined to agree. I think that's a quite useful stat when looking\\n> at an installation one previously didn't have a lot of interaction with.\\n\\nWell, let's at least have an \\\"other\\\" category so you can add up the\\ncounters and get a meaningful total.\\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, 22 Aug 2016 16:27:06 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On 23/08/16 08:27, Tom Lane wrote:\\n> Andres Freund <andres@anarazel.de> writes:\\n>> On 2016-08-22 13:54:43 -0400, Robert Haas wrote:\\n>>> On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>>>> I'm inclined to suggest you forget this approach and propose a single\\n>>>> counter for \\\"SQL commands executed\\\", which avoids all of the above\\n>>>> definitional problems. People who need more detail than that are\\n>>>> probably best advised to look to contrib/pg_stat_statements, anyway.\\n>>> I disagree. I think SQL commands executed, lumping absolutely\\n>>> everything together, really isn't much use.\\n>> I'm inclined to agree. I think that's a quite useful stat when looking\\n>> at an installation one previously didn't have a lot of interaction with.\\n> Well, let's at least have an \\\"other\\\" category so you can add up the\\n> counters and get a meaningful total.\\n>\\n> \\t\\t\\tregards, tom lane\\n>\\n>\\nInitially I thought of something I thought was factious, but then \\nrealized it might actually be both practicable & useful...\\n\\nHow about 2 extra categories (if appropriate!!!):\\n\\n 1. Things that actually might be sort of as fitting in 2 or more of the\\n existing categories, or there is an ambiguity as to which category\\n is appropriate.\\n\\n 2. Things that don't fit into any existing category\\n\\nThis was inspired by a real use case, in a totally unrelated area - but \\nwhere I attempted to ensure counts were in the most useful categories I \\nwas able to provide. The user had listed categories, but I found that \\nthe data didn't always fit neatly into them as specified.\\n\\n\\nCheers,\\n\\nGavin\\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, 23 Aug 2016 09:31:36 +1200\",\n \"msg_from\": \"Gavin Flower <GavinFlower@archidevsys.co.nz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \">Andres Freund <andres@anarazel.de> writes:\\n>> On 2016-08-22 13:54:43 -0400, Robert Haas wrote:\\n>> On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>>>> I'm inclined to suggest you forget this approach and propose a single\\n>>>> counter for \\\"SQL commands executed\\\", which avoids all of the above\\n>>>> definitional problems. People who need more detail than that are\\n>>>> probably best advised to look to contrib/pg_stat_statements, anyway.\\n\\n>>> I disagree. I think SQL commands executed, lumping absolutely\\n>>> everything together, really isn't much use.\\n\\n>> I'm inclined to agree. I think that's a quite useful stat when looking\\n>> at an installation one previously didn't have a lot of interaction with.\\n\\n>Well, let's at least have an \\\"other\\\" category so you can add up the\\n>counters and get a meaningful total.\\n\\nHow would that meaningful total might help a user. What can a user might\\nanalyse with the counter in 'other' category.\\n\\n\\nNeha\\n\\n>Andres Freund <andres@anarazel.de> writes:>> On 2016-08-22 13:54:43 -0400, Robert Haas wrote:>> On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:>>>> I'm inclined to suggest you forget this approach and propose a single>>>> counter for \\\"SQL commands executed\\\", which avoids all of the above>>>> definitional problems. People who need more detail than that are>>>> probably best advised to look to contrib/pg_stat_statements, anyway.>>> I disagree. I think SQL commands executed, lumping absolutely>>> everything together, really isn't much use.>> I'm inclined to agree. I think that's a quite useful stat when looking>> at an installation one previously didn't have a lot of interaction with.>Well, let's at least have an \\\"other\\\" category so you can add up the>counters and get a meaningful total.How would that meaningful total might help a user. What can a user might analyse with the counter in 'other' category.Neha\",\n \"msg_date\": \"Wed, 24 Aug 2016 10:02:26 +1000\",\n \"msg_from\": \"neha khatri <nehakhatri5@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On 24/08/16 12:02, neha khatri wrote:\\n> >Andres Freund <andres@anarazel.de <mailto:andres@anarazel.de>> writes:\\n> >> On 2016-08-22 13:54:43 -0400, Robert Haas wrote:\\n> >> On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us \\n> <mailto:tgl@sss.pgh.pa.us>> wrote:\\n> >>>> I'm inclined to suggest you forget this approach and propose a single\\n> >>>> counter for \\\"SQL commands executed\\\", which avoids all of the above\\n> >>>> definitional problems. People who need more detail than that are\\n> >>>> probably best advised to look to contrib/pg_stat_statements, anyway.\\n>\\n> >>> I disagree. I think SQL commands executed, lumping absolutely\\n> >>> everything together, really isn't much use.\\n>\\n> >> I'm inclined to agree. I think that's a quite useful stat when looking\\n> >> at an installation one previously didn't have a lot of interaction \\n> with.\\n>\\n> >Well, let's at least have an \\\"other\\\" category so you can add up the\\n> >counters and get a meaningful total.\\n>\\n> How would that meaningful total might help a user. What can a user \\n> might analyse with the counter in 'other' category.\\n>\\n>\\n> Neha\\n>\\nThe user could then judge if there were a significant number of examples \\nnot covered in the other categories - this may, or may not, be a \\nproblem; depending on the use case.\\n\\n\\nCheers,\\nGavin\\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 Aug 2016 12:07:34 +1200\",\n \"msg_from\": \"Gavin Flower <GavinFlower@archidevsys.co.nz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 24, 2016 at 10:07 AM, Gavin Flower <\\nGavinFlower@archidevsys.co.nz> wrote:\\n\\n> On 24/08/16 12:02, neha khatri wrote:\\n>\\n>> >Andres Freund <andres@anarazel.de <mailto:andres@anarazel.de>> writes:\\n>> >> On 2016-08-22 13:54:43 -0400, Robert Haas wrote:\\n>> >> On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us <mailto:\\n>> tgl@sss.pgh.pa.us>> wrote:\\n>> >>>> I'm inclined to suggest you forget this approach and propose a single\\n>> >>>> counter for \\\"SQL commands executed\\\", which avoids all of the above\\n>> >>>> definitional problems. People who need more detail than that are\\n>> >>>> probably best advised to look to contrib/pg_stat_statements, anyway.\\n>>\\n>> >>> I disagree. I think SQL commands executed, lumping absolutely\\n>> >>> everything together, really isn't much use.\\n>>\\n>> >> I'm inclined to agree. I think that's a quite useful stat when looking\\n>> >> at an installation one previously didn't have a lot of interaction\\n>> with.\\n>>\\n>> >Well, let's at least have an \\\"other\\\" category so you can add up the\\n>> >counters and get a meaningful total.\\n>>\\n>> How would that meaningful total might help a user. What can a user might\\n>> analyse with the counter in 'other' category.\\n>>\\n>>\\n>>\\n>> The user could then judge if there were a significant number of examples\\n> not covered in the other categories - this may, or may not, be a problem;\\n> depending on the use case.\\n>\\n>\\n> Cheers,\\n> Gavin\\n>\\n> For the user to be able to judge that whether the number in the 'other'\\ncategory is a problem or not, the user is also required to know what all\\nmight fall under the 'other' category. It may not be good to say that\\n_anything_ that is not part of the already defined category is part of\\n'other'. Probably, 'other' should also be a set of predefined operations.\\n\\nNeha\\n\\nOn Wed, Aug 24, 2016 at 10:07 AM, Gavin Flower <GavinFlower@archidevsys.co.nz> wrote:On 24/08/16 12:02, neha khatri wrote:\\n\\n>Andres Freund <andres@anarazel.de <mailto:andres@anarazel.de>> writes:\\n>> On 2016-08-22 13:54:43 -0400, Robert Haas wrote:\\n>> On Sat, Aug 20, 2016 at 11:17 AM, Tom Lane <tgl@sss.pgh.pa.us <mailto:tgl@sss.pgh.pa.us>> wrote:\\n>>>> I'm inclined to suggest you forget this approach and propose a single\\n>>>> counter for \\\"SQL commands executed\\\", which avoids all of the above\\n>>>> definitional problems. People who need more detail than that are\\n>>>> probably best advised to look to contrib/pg_stat_statements, anyway.\\n\\n>>> I disagree. I think SQL commands executed, lumping absolutely\\n>>> everything together, really isn't much use.\\n\\n>> I'm inclined to agree. I think that's a quite useful stat when looking\\n>> at an installation one previously didn't have a lot of interaction with.\\n\\n>Well, let's at least have an \\\"other\\\" category so you can add up the\\n>counters and get a meaningful total.\\n\\nHow would that meaningful total might help a user. What can a user might analyse with the counter in 'other' category.\\n\\n\\n\\nThe user could then judge if there were a significant number of examples not covered in the other categories - this may, or may not, be a problem; depending on the use case.\\n\\n\\nCheers,\\nGavin\\n\\nFor the user to be able to judge that whether the number in the 'other' category is a problem or not, the user is also required to know what all might fall under the 'other' category. It may not be good to say that _anything_ that is not part of the already defined category is part of 'other'. Probably, 'other' should also be a set of predefined operations.Neha\",\n \"msg_date\": \"Wed, 24 Aug 2016 11:40:31 +1000\",\n \"msg_from\": \"neha khatri <nehakhatri5@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 23, 2016 at 3:59 AM, Andres Freund <andres@anarazel.de> wrote:\\n>> Haribabu's categorization\\n>> scheme seems to need some work, but the idea of categorizing\\n>> statements by type and counting executions per type seems very\\n>> reasonable.\\n>\\n> I'd consider instead using something like COALESCE(commandType,\\n> nodeTag(Query->utilityStmt)) as the categories. Not sure if I'd even\\n> pivot that.\\n\\nCommandType means to use select, insert, update, delete and utility\\nas the categorization to display the counters? I think I am not getting\\nyour point correctly.\\n\\nApart from the above, here are the following list of command tags that\\nare generated in the code, I took only the first word of the command tag\\njust to see how many categories present. The number indicates the\\nsubset of operations or number of types it is used. Like create table,\\ncreate function and etc.\\n\\ninsert\\ndelete\\nupdate\\nselect (6)\\ntransaction (10)\\ndeclare\\nclose (2)\\nmove\\nfetch\\ncreate (37)\\ndrop (36)\\nAlter (56)\\nimport\\ntruncate\\ncomment\\nsecurity\\ncopy\\ngrant\\nrevoke\\nnotify\\nlisten\\nunlisten\\nload\\ncluster\\nvacuum\\nanalyze\\nexplain\\nrefresh\\nset (2)\\nreset\\nshow\\ndiscard (4)\\nreassign\\nlock\\ncheckpoint\\nreindex\\nprepare\\nexecute\\ndeallocate\\n\\nwe can try to pick something from the above list also for main\\ncategories and rest will fall under\\nother.\\n\\n\\nRegards,\\nHari Babu\\nFujitsu Australia\\n\\n\\n-- \\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 Aug 2016 18:07:54 +1000\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"Haribabu Kommi wrote:\\n\\n> Apart from the above, here are the following list of command tags that\\n> are generated in the code, I took only the first word of the command tag\\n> just to see how many categories present. The number indicates the\\n> subset of operations or number of types it is used. Like create table,\\n> create function and etc.\\n\\nSounds about right. I suppose all those cases that you aggregated here\\nwould expand to full tags in the actual code. I furthermore suppose\\nthat some of these could be ignored, such as the transaction ones and\\nthings like load, lock, move, fetch, discard, deallocate (maybe lump\\nthem all together under \\\"other\\\", or some other rough categorization, as\\nTom suggests). Also, for many of these commands it's probably relevant\\nwhether they are acting on a temporary object or not; we should either\\ncount these separately, or not count the temp ones at all.\\n\\n> insert\\n> delete\\n> update\\n> select (6)\\n> transaction (10)\\n> declare\\n> close (2)\\n> move\\n> fetch\\n> create (37)\\n> drop (36)\\n> Alter (56)\\n> import\\n> truncate\\n> comment\\n> security\\n> copy\\n> grant\\n> revoke\\n> notify\\n> listen\\n> unlisten\\n> load\\n> cluster\\n> vacuum\\n> analyze\\n> explain\\n> refresh\\n> set (2)\\n> reset\\n> show\\n> discard (4)\\n> reassign\\n> lock\\n> checkpoint\\n> reindex\\n> prepare\\n> execute\\n> deallocate\\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\": \"Tue, 30 Aug 2016 14:19:24 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 31, 2016 at 3:19 AM, Alvaro Herrera <alvherre@2ndquadrant.com>\\nwrote:\\n> Haribabu Kommi wrote:\\n>\\n>> Apart from the above, here are the following list of command tags that\\n>> are generated in the code, I took only the first word of the command tag\\n>> just to see how many categories present. The number indicates the\\n>> subset of operations or number of types it is used. Like create table,\\n>> create function and etc.\\n>\\n> Sounds about right. I suppose all those cases that you aggregated here\\n> would expand to full tags in the actual code. I furthermore suppose\\n> that some of these could be ignored, such as the transaction ones and\\n> things like load, lock, move, fetch, discard, deallocate (maybe lump\\n> them all together under \\\"other\\\", or some other rough categorization, as\\n> Tom suggests).\\n\\nFollowing is the pg_stat_sql view with the SQL categories that I considered\\nthat are important. Rest of the them will be shown under others category.\\n\\npostgres=# \\\\d pg_stat_sql\\n View \\\"pg_catalog.pg_stat_sql\\\"\\n Column | Type | Modifiers\\n-----------------+--------------------------+-----------\\n inserts | bigint |\\n deletes | bigint |\\n updates | bigint |\\n selects | bigint |\\n declare_cursors | bigint |\\n closes | bigint |\\n creates | bigint |\\n drops | bigint |\\n alters | bigint |\\n imports | bigint |\\n truncates | bigint |\\n copies | bigint |\\n grants | bigint |\\n revokes | bigint |\\n clusters | bigint |\\n vacuums | bigint |\\n analyzes | bigint |\\n refreshs | bigint |\\n locks | bigint |\\n checkpoints | bigint |\\n reindexes | bigint |\\n deallocates | bigint |\\n others | bigint |\\n stats_reset | timestamp with time zone |\\n\\n\\nIf any additions/deletions, I can accommodate them.\\n\\nThe stats data gets updated in exec_simple_query and exec_execute_message\\nfunctions and the collected stats will be sent to stats collector similar\\nlike function usage stats in pgstat_report_stat function.\\n\\nThese SQL statistics data is stored in the stats file similar like global\\nstatistics. The STAT file format is updated to accommodate the new stats.\\n\\nA new GUC \\\"track_sql\\\" is added to track the SQL statement\\nstatistics, by default this is off. Only superuser can change this\\nparameter.\\n\\nAttached a patch for the same.\\n\\n> Also, for many of these commands it's probably relevant\\n> whether they are acting on a temporary object or not; we should either\\n> count these separately, or not count the temp ones at all.\\n\\nCurrently the SQL stats are not checking any object level that is a temp\\none or not? The temp objects are specific to a backend only. But what\\nI feel, any way that is an SQL query that was executed on a temp object,\\nso we need to count that operation.\\n\\nI feel this SQL stats should not worry about the object type. May be I am\\nwrong.\\n\\nRegards,\\nHari Babu\\nFujitsu Australia\\n\\n\\n-- \\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, 2 Sep 2016 16:33:53 +1000\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Fri, Sep 2, 2016 at 2:33 AM, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:\\n> On Wed, Aug 31, 2016 at 3:19 AM, Alvaro Herrera <alvherre@2ndquadrant.com>\\n> wrote:\\n>> Haribabu Kommi wrote:\\n>>\\n>>> Apart from the above, here are the following list of command tags that\\n>>> are generated in the code, I took only the first word of the command tag\\n>>> just to see how many categories present. The number indicates the\\n>>> subset of operations or number of types it is used. Like create table,\\n>>> create function and etc.\\n>>\\n>> Sounds about right. I suppose all those cases that you aggregated here\\n>> would expand to full tags in the actual code. I furthermore suppose\\n>> that some of these could be ignored, such as the transaction ones and\\n>> things like load, lock, move, fetch, discard, deallocate (maybe lump\\n>> them all together under \\\"other\\\", or some other rough categorization, as\\n>> Tom suggests).\\n>\\n> Following is the pg_stat_sql view with the SQL categories that I considered\\n> that are important. Rest of the them will be shown under others category.\\n>\\n> postgres=# \\\\d pg_stat_sql\\n> View \\\"pg_catalog.pg_stat_sql\\\"\\n> Column | Type | Modifiers\\n> -----------------+--------------------------+-----------\\n> inserts | bigint |\\n> deletes | bigint |\\n> updates | bigint |\\n> selects | bigint |\\n> declare_cursors | bigint |\\n> closes | bigint |\\n> creates | bigint |\\n> drops | bigint |\\n> alters | bigint |\\n> imports | bigint |\\n> truncates | bigint |\\n> copies | bigint |\\n> grants | bigint |\\n> revokes | bigint |\\n> clusters | bigint |\\n> vacuums | bigint |\\n> analyzes | bigint |\\n> refreshs | bigint |\\n> locks | bigint |\\n> checkpoints | bigint |\\n> reindexes | bigint |\\n> deallocates | bigint |\\n> others | bigint |\\n> stats_reset | timestamp with time zone |\\n>\\n>\\n> If any additions/deletions, I can accommodate them.\\n\\nI think it is not a good idea to make the command names used here the\\nplural forms of the command tags. Instead of \\\"inserts\\\", \\\"updates\\\",\\n\\\"imports\\\", etc. just use \\\"INSERT\\\", \\\"UPDATE\\\", \\\"IMPORT\\\". That's simpler\\nand less error prone - e.g. you won't end up with things like\\n\\\"refreshs\\\", which is not a word.\\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 Sep 2016 16:01:25 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Thu, Sep 15, 2016 at 6:01 AM, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On Fri, Sep 2, 2016 at 2:33 AM, Haribabu Kommi <kommi.haribabu@gmail.com>\\n> wrote:\\n> > On Wed, Aug 31, 2016 at 3:19 AM, Alvaro Herrera <\\n> alvherre@2ndquadrant.com>\\n> > wrote:\\n> >> Haribabu Kommi wrote:\\n> >>\\n> >>> Apart from the above, here are the following list of command tags that\\n> >>> are generated in the code, I took only the first word of the command\\n> tag\\n> >>> just to see how many categories present. The number indicates the\\n> >>> subset of operations or number of types it is used. Like create table,\\n> >>> create function and etc.\\n> >>\\n> >> Sounds about right. I suppose all those cases that you aggregated here\\n> >> would expand to full tags in the actual code. I furthermore suppose\\n> >> that some of these could be ignored, such as the transaction ones and\\n> >> things like load, lock, move, fetch, discard, deallocate (maybe lump\\n> >> them all together under \\\"other\\\", or some other rough categorization, as\\n> >> Tom suggests).\\n> >\\n> > Following is the pg_stat_sql view with the SQL categories that I\\n> considered\\n> > that are important. Rest of the them will be shown under others category.\\n> >\\n> > postgres=# \\\\d pg_stat_sql\\n> > View \\\"pg_catalog.pg_stat_sql\\\"\\n> > Column | Type | Modifiers\\n> > -----------------+--------------------------+-----------\\n> > inserts | bigint |\\n> > deletes | bigint |\\n> > updates | bigint |\\n> > selects | bigint |\\n> > declare_cursors | bigint |\\n> > closes | bigint |\\n> > creates | bigint |\\n> > drops | bigint |\\n> > alters | bigint |\\n> > imports | bigint |\\n> > truncates | bigint |\\n> > copies | bigint |\\n> > grants | bigint |\\n> > revokes | bigint |\\n> > clusters | bigint |\\n> > vacuums | bigint |\\n> > analyzes | bigint |\\n> > refreshs | bigint |\\n> > locks | bigint |\\n> > checkpoints | bigint |\\n> > reindexes | bigint |\\n> > deallocates | bigint |\\n> > others | bigint |\\n> > stats_reset | timestamp with time zone |\\n> >\\n> >\\n> > If any additions/deletions, I can accommodate them.\\n>\\n> I think it is not a good idea to make the command names used here the\\n> plural forms of the command tags. Instead of \\\"inserts\\\", \\\"updates\\\",\\n> \\\"imports\\\", etc. just use \\\"INSERT\\\", \\\"UPDATE\\\", \\\"IMPORT\\\". That's simpler\\n> and less error prone - e.g. you won't end up with things like\\n> \\\"refreshs\\\", which is not a word.\\n\\n\\nThanks for your suggestion. I also thought of changing the name while\\nwriting\\n\\\"refreshs\\\" as a column name of the view originally. A small restriction\\nwith the\\nchange of names to command names is that, user needs to execute the query\\nas follows.\\n\\nselect pg_stat_sql.select from pg_stat_sql;\\n\\nUpdated patch is attached with the corrections. Apart from name change,\\nadded documentation and tests.\\n\\nRegards,\\nHari Babu\\nFujitsu Australia\\n\\n\\n-- \\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, 21 Sep 2016 12:23:02 +1000\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On 9/14/16 4:01 PM, Robert Haas wrote:\\n> I think it is not a good idea to make the command names used here the\\n> plural forms of the command tags. Instead of \\\"inserts\\\", \\\"updates\\\",\\n> \\\"imports\\\", etc. just use \\\"INSERT\\\", \\\"UPDATE\\\", \\\"IMPORT\\\". That's simpler\\n> and less error prone - e.g. you won't end up with things like\\n> \\\"refreshs\\\", which is not a word.\\n\\nHow about having the tag not be a column name but a row entry. So you'd\\ndo something like\\n\\nSELECT * FROM pg_stat_sql WHERE tag = 'ALTER VIEW';\\n\\nThat way, we don't have to keep updating (and re-debating) this when new\\ncommand types or subtypes are added. And queries written for future\\nversions will not fail when run against old servers.\\n\\n-- \\nPeter Eisentraut 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, 21 Sep 2016 11:25:14 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Wed, Sep 21, 2016 at 11:25:14AM -0400, Peter Eisentraut wrote:\\n> On 9/14/16 4:01 PM, Robert Haas wrote:\\n> > I think it is not a good idea to make the command names used here the\\n> > plural forms of the command tags. Instead of \\\"inserts\\\", \\\"updates\\\",\\n> > \\\"imports\\\", etc. just use \\\"INSERT\\\", \\\"UPDATE\\\", \\\"IMPORT\\\". That's simpler\\n> > and less error prone - e.g. you won't end up with things like\\n> > \\\"refreshs\\\", which is not a word.\\n> \\n> How about having the tag not be a column name but a row entry. So you'd\\n> do something like\\n> \\n> SELECT * FROM pg_stat_sql WHERE tag = 'ALTER VIEW';\\n\\n+1 for this. It's MUCH easier to deal with changes in row counts than\\nchanges in row type.\\n\\nBest,\\nDavid.\\n-- \\nDavid Fetter <david(at)fetter(dot)org> http://fetter.org/\\nPhone: +1 415 235 3778 AIM: dfetter666 Yahoo!: dfetter\\nSkype: davidfetter XMPP: david(dot)fetter(at)gmail(dot)com\\n\\nRemember to vote!\\nConsider donating to Postgres: http://www.postgresql.org/about/donate\\n\\n\\n-- \\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, 21 Sep 2016 09:22:59 -0700\",\n \"msg_from\": \"David Fetter <david@fetter.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut wrote:\\n\\n> How about having the tag not be a column name but a row entry. So you'd\\n> do something like\\n> \\n> SELECT * FROM pg_stat_sql WHERE tag = 'ALTER VIEW';\\n> \\n> That way, we don't have to keep updating (and re-debating) this when new\\n> command types or subtypes are added. And queries written for future\\n> versions will not fail when run against old servers.\\n\\nYeah, good idea.\\n\\nLet's also discuss the interface from the stats collector. Currently we\\nhave some 20 new SQL functions, all alike, each loading the whole data\\nand returning a single counter, and then the view invokes each function\\nseparately. That doesn't seem great to me. How about having a single C\\nfunction that returns the whole thing as a SRF instead, and the view is\\njust a single function invocation -- something like pg_lock_status\\nfilling pg_locks in one go.\\n\\nAnother consideration is that the present patch lumps together all ALTER\\ncases in a single counter. This isn't great, but at the same time we\\ndon't want to bloat the stat files by having hundreds of counters per\\ndatabase, do we?\\n\\n-- \\n�lvaro Herrera https://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, 21 Sep 2016 14:05:24 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Wed, Sep 21, 2016 at 02:05:24PM -0300, Alvaro Herrera wrote:\\n> Another consideration is that the present patch lumps together all\\n> ALTER cases in a single counter. This isn't great, but at the same\\n> time we don't want to bloat the stat files by having hundreds of\\n> counters per database, do we?\\n\\nI count 37 documented versions of ALTER as of git master. Is there\\nsome multiplier I'm missing?\\n\\nBest,\\nDavid.\\n-- \\nDavid Fetter <david(at)fetter(dot)org> http://fetter.org/\\nPhone: +1 415 235 3778 AIM: dfetter666 Yahoo!: dfetter\\nSkype: davidfetter XMPP: david(dot)fetter(at)gmail(dot)com\\n\\nRemember to vote!\\nConsider donating to Postgres: http://www.postgresql.org/about/donate\\n\\n\\n-- \\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, 21 Sep 2016 11:46:24 -0700\",\n \"msg_from\": \"David Fetter <david@fetter.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Thu, Sep 22, 2016 at 3:05 AM, Alvaro Herrera <alvherre@2ndquadrant.com>\\nwrote:\\n\\n> Peter Eisentraut wrote:\\n>\\n> > How about having the tag not be a column name but a row entry. So you'd\\n> > do something like\\n> >\\n> > SELECT * FROM pg_stat_sql WHERE tag = 'ALTER VIEW';\\n> >\\n> > That way, we don't have to keep updating (and re-debating) this when new\\n> > command types or subtypes are added. And queries written for future\\n> > versions will not fail when run against old servers.\\n>\\n> Yeah, good idea.\\n>\\n\\nYes, Having it as a row entry is good.\\n\\n\\n\\n> Let's also discuss the interface from the stats collector. Currently we\\n> have some 20 new SQL functions, all alike, each loading the whole data\\n> and returning a single counter, and then the view invokes each function\\n> separately. That doesn't seem great to me. How about having a single C\\n> function that returns the whole thing as a SRF instead, and the view is\\n> just a single function invocation -- something like pg_lock_status\\n> filling pg_locks in one go.\\n>\\n> Another consideration is that the present patch lumps together all ALTER\\n> cases in a single counter. This isn't great, but at the same time we\\n> don't want to bloat the stat files by having hundreds of counters per\\n> database, do we?\\n\\n\\nCurrently, The SQL stats is a fixed size counter to track the all the ALTER\\ncases as single counter. So while sending the stats from the backend to\\nstats collector at the end of the transaction, the cost is same, because of\\nit's fixed size. This approach adds overhead to send and read the stats\\nis minimal.\\n\\nWith the following approach, I feel it is possible to support the counter at\\ncommand tag level.\\n\\nAdd a Global and local Hash to keep track of the counters by using the\\ncommand tag as the key, this hash table increases dynamically whenever\\na new type of SQL command gets executed. The Local Hash data is passed\\nto stats collector whenever the transaction gets committed.\\n\\nThe problem I am thinking is that, Sending data from Hash and populating\\nthe Hash from stats file for all the command tags adds some overhead.\\n\\n\\nRegards,\\nHari Babu\\nFujitsu Australia\\n\\nOn Thu, Sep 22, 2016 at 3:05 AM, Alvaro Herrera <alvherre@2ndquadrant.com> wrote:Peter Eisentraut wrote:\\n\\n> How about having the tag not be a column name but a row entry. So you'd\\n> do something like\\n>\\n> SELECT * FROM pg_stat_sql WHERE tag = 'ALTER VIEW';\\n>\\n> That way, we don't have to keep updating (and re-debating) this when new\\n> command types or subtypes are added. And queries written for future\\n> versions will not fail when run against old servers.\\n\\nYeah, good idea.Yes, Having it as a row entry is good. \\nLet's also discuss the interface from the stats collector. Currently we\\nhave some 20 new SQL functions, all alike, each loading the whole data\\nand returning a single counter, and then the view invokes each function\\nseparately. That doesn't seem great to me. How about having a single C\\nfunction that returns the whole thing as a SRF instead, and the view is\\njust a single function invocation -- something like pg_lock_status\\nfilling pg_locks in one go.\\n\\nAnother consideration is that the present patch lumps together all ALTER\\ncases in a single counter. This isn't great, but at the same time we\\ndon't want to bloat the stat files by having hundreds of counters per\\ndatabase, do we?Currently, The SQL stats is a fixed size counter to track the all the ALTER cases as single counter. So while sending the stats from the backend tostats collector at the end of the transaction, the cost is same, because ofit's fixed size. This approach adds overhead to send and read the statsis minimal.With the following approach, I feel it is possible to support the counter atcommand tag level.Add a Global and local Hash to keep track of the counters by using thecommand tag as the key, this hash table increases dynamically whenevera new type of SQL command gets executed. The Local Hash data is passedto stats collector whenever the transaction gets committed.The problem I am thinking is that, Sending data from Hash and populatingthe Hash from stats file for all the command tags adds some overhead.Regards,Hari BabuFujitsu Australia\",\n \"msg_date\": \"Thu, 29 Sep 2016 15:45:35 +1000\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Thu, Sep 29, 2016 at 3:45 PM, Haribabu Kommi <kommi.haribabu@gmail.com>\\nwrote:\\n\\n>\\n>\\n> On Thu, Sep 22, 2016 at 3:05 AM, Alvaro Herrera <alvherre@2ndquadrant.com>\\n> wrote:\\n>\\n>> Peter Eisentraut wrote:\\n>>\\n>> > How about having the tag not be a column name but a row entry. So you'd\\n>> > do something like\\n>> >\\n>> > SELECT * FROM pg_stat_sql WHERE tag = 'ALTER VIEW';\\n>> >\\n>> > That way, we don't have to keep updating (and re-debating) this when new\\n>> > command types or subtypes are added. And queries written for future\\n>> > versions will not fail when run against old servers.\\n>>\\n>> Yeah, good idea.\\n>>\\n>\\n> Yes, Having it as a row entry is good.\\n>\\n>\\n>\\n>> Let's also discuss the interface from the stats collector. Currently we\\n>> have some 20 new SQL functions, all alike, each loading the whole data\\n>> and returning a single counter, and then the view invokes each function\\n>> separately. That doesn't seem great to me. How about having a single C\\n>> function that returns the whole thing as a SRF instead, and the view is\\n>> just a single function invocation -- something like pg_lock_status\\n>> filling pg_locks in one go.\\n>>\\n>> Another consideration is that the present patch lumps together all ALTER\\n>> cases in a single counter. This isn't great, but at the same time we\\n>> don't want to bloat the stat files by having hundreds of counters per\\n>> database, do we?\\n>\\n>\\n> Currently, The SQL stats is a fixed size counter to track the all the\\n> ALTER\\n> cases as single counter. So while sending the stats from the backend to\\n> stats collector at the end of the transaction, the cost is same, because of\\n> it's fixed size. This approach adds overhead to send and read the stats\\n> is minimal.\\n>\\n> With the following approach, I feel it is possible to support the counter\\n> at\\n> command tag level.\\n>\\n> Add a Global and local Hash to keep track of the counters by using the\\n> command tag as the key, this hash table increases dynamically whenever\\n> a new type of SQL command gets executed. The Local Hash data is passed\\n> to stats collector whenever the transaction gets committed.\\n>\\n> The problem I am thinking is that, Sending data from Hash and populating\\n> the Hash from stats file for all the command tags adds some overhead.\\n>\\n\\n\\nI tried changing the pg_stat_sql into row mode and ran the regress suite to\\nadd different type of SQL commands to the view and ran the pgbench test\\non my laptop to find out any performance impact with this patch.\\n\\n HEAD PATCH\\npgbench - select 828 816\\n\\nHere I attached the pg_stat_sql patch to keep track of all SQL commands\\nbased on the commandTag and their counts. I attached the result of this\\nview that is run on the database after \\\"make installcheck\\\" just for\\nreference.\\n\\nRegards,\\nHari Babu\\nFujitsu Australia\\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, 12 Oct 2016 14:21:25 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On 2016/10/12 12:21, Haribabu Kommi wrote:\\n>\\n>\\n> On Thu, Sep 29, 2016 at 3:45 PM, Haribabu Kommi \\n> <kommi.haribabu@gmail.com <mailto:kommi.haribabu@gmail.com>> wrote:\\n>\\n>\\n>\\n> On Thu, Sep 22, 2016 at 3:05 AM, Alvaro Herrera\\n> <alvherre@2ndquadrant.com <mailto:alvherre@2ndquadrant.com>> wrote:\\n>\\n> Peter Eisentraut wrote:\\n>\\n> > How about having the tag not be a column name but a row\\n> entry. So you'd\\n> > do something like\\n> >\\n> > SELECT * FROM pg_stat_sql WHERE tag = 'ALTER VIEW';\\n> >\\n> > That way, we don't have to keep updating (and re-debating)\\n> this when new\\n> > command types or subtypes are added. And queries written\\n> for future\\n> > versions will not fail when run against old servers.\\n>\\n> Yeah, good idea.\\n>\\n>\\n> Yes, Having it as a row entry is good.\\n>\\n> Let's also discuss the interface from the stats collector. \\n> Currently we\\n> have some 20 new SQL functions, all alike, each loading the\\n> whole data\\n> and returning a single counter, and then the view invokes each\\n> function\\n> separately. That doesn't seem great to me. How about having\\n> a single C\\n> function that returns the whole thing as a SRF instead, and\\n> the view is\\n> just a single function invocation -- something like pg_lock_status\\n> filling pg_locks in one go.\\n>\\n> Another consideration is that the present patch lumps together\\n> all ALTER\\n> cases in a single counter. This isn't great, but at the same\\n> time we\\n> don't want to bloat the stat files by having hundreds of\\n> counters per\\n> database, do we?\\n>\\n>\\n> Currently, The SQL stats is a fixed size counter to track the all\\n> the ALTER\\n> cases as single counter. So while sending the stats from the\\n> backend to\\n> stats collector at the end of the transaction, the cost is same,\\n> because of\\n> it's fixed size. This approach adds overhead to send and read the\\n> stats\\n> is minimal.\\n>\\n> With the following approach, I feel it is possible to support the\\n> counter at\\n> command tag level.\\n>\\n> Add a Global and local Hash to keep track of the counters by using the\\n> command tag as the key, this hash table increases dynamically whenever\\n> a new type of SQL command gets executed. The Local Hash data is passed\\n> to stats collector whenever the transaction gets committed.\\n>\\n> The problem I am thinking is that, Sending data from Hash and\\n> populating\\n> the Hash from stats file for all the command tags adds some overhead.\\n>\\n>\\n>\\n> I tried changing the pg_stat_sql into row mode and ran the regress \\n> suite to\\n> add different type of SQL commands to the view and ran the pgbench test\\n> on my laptop to find out any performance impact with this patch.\\n>\\n> HEAD PATCH\\n> pgbench - select 828 816\\n>\\n> Here I attached the pg_stat_sql patch to keep track of all SQL commands\\n> based on the commandTag and their counts. I attached the result of this\\n> view that is run on the database after \\\"make installcheck\\\" just for \\n> reference.\\n>\\nThank you for the patch.\\n\\nTest: Commands with uppercase and lowercase\\n====\\nIf the tag='select' then it returns the 0 rows but count is actually \\nincrement by 1.\\n\\ntag='select' vs tag='SELECT'\\n\\npostgres=# SET track_sql TO ON;\\nSET\\npostgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\\n tag | count\\n--------+-------\\n SELECT | 12\\n(1 row)\\n\\npostgres=# SELECT * FROM pg_stat_sql where tag=*'SELECT'*;\\n tag | count\\n--------+-------\\n* SELECT | 13*\\n(1 row)\\n\\npostgres=# SELECT * FROM pg_stat_sql where tag=*'select'*;\\n tag | count\\n-----+-------\\n*(0 rows)*\\n\\npostgres=# SELECT * FROM pg_stat_sql where tag=*'SELECT'*;\\n tag | count\\n--------+-------\\n* SELECT | 15*\\n(1 row)\\n\\nI think all command works same as above.\\n\\nRegards,\\nVinayak Pokale\\nNTT Open Source Software Center\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nOn 2016/10/12 12:21, Haribabu Kommi\\n wrote:\\n\\n\\n\\n\\nOn Thu, Sep 29, 2016 at 3:45 PM,\\n Haribabu Kommi <kommi.haribabu@gmail.com>\\n wrote:\\n\\n\\n\\nOn Thu, Sep\\n 22, 2016 at 3:05 AM, Alvaro Herrera <alvherre@2ndquadrant.com>\\n wrote:\\nPeter Eisentraut\\n wrote:\\n\\n > How about having the tag not be a column\\n name but a row entry.� So you'd\\n > do something like\\n >\\n > SELECT * FROM pg_stat_sql WHERE tag =\\n 'ALTER VIEW';\\n >\\n > That way, we don't have to keep updating\\n (and re-debating) this when new\\n > command types or subtypes are added.� And\\n queries written for future\\n > versions will not fail when run against\\n old servers.\\n\\nYeah, good idea.\\n\\n\\n\\n\\nYes, Having it as a row entry is good.\\n\\n�\\n�\\n\\n Let's also discuss the interface from the stats\\n collector.� Currently we\\n have some 20 new SQL functions, all alike, each\\n loading the whole data\\n and returning a single counter, and then the\\n view invokes each function\\n separately.� That doesn't seem great to me.� How\\n about having a single C\\n function that returns the whole thing as a SRF\\n instead, and the view is\\n just a single function invocation -- something\\n like pg_lock_status\\n filling pg_locks in one go.\\n\\n Another consideration is that the present patch\\n lumps together all ALTER\\n cases in a single counter.� This isn't great,\\n but at the same time we\\n don't want to bloat the stat files by having\\n hundreds of counters per\\n database, do we?\\n\\n\\nCurrently, The SQL stats is a fixed size counter\\n to track the all the ALTER�\\ncases as single counter. So while sending the\\n stats from the backend to\\nstats collector at the end of the transaction,\\n the cost is same, because of\\nit's fixed size. This approach adds overhead to\\n send and read the stats\\nis minimal.\\n\\n\\nWith the following approach, I feel it is\\n possible to support the counter at\\ncommand tag level.\\n\\n\\nAdd a Global and local Hash to keep track of the\\n counters by using the\\ncommand tag as the key, this hash table increases\\n dynamically whenever\\na new type of SQL command gets executed. The\\n Local Hash data is passed\\nto stats collector whenever the transaction gets\\n committed.\\n\\n\\nThe problem I am thinking is\\n that, Sending data from Hash and populating\\nthe Hash from stats file for\\n all the command tags adds some overhead.\\n\\n\\n\\n\\n\\n\\nI tried changing the pg_stat_sql into row mode and ran\\n the regress suite to�\\nadd different type of SQL commands to the view and ran\\n the pgbench test\\non my laptop to find out any performance impact with this\\n patch.\\n\\n\\n� � � � � � � � � � � � � � � � HEAD � � �PATCH\\npgbench - select � � �828 � � � � �816\\n\\n\\nHere I attached the pg_stat_sql patch\\n to keep track of all SQL commands\\nbased on the commandTag and their\\n counts. I attached the result of this\\nview that is run on the database\\n after \\\"make installcheck\\\" just for reference.\\n\\n\\n\\n\\n Thank you for the patch.\\n\\n Test: Commands with uppercase and lowercase \\n ====\\n If the tag='select' then it returns the 0 rows but count is actually\\n increment by 1.\\n\\n tag='select' vs tag='SELECT'\\n\\n postgres=# SET track_sql TO ON;\\n SET\\n postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\\n � tag�� | count\\n --------+-------\\n �SELECT |��� 12\\n (1 row)\\n\\n postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\\n � tag�� | count\\n --------+-------\\n�SELECT |��� 13\\n (1 row)\\n\\n postgres=# SELECT * FROM pg_stat_sql where tag='select';\\n �tag | count\\n -----+-------\\n(0 rows)\\n\\n postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\\n � tag�� | count\\n --------+-------\\n�SELECT |��� 15\\n (1 row)\\n\\n I think all command works same as above.\\n\\n Regards,\\n Vinayak Pokale\\n NTT Open Source Software Center\",\n \"msg_date\": \"Wed, 12 Oct 2016 14:06:34 +0900\",\n \"msg_from\": \"vinayak <Pokale_Vinayak_q3@lab.ntt.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On 2016/10/12 12:21, Haribabu Kommi wrote:\\n>\\n>\\n> On Thu, Sep 29, 2016 at 3:45 PM, Haribabu Kommi \\n> <kommi.haribabu@gmail.com <mailto:kommi.haribabu@gmail.com>> wrote:\\n>\\n>\\n>\\n> On Thu, Sep 22, 2016 at 3:05 AM, Alvaro Herrera\\n> <alvherre@2ndquadrant.com <mailto:alvherre@2ndquadrant.com>> wrote:\\n>\\n> Peter Eisentraut wrote:\\n>\\n> > How about having the tag not be a column name but a row\\n> entry. So you'd\\n> > do something like\\n> >\\n> > SELECT * FROM pg_stat_sql WHERE tag = 'ALTER VIEW';\\n> >\\n> > That way, we don't have to keep updating (and re-debating)\\n> this when new\\n> > command types or subtypes are added. And queries written\\n> for future\\n> > versions will not fail when run against old servers.\\n>\\n> Yeah, good idea.\\n>\\n>\\n> Yes, Having it as a row entry is good.\\n>\\n> Let's also discuss the interface from the stats collector. \\n> Currently we\\n> have some 20 new SQL functions, all alike, each loading the\\n> whole data\\n> and returning a single counter, and then the view invokes each\\n> function\\n> separately. That doesn't seem great to me. How about having\\n> a single C\\n> function that returns the whole thing as a SRF instead, and\\n> the view is\\n> just a single function invocation -- something like pg_lock_status\\n> filling pg_locks in one go.\\n>\\n> Another consideration is that the present patch lumps together\\n> all ALTER\\n> cases in a single counter. This isn't great, but at the same\\n> time we\\n> don't want to bloat the stat files by having hundreds of\\n> counters per\\n> database, do we?\\n>\\n>\\n> Currently, The SQL stats is a fixed size counter to track the all\\n> the ALTER\\n> cases as single counter. So while sending the stats from the\\n> backend to\\n> stats collector at the end of the transaction, the cost is same,\\n> because of\\n> it's fixed size. This approach adds overhead to send and read the\\n> stats\\n> is minimal.\\n>\\n> With the following approach, I feel it is possible to support the\\n> counter at\\n> command tag level.\\n>\\n> Add a Global and local Hash to keep track of the counters by using the\\n> command tag as the key, this hash table increases dynamically whenever\\n> a new type of SQL command gets executed. The Local Hash data is passed\\n> to stats collector whenever the transaction gets committed.\\n>\\n> The problem I am thinking is that, Sending data from Hash and\\n> populating\\n> the Hash from stats file for all the command tags adds some overhead.\\n>\\n>\\n>\\n> I tried changing the pg_stat_sql into row mode and ran the regress \\n> suite to\\n> add different type of SQL commands to the view and ran the pgbench test\\n> on my laptop to find out any performance impact with this patch.\\n>\\n> HEAD PATCH\\n> pgbench - select 828 816\\n>\\n> Here I attached the pg_stat_sql patch to keep track of all SQL commands\\n> based on the commandTag and their counts. I attached the result of this\\n> view that is run on the database after \\\"make installcheck\\\" just for \\n> reference.\\nSome comments:\\nI think we can use pgstat_* instead of pgstat* for code consistency.\\n\\n+static HTAB *pgStatBackendSql = NULL;\\nHow about *pgstat_backend_sql\\n\\n+HTAB *pgStatSql = NULL;\\nHow about *pgstat_sql\\n\\n+static void pgstat_recv_sqlstmt(PgStat_MsgSqlstmt * msg, int len);\\nHow about PgStat_MsgSqlstmt *msg instead of PgStat_MsgSqlstmt * msg\\n\\n+typedef struct PgStatSqlstmtEntry\\nHow about PgStat_SqlstmtEntry\\n\\n+typedef struct PgStatMsgSqlstmt\\nHow about PgStat_MsgSqlstmt\\n\\nI have observed below behavior.\\nI have SET track_sql to ON and then execute the SELECT command and it \\nreturn 0 rows.\\nIt start counting from the second command.\\npostgres=# SET track_sql TO ON;\\nSET\\npostgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\\n tag | count\\n-----+-------\\n(0 rows)\\n\\npostgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\\n tag | count\\n--------+-------\\n SELECT | 1\\n(1 row)\\nIs this a correct behavior?\\n\\nRegards,\\nVinayak Pokale\\nNTT Open Source Software Center\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nOn 2016/10/12 12:21, Haribabu Kommi\\n wrote:\\n\\n\\n\\n\\nOn Thu, Sep 29, 2016 at 3:45 PM,\\n Haribabu Kommi <kommi.haribabu@gmail.com>\\n wrote:\\n\\n\\n\\nOn Thu, Sep\\n 22, 2016 at 3:05 AM, Alvaro Herrera <alvherre@2ndquadrant.com>\\n wrote:\\nPeter Eisentraut\\n wrote:\\n\\n > How about having the tag not be a column\\n name but a row entry.� So you'd\\n > do something like\\n >\\n > SELECT * FROM pg_stat_sql WHERE tag =\\n 'ALTER VIEW';\\n >\\n > That way, we don't have to keep updating\\n (and re-debating) this when new\\n > command types or subtypes are added.� And\\n queries written for future\\n > versions will not fail when run against\\n old servers.\\n\\nYeah, good idea.\\n\\n\\n\\n\\nYes, Having it as a row entry is good.\\n\\n�\\n�\\n\\n Let's also discuss the interface from the stats\\n collector.� Currently we\\n have some 20 new SQL functions, all alike, each\\n loading the whole data\\n and returning a single counter, and then the\\n view invokes each function\\n separately.� That doesn't seem great to me.� How\\n about having a single C\\n function that returns the whole thing as a SRF\\n instead, and the view is\\n just a single function invocation -- something\\n like pg_lock_status\\n filling pg_locks in one go.\\n\\n Another consideration is that the present patch\\n lumps together all ALTER\\n cases in a single counter.� This isn't great,\\n but at the same time we\\n don't want to bloat the stat files by having\\n hundreds of counters per\\n database, do we?\\n\\n\\nCurrently, The SQL stats is a fixed size counter\\n to track the all the ALTER�\\ncases as single counter. So while sending the\\n stats from the backend to\\nstats collector at the end of the transaction,\\n the cost is same, because of\\nit's fixed size. This approach adds overhead to\\n send and read the stats\\nis minimal.\\n\\n\\nWith the following approach, I feel it is\\n possible to support the counter at\\ncommand tag level.\\n\\n\\nAdd a Global and local Hash to keep track of the\\n counters by using the\\ncommand tag as the key, this hash table increases\\n dynamically whenever\\na new type of SQL command gets executed. The\\n Local Hash data is passed\\nto stats collector whenever the transaction gets\\n committed.\\n\\n\\nThe problem I am thinking is\\n that, Sending data from Hash and populating\\nthe Hash from stats file for\\n all the command tags adds some overhead.\\n\\n\\n\\n\\n\\n\\nI tried changing the pg_stat_sql into row mode and ran\\n the regress suite to�\\nadd different type of SQL commands to the view and ran\\n the pgbench test\\non my laptop to find out any performance impact with this\\n patch.\\n\\n\\n� � � � � � � � � � � � � � � � HEAD � � �PATCH\\npgbench - select � � �828 � � � � �816\\n\\n\\nHere I attached the pg_stat_sql patch\\n to keep track of all SQL commands\\nbased on the commandTag and their\\n counts. I attached the result of this\\nview that is run on the database\\n after \\\"make installcheck\\\" just for reference.\\n\\n\\n\\n Some comments:\\n I think we can use pgstat_* instead of pgstat* for code consistency.\\n �\\n +static HTAB *pgStatBackendSql = NULL;\\n How about *pgstat_backend_sql\\n\\n +HTAB������ *pgStatSql = NULL;\\n How about *pgstat_sql\\n\\n +static void pgstat_recv_sqlstmt(PgStat_MsgSqlstmt * msg, int len);\\n How about PgStat_MsgSqlstmt *msg instead of PgStat_MsgSqlstmt * msg\\n\\n +typedef struct PgStatSqlstmtEntry\\n How about PgStat_SqlstmtEntry\\n\\n +typedef struct PgStatMsgSqlstmt\\n How about PgStat_MsgSqlstmt\\n\\n I have observed below behavior.\\n I have SET track_sql to ON and then execute the SELECT command and\\n it return 0 rows.\\n It start counting from the second command. \\n postgres=# SET track_sql TO ON;\\n SET\\n postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\\n �tag | count\\n -----+-------\\n (0 rows)\\n\\n postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\\n � tag�� | count\\n --------+-------\\n �SELECT |���� 1\\n (1 row)\\n Is this a correct behavior?\\n\\n Regards,\\n Vinayak Pokale\\n NTT Open Source Software Center\",\n \"msg_date\": \"Fri, 14 Oct 2016 17:48:27 +0900\",\n \"msg_from\": \"vinayak <Pokale_Vinayak_q3@lab.ntt.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 12, 2016 at 4:06 PM, vinayak <Pokale_Vinayak_q3@lab.ntt.co.jp>\\nwrote:\\n>\\n>\\n> Thank you for the patch.\\n>\\n> Test: Commands with uppercase and lowercase\\n> ====\\n> If the tag='select' then it returns the 0 rows but count is actually\\n> increment by 1.\\n>\\n> tag='select' vs tag='SELECT'\\n>\\n> postgres=# SET track_sql TO ON;\\n> SET\\n> postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\\n> tag | count\\n> --------+-------\\n> SELECT | 12\\n> (1 row)\\n>\\n> postgres=# SELECT * FROM pg_stat_sql where tag=*'SELECT'*;\\n> tag | count\\n> --------+-------\\n> * SELECT | 13*\\n> (1 row)\\n>\\n> postgres=# SELECT * FROM pg_stat_sql where tag=*'select'*;\\n> tag | count\\n> -----+-------\\n> *(0 rows)*\\n>\\n> postgres=# SELECT * FROM pg_stat_sql where tag=*'SELECT'*;\\n> tag | count\\n> --------+-------\\n> * SELECT | 15*\\n> (1 row)\\n>\\n> I think all command works same as above.\\n>\\n\\nThanks for checking the patch.\\n\\nYes, that's correct. Currently the CAPS letters are used as tag names\\nthat are getting displayed whenever any SQL command is executed.\\nSo I used the same names as entries to store the details of stats of\\nSQL statements.\\n\\nIf anyone feels the other way is better, I am fine for it.\\n\\n\\nRegards,\\nHari Babu\\nFujitsu Australia\\n\\nOn Wed, Oct 12, 2016 at 4:06 PM, vinayak <Pokale_Vinayak_q3@lab.ntt.co.jp> wrote:\\n\\n\\n Thank you for the patch.\\n\\n Test: Commands with uppercase and lowercase \\n ====\\n If the tag='select' then it returns the 0 rows but count is actually\\n increment by 1.\\n\\n tag='select' vs tag='SELECT'\\n\\n postgres=# SET track_sql TO ON;\\n SET\\n postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\\n tag | count\\n --------+-------\\n SELECT | 12\\n (1 row)\\n\\n postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\\n tag | count\\n --------+-------\\n SELECT | 13\\n (1 row)\\n\\n postgres=# SELECT * FROM pg_stat_sql where tag='select';\\n tag | count\\n -----+-------\\n(0 rows)\\n\\n postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\\n tag | count\\n --------+-------\\n SELECT | 15\\n (1 row)\\n\\n I think all command works same as above.Thanks for checking the patch.Yes, that's correct. Currently the CAPS letters are used as tag namesthat are getting displayed whenever any SQL command is executed.So I used the same names as entries to store the details of stats ofSQL statements.If anyone feels the other way is better, I am fine for it.Regards,Hari BabuFujitsu Australia\",\n \"msg_date\": \"Mon, 17 Oct 2016 12:22:20 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 14, 2016 at 7:48 PM, vinayak <Pokale_Vinayak_q3@lab.ntt.co.jp>\\nwrote:\\n\\n>\\n> On 2016/10/12 12:21, Haribabu Kommi wrote:\\n>\\n> I tried changing the pg_stat_sql into row mode and ran the regress suite\\n> to\\n> add different type of SQL commands to the view and ran the pgbench test\\n> on my laptop to find out any performance impact with this patch.\\n>\\n> HEAD PATCH\\n> pgbench - select 828 816\\n>\\n> Here I attached the pg_stat_sql patch to keep track of all SQL commands\\n> based on the commandTag and their counts. I attached the result of this\\n> view that is run on the database after \\\"make installcheck\\\" just for\\n> reference.\\n>\\n> Some comments:\\n> I think we can use pgstat_* instead of pgstat* for code consistency.\\n>\\n> +static HTAB *pgStatBackendSql = NULL;\\n> How about *pgstat_backend_sql\\n>\\n> +HTAB *pgStatSql = NULL;\\n> How about *pgstat_sql\\n>\\n\\nChanged.\\n\\n\\n> +static void pgstat_recv_sqlstmt(PgStat_MsgSqlstmt * msg, int len);\\n> How about PgStat_MsgSqlstmt *msg instead of PgStat_MsgSqlstmt * msg\\n>\\n\\nAdded the typdef into typdef.list file so this problem never occurs.\\n\\n\\n> +typedef struct PgStatSqlstmtEntry\\n> How about PgStat_SqlstmtEntry\\n>\\n> +typedef struct PgStatMsgSqlstmt\\n> How about PgStat_MsgSqlstmt\\n>\\n\\n\\n changed.\\n\\n\\n> I have observed below behavior.\\n> I have SET track_sql to ON and then execute the SELECT command and it\\n> return 0 rows.\\n> It start counting from the second command.\\n> postgres=# SET track_sql TO ON;\\n> SET\\n> postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\\n> tag | count\\n> -----+-------\\n> (0 rows)\\n>\\n> postgres=# SELECT * FROM pg_stat_sql where tag='SELECT';\\n> tag | count\\n> --------+-------\\n> SELECT | 1\\n> (1 row)\\n> Is this a correct behavior?\\n>\\n\\n\\nYes. This is because the stats of the SQL statements that are collected are\\nsent to\\nthe stats collector at the end of the SQL statement execution. The current\\nSQL statement\\nis not counted in the result.\\n\\nUpdated patch is attached.\\n\\nRegards,\\nHari Babu\\nFujitsu Australia\\n\\n\\n-- \\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, 17 Oct 2016 12:22:51 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On 2016/10/17 10:22, Haribabu Kommi wrote:\\n>\\n>\\n> On Fri, Oct 14, 2016 at 7:48 PM, vinayak \\n> <Pokale_Vinayak_q3@lab.ntt.co.jp \\n> <mailto:Pokale_Vinayak_q3@lab.ntt.co.jp>> wrote:\\n>\\n>\\n> On 2016/10/12 12:21, Haribabu Kommi wrote:\\n>> I tried changing the pg_stat_sql into row mode and ran the\\n>> regress suite to\\n>> add different type of SQL commands to the view and ran the\\n>> pgbench test\\n>> on my laptop to find out any performance impact with this patch.\\n>>\\n>> HEAD PATCH\\n>> pgbench - select 828 816\\n>>\\n>> Here I attached the pg_stat_sql patch to keep track of all SQL\\n>> commands\\n>> based on the commandTag and their counts. I attached the result\\n>> of this\\n>> view that is run on the database after \\\"make installcheck\\\" just\\n>> for reference.\\n> Some comments:\\n> I think we can use pgstat_* instead of pgstat* for code consistency.\\n>\\n> +static HTAB *pgStatBackendSql = NULL;\\n> How about *pgstat_backend_sql\\n>\\n> +HTAB *pgStatSql = NULL;\\n> How about *pgstat_sql\\n>\\n>\\n> Changed.\\n>\\n> +static void pgstat_recv_sqlstmt(PgStat_MsgSqlstmt * msg, int len);\\n> How about PgStat_MsgSqlstmt *msg instead of PgStat_MsgSqlstmt * msg\\n>\\n>\\n> Added the typdef into typdef.list file so this problem never occurs.\\n>\\n> +typedef struct PgStatSqlstmtEntry\\n> How about PgStat_SqlstmtEntry\\n>\\n> +typedef struct PgStatMsgSqlstmt\\n> How about PgStat_MsgSqlstmt\\n>\\n>\\n>\\n> changed.\\n>\\n> Updated patch is attached.\\nThe updated patch gives following error.\\nError:\\n../../../../src/include/pgstat.h:557: error: ‘PgStatSqlstmtEntry’ \\nundeclared here (not in a function)\\n\\n- ((PGSTAT_MSG_PAYLOAD - sizeof(int)) / sizeof(PgStatSqlstmtEntry))\\n+ ((PGSTAT_MSG_PAYLOAD - sizeof(int)) / sizeof(PgStat_SqlstmtEntry))\\nThe attached patch fixed the error.\\n\\nRegards,\\nVinayak Pokale\\n\\n\\n-- \\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, 17 Oct 2016 14:35:39 +0900\",\n \"msg_from\": \"vinayak <Pokale_Vinayak_q3@lab.ntt.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Thu, Sep 29, 2016 at 1:45 AM, Haribabu Kommi\\n<kommi.haribabu@gmail.com> wrote:\\n> Currently, The SQL stats is a fixed size counter to track the all the ALTER\\n> cases as single counter. So while sending the stats from the backend to\\n> stats collector at the end of the transaction, the cost is same, because of\\n> it's fixed size. This approach adds overhead to send and read the stats\\n> is minimal.\\n>\\n> With the following approach, I feel it is possible to support the counter at\\n> command tag level.\\n>\\n> Add a Global and local Hash to keep track of the counters by using the\\n> command tag as the key, this hash table increases dynamically whenever\\n> a new type of SQL command gets executed. The Local Hash data is passed\\n> to stats collector whenever the transaction gets committed.\\n>\\n> The problem I am thinking is that, Sending data from Hash and populating\\n> the Hash from stats file for all the command tags adds some overhead.\\n\\nYeah, I'm not very excited about that overhead. This seems useful as\\nfar as it goes, but I don't really want to incur measurable overhead\\nwhen it's in use. Having a hash table rather than a fixed array of\\nslots means that you have to pass this through the stats collector\\nrather than updating shared memory directly, which is fairly heavy\\nweight. If each backend could have its own copy of the slot array and\\njust update that, and readers added up the values across the whole\\narray, this could be done without any locking at all, and it would\\ngenerally be much lighter-weight than this approach.\\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, 18 Oct 2016 14:11:57 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 19, 2016 at 5:11 AM, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On Thu, Sep 29, 2016 at 1:45 AM, Haribabu Kommi\\n> <kommi.haribabu@gmail.com> wrote:\\n> > Currently, The SQL stats is a fixed size counter to track the all the\\n> ALTER\\n> > cases as single counter. So while sending the stats from the backend to\\n> > stats collector at the end of the transaction, the cost is same, because\\n> of\\n> > it's fixed size. This approach adds overhead to send and read the stats\\n> > is minimal.\\n> >\\n> > With the following approach, I feel it is possible to support the\\n> counter at\\n> > command tag level.\\n> >\\n> > Add a Global and local Hash to keep track of the counters by using the\\n> > command tag as the key, this hash table increases dynamically whenever\\n> > a new type of SQL command gets executed. The Local Hash data is passed\\n> > to stats collector whenever the transaction gets committed.\\n> >\\n> > The problem I am thinking is that, Sending data from Hash and populating\\n> > the Hash from stats file for all the command tags adds some overhead.\\n>\\n> Yeah, I'm not very excited about that overhead. This seems useful as\\n> far as it goes, but I don't really want to incur measurable overhead\\n> when it's in use. Having a hash table rather than a fixed array of\\n> slots means that you have to pass this through the stats collector\\n> rather than updating shared memory directly, which is fairly heavy\\n> weight. If each backend could have its own copy of the slot array and\\n> just update that, and readers added up the values across the whole\\n> array, this could be done without any locking at all, and it would\\n> generally be much lighter-weight than this approach.\\n\\n\\nUsing limited information like combining all ALTER XXX into Alter counter,\\nthe fixed array logic works fine without having any problems. But people\\nare suggesting to provide more details like ALTER VIEW and etc, so I\\nchecked the following ways,\\n\\n1. Using of nodetag as an array of index to update the counter. But this\\napproach doesn't work for some cases like T_DropStmt where the tag\\nvaries based on the removeType of the DropStmt. So I decide to drop\\nthis approach.\\n\\n2. Using of tag name for searching the fixed size array that is sorted with\\nall command tags that are possible in PostgreSQL. Using a binary search,\\nfind out the location and update the counter. In this approach, first the\\narray\\nneeds to be filed with TAG information in the sorted order.\\n\\n3. Using of Hash table to store the counters information based on the TAG\\nname as key.\\n\\nI choose the approach-3 as it can scale for any additional commands.\\n\\nAny better alternatives with that directly it can point to the array index?\\n\\nif we are going with ALTER XXX into single Alter counter is approach, then\\nI will change the code with a fixed array approach.\\n\\nRegards,\\nHari Babu\\nFujitsu Australia\\n\\nOn Wed, Oct 19, 2016 at 5:11 AM, Robert Haas <robertmhaas@gmail.com> wrote:On Thu, Sep 29, 2016 at 1:45 AM, Haribabu Kommi\\n<kommi.haribabu@gmail.com> wrote:\\n> Currently, The SQL stats is a fixed size counter to track the all the ALTER\\n> cases as single counter. So while sending the stats from the backend to\\n> stats collector at the end of the transaction, the cost is same, because of\\n> it's fixed size. This approach adds overhead to send and read the stats\\n> is minimal.\\n>\\n> With the following approach, I feel it is possible to support the counter at\\n> command tag level.\\n>\\n> Add a Global and local Hash to keep track of the counters by using the\\n> command tag as the key, this hash table increases dynamically whenever\\n> a new type of SQL command gets executed. The Local Hash data is passed\\n> to stats collector whenever the transaction gets committed.\\n>\\n> The problem I am thinking is that, Sending data from Hash and populating\\n> the Hash from stats file for all the command tags adds some overhead.\\n\\nYeah, I'm not very excited about that overhead. This seems useful as\\nfar as it goes, but I don't really want to incur measurable overhead\\nwhen it's in use. Having a hash table rather than a fixed array of\\nslots means that you have to pass this through the stats collector\\nrather than updating shared memory directly, which is fairly heavy\\nweight. If each backend could have its own copy of the slot array and\\njust update that, and readers added up the values across the whole\\narray, this could be done without any locking at all, and it would\\ngenerally be much lighter-weight than this approach.Using limited information like combining all ALTER XXX into Alter counter,the fixed array logic works fine without having any problems. But peopleare suggesting to provide more details like ALTER VIEW and etc, so Ichecked the following ways,1. Using of nodetag as an array of index to update the counter. But thisapproach doesn't work for some cases like T_DropStmt where the tagvaries based on the removeType of the DropStmt. So I decide to dropthis approach.2. Using of tag name for searching the fixed size array that is sorted withall command tags that are possible in PostgreSQL. Using a binary search,find out the location and update the counter. In this approach, first the arrayneeds to be filed with TAG information in the sorted order.3. Using of Hash table to store the counters information based on the TAGname as key.I choose the approach-3 as it can scale for any additional commands.Any better alternatives with that directly it can point to the array index?if we are going with ALTER XXX into single Alter counter is approach, thenI will change the code with a fixed array approach.Regards,Hari BabuFujitsu Australia\",\n \"msg_date\": \"Wed, 19 Oct 2016 11:36:11 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"Hi Vinayak,\\n\\nThis is a gentle reminder.\\n\\nyou assigned as reviewer to the current patch in the 11-2016 commitfest.\\nIf you have any more review comments / performance results regarding the\\napproach of the patch, please share it. Otherwise, you can mark the patch\\nas \\\"Ready for committer\\\" for committer feedback. This will help us in\\nsmoother\\noperation of the commitfest.\\n\\n\\nRegards,\\nHari Babu\\nFujitsu Australia\\n\\nHi Vinayak,This is a gentle reminder.you assigned as reviewer to the current patch in the 11-2016 commitfest.If you have any more review comments / performance results regarding theapproach of the patch, please share it. Otherwise, you can mark the patchas \\\"Ready for committer\\\" for committer feedback. This will help us in smootheroperation of the commitfest. Regards,Hari BabuFujitsu Australia\",\n \"msg_date\": \"Tue, 22 Nov 2016 17:57:55 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 22, 2016 at 5:57 PM, Haribabu Kommi <kommi.haribabu@gmail.com>\\nwrote:\\n\\n> Hi Vinayak,\\n>\\n> This is a gentle reminder.\\n>\\n> you assigned as reviewer to the current patch in the 11-2016 commitfest.\\n> If you have any more review comments / performance results regarding the\\n> approach of the patch, please share it. Otherwise, you can mark the patch\\n> as \\\"Ready for committer\\\" for committer feedback. This will help us in\\n> smoother\\n> operation of the commitfest.\\n>\\n\\nMoved to next CF with \\\"needs review\\\" status.\\n\\n\\nRegards,\\nHari Babu\\nFujitsu Australia\\n\\nOn Tue, Nov 22, 2016 at 5:57 PM, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:Hi Vinayak,This is a gentle reminder.you assigned as reviewer to the current patch in the 11-2016 commitfest.If you have any more review comments / performance results regarding theapproach of the patch, please share it. Otherwise, you can mark the patchas \\\"Ready for committer\\\" for committer feedback. This will help us in smootheroperation of the commitfest.\\n\\nMoved to next CF with \\\"needs review\\\" status.Regards,Hari BabuFujitsu Australia\",\n \"msg_date\": \"Mon, 5 Dec 2016 13:31:20 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 5, 2016 at 8:01 AM, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:\\n> Moved to next CF with \\\"needs review\\\" status.\\n\\nI have started reviewing the patch, Some initial comments.\\n\\n$ git apply ../pg_stat_sql_row_mode_2.patch\\nerror: patch failed: src/include/catalog/pg_proc.h:2893\\nerror: src/include/catalog/pg_proc.h: patch does not apply\\n\\nPatch is not applying on the head, I think it needs to be rebased.\\n\\n+void\\n+pgstat_count_sqlstmt(const char *commandTag)\\n+{\\n+ PgStat_SqlstmtEntry *htabent;\\n+ bool found;\\n+\\n+ if (!pgstat_track_sql)\\n+ return\\n\\nCallers of pgstat_count_sqlstmt are always ensuring that\\npgstat_track_sql is true, seems it's redundant here.\\n\\n-- \\nRegards,\\nDilip Kumar\\nEnterpriseDB: http://www.enterprisedb.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, 11 Jan 2017 19:47:35 +0530\",\n \"msg_from\": \"Dilip Kumar <dilipbalaut@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 11, 2017 at 9:17 AM, Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n> On Mon, Dec 5, 2016 at 8:01 AM, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:\\n>> Moved to next CF with \\\"needs review\\\" status.\\n>\\n> I have started reviewing the patch, Some initial comments.\\n>\\n> $ git apply ../pg_stat_sql_row_mode_2.patch\\n> error: patch failed: src/include/catalog/pg_proc.h:2893\\n> error: src/include/catalog/pg_proc.h: patch does not apply\\n\\nI suggest using 'patch' directly. 'git apply' is far too picky.\\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, 11 Jan 2017 15:06:26 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 11, 2017 at 7:47 PM, Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n> +void\\n> +pgstat_count_sqlstmt(const char *commandTag)\\n> +{\\n> + PgStat_SqlstmtEntry *htabent;\\n> + bool found;\\n> +\\n> + if (!pgstat_track_sql)\\n> + return\\n>\\n> Callers of pgstat_count_sqlstmt are always ensuring that\\n> pgstat_track_sql is true, seems it's redundant here.\\n\\nI have done testing of the feature, it's mostly working as per the expectation.\\n\\nI have few comments/questions.\\n\\n--------\\nIf you execute the query with EXECUTE then commandTag will be EXECUTE\\nthat way we will not show the actual query type, I mean all the\\nstatements will get the common tag \\\"EXECUTE\\\".\\n\\nYou can try this.\\nPREPARE fooplan (int) AS INSERT INTO t VALUES($1);\\nEXECUTE fooplan(1);\\n\\n------\\n\\n+ /* Destroy the SQL statement counter stats HashTable */\\n+ hash_destroy(pgstat_sql);\\n+\\n+ /* Create SQL statement counter Stats HashTable */\\n\\nI think in the patch we have used mixed of \\\"Stats HashTable\\\" and\\n\\\"stats HashTable\\\", I think better\\nto be consistent throughout the patch. Check other similar instances.\\n\\n----------------\\n\\n@@ -1102,6 +1102,12 @@ exec_simple_query(const char *query_string)\\n\\n PortalDrop(portal, false);\\n\\n+ /*\\n+ * Count SQL statement for pg_stat_sql view\\n+ */\\n+ if (pgstat_track_sql)\\n+ pgstat_count_sqlstmt(commandTag);\\n\\nWe are only counting the successful SQL statement, Is that intentional?\\n\\n------\\nI have noticed that pgstat_count_sqlstmt is called from\\nexec_simple_query and exec_execute_message. Don't we want to track the\\nstatement executed from functions?\\n-------\\n\\n-- \\nRegards,\\nDilip Kumar\\nEnterpriseDB: http://www.enterprisedb.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, 13 Jan 2017 21:28:43 +0530\",\n \"msg_from\": \"Dilip Kumar <dilipbalaut@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Sat, Jan 14, 2017 at 12:58 AM, Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n> On Wed, Jan 11, 2017 at 7:47 PM, Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n>> +void\\n>> +pgstat_count_sqlstmt(const char *commandTag)\\n>> +{\\n>> + PgStat_SqlstmtEntry *htabent;\\n>> + bool found;\\n>> +\\n>> + if (!pgstat_track_sql)\\n>> + return\\n>>\\n>> Callers of pgstat_count_sqlstmt are always ensuring that\\n>> pgstat_track_sql is true, seems it's redundant here.\\n>\\n> I have done testing of the feature, it's mostly working as per the expectation.\\n>\\n> I have few comments/questions.\\n>\\n> --------\\n> If you execute the query with EXECUTE then commandTag will be EXECUTE\\n> that way we will not show the actual query type, I mean all the\\n> statements will get the common tag \\\"EXECUTE\\\".\\n>\\n> You can try this.\\n> PREPARE fooplan (int) AS INSERT INTO t VALUES($1);\\n> EXECUTE fooplan(1);\\n>\\n> ------\\n>\\n> + /* Destroy the SQL statement counter stats HashTable */\\n> + hash_destroy(pgstat_sql);\\n> +\\n> + /* Create SQL statement counter Stats HashTable */\\n>\\n> I think in the patch we have used mixed of \\\"Stats HashTable\\\" and\\n> \\\"stats HashTable\\\", I think better\\n> to be consistent throughout the patch. Check other similar instances.\\n>\\n> ----------------\\n>\\n> @@ -1102,6 +1102,12 @@ exec_simple_query(const char *query_string)\\n>\\n> PortalDrop(portal, false);\\n>\\n> + /*\\n> + * Count SQL statement for pg_stat_sql view\\n> + */\\n> + if (pgstat_track_sql)\\n> + pgstat_count_sqlstmt(commandTag);\\n>\\n> We are only counting the successful SQL statement, Is that intentional?\\n>\\n> ------\\n> I have noticed that pgstat_count_sqlstmt is called from\\n> exec_simple_query and exec_execute_message. Don't we want to track the\\n> statement executed from functions?\\n> -------\\n\\nThe latest patch available has rotten, could you rebase please?\\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\": \"Tue, 17 Jan 2017 16:38:48 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 13, 2017 at 10:58 AM, Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n> If you execute the query with EXECUTE then commandTag will be EXECUTE\\n> that way we will not show the actual query type, I mean all the\\n> statements will get the common tag \\\"EXECUTE\\\".\\n\\nSomebody might say that's a feature, not a bug. But then again,\\nsomeone else might say the opposite.\\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, 17 Jan 2017 16:14:19 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Sat, Jan 14, 2017 at 2:58 AM, Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n\\n> On Wed, Jan 11, 2017 at 7:47 PM, Dilip Kumar <dilipbalaut@gmail.com>\\n> wrote:\\n> > +void\\n> > +pgstat_count_sqlstmt(const char *commandTag)\\n> > +{\\n> > + PgStat_SqlstmtEntry *htabent;\\n> > + bool found;\\n> > +\\n> > + if (!pgstat_track_sql)\\n> > + return\\n> >\\n> > Callers of pgstat_count_sqlstmt are always ensuring that\\n> > pgstat_track_sql is true, seems it's redundant here.\\n>\\n> I have done testing of the feature, it's mostly working as per the\\n> expectation.\\n>\\n\\nThanks for the review and test.\\n\\nThe use case for this view is to find out the number of different SQL\\nstatements\\nthat are getting executed successfully on an instance by the\\napplication/user.\\n\\nI have few comments/questions.\\n>\\n> --------\\n> If you execute the query with EXECUTE then commandTag will be EXECUTE\\n> that way we will not show the actual query type, I mean all the\\n> statements will get the common tag \\\"EXECUTE\\\".\\n>\\n> You can try this.\\n> PREPARE fooplan (int) AS INSERT INTO t VALUES($1);\\n> EXECUTE fooplan(1);\\n>\\n> ------\\n>\\n\\nYes, that's correct. Currently the count is incremented based on the base\\ncommand,\\nbecause of this reason, the EXECUTE is increased, instead of the actual\\noperation.\\n\\n\\n\\n> + /* Destroy the SQL statement counter stats HashTable */\\n> + hash_destroy(pgstat_sql);\\n> +\\n> + /* Create SQL statement counter Stats HashTable */\\n>\\n> I think in the patch we have used mixed of \\\"Stats HashTable\\\" and\\n> \\\"stats HashTable\\\", I think better\\n> to be consistent throughout the patch. Check other similar instances.\\n>\\n> ----------------\\n>\\n\\nCorrected.\\n\\n\\n>\\n> @@ -1102,6 +1102,12 @@ exec_simple_query(const char *query_string)\\n>\\n> PortalDrop(portal, false);\\n>\\n> + /*\\n> + * Count SQL statement for pg_stat_sql view\\n> + */\\n> + if (pgstat_track_sql)\\n> + pgstat_count_sqlstmt(commandTag);\\n>\\n> We are only counting the successful SQL statement, Is that intentional?\\n>\\n\\nYes, I thought of counting the successful SQL operations that changed the\\ndatabase over a period of time. I am not sure whether there will be many\\nfailure operations that can occur on an instance to be counted.\\n\\n\\n> ------\\n> I have noticed that pgstat_count_sqlstmt is called from\\n> exec_simple_query and exec_execute_message. Don't we want to track the\\n> statement executed from functions?\\n> -------\\n\\n\\nThe view is currently designed to count user/application initiated SQL\\nstatements,\\nbut not the internal SQL queries that are getting executed from functions\\nand etc.\\n\\n>>+void\\n>>+pgstat_count_sqlstmt(const char *commandTag)\\n>>+{\\n>>+ PgStat_SqlstmtEntry *htabent;\\n>>+ bool found;\\n>>+\\n>>+ if (!pgstat_track_sql)\\n>>+ return\\n>\\n>Callers of pgstat_count_sqlstmt are always ensuring that\\n>pgstat_track_sql is true, seems it's redundant here.\\n\\nRemoved the check in pgstat_count_sqlstmt statement and add it\\nexec_execute_message\\nfunction where the check is missed.\\n\\nUpdated patch attached.\\n\\nRegards,\\nHari Babu\\nFujitsu Australia\\n\\n\\n-- \\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, 18 Jan 2017 16:15:30 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On 2017/01/18 14:15, Haribabu Kommi wrote:\\n>\\n>\\n> On Sat, Jan 14, 2017 at 2:58 AM, Dilip Kumar <dilipbalaut@gmail.com \\n> <mailto:dilipbalaut@gmail.com>> wrote:\\n>\\n> On Wed, Jan 11, 2017 at 7:47 PM, Dilip Kumar\\n> <dilipbalaut@gmail.com <mailto:dilipbalaut@gmail.com>> wrote:\\n> > +void\\n> > +pgstat_count_sqlstmt(const char *commandTag)\\n> > +{\\n> > + PgStat_SqlstmtEntry *htabent;\\n> > + bool found;\\n> > +\\n> > + if (!pgstat_track_sql)\\n> > + return\\n> >\\n> > Callers of pgstat_count_sqlstmt are always ensuring that\\n> > pgstat_track_sql is true, seems it's redundant here.\\n>\\n> I have done testing of the feature, it's mostly working as per the\\n> expectation.\\n>\\n>\\n> Thanks for the review and test.\\n>\\n> The use case for this view is to find out the number of different SQL \\n> statements\\n> that are getting executed successfully on an instance by the \\n> application/user.\\n>\\n> I have few comments/questions.\\n>\\n> --------\\n> If you execute the query with EXECUTE then commandTag will be EXECUTE\\n> that way we will not show the actual query type, I mean all the\\n> statements will get the common tag \\\"EXECUTE\\\".\\n>\\n> You can try this.\\n> PREPARE fooplan (int) AS INSERT INTO t VALUES($1);\\n> EXECUTE fooplan(1);\\n>\\n> ------\\n>\\n>\\n> Yes, that's correct. Currently the count is incremented based on the \\n> base command,\\n> because of this reason, the EXECUTE is increased, instead of the \\n> actual operation.\\n>\\n> + /* Destroy the SQL statement counter stats HashTable */\\n> + hash_destroy(pgstat_sql);\\n> +\\n> + /* Create SQL statement counter Stats HashTable */\\n>\\n> I think in the patch we have used mixed of \\\"Stats HashTable\\\" and\\n> \\\"stats HashTable\\\", I think better\\n> to be consistent throughout the patch. Check other similar instances.\\n>\\n> ----------------\\n>\\n>\\n> Corrected.\\n>\\n>\\n> @@ -1102,6 +1102,12 @@ exec_simple_query(const char *query_string)\\n>\\n> PortalDrop(portal, false);\\n>\\n> + /*\\n> + * Count SQL statement for pg_stat_sql view\\n> + */\\n> + if (pgstat_track_sql)\\n> + pgstat_count_sqlstmt(commandTag);\\n>\\n> We are only counting the successful SQL statement, Is that\\n> intentional?\\n>\\n>\\n> Yes, I thought of counting the successful SQL operations that changed the\\n> database over a period of time. I am not sure whether there will be many\\n> failure operations that can occur on an instance to be counted.\\n>\\n> ------\\n> I have noticed that pgstat_count_sqlstmt is called from\\n> exec_simple_query and exec_execute_message. Don't we want to track the\\n> statement executed from functions?\\n> -------\\n>\\n> The view is currently designed to count user/application initiated SQL \\n> statements,\\n> but not the internal SQL queries that are getting executed from \\n> functions and etc.\\n>\\n> >>+void\\n> >>+pgstat_count_sqlstmt(const char *commandTag)\\n> >>+{\\n> >>+ PgStat_SqlstmtEntry *htabent;\\n> >>+ bool found;\\n> >>+\\n> >>+ if (!pgstat_track_sql)\\n> >>+ return\\n> >\\n> >Callers of pgstat_count_sqlstmt are always ensuring that\\n> >pgstat_track_sql is true, seems it's redundant here.\\n>\\n> Removed the check in pgstat_count_sqlstmt statement and add it \\n> exec_execute_message\\n> function where the check is missed.\\n>\\n> Updated patch attached.\\nI have reviewed the patch. All the test cases works fine.\\n\\nRegards,\\nVinayak Pokale\\nNTT Open Source Software Center\\n\\n\\n\\n\\n\\n\\n\\nOn 2017/01/18 14:15, Haribabu Kommi\\n wrote:\\n\\n\\n\\n\\nOn Sat, Jan 14, 2017 at 2:58 AM,\\n Dilip Kumar <dilipbalaut@gmail.com>\\n wrote:\\nOn Wed, Jan 11,\\n 2017 at 7:47 PM, Dilip Kumar <dilipbalaut@gmail.com>\\n wrote:\\n > +void\\n > +pgstat_count_sqlstmt(const char *commandTag)\\n > +{\\n > + PgStat_SqlstmtEntry *htabent;\\n > + bool found;\\n > +\\n > + if (!pgstat_track_sql)\\n > + return\\n >\\n > Callers of pgstat_count_sqlstmt are always ensuring\\n that\\n > pgstat_track_sql is true, seems it's redundant\\n here.\\n\\nI have done testing of the feature, it's mostly\\n working as per the expectation.\\n\\n\\n\\nThanks for the review and test.\\n\\n\\nThe use case for this view is to find out the number of\\n different SQL statements\\nthat are getting executed successfully on an instance\\n by the application/user. \\n\\n\\n\\n I have few comments/questions.\\n\\n --------\\n If you execute the query with EXECUTE then commandTag will\\n be EXECUTE\\n that way we will not show the actual query type, I mean\\n all the\\n statements will get the common tag \\\"EXECUTE\\\".\\n\\n You can try this.\\n PREPARE fooplan (int) AS INSERT INTO t VALUES($1);\\n EXECUTE fooplan(1);\\n\\n ------\\n\\n\\n\\nYes, that's correct. Currently the count is incremented\\n based on the base command,\\nbecause of this reason, the EXECUTE is increased,\\n instead of the actual operation.\\n\\n\\n \\n\\n + /* Destroy the SQL statement counter stats HashTable */\\n + hash_destroy(pgstat_sql);\\n +\\n + /* Create SQL statement counter Stats HashTable */\\n\\n I think in the patch we have used mixed of \\\"Stats\\n HashTable\\\" and\\n \\\"stats HashTable\\\", I think better\\n to be consistent throughout the patch. Check other similar\\n instances.\\n\\n ----------------\\n\\n\\n\\nCorrected.\\n \\n\\n\\n @@ -1102,6 +1102,12 @@ exec_simple_query(const char\\n *query_string)\\n\\n PortalDrop(portal, false);\\n\\n + /*\\n + * Count SQL statement for pg_stat_sql view\\n + */\\n + if (pgstat_track_sql)\\n + pgstat_count_sqlstmt(commandTag);\\n\\n We are only counting the successful SQL statement, Is that\\n intentional?\\n\\n\\n\\nYes, I thought of counting the successful SQL\\n operations that changed the\\ndatabase over a period of time. I am not sure whether\\n there will be many\\nfailure operations that can occur on an instance to be\\n counted.\\n \\n\\n ------\\n I have noticed that pgstat_count_sqlstmt is called from\\n exec_simple_query and exec_execute_message. Don't we want\\n to track the\\n statement executed from functions?\\n -------\\n \\n\\n The view is currently designed to count user/application\\n initiated SQL statements,\\nbut not the internal SQL queries that\\n are getting executed from functions and etc.\\n\\n\\n>>+void\\n>>+pgstat_count_sqlstmt(const\\n char *commandTag)\\n>>+{\\n>>+ PgStat_SqlstmtEntry\\n *htabent;\\n>>+ bool found;\\n>>+\\n>>+ if\\n (!pgstat_track_sql)\\n>>+ return\\n >\\n>Callers of\\n pgstat_count_sqlstmt are always ensuring that\\n>pgstat_track_sql is true,\\n seems it's redundant here.\\n\\n\\n\\n Removed the check in pgstat_count_sqlstmt\\n statement and add it exec_execute_message\\nfunction\\n where the check is missed.\\n\\n\\nUpdated\\n patch attached.\\n\\n\\n I have reviewed the patch. All the test cases works fine.\\n\\n Regards,\\n Vinayak Pokale\\n NTT Open Source Software Center\",\n \"msg_date\": \"Fri, 20 Jan 2017 13:31:43 +0900\",\n \"msg_from\": \"vinayak <Pokale_Vinayak_q3@lab.ntt.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 18, 2017 at 10:45 AM, Haribabu Kommi\\n<kommi.haribabu@gmail.com> wrote:\\n> The use case for this view is to find out the number of different SQL\\n> statements\\n> that are getting executed successfully on an instance by the\\n> application/user.\\n>\\n>> I have few comments/questions.\\n>>\\n>> --------\\n>> If you execute the query with EXECUTE then commandTag will be EXECUTE\\n>> that way we will not show the actual query type, I mean all the\\n>> statements will get the common tag \\\"EXECUTE\\\".\\n>>\\n>> You can try this.\\n>> PREPARE fooplan (int) AS INSERT INTO t VALUES($1);\\n>> EXECUTE fooplan(1);\\n>>\\n>> ------\\n>\\n>\\n> Yes, that's correct. Currently the count is incremented based on the base\\n> command,\\n> because of this reason, the EXECUTE is increased, instead of the actual\\n> operation.\\n\\nOkay, As per your explaination, and Robert also pointed out that it\\ncan be a feature for someone and bug for others. I would have\\npreferred it shows actual SQL statement. But I have no objection to\\nwhat you are doing.\\n>\\n>\\n>>\\n>> + /* Destroy the SQL statement counter stats HashTable */\\n>> + hash_destroy(pgstat_sql);\\n>> +\\n>> + /* Create SQL statement counter Stats HashTable */\\n>>\\n>> I think in the patch we have used mixed of \\\"Stats HashTable\\\" and\\n>> \\\"stats HashTable\\\", I think better\\n>> to be consistent throughout the patch. Check other similar instances.\\n>>\\n>> ----------------\\n>\\n>\\n> Corrected.\\n>\\n>>\\n>>\\n>> @@ -1102,6 +1102,12 @@ exec_simple_query(const char *query_string)\\n>>\\n>> PortalDrop(portal, false);\\n>>\\n>> + /*\\n>> + * Count SQL statement for pg_stat_sql view\\n>> + */\\n>> + if (pgstat_track_sql)\\n>> + pgstat_count_sqlstmt(commandTag);\\n>>\\n>> We are only counting the successful SQL statement, Is that intentional?\\n>\\n>\\n> Yes, I thought of counting the successful SQL operations that changed the\\n> database over a period of time. I am not sure whether there will be many\\n> failure operations that can occur on an instance to be counted.\\n\\nOkay..\\n>\\n>>\\n>> ------\\n>> I have noticed that pgstat_count_sqlstmt is called from\\n>> exec_simple_query and exec_execute_message. Don't we want to track the\\n>> statement executed from functions?\\n>> -------\\n> The view is currently designed to count user/application initiated SQL\\n> statements,\\n> but not the internal SQL queries that are getting executed from functions\\n> and etc.\\n\\nOkay..\\n>\\n>>>+void\\n>>>+pgstat_count_sqlstmt(const char *commandTag)\\n>>>+{\\n>>>+ PgStat_SqlstmtEntry *htabent;\\n>>>+ bool found;\\n>>>+\\n>>>+ if (!pgstat_track_sql)\\n>>>+ return\\n>>\\n>>Callers of pgstat_count_sqlstmt are always ensuring that\\n>>pgstat_track_sql is true, seems it's redundant here.\\n>\\n> Removed the check in pgstat_count_sqlstmt statement and add it\\n> exec_execute_message\\n> function where the check is missed.\\n>\\n> Updated patch attached.\\n\\nOverall patch looks fine to me and marking it \\\"ready for committer\\\".\\n\\nThere is two design decision, which I leave it for committer's decision.\\n\\n1. EXECUTE statement should show only as EXECUTE count, not the\\nactual SQL query.\\n2. should we count statement execute from Functions or only statement\\nwhat is executed directly by the user as it's doing now?\\n\\n\\n-- \\nRegards,\\nDilip Kumar\\nEnterpriseDB: http://www.enterprisedb.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, 24 Jan 2017 10:29:32 +0530\",\n \"msg_from\": \"Dilip Kumar <dilipbalaut@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 24, 2017 at 3:59 PM, Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n\\n>\\n> Overall patch looks fine to me and marking it \\\"ready for committer\\\".\\n>\\n> There is two design decision, which I leave it for committer's decision.\\n>\\n> 1. EXECUTE statement should show only as EXECUTE count, not the\\n> actual SQL query.\\n> 2. should we count statement execute from Functions or only statement\\n> what is executed directly by the user as it's doing now?\\n\\n\\nThanks for the review.\\nLet's wait for the committer's opinion.\\n\\nRegards,\\nHari Babu\\nFujitsu Australia\\n\\nOn Tue, Jan 24, 2017 at 3:59 PM, Dilip Kumar <dilipbalaut@gmail.com> wrote:\\nOverall patch looks fine to me and marking it \\\"ready for committer\\\".\\n\\nThere is two design decision, which I leave it for committer's decision.\\n\\n1. EXECUTE statement should show only as EXECUTE count, not the\\nactual SQL query.\\n2. should we count statement execute from Functions or only statement\\nwhat is executed directly by the user as it's doing now?Thanks for the review.Let's wait for the committer's opinion.Regards,Hari BabuFujitsu Australia\",\n \"msg_date\": \"Fri, 27 Jan 2017 12:26:01 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 27, 2017 at 10:26 AM, Haribabu Kommi\\n<kommi.haribabu@gmail.com> wrote:\\n> Thanks for the review.\\n> Let's wait for the committer's opinion.\\n\\nI have moved this patch to CF 2017-03 to wait for this to happen.\\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, 1 Feb 2017 13:13:11 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Wed, Feb 1, 2017 at 3:13 PM, Michael Paquier <michael.paquier@gmail.com>\\nwrote:\\n\\n> On Fri, Jan 27, 2017 at 10:26 AM, Haribabu Kommi\\n> <kommi.haribabu@gmail.com> wrote:\\n> > Thanks for the review.\\n> > Let's wait for the committer's opinion.\\n>\\n> I have moved this patch to CF 2017-03 to wait for this to happen.\\n>\\n\\nAttached a rebased patch to resolve the OID conflict.\\n\\nRegards,\\nHari Babu\\nFujitsu Australia\\n\\n\\n-- \\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, 8 Mar 2017 14:39:23 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\n\\nI'm somewhat inclined to think that this really would be better done in\\nan extension like pg_stat_statements.\\n\\n\\nOn 2017-03-08 14:39:23 +1100, Haribabu Kommi wrote:\\n> On Wed, Feb 1, 2017 at 3:13 PM, Michael Paquier <michael.paquier@gmail.com>\\n\\n> + <varlistentry id=\\\"guc-track-sql\\\" xreflabel=\\\"track_sql\\\">\\n> + <term><varname>track_sql</varname> (<type>boolean</type>)\\n> + <indexterm>\\n> + <primary><varname>track_sql</> configuration parameter</primary>\\n> + </indexterm>\\n> + </term>\\n> + <listitem>\\n> + <para>\\n> + Enables collection of different SQL statement statistics that are\\n> + executed on the instance. This parameter is off by default. Only\\n> + superusers can change this setting.\\n> + </para>\\n> + </listitem>\\n> + </varlistentry>\\n> +\\n\\nI don't like this name much, seems a bit too generic to\\nme. 'track_activities', 'track_io_timings' are at least a bit clearer.\\nHow about track_statement_statistics + corresponding view/function renaming?\\n\\n\\n> + <table id=\\\"pg-stat-sql-view\\\" xreflabel=\\\"pg_stat_sql\\\">\\n> + <title><structname>pg_stat_sql</structname> View</title>\\n> + <tgroup cols=\\\"3\\\">\\n> + <thead>\\n> + <row>\\n> + <entry>Column</entry>\\n> + <entry>Type</entry>\\n> + <entry>Description</entry>\\n> + </row>\\n> + </thead>\\n> +\\n> + <tbody>\\n> + <row>\\n> + <entry><structfield>tag</></entry>\\n> + <entry><type>text</></entry>\\n> + <entry>Name of the SQL statement</entry>\\n> + </row>\\n\\nIt's not really the \\\"name\\\" of a statement. Internally and IIRC elsewhere\\nin the docs we describe something like this as tag?\\n\\n\\n> +/* ----------\\n> + * pgstat_send_sqlstmt(void)\\n> + *\\n> + *\\t\\tSend SQL statement statistics to the collector\\n> + * ----------\\n> + */\\n> +static void\\n> +pgstat_send_sqlstmt(void)\\n> +{\\n> +\\tPgStat_MsgSqlstmt msg;\\n> +\\tPgStat_SqlstmtEntry *entry;\\n> +\\tHASH_SEQ_STATUS hstat;\\n> +\\n> +\\tif (pgstat_backend_sql == NULL)\\n> +\\t\\treturn;\\n> +\\n> +\\tpgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_SQLSTMT);\\n> +\\tmsg.m_nentries = 0;\\n> +\\n> +\\thash_seq_init(&hstat, pgstat_backend_sql);\\n> +\\twhile ((entry = (PgStat_SqlstmtEntry *) hash_seq_search(&hstat)) != NULL)\\n> +\\t{\\n> +\\t\\tPgStat_SqlstmtEntry *m_ent;\\n> +\\n> +\\t\\t/* Skip it if no counts accumulated since last time */\\n> +\\t\\tif (entry->count == 0)\\n> +\\t\\t\\tcontinue;\\n> +\\n> +\\t\\t/* need to convert format of time accumulators */\\n> +\\t\\tm_ent = &msg.m_entry[msg.m_nentries];\\n> +\\t\\tmemcpy(m_ent, entry, sizeof(PgStat_SqlstmtEntry));\\n> +\\n> +\\t\\tif (++msg.m_nentries >= PGSTAT_NUM_SQLSTMTS)\\n> +\\t\\t{\\n> +\\t\\t\\tpgstat_send(&msg, offsetof(PgStat_MsgSqlstmt, m_entry[0]) +\\n> +\\t\\t\\t\\t\\t\\tmsg.m_nentries * sizeof(PgStat_SqlstmtEntry));\\n> +\\t\\t\\tmsg.m_nentries = 0;\\n> +\\t\\t}\\n> +\\n> +\\t\\t/* reset the entry's count */\\n> +\\t\\tentry->count = 0;\\n> +\\t}\\n> +\\n> +\\tif (msg.m_nentries > 0)\\n> +\\t\\tpgstat_send(&msg, offsetof(PgStat_MsgSqlstmt, m_entry[0]) +\\n> +\\t\\t\\t\\t\\tmsg.m_nentries * sizeof(PgStat_SqlstmtEntry));\\n\\nHm. So pgstat_backend_sql is never deleted, which'll mean that if a\\nbackend lives long we'll continually iterate over a lot of 0 entries. I\\nthink performance evaluation of this feature should take that into\\naccount.\\n\\n\\n> +}\\n> +\\n> +/*\\n> + * Count SQL statement for pg_stat_sql view\\n> + */\\n> +void\\n> +pgstat_count_sqlstmt(const char *commandTag)\\n> +{\\n> +\\tPgStat_SqlstmtEntry *htabent;\\n> +\\tbool\\t\\tfound;\\n> +\\n> +\\tif (!pgstat_backend_sql)\\n> +\\t{\\n> +\\t\\t/* First time through - initialize SQL statement stat table */\\n> +\\t\\tHASHCTL\\t\\thash_ctl;\\n> +\\n> +\\t\\tmemset(&hash_ctl, 0, sizeof(hash_ctl));\\n> +\\t\\thash_ctl.keysize = NAMEDATALEN;\\n> +\\t\\thash_ctl.entrysize = sizeof(PgStat_SqlstmtEntry);\\n> +\\t\\tpgstat_backend_sql = hash_create(\\\"SQL statement stat entries\\\",\\n> +\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t PGSTAT_SQLSTMT_HASH_SIZE,\\n> +\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t &hash_ctl,\\n> +\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t HASH_ELEM | HASH_BLOBS);\\n> +\\t}\\n> +\\n> +\\t/* Get the stats entry for this SQL statement, create if necessary */\\n> +\\thtabent = hash_search(pgstat_backend_sql, commandTag,\\n> +\\t\\t\\t\\t\\t\\t HASH_ENTER, &found);\\n> +\\tif (!found)\\n> +\\t\\thtabent->count = 1;\\n> +\\telse\\n> +\\t\\thtabent->count++;\\n> +}\\n\\n\\nThat's not really something in this patch, but a lot of this would be\\nbetter if we didn't internally have command tags as strings, but as an\\nenum. We should really only convert to a string with needed. That\\nwe're doing string comparisons on Portal->commandTag is just plain bad.\\n\\nIf so, we could also make this whole patch a lot cheaper - have a fixed\\nsize array that has an entry for every possible tag (possibly even in\\nshared memory, and then use atomics there).\\n\\n> +++ b/src/backend/tcop/postgres.c\\n> @@ -1109,6 +1109,12 @@ exec_simple_query(const char *query_string)\\n> \\n> \\t\\tPortalDrop(portal, false);\\n> \\n> +\\t\\t/*\\n> +\\t\\t * Count SQL statement for pg_stat_sql view\\n> +\\t\\t */\\n> +\\t\\tif (pgstat_track_sql)\\n> +\\t\\t\\tpgstat_count_sqlstmt(commandTag);\\n> +\\n\\nIsn't doing this at the message level quite wrong? What about\\nstatements executed from functions and such? Shouldn't this integrate\\nat the executor level instead?\\n\\n\\n> \\t\\tif (IsA(parsetree->stmt, TransactionStmt))\\n> \\t\\t{\\n> \\t\\t\\t/*\\n> @@ -1991,6 +1997,29 @@ exec_execute_message(const char *portal_name, long max_rows)\\n> \\n> \\t(*receiver->rDestroy) (receiver);\\n> \\n> +\\t/*\\n> +\\t * Count SQL Statement for pgx_stat_sql\\n> +\\t */\\n> +\\tif (pgstat_track_sql)\\n> +\\t{\\n> +\\t\\tCachedPlanSource *psrc = NULL;\\n> +\\n> +\\t\\tif (portal->prepStmtName)\\n> +\\t\\t{\\n> +\\t\\t\\tPreparedStatement *pstmt;\\n> +\\n> +\\t\\t\\tpstmt = FetchPreparedStatement(portal->prepStmtName, false);\\n> +\\t\\t\\tif (pstmt)\\n> +\\t\\t\\t\\tpsrc = pstmt->plansource;\\n> +\\t\\t}\\n> +\\t\\telse\\n> +\\t\\t\\tpsrc = unnamed_stmt_psrc;\\n> +\\n> +\\t\\t/* psrc should not be NULL here */\\n> +\\t\\tif (psrc && psrc->commandTag && !execute_is_fetch && pgstat_track_sql)\\n> +\\t\\t\\tpgstat_count_sqlstmt(psrc->commandTag);\\n\\nWait, we're re-fetching the statement here? That doesn't sound\\nalright.\\n\\n\\n> +Datum\\n> +pg_stat_sql(PG_FUNCTION_ARGS)\\n> +{\\n> +\\tTupleDesc\\ttupdesc;\\n> +\\tDatum\\t\\tvalues[2];\\n> +\\tbool\\t\\tnulls[2];\\n> +\\tReturnSetInfo *rsi;\\n> +\\tMemoryContext old_cxt;\\n> +\\tTuplestorestate *tuple_store;\\n> +\\tPgStat_SqlstmtEntry *sqlstmtentry;\\n> +\\tHASH_SEQ_STATUS hstat;\\n> +\\n> +\\t/* Function call to let the backend read the stats file */\\n> +\\tpgstat_fetch_global();\\n> +\\n> +\\t/* Initialize values and nulls arrays */\\n> +\\tMemSet(values, 0, sizeof(values));\\n> +\\tMemSet(nulls, 0, sizeof(nulls));\\n\\nwhy not instead declare values[2] = {0}? Obviously not important.\\n\\n\\n> +\\ttuple_store =\\n> +\\t\\ttuplestore_begin_heap(rsi->allowedModes & SFRM_Materialize_Random,\\n> +\\t\\t\\t\\t\\t\\t\\t false, work_mem);\\n\\nHuh, why do we need random?\\n\\n\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n-- \\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 Apr 2017 12:17:57 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Thu, Apr 6, 2017 at 5:17 AM, Andres Freund <andres@anarazel.de> wrote:\\n\\n> Hi,\\n>\\n>\\n> I'm somewhat inclined to think that this really would be better done in\\n> an extension like pg_stat_statements.\\n>\\n\\nThanks for the review.\\n\\n\\n\\n>\\n> > +}\\n> > +\\n> > +/*\\n> > + * Count SQL statement for pg_stat_sql view\\n> > + */\\n> > +void\\n> > +pgstat_count_sqlstmt(const char *commandTag)\\n> > +{\\n> > + PgStat_SqlstmtEntry *htabent;\\n> > + bool found;\\n> > +\\n> > + if (!pgstat_backend_sql)\\n> > + {\\n> > + /* First time through - initialize SQL statement stat\\n> table */\\n> > + HASHCTL hash_ctl;\\n> > +\\n> > + memset(&hash_ctl, 0, sizeof(hash_ctl));\\n> > + hash_ctl.keysize = NAMEDATALEN;\\n> > + hash_ctl.entrysize = sizeof(PgStat_SqlstmtEntry);\\n> > + pgstat_backend_sql = hash_create(\\\"SQL statement stat\\n> entries\\\",\\n> > +\\n> PGSTAT_SQLSTMT_HASH_SIZE,\\n> > +\\n> &hash_ctl,\\n> > +\\n> HASH_ELEM | HASH_BLOBS);\\n> > + }\\n> > +\\n> > + /* Get the stats entry for this SQL statement, create if necessary\\n> */\\n> > + htabent = hash_search(pgstat_backend_sql, commandTag,\\n> > + HASH_ENTER, &found);\\n> > + if (!found)\\n> > + htabent->count = 1;\\n> > + else\\n> > + htabent->count++;\\n> > +}\\n>\\n>\\n> That's not really something in this patch, but a lot of this would be\\n> better if we didn't internally have command tags as strings, but as an\\n> enum. We should really only convert to a string with needed. That\\n> we're doing string comparisons on Portal->commandTag is just plain bad.\\n>\\n> If so, we could also make this whole patch a lot cheaper - have a fixed\\n> size array that has an entry for every possible tag (possibly even in\\n> shared memory, and then use atomics there).\\n>\\n\\nDuring the development, I thought of using an array of all command tags\\nand update the counters using the tag name, but not like the enum values.\\nNow I have to create a mapping array with enums and tag names for easier\\ncounter updates. The problem in this approach is, whenever any addition\\nof new commands, this mapping array needs to be updated with both\\nnew enum and new tag name, whereas with hash table approach, it works\\nfuture command additions also, but there is some performance overhead\\ncompared to the array approach.\\n\\nI will modify the patch to use the array approach and provide it to the next\\ncommitfest.\\n\\n\\n\\n> > +++ b/src/backend/tcop/postgres.c\\n> > @@ -1109,6 +1109,12 @@ exec_simple_query(const char *query_string)\\n> >\\n> > PortalDrop(portal, false);\\n> >\\n> > + /*\\n> > + * Count SQL statement for pg_stat_sql view\\n> > + */\\n> > + if (pgstat_track_sql)\\n> > + pgstat_count_sqlstmt(commandTag);\\n> > +\\n>\\n> Isn't doing this at the message level quite wrong? What about\\n> statements executed from functions and such? Shouldn't this integrate\\n> at the executor level instead?\\n>\\n\\nCurrently the patch is designed to find out only the user executed\\nstatements\\nthat are successfully finished, (no internal statements that are executed\\nfrom\\nfunctions and etc).\\n\\nThe same question was asked by dilip also in earlier mails, may be now it is\\nthe time that we can decide the approach of statement counts.\\n\\n\\nRegards,\\nHari Babu\\nFujitsu Australia\\n\\nOn Thu, Apr 6, 2017 at 5:17 AM, Andres Freund <andres@anarazel.de> wrote:Hi,\\n\\n\\nI'm somewhat inclined to think that this really would be better done in\\nan extension like pg_stat_statements.Thanks for the review. \\n> +}\\n> +\\n> +/*\\n> + * Count SQL statement for pg_stat_sql view\\n> + */\\n> +void\\n> +pgstat_count_sqlstmt(const char *commandTag)\\n> +{\\n> + PgStat_SqlstmtEntry *htabent;\\n> + bool found;\\n> +\\n> + if (!pgstat_backend_sql)\\n> + {\\n> + /* First time through - initialize SQL statement stat table */\\n> + HASHCTL hash_ctl;\\n> +\\n> + memset(&hash_ctl, 0, sizeof(hash_ctl));\\n> + hash_ctl.keysize = NAMEDATALEN;\\n> + hash_ctl.entrysize = sizeof(PgStat_SqlstmtEntry);\\n> + pgstat_backend_sql = hash_create(\\\"SQL statement stat entries\\\",\\n> + PGSTAT_SQLSTMT_HASH_SIZE,\\n> + &hash_ctl,\\n> + HASH_ELEM | HASH_BLOBS);\\n> + }\\n> +\\n> + /* Get the stats entry for this SQL statement, create if necessary */\\n> + htabent = hash_search(pgstat_backend_sql, commandTag,\\n> + HASH_ENTER, &found);\\n> + if (!found)\\n> + htabent->count = 1;\\n> + else\\n> + htabent->count++;\\n> +}\\n\\n\\nThat's not really something in this patch, but a lot of this would be\\nbetter if we didn't internally have command tags as strings, but as an\\nenum. We should really only convert to a string with needed. That\\nwe're doing string comparisons on Portal->commandTag is just plain bad.\\n\\nIf so, we could also make this whole patch a lot cheaper - have a fixed\\nsize array that has an entry for every possible tag (possibly even in\\nshared memory, and then use atomics there).During the development, I thought of using an array of all command tagsand update the counters using the tag name, but not like the enum values.Now I have to create a mapping array with enums and tag names for easiercounter updates. The problem in this approach is, whenever any additionof new commands, this mapping array needs to be updated with bothnew enum and new tag name, whereas with hash table approach, it worksfuture command additions also, but there is some performance overheadcompared to the array approach.I will modify the patch to use the array approach and provide it to the nextcommitfest. \\n> +++ b/src/backend/tcop/postgres.c\\n> @@ -1109,6 +1109,12 @@ exec_simple_query(const char *query_string)\\n>\\n> PortalDrop(portal, false);\\n>\\n> + /*\\n> + * Count SQL statement for pg_stat_sql view\\n> + */\\n> + if (pgstat_track_sql)\\n> + pgstat_count_sqlstmt(commandTag);\\n> +\\n\\nIsn't doing this at the message level quite wrong? What about\\nstatements executed from functions and such? Shouldn't this integrate\\nat the executor level instead?Currently the patch is designed to find out only the user executed statementsthat are successfully finished, (no internal statements that are executed fromfunctions and etc).The same question was asked by dilip also in earlier mails, may be now it isthe time that we can decide the approach of statement counts.Regards,Hari BabuFujitsu Australia\",\n \"msg_date\": \"Fri, 7 Apr 2017 15:19:28 +1000\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"Dear Hackers,\\n\\nWe have resurrected this 2 year old \\\"SQL statements statistics counter\\\" proposal from Hari.\\n\\nThe attached patch addresses the previous review issues.\\n\\nThe commit-fest entry has been moved forward to the next commit-fest\\nhttps://commitfest.postgresql.org/25/790/\\n\\nPlease review again, and consider if this is OK for \\\"Ready for Committer\\\" status.\\n\\nKind Regards\\n--\\nPeter Smith\\nFujitsu Australia\\n\\n\\n\\n-----Original Message-----\\nFrom: pgsql-hackers-owner@postgresql.org <pgsql-hackers-owner@postgresql.org> On Behalf Of Andres Freund\\nSent: Thursday, 6 April 2017 5:18 AM\\nTo: Haribabu Kommi <kommi.haribabu@gmail.com>\\nCc: Michael Paquier <michael.paquier@gmail.com>; Dilip Kumar <dilipbalaut@gmail.com>; vinayak <Pokale_Vinayak_q3@lab.ntt.co.jp>; pgsql-hackers@postgresql.org\\nSubject: Re: New SQL counter statistics view (pg_stat_sql)\\n\\nHi,\\n\\n\\nI'm somewhat inclined to think that this really would be better done in an extension like pg_stat_statements.\\n\\n\\nOn 2017-03-08 14:39:23 +1100, Haribabu Kommi wrote:\\n> On Wed, Feb 1, 2017 at 3:13 PM, Michael Paquier \\n> <michael.paquier@gmail.com>\\n\\n> + <varlistentry id=\\\"guc-track-sql\\\" xreflabel=\\\"track_sql\\\">\\n> + <term><varname>track_sql</varname> (<type>boolean</type>)\\n> + <indexterm>\\n> + <primary><varname>track_sql</> configuration parameter</primary>\\n> + </indexterm>\\n> + </term>\\n> + <listitem>\\n> + <para>\\n> + Enables collection of different SQL statement statistics that are\\n> + executed on the instance. This parameter is off by default. Only\\n> + superusers can change this setting.\\n> + </para>\\n> + </listitem>\\n> + </varlistentry>\\n> +\\n\\nI don't like this name much, seems a bit too generic to me. 'track_activities', 'track_io_timings' are at least a bit clearer.\\nHow about track_statement_statistics + corresponding view/function renaming?\\n\\n\\n> + <table id=\\\"pg-stat-sql-view\\\" xreflabel=\\\"pg_stat_sql\\\">\\n> + <title><structname>pg_stat_sql</structname> View</title>\\n> + <tgroup cols=\\\"3\\\">\\n> + <thead>\\n> + <row>\\n> + <entry>Column</entry>\\n> + <entry>Type</entry>\\n> + <entry>Description</entry>\\n> + </row>\\n> + </thead>\\n> +\\n> + <tbody>\\n> + <row>\\n> + <entry><structfield>tag</></entry>\\n> + <entry><type>text</></entry>\\n> + <entry>Name of the SQL statement</entry>\\n> + </row>\\n\\nIt's not really the \\\"name\\\" of a statement. Internally and IIRC elsewhere in the docs we describe something like this as tag?\\n\\n\\n> +/* ----------\\n> + * pgstat_send_sqlstmt(void)\\n> + *\\n> + *\\t\\tSend SQL statement statistics to the collector\\n> + * ----------\\n> + */\\n> +static void\\n> +pgstat_send_sqlstmt(void)\\n> +{\\n> +\\tPgStat_MsgSqlstmt msg;\\n> +\\tPgStat_SqlstmtEntry *entry;\\n> +\\tHASH_SEQ_STATUS hstat;\\n> +\\n> +\\tif (pgstat_backend_sql == NULL)\\n> +\\t\\treturn;\\n> +\\n> +\\tpgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_SQLSTMT);\\n> +\\tmsg.m_nentries = 0;\\n> +\\n> +\\thash_seq_init(&hstat, pgstat_backend_sql);\\n> +\\twhile ((entry = (PgStat_SqlstmtEntry *) hash_seq_search(&hstat)) != NULL)\\n> +\\t{\\n> +\\t\\tPgStat_SqlstmtEntry *m_ent;\\n> +\\n> +\\t\\t/* Skip it if no counts accumulated since last time */\\n> +\\t\\tif (entry->count == 0)\\n> +\\t\\t\\tcontinue;\\n> +\\n> +\\t\\t/* need to convert format of time accumulators */\\n> +\\t\\tm_ent = &msg.m_entry[msg.m_nentries];\\n> +\\t\\tmemcpy(m_ent, entry, sizeof(PgStat_SqlstmtEntry));\\n> +\\n> +\\t\\tif (++msg.m_nentries >= PGSTAT_NUM_SQLSTMTS)\\n> +\\t\\t{\\n> +\\t\\t\\tpgstat_send(&msg, offsetof(PgStat_MsgSqlstmt, m_entry[0]) +\\n> +\\t\\t\\t\\t\\t\\tmsg.m_nentries * sizeof(PgStat_SqlstmtEntry));\\n> +\\t\\t\\tmsg.m_nentries = 0;\\n> +\\t\\t}\\n> +\\n> +\\t\\t/* reset the entry's count */\\n> +\\t\\tentry->count = 0;\\n> +\\t}\\n> +\\n> +\\tif (msg.m_nentries > 0)\\n> +\\t\\tpgstat_send(&msg, offsetof(PgStat_MsgSqlstmt, m_entry[0]) +\\n> +\\t\\t\\t\\t\\tmsg.m_nentries * sizeof(PgStat_SqlstmtEntry));\\n\\nHm. So pgstat_backend_sql is never deleted, which'll mean that if a backend lives long we'll continually iterate over a lot of 0 entries. I think performance evaluation of this feature should take that into account.\\n\\n\\n> +}\\n> +\\n> +/*\\n> + * Count SQL statement for pg_stat_sql view */ void \\n> +pgstat_count_sqlstmt(const char *commandTag) {\\n> +\\tPgStat_SqlstmtEntry *htabent;\\n> +\\tbool\\t\\tfound;\\n> +\\n> +\\tif (!pgstat_backend_sql)\\n> +\\t{\\n> +\\t\\t/* First time through - initialize SQL statement stat table */\\n> +\\t\\tHASHCTL\\t\\thash_ctl;\\n> +\\n> +\\t\\tmemset(&hash_ctl, 0, sizeof(hash_ctl));\\n> +\\t\\thash_ctl.keysize = NAMEDATALEN;\\n> +\\t\\thash_ctl.entrysize = sizeof(PgStat_SqlstmtEntry);\\n> +\\t\\tpgstat_backend_sql = hash_create(\\\"SQL statement stat entries\\\",\\n> +\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t PGSTAT_SQLSTMT_HASH_SIZE,\\n> +\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t &hash_ctl,\\n> +\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t HASH_ELEM | HASH_BLOBS);\\n> +\\t}\\n> +\\n> +\\t/* Get the stats entry for this SQL statement, create if necessary */\\n> +\\thtabent = hash_search(pgstat_backend_sql, commandTag,\\n> +\\t\\t\\t\\t\\t\\t HASH_ENTER, &found);\\n> +\\tif (!found)\\n> +\\t\\thtabent->count = 1;\\n> +\\telse\\n> +\\t\\thtabent->count++;\\n> +}\\n\\n\\nThat's not really something in this patch, but a lot of this would be better if we didn't internally have command tags as strings, but as an enum. We should really only convert to a string with needed. That we're doing string comparisons on Portal->commandTag is just plain bad.\\n\\nIf so, we could also make this whole patch a lot cheaper - have a fixed size array that has an entry for every possible tag (possibly even in shared memory, and then use atomics there).\\n\\n> +++ b/src/backend/tcop/postgres.c\\n> @@ -1109,6 +1109,12 @@ exec_simple_query(const char *query_string)\\n> \\n> \\t\\tPortalDrop(portal, false);\\n> \\n> +\\t\\t/*\\n> +\\t\\t * Count SQL statement for pg_stat_sql view\\n> +\\t\\t */\\n> +\\t\\tif (pgstat_track_sql)\\n> +\\t\\t\\tpgstat_count_sqlstmt(commandTag);\\n> +\\n\\nIsn't doing this at the message level quite wrong? What about statements executed from functions and such? Shouldn't this integrate at the executor level instead?\\n\\n\\n> \\t\\tif (IsA(parsetree->stmt, TransactionStmt))\\n> \\t\\t{\\n> \\t\\t\\t/*\\n> @@ -1991,6 +1997,29 @@ exec_execute_message(const char *portal_name, \\n> long max_rows)\\n> \\n> \\t(*receiver->rDestroy) (receiver);\\n> \\n> +\\t/*\\n> +\\t * Count SQL Statement for pgx_stat_sql\\n> +\\t */\\n> +\\tif (pgstat_track_sql)\\n> +\\t{\\n> +\\t\\tCachedPlanSource *psrc = NULL;\\n> +\\n> +\\t\\tif (portal->prepStmtName)\\n> +\\t\\t{\\n> +\\t\\t\\tPreparedStatement *pstmt;\\n> +\\n> +\\t\\t\\tpstmt = FetchPreparedStatement(portal->prepStmtName, false);\\n> +\\t\\t\\tif (pstmt)\\n> +\\t\\t\\t\\tpsrc = pstmt->plansource;\\n> +\\t\\t}\\n> +\\t\\telse\\n> +\\t\\t\\tpsrc = unnamed_stmt_psrc;\\n> +\\n> +\\t\\t/* psrc should not be NULL here */\\n> +\\t\\tif (psrc && psrc->commandTag && !execute_is_fetch && pgstat_track_sql)\\n> +\\t\\t\\tpgstat_count_sqlstmt(psrc->commandTag);\\n\\nWait, we're re-fetching the statement here? That doesn't sound alright.\\n\\n\\n> +Datum\\n> +pg_stat_sql(PG_FUNCTION_ARGS)\\n> +{\\n> +\\tTupleDesc\\ttupdesc;\\n> +\\tDatum\\t\\tvalues[2];\\n> +\\tbool\\t\\tnulls[2];\\n> +\\tReturnSetInfo *rsi;\\n> +\\tMemoryContext old_cxt;\\n> +\\tTuplestorestate *tuple_store;\\n> +\\tPgStat_SqlstmtEntry *sqlstmtentry;\\n> +\\tHASH_SEQ_STATUS hstat;\\n> +\\n> +\\t/* Function call to let the backend read the stats file */\\n> +\\tpgstat_fetch_global();\\n> +\\n> +\\t/* Initialize values and nulls arrays */\\n> +\\tMemSet(values, 0, sizeof(values));\\n> +\\tMemSet(nulls, 0, sizeof(nulls));\\n\\nwhy not instead declare values[2] = {0}? Obviously not important.\\n\\n\\n> +\\ttuple_store =\\n> +\\t\\ttuplestore_begin_heap(rsi->allowedModes & SFRM_Materialize_Random,\\n> +\\t\\t\\t\\t\\t\\t\\t false, work_mem);\\n\\nHuh, why do we need random?\\n\\n\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n--\\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Thu, 17 Oct 2019 02:22:32 +0000\",\n \"msg_from\": \"\\\"Smith, Peter\\\" <peters@fast.au.fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 17, 2019 at 3:22 PM Smith, Peter <peters@fast.au.fujitsu.com> wrote:\\n> We have resurrected this 2 year old \\\"SQL statements statistics counter\\\" proposal from Hari.\\n>\\n> The attached patch addresses the previous review issues.\\n\\nHi,\\n\\nNo comment on the patch but I noticed that the documentation changes\\ndon't build. Please make sure you can \\\"make docs\\\" successfully,\\nhaving installed the documentation tools[1].\\n\\n[1] https://www.postgresql.org/docs/devel/docguide-toolsets.html\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Nov 2019 15:42:39 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"From: Thomas Munro <thomas.munro@gmail.com> Sent: Monday, 4 November 2019 1:43 PM\\r\\n\\r\\n> No comment on the patch but I noticed that the documentation changes don't build. Please make sure you can \\\"make docs\\\" successfully, having installed the documentation tools[1].\\r\\n\\r\\nThanks for the feedback. An updated patch which fixes the docs issue is attached.\\r\\n\\r\\nKind Regards.\\r\\nPeter Smith\\r\\n---\\r\\nFujitsu Australia\",\n \"msg_date\": \"Wed, 13 Nov 2019 03:51:05 +0000\",\n \"msg_from\": \"\\\"Smith, Peter\\\" <peters@fast.au.fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On 2019-Nov-13, Smith, Peter wrote:\\n\\n> From: Thomas Munro <thomas.munro@gmail.com> Sent: Monday, 4 November 2019 1:43 PM\\n> \\n> > No comment on the patch but I noticed that the documentation changes don't build. Please make sure you can \\\"make docs\\\" successfully, having installed the documentation tools[1].\\n> \\n> Thanks for the feedback. An updated patch which fixes the docs issue is attached.\\n\\nSo, who is updating this patch for the new cmdtaglist.h stuff?\\n\\n-- \\n�lvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Tue, 3 Mar 2020 23:50:04 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Mar 3, 2020, at 6:50 PM, Alvaro Herrera <alvherre@2ndquadrant.com> wrote:\\n> \\n> On 2019-Nov-13, Smith, Peter wrote:\\n> \\n>> From: Thomas Munro <thomas.munro@gmail.com> Sent: Monday, 4 November 2019 1:43 PM\\n>> \\n>>> No comment on the patch but I noticed that the documentation changes don't build. Please make sure you can \\\"make docs\\\" successfully, having installed the documentation tools[1].\\n>> \\n>> Thanks for the feedback. An updated patch which fixes the docs issue is attached.\\n> \\n> So, who is updating this patch for the new cmdtaglist.h stuff?\\n\\nI have a much altered (competing?) patch inspired by pg_stat_sql. I am working to integrate the changes you made to my commandtag patch into my commandstats patch before posting it.\\n\\nIt might be good if somebody else tackled the pg_stat_sql patch in case my version is rejected.\\n\\nI don't want to highjack this thread to talk in detail about the other patch, so I'll just give an overview that might be useful for anybody thinking about putting effort into committing this patch first. \\n\\nThe motivation for the commandtag patch that you've committed (thanks again!) was to make the commandstats patch possible. I didn't like the way pg_stat_sql stored the tags as strings, but I also still don't like the way it needs to take a lock every time a backend increments the counter for a command. That is bad, I think, because the lock overhead for an otherwise fast command (like moving a cursor) may be high enough to discourage users from turning this feature on in production.\\n\\nI solve that by trading off speed and memory. For most commandtags, there is a single uint64 counter, and incrementing the counter requires taking a lock. This is very similar to how the other patch works. For the few commandtags that we deem worthy of the special treatment, there is an additional uint32 counter *per backend* reserved in shared memory. Backends update this counter using atomics, but do not require locks. When the counter is in danger of overflow, locks are taken to roll the count into the uint64 counters and zero out the backend-specific uint32 counter. I have roughly 20 command tags receiving this special treatment, but the design of the patch doesn't make it too hard to change the exact list of tags worthy of it, so I don't plan to stress too much about the exact set of tags. I'm sure you'll all tell me which ones you think do or do not deserve the treatment.\\n\\nIf it isn't obvious, the logic of having this treatment for only a subset of tags is that 192 tags * 4 bytes per counter * MaxBackends gets expensive for folks running a high number of backends. Taking that from 192 tags down to one or two dozen seems worthwhile.\\n\\nI expect to post on a separate thread before the day is over.\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 4 Mar 2020 13:52:56 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"Hackers,\\n\\nas mentioned in [1], I have created an implementation of command counter statistics very similar in purpose to the one already pending in the commitfest going by the name \\\"pg_stat_sql\\\". I don't really care if this implementation is seen as building on that one or as separate, but I was worried about hijacking that thread, so I'm starting this thead. There are elements of this patch that borrowed from that one, so if this is accepted, authorship should reflect that.\\n\\nSee the work by Haribabu Kommi (haribabu) at https://commitfest.postgresql.org/27/790/\\n\\nThe two main differences are that\\n\\n(1) This implementation is based on commandtags as enums, not strings and\\n(2) This implementation uses techniques to reduce lock contention\\n\\nI think (2) is the more important part.\\n\\n[1] https://www.postgresql.org/message-id/28F46159-3764-421B-904B-004DEA339310%40enterprisedb.com\\n\\n\\n\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Wed, 4 Mar 2020 19:43:25 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Command statistics system (cmdstats)\"\n },\n {\n \"msg_contents\": \"> On Mar 4, 2020, at 7:43 PM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n> \\n> Hackers,\\n> \\n> as mentioned in [1], I have created an implementation of command counter statistics very similar in purpose to the one already pending in the commitfest going by the name \\\"pg_stat_sql\\\". I don't really care if this implementation is seen as building on that one or as separate, but I was worried about hijacking that thread, so I'm starting this thead. There are elements of this patch that borrowed from that one, so if this is accepted, authorship should reflect that.\\n> \\n> See the work by Haribabu Kommi (haribabu) at https://commitfest.postgresql.org/27/790/\\n> \\n> The two main differences are that\\n> \\n> (1) This implementation is based on commandtags as enums, not strings and\\n> (2) This implementation uses techniques to reduce lock contention\\n> \\n> I think (2) is the more important part.\\n> \\n> [1] https://www.postgresql.org/message-id/28F46159-3764-421B-904B-004DEA339310%40enterprisedb.com\\n> \\n> <v1-0001-Implementing-the-cmdstats-subsystem.patch>\\n> —\\n\\nRebased patch against master caa3c4242cf86322e2ed0c86199e6462a2c41565.\\n\\n\\n\\n\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Tue, 2 Jun 2020 09:58:25 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Command statistics system (cmdstats)\"\n },\n {\n \"msg_contents\": \"> On Jun 2, 2020, at 9:58 AM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n> \\n>> \\n>> On Mar 4, 2020, at 7:43 PM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n>> \\n>> Hackers,\\n>> \\n>> as mentioned in [1], I have created an implementation of command counter statistics very similar in purpose to the one already pending in the commitfest going by the name \\\"pg_stat_sql\\\". I don't really care if this implementation is seen as building on that one or as separate, but I was worried about hijacking that thread, so I'm starting this thead. There are elements of this patch that borrowed from that one, so if this is accepted, authorship should reflect that.\\n>> \\n>> See the work by Haribabu Kommi (haribabu) at https://commitfest.postgresql.org/27/790/\\n>> \\n>> The two main differences are that\\n>> \\n>> (1) This implementation is based on commandtags as enums, not strings and\\n>> (2) This implementation uses techniques to reduce lock contention\\n>> \\n>> I think (2) is the more important part.\\n>> \\n>> [1] https://www.postgresql.org/message-id/28F46159-3764-421B-904B-004DEA339310%40enterprisedb.com\\n>> \\n>> <v1-0001-Implementing-the-cmdstats-subsystem.patch>\\n>> —\\n> \\n> Rebased patch against master caa3c4242cf86322e2ed0c86199e6462a2c41565.\\n\\n\\n\\n\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Tue, 2 Jun 2020 10:02:24 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Command statistics system (cmdstats)\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 4, 2020 at 10:43 PM Mark Dilger\\n<mark.dilger@enterprisedb.com> wrote:\\n> The two main differences are that\\n>\\n> (1) This implementation is based on commandtags as enums, not strings and\\n> (2) This implementation uses techniques to reduce lock contention\\n>\\n> I think (2) is the more important part.\\n\\nMy spidey sense is tingling here, telling me that we need some actual\\nbenchmarking. Like, suppose we test the two patches under normal cases\\nand under cases that are constructed to be as bad as possible for each\\nof them. Or suppose we test this patch with the lock mitigation\\nstrategies and then remove the mitigations for some inexpensive\\ncommand (e.g. SHOW) and then use pgbench to spam that command. Like\\nyou, I suspect that the locking mitigations are important in some\\nworkloads, but it would be good to have some figures to back that out,\\nas well as to find out whether there's still too much overhead.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Jun 2020 13:45:02 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Command statistics system (cmdstats)\"\n },\n {\n \"msg_contents\": \"On Wed, Jun 10, 2020 at 01:45:02PM -0400, Robert Haas wrote:\\n> My spidey sense is tingling here, telling me that we need some actual\\n> benchmarking. Like, suppose we test the two patches under normal cases\\n> and under cases that are constructed to be as bad as possible for each\\n> of them. Or suppose we test this patch with the lock mitigation\\n> strategies and then remove the mitigations for some inexpensive\\n> command (e.g. SHOW) and then use pgbench to spam that command. Like\\n> you, I suspect that the locking mitigations are important in some\\n> workloads, but it would be good to have some figures to back that out,\\n> as well as to find out whether there's still too much overhead.\\n\\nThis patch has not received any replies after this comment for three\\nmonths, so I am marking it as returned with feedback. I agree that\\nthis should be benchmarked carefully.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 17 Sep 2020 13:51:23 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Command statistics system (cmdstats)\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 03, 2020 at 11:50:04PM -0300, Alvaro Herrera wrote:\\n> So, who is updating this patch for the new cmdtaglist.h stuff?\\n\\nThis patch had no activity for months, so I am marking it as returned\\nwith feedback.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 17 Sep 2020 14:01:19 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New SQL counter statistics view (pg_stat_sql)\"\n },\n {\n \"msg_contents\": \"On 2020-Sep-17, Michael Paquier wrote:\\n\\n> On Wed, Jun 10, 2020 at 01:45:02PM -0400, Robert Haas wrote:\\n> > My spidey sense is tingling here, telling me that we need some actual\\n> > benchmarking.\\n> \\n> This patch has not received any replies after this comment for three\\n> months, so I am marking it as returned with feedback. I agree that\\n> this should be benchmarked carefully.\\n\\nIt seems fine to mark the patch as RwF at the end of commitfest, but\\nthat's two weeks away. I don't understand what is accomplished by doing\\nit ahead of time, other than alienating the patch authors.\\n\\n-- \\n�lvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Sep 2020 09:40:20 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Command statistics system (cmdstats)\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Sep 17, 2020, at 5:40 AM, Alvaro Herrera <alvherre@2ndquadrant.com> wrote:\\n> \\n> On 2020-Sep-17, Michael Paquier wrote:\\n> \\n>> On Wed, Jun 10, 2020 at 01:45:02PM -0400, Robert Haas wrote:\\n>>> My spidey sense is tingling here, telling me that we need some actual\\n>>> benchmarking.\\n>> \\n>> This patch has not received any replies after this comment for three\\n>> months, so I am marking it as returned with feedback. I agree that\\n>> this should be benchmarked carefully.\\n\\nYes, I have prioritized a couple other patches over this one, with benchmarking this patch lower down my priority list. Thank you, Michael, for the reminder!\\n\\n> \\n> It seems fine to mark the patch as RwF at the end of commitfest, but\\n> that's two weeks away. I don't understand what is accomplished by doing\\n> it ahead of time, other than alienating the patch authors.\\n\\nThanks, Álvaro! I do not feel alienated, and I hope Haribabu Kommi does not either.\\n\\nHaribabu, do you have an opinion on which of the two patches should proceed? I don't recall seeing a response from you about whether you liked what I did with your patch. Perhaps you were waiting on the benchmarking results?\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Sep 2020 08:03:55 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Command statistics system (cmdstats)\"\n },\n {\n \"msg_contents\": \"On Thu, Sep 17, 2020 at 08:03:55AM -0700, Mark Dilger wrote:\\n> Yes, I have prioritized a couple other patches over this one, with\\n> benchmarking this patch lower down my priority list. Thank you,\\n> Michael, for the reminder!\\n\\nBased on the low level of activity and the fact that the patch was\\nmarked as waiting on author for a couple of weeks, it looks like\\nlittle could be achieved by the end of the CF, and the attention was\\nelsewhere, so it looked better (and it still does IMO) to give more\\nattention to the remaining 170~ patches that are still lying on the CF\\napp.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 18 Sep 2020 14:55:43 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Command statistics system (cmdstats)\"\n },\n {\n \"msg_contents\": \"On 2020-Sep-18, Michael Paquier wrote:\\n\\n> Based on the low level of activity and the fact that the patch was\\n> marked as waiting on author for a couple of weeks, it looks like\\n> little could be achieved by the end of the CF, and the attention was\\n> elsewhere, so it looked better (and it still does IMO) to give more\\n> attention to the remaining 170~ patches that are still lying on the CF\\n> app.\\n\\n\\\"A couple of weeks\\\" of inactivity is not sufficient, in my view, to boot\\na patch out of the commitfest process. Whenever the patch is\\nresurrected, it will be a new entry which won't have the history that it\\nhad accumulated in the long time since it was created -- which biases\\nit against other new patches.\\n\\nI'm not really arguing about this one patch only, but more generally\\nabout the process you're following. The problem is that you as CFM are\\nimposing your personal priorities on the whole process by punting\\npatches ahead of time -- you are saying that nobody else should be\\nlooking at updating this patch because it is now closed. It seems fair\\nto do so at the end of the commitfest, but it does not seem fair to do\\nit when it's still mid-cycle.\\n\\nPutting also in perspective the history that the patch had prior to your\\nunfairly closing it, there was a lot of effort in getting it reviewed to\\na good state and updated by the author -- yet a single unsubstantiated\\ncomment, without itself a lot of effort on the reviewer's part, that\\n\\\"maybe it has a perf drawback\\\" is sufficient to get it booted? It\\ndoesn't seem appropriate.\\n\\n-- \\n�lvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Mon, 21 Sep 2020 10:41:46 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Command statistics system (cmdstats)\"\n },\n {\n \"msg_contents\": \"On Mon, Sep 21, 2020 at 9:41 AM Alvaro Herrera <alvherre@2ndquadrant.com> wrote:\\n> \\\"A couple of weeks\\\" of inactivity is not sufficient, in my view, to boot\\n> a patch out of the commitfest process. Whenever the patch is\\n> resurrected, it will be a new entry which won't have the history that it\\n> had accumulated in the long time since it was created -- which biases\\n> it against other new patches.\\n\\nAs Michael said, it had been inactive for three *months*. I don't\\nthink there's a big problem here. I think that the redesign of the\\nCommitFest application encourages carrying things along from CF to CF\\nforever, instead of getting rid of things that aren't wanted or aren't\\nmaking any progress. That's work we don't need. There ought to be a\\nway to mark a patch RwF when nothing's happen that lets it be revived\\nlater if and when someone gets around to resubmitting, but I think\\nthat's not possible now.\\n\\nThen, too, since we have email integration now, maybe we ought to do\\nthat automatically if the thread isn't getting updated and the patch\\nis setting there waiting on author. It's a real waste of CFM time to\\nchase down and kick out obviously-inactive patches, and if the CFM is\\ngoing to get flack for it then that's even worse. Like, do we have 170\\npatches now because we have more activity than a few years ago, or\\njust because we've become more reluctant to boot things? I don't know,\\nbut to the extent it's from unwillingness to boot things, I think\\nthat's bad. Discouraging contributors is not good, but wasting CFM and\\ncommiter time is also bad. You've got to draw a line someplace or\\nyou'll go nuts.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Mon, 21 Sep 2020 17:19:07 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Command statistics system (cmdstats)\"\n },\n {\n \"msg_contents\": \"On Mon, Sep 21, 2020 at 10:41:46AM -0300, Alvaro Herrera wrote:\\n> \\\"A couple of weeks\\\" of inactivity is not sufficient, in my view, to boot\\n> a patch out of the commitfest process. Whenever the patch is\\n> resurrected, it will be a new entry which won't have the history that it\\n> had accumulated in the long time since it was created -- which biases\\n> it against other new patches.\\n\\nAny of the patches that have been marked as RwF last week have been\\nwaiting on author since at least the end of the commit fest of July,\\nwith zero actions taken on them. In my experience, this points\\nstrongly to the fact that these are unlikely going to be worked on\\nactively in a short time frame.\\n\\n> I'm not really arguing about this one patch only, but more generally\\n> about the process you're following. The problem is that you as CFM are\\n> imposing your personal priorities on the whole process by punting\\n> patches ahead of time -- you are saying that nobody else should be\\n> looking at updating this patch because it is now closed. It seems fair\\n> to do so at the end of the commitfest, but it does not seem fair to do\\n> it when it's still mid-cycle.\\n\\nI think that this remark is a bit unfair. When it comes to any patch\\nin the commit fest, and I have managed a couple of them over the\\nyears, I have tried to keep a neutral view for all of them, meaning\\nthat I only let the numbers speak by themselves, mostly:\\n- When has a patch lastly been updated.\\n- What's the current state in the CF app If the patch had no reviews\\nand still in \\\"Needs review\\\", simply switch it to \\\"waiting on author\\\".\\nIf the patch has been waiting on author for more than two weeks, it\\nwould get marked as RwF at the end of the CF. There are also cases\\nwhere the status of the patch is incorrect, mostly that a patch\\nwaiting on author because of a review was not updated as such in the\\nCF app. In this case, and in the middle of a CF, this would get get\\nmarked as Rwf, but the author would get a reminder of that.\\n\\nAny of the patches that have been discarded by me last week were\\nwaiting in the stack for much, much, longer than that, and I think\\nthat it does not help reviewers and committers in keeping this stuff\\nlonger than necessary because this mainly bloats the CF app in a\\nnon-necessary way, hiding patches that should deserve more attention.\\nSome of those patches have been left idle for half a year.\\n\\nAnother point that I'd like to make here is that my intention was to\\nlift the curve here (a term quite fashionable this year particularly,\\nisn't it?), due to the high number of patches to handle in a single\\nCF, and the shortage of hands actually doing this kind of work.\\n\\nThen, let's be clear about one last thing. If at *any* moment people\\nhere feel that at least the state of *one* patch of this CF has been\\nchanged based on some of my \\\"personal\\\" priority views, I will\\nimmediately stop any action as CFM and resign from it. So, if that's\\nthe case, please feel free to speak up here or on any of the threads\\nrelated to those patches.\\n\\n> Putting also in perspective the history that the patch had prior to your\\n> unfairly closing it, there was a lot of effort in getting it reviewed to\\n> a good state and updated by the author -- yet a single unsubstantiated\\n> comment, without itself a lot of effort on the reviewer's part, that\\n> \\\"maybe it has a perf drawback\\\" is sufficient to get it booted? It\\n> doesn't seem appropriate.\\n\\nI agree on the fact that not being able to recreate a new CF entry\\nthat keeps the history of an old one, with all the past threads or\\nannotations, could be helpful in some cases. Now, I have seen as well \\ncases where a patch got around for so long with a very long thread\\nmade things confusing, and that people also like starting a new thread\\nwith a summary of the past state as a first message to attempt\\nclarifying things. So both have advantages.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 23 Sep 2020 10:41:10 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Command statistics system (cmdstats)\"\n },\n {\n \"msg_contents\": \"On Wed, Sep 23, 2020 at 10:41:10AM +0900, Michael Paquier wrote:\\n> I think that this remark is a bit unfair. When it comes to any patch\\n> in the commit fest, and I have managed a couple of them over the\\n> years, I have tried to keep a neutral view for all of them, meaning\\n> that I only let the numbers speak by themselves, mostly:\\n> - When has a patch lastly been updated.\\n> - What's the current state in the CF app If the patch had no reviews\\n> and still in \\\"Needs review\\\", simply switch it to \\\"waiting on author\\\".\\n\\nSorry, forgot to mention here that this is in the case where a patch\\ndoes not apply anymore or that the CF bot complains, requiring a check\\nof the patch from the author.\\n\\n> If the patch has been waiting on author for more than two weeks, it\\n> would get marked as RwF at the end of the CF. There are also cases\\n> where the status of the patch is incorrect, mostly that a patch\\n> waiting on author because of a review was not updated as such in the\\n> CF app. In this case, and in the middle of a CF, this would get get\\n> marked as Rwf, but the author would get a reminder of that.\\n\\nLast sentence here is incorrect as well: s/get get/not get/, and the\\nauthor would be reminded to update his/her patch.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 23 Sep 2020 11:12:14 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Command statistics system (cmdstats)\"\n },\n {\n \"msg_contents\": \"> >> On Mar 4, 2020, at 7:43 PM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n\\n> >> as mentioned in [1], I have created an implementation of command\\n> >> counter statistics very similar in purpose to the one already\\n> >> pending in the commitfest going by the name \\\"pg_stat_sql\\\". I don't\\n> >> really care if this implementation is seen as building on that one\\n> >> or as separate, but I was worried about hijacking that thread, so\\n> >> I'm starting this thead. There are elements of this patch that\\n> >> borrowed from that one, so if this is accepted, authorship should\\n> >> reflect that.\\n\\nHello, was this abandoned or is there another thread somewhere? The\\npatch in the email I'm replying to doesn't apply, but the conflicts\\ndon't look very bad.\\n\\n-- \\n�lvaro Herrera Valdivia, Chile\\n\\n\\n\",\n \"msg_date\": \"Fri, 28 May 2021 09:42:24 -0400\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Command statistics system (cmdstats)\"\n },\n {\n \"msg_contents\": \"\\n\\n> On May 28, 2021, at 6:42 AM, Alvaro Herrera <alvherre@2ndquadrant.com> wrote:\\n> \\n>>>> On Mar 4, 2020, at 7:43 PM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n> \\n>>>> as mentioned in [1], I have created an implementation of command\\n>>>> counter statistics very similar in purpose to the one already\\n>>>> pending in the commitfest going by the name \\\"pg_stat_sql\\\". I don't\\n>>>> really care if this implementation is seen as building on that one\\n>>>> or as separate, but I was worried about hijacking that thread, so\\n>>>> I'm starting this thead. There are elements of this patch that\\n>>>> borrowed from that one, so if this is accepted, authorship should\\n>>>> reflect that.\\n> \\n> Hello, was this abandoned or is there another thread somewhere? The\\n> patch in the email I'm replying to doesn't apply, but the conflicts\\n> don't look very bad.\\n\\nI abandoned it for v14, choosing instead to work on amcheck features. Would you like me to take this up again for v15?\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 28 May 2021 08:30:49 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Command statistics system (cmdstats)\"\n }\n]","isConversational":false}}},{"rowIdx":42995,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi all\n\nFollowing on from\n\nbigint txids vs 'xid' type, new txid_recent(bigint) => xid\n\nhttps://www.postgresql.org/message-id/CAMsr+YFDZMN_iZ7KrRoe+j0KVLQvFVgvZxbcVxR-MLjgtoZugA@mail.gmail.com\n\n\nhere's a patch that implements a txid_status(bigint) function to report the\ncommit status of a function.\n\nIf an application is disconnected while a COMMIT request is in flight,\nthe backend crashes mid-commit, etc, then an application may not be\nsure whether or not a commit completed successfully or was rolled\nback. While two-phase commit solves this it does so at a considerable\noverhead, so introduce a lighter alternative.\n\ntxid_status(bigint) lets an application determine the status of a\na commit based on an xid-with-epoch as returned by txid_current()\nor similar. Status may be committed, aborted, in-progress (including\nprepared xacts) or null if the xact is too old for its commit status\nto still be retained because it has passed the wrap-around epoch\nboundary.\n\nApplications must call txid_current() in their transactions to make\nmuch use of this since PostgreSQL does not automatically report an xid\nto the client when one is assigned.\n\nA future protocol enhancement to report txid assignment would be very\nuseful, but quite separate to this.\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","msg_date":"Sat, 20 Aug 2016 21:24:47 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"[PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 20 August 2016 at 21:24, Craig Ringer <craig@2ndquadrant.com> wrote:\n\n> Hi all\n>\n> Following on from\n>\n> bigint txids vs 'xid' type, new txid_recent(bigint) => xid\n>\n>\nAhem. Forgot to squash in a fixup commit. Updated patch of\ntxid_status(bigint) attachd.\n\nA related patch follows, adding a new txid_current_ifassigned(bigint)\nfunction as suggested by Jim Nasby. It's usefully connected to\ntxid_status() and might as well be added at the same time.\n\nSince it builds on the same history I've also attached an updated version\nof txid_recent(bigint) now called txid_convert_ifrecent(bigint), per the\nabove-linked thread.\n\nFinally, and not intended for commit, is a useful test function I wrote to\ncause extremely rapid xid wraparound, bundled up into a src/test/regress\ntest case. txid_incinerate() can jump the server about UINT32/2 xids in ~2\nseconds if fsync is off, making it handy for testing. Posting so others\ncan use it for their own test needs later and because it's useful for\ntesting these patches that touch on the xid epoch.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\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","msg_date":"Sat, 20 Aug 2016 21:46:55 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Sat, Aug 20, 2016 at 9:46 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> Ahem. Forgot to squash in a fixup commit. Updated patch of\n> txid_status(bigint) attachd.\n>\n> A related patch follows, adding a new txid_current_ifassigned(bigint)\n> function as suggested by Jim Nasby. It's usefully connected to txid_status()\n> and might as well be added at the same time.\n>\n> Since it builds on the same history I've also attached an updated version of\n> txid_recent(bigint) now called txid_convert_ifrecent(bigint), per the\n> above-linked thread.\n>\n> Finally, and not intended for commit, is a useful test function I wrote to\n> cause extremely rapid xid wraparound, bundled up into a src/test/regress\n> test case. txid_incinerate() can jump the server about UINT32/2 xids in ~2\n> seconds if fsync is off, making it handy for testing. Posting so others can\n> use it for their own test needs later and because it's useful for testing\n> these patches that touch on the xid epoch.\n\nI think you should use underscores to separate all of the words\ninstead of only some of them.\n\nAlso, note that the corresponding internal function is\nGetTopTransactionIdIfAny(), which might suggest txid_current_if_any()\nrather than txid_current_if_assigned(), but you could argue that your\nnaming is better...\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, 22 Aug 2016 13:03:25 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 23 August 2016 at 01:03, Robert Haas <robertmhaas@gmail.com> wrote:\n\n\n> I think you should use underscores to separate all of the words\n> instead of only some of them.\n>\n\nRight. Will fix.\n\nThanks for taking a look.\n\n\n> Also, note that the corresponding internal function is\n> GetTopTransactionIdIfAny(), which might suggest txid_current_if_any()\n> rather than txid_current_if_assigned(), but you could argue that your\n> naming is better.\n\n\nYeah, I do argue that in this case. Not a hugely strong opinion, but we\nrefer to txid_current() in the docs as:\n\n\"get current transaction ID, assigning a new one if the current transaction\ndoes not have one\"\n\nso I thought it'd be worth being consistent with that. It's not like\ntxid_current() mirrors the name of GetTopTransactionId() after all ;) and I\nthink it's more important to be consistent with what the user sees than\nwith the code.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\nOn 23 August 2016 at 01:03, Robert Haas <robertmhaas@gmail.com> wrote: I think you should use underscores to separate all of the words\ninstead of only some of them.Right. Will fix.Thanks for taking a look. Also, note that the corresponding internal function is\nGetTopTransactionIdIfAny(), which might suggest txid_current_if_any()\nrather than txid_current_if_assigned(), but you could argue that your\nnaming is better.Yeah, I do argue that in this case. Not a hugely strong opinion, but we refer to txid_current() in the docs as:\"get current transaction ID, assigning a new one if the current transaction does not have one\"so I thought it'd be worth being consistent with that. It's not like txid_current() mirrors the name of GetTopTransactionId() after all ;) and I think it's more important to be consistent with what the user sees than with the code.-- Craig Ringer http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services","msg_date":"Tue, 23 Aug 2016 08:31:26 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 23 August 2016 at 01:03, Robert Haas <robertmhaas@gmail.com> wrote:\n\n\n>\n> I think you should use underscores to separate all of the words\n> instead of only some of them.\n>\n>\nifassigned => if_assigned\n\nifrecent=> if_recent\n\nUpdated patch series attached. As before, 0-4 intended for commit, 5 just\nbecause it'll be handy to have around for people doing wraparound related\ntesting.\n\nAgain, thanks for taking a look.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\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","msg_date":"Tue, 23 Aug 2016 08:55:00 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 23/08/16 02:55, Craig Ringer wrote:\n> On 23 August 2016 at 01:03, Robert Haas <robertmhaas@gmail.com\n> <mailto:robertmhaas@gmail.com>> wrote:\n>\n>\n>\n> I think you should use underscores to separate all of the words\n> instead of only some of them.\n>\n>\n> ifassigned => if_assigned\n>\n> ifrecent=> if_recent\n>\n> Updated patch series attached. As before, 0-4 intended for commit, 5\n> just because it'll be handy to have around for people doing wraparound\n> related testing.\n\nI guess you mean 0-3 for commit and 4 is just handy?\n\n From the point of code this patch seems good to me.\n\nI do wonder about the 3rd patch though. I wonder if it would not be \nbetter to have the opposite function instead - converting xid to txid as \nthat will always work and does not have to have the NULL case and would \nbe simpler in terms of code.\n\n-- \n Petr Jelinek 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, 23 Aug 2016 10:02:16 +0200","msg_from":"Petr Jelinek <petr@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 23 Aug 2016 16:02, \"Petr Jelinek\" <petr@2ndquadrant.com> wrote:\n>\n> On 23/08/16 02:55, Craig Ringer wrote:\n>>\n>> On 23 August 2016 at 01:03, Robert Haas <robertmhaas@gmail.com\n>> <mailto:robertmhaas@gmail.com>> wrote:\n>>\n>>\n>>\n>> I think you should use underscores to separate all of the words\n>> instead of only some of them.\n>>\n>>\n>> ifassigned => if_assigned\n>>\n>> ifrecent=> if_recent\n>>\n>> Updated patch series attached. As before, 0-4 intended for commit, 5\n>> just because it'll be handy to have around for people doing wraparound\n>> related testing.\n>\n>\n> I guess you mean 0-3 for commit and 4 is just handy?\n\nEr. Right. 1-3. 4 just as handy test/tool.\n\n1 most important and useful. Then 2. Then 3.\n\n> From the point of code this patch seems good to me.\n\nThanks.\n\n> I do wonder about the 3rd patch though. I wonder if it would not be\nbetter to have the opposite function instead - converting xid to txid as\nthat will always work and does not have to have the NULL case and would be\nsimpler in terms of code.\n\nYeah, but it wouldn't solve the need to take txid_current() output and do\nstuff with it other than ordinal comparison. Like pass to commit ts\nfunctions and others that take xid. If we extend all funcs that take xid to\ntake bigint instead, they just get to use the same epoch logic in them,\ncomplete with some way to deal with wrapped xids sensibly. It has to be\ndone somewhere. Though it's prettier if hidden from the user.\n\nMore importantly imo, txid => bigint has to assume the current epoch. We\nhave no way to make sure the user doesn't try to use something already\nwrapped.\n\nI don't mind if everyone decides it's better to make xid go away and use\nbigint everywhere user facing. Or even a new bigxid type. More work than I\ncan really afford but can manage; shouldn't block #1 and #2 though as they\nalready use bigint.\n\nOn 23 Aug 2016 16:02, \"Petr Jelinek\" <petr@2ndquadrant.com> wrote:\n>\n> On 23/08/16 02:55, Craig Ringer wrote:\n>>\n>> On 23 August 2016 at 01:03, Robert Haas <robertmhaas@gmail.com\n>> <mailto:robertmhaas@gmail.com>> wrote:\n>>\n>>\n>>\n>> I think you should use underscores to separate all of the words\n>> instead of only some of them.\n>>\n>>\n>> ifassigned => if_assigned\n>>\n>> ifrecent=> if_recent\n>>\n>> Updated patch series attached. As before, 0-4 intended for commit, 5\n>> just because it'll be handy to have around for people doing wraparound\n>> related testing.\n>\n>\n> I guess you mean 0-3 for commit and 4 is just handy?\nEr. Right. 1-3. 4 just as handy test/tool.\n1 most important and useful. Then 2. Then 3.\n> From the point of code this patch seems good to me.\nThanks.\n> I do wonder about the 3rd patch though. I wonder if it would not be better to have the opposite function instead - converting xid to txid as that will always work and does not have to have the NULL case and would be simpler in terms of code.\nYeah, but it wouldn't solve the need to take txid_current() output and do stuff with it other than ordinal comparison. Like pass to commit ts functions and others that take xid. If we extend all funcs that take xid to take bigint instead, they just get to use the same epoch logic in them, complete with some way to deal with wrapped xids sensibly. It has to be done somewhere. Though it's prettier if hidden from the user.\nMore importantly imo, txid => bigint has to assume the current epoch. We have no way to make sure the user doesn't try to use something already wrapped.\nI don't mind if everyone decides it's better to make xid go away and use bigint everywhere user facing. Or even a new bigxid type. More work than I can really afford but can manage; shouldn't block #1 and #2 though as they already use bigint.","msg_date":"Tue, 23 Aug 2016 17:27:02 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 23/08/16 11:27, Craig Ringer wrote:\n> On 23 Aug 2016 16:02, \"Petr Jelinek\" <petr@2ndquadrant.com\n> <mailto:petr@2ndquadrant.com>> wrote:\n>>\n>> On 23/08/16 02:55, Craig Ringer wrote:\n>>>\n>>> On 23 August 2016 at 01:03, Robert Haas <robertmhaas@gmail.com\n> <mailto:robertmhaas@gmail.com>\n>>> <mailto:robertmhaas@gmail.com <mailto:robertmhaas@gmail.com>>> wrote:\n>>>\n>>>\n>>>\n>>> I think you should use underscores to separate all of the words\n>>> instead of only some of them.\n>>>\n>>>\n>>> ifassigned => if_assigned\n>>>\n>>> ifrecent=> if_recent\n>>>\n>>> Updated patch series attached. As before, 0-4 intended for commit, 5\n>>> just because it'll be handy to have around for people doing wraparound\n>>> related testing.\n>>\n>>\n>> I guess you mean 0-3 for commit and 4 is just handy?\n>\n> Er. Right. 1-3. 4 just as handy test/tool.\n>\n> 1 most important and useful. Then 2. Then 3.\n>\n>> From the point of code this patch seems good to me.\n>\n> Thanks.\n>\n>> I do wonder about the 3rd patch though. I wonder if it would not be\n> better to have the opposite function instead - converting xid to txid as\n> that will always work and does not have to have the NULL case and would\n> be simpler in terms of code.\n>\n> Yeah, but it wouldn't solve the need to take txid_current() output and\n> do stuff with it other than ordinal comparison. Like pass to commit ts\n> functions and others that take xid. If we extend all funcs that take xid\n> to take bigint instead, they just get to use the same epoch logic in\n> them, complete with some way to deal with wrapped xids sensibly. It has\n> to be done somewhere. Though it's prettier if hidden from the user.\n>\n> More importantly imo, txid => bigint has to assume the current epoch. We\n> have no way to make sure the user doesn't try to use something already\n> wrapped.\n>\n\nOkay, fair points.\n\n> I don't mind if everyone decides it's better to make xid go away and use\n> bigint everywhere user facing. Or even a new bigxid type. More work than\n> I can really afford but can manage; shouldn't block #1 and #2 though as\n> they already use bigint.\n>\n\nI don't think that would be very straightforward to be honest. I guess \nfor what you want to achieve the approach chosen is the best one then.\n\n-- \n Petr Jelinek 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, 23 Aug 2016 12:04:21 +0200","msg_from":"Petr Jelinek <petr@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Mon, Aug 22, 2016 at 8:55 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> Updated patch series attached. As before, 0-4 intended for commit, 5 just\n> because it'll be handy to have around for people doing wraparound related\n> testing.\n>\n> Again, thanks for taking a look.\n\n/me reviews a bit more deeply.\n\nIn 0001, it seems to me that \"in-progress\" should be \"in progress\". I\ndon't think it's normal to hyphenate that. We have admittedly\nsometimes done so, but:\n\n[rhaas pgsql]$ git grep 'in-progress' | wc -l\n 63\n[rhaas pgsql]$ git grep 'in progress' | wc -l\n 346\n\nIt may make sense to speak of an \"in-progress transaction\" but I would\nsay \"the transaction is in progress\" not \"the transaction is\nin-progress\", which seems to me to argue for a space as the proper\nseparator here.\n\nAlso:\n\n+CREATE TYPE txid_status AS ENUM ('committed', 'in-progress', 'aborted');\n+\n+CREATE FUNCTION\n+ txid_status(txid bigint)\n+RETURNS txid_status\n+LANGUAGE sql\n+VOLATILE PARALLEL SAFE\n+AS $$\n+SELECT CASE\n+ WHEN s IS NULL THEN NULL::txid_status\n+ WHEN s = -1 THEN 'aborted'::txid_status\n+ WHEN s = 0 THEN 'in-progress'::txid_status\n+ WHEN s = 1 THEN 'committed'::txid_status\n+END\n+FROM pg_catalog.txid_status_internal($1) s;\n+$$;\n+\n+COMMENT ON FUNCTION txid_status(bigint)\n+IS 'get commit status of given recent xid or null if too old';\n\nI'm not really that keen on this approach. I don't think we need to\nintroduce a new data type for this, and I would rather not use SQL,\neither. It would be faster and simpler to just return the appropriate\nstring from a C function defined directly.\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 Aug 2016 10:18:47 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Tue, Aug 23, 2016 at 10:18 AM, Robert Haas <robertmhaas@gmail.com> wrote:\n> On Mon, Aug 22, 2016 at 8:55 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n>> Updated patch series attached. As before, 0-4 intended for commit, 5 just\n>> because it'll be handy to have around for people doing wraparound related\n>> testing.\n>>\n>> Again, thanks for taking a look.\n>\n> /me reviews a bit more deeply.\n>\n> In 0001, ...\n\n0002 looks good, so I committed it. You forgot a function prototype\nfor the new SQL-callable function, though, so I added that. For me,\nit generates a compiler warning if that's missing; you might want to\ntry to achieve a similar setup.\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 Aug 2016 10:34:17 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 23 August 2016 at 22:18, Robert Haas <robertmhaas@gmail.com> wrote:\n\n> On Mon, Aug 22, 2016 at 8:55 PM, Craig Ringer <craig@2ndquadrant.com>\n> wrote:\n> > Updated patch series attached. As before, 0-4 intended for commit, 5 just\n> > because it'll be handy to have around for people doing wraparound related\n> > testing.\n> >\n> > Again, thanks for taking a look.\n>\n> /me reviews a bit more deeply.\n>\n> In 0001, it seems to me that \"in-progress\" should be \"in progress\".\n>\n\nFine by me. I was on the fence about it anyway.\n\n+CREATE TYPE txid_status AS ENUM ('committed', 'in-progress', 'aborted');\n>\n\n\n> I'm not really that keen on this approach. I don't think we need to\n> introduce a new data type for this, and I would rather not use SQL,\n> either. It would be faster and simpler to just return the appropriate\n> string from a C function defined directly.\n>\n\nAlso fine by me. You're right, keep it simple. It means the potential set\nof values isn't discoverable the same way, but ... meh. Using it usefully\nmeans reading the docs anyway.\n\nThe remaining 2 patches of interest are attached - txid_status() and\ntxid_convert_if_recent(). Thanks for committing txid_current_if_assigned().\n\nNow I'd best stop pretending I'm in a sensible timezone.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, 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":"Wed, 24 Aug 2016 00:59:35 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Tue, Aug 23, 2016 at 12:59 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> Also fine by me. You're right, keep it simple. It means the potential set of\n> values isn't discoverable the same way, but ... meh. Using it usefully means\n> reading the docs anyway.\n>\n> The remaining 2 patches of interest are attached - txid_status() and\n> txid_convert_if_recent(). Thanks for committing txid_current_if_assigned().\n>\n> Now I'd best stop pretending I'm in a sensible timezone.\n\nI reviewed this version some more and found some more problems.\n\n+ uint32 xid_epoch = (uint32)(xid_with_epoch >>32);\n+ TransactionId xid = (TransactionId)(xid_with_epoch);\n\nI think this is not project style. In particular, I think that the\nfirst one needs a space after the cast and another space before the\n32; and I think the second one has an unnecessary set of parentheses\nand needs a space added.\n\n+/*\n+ * Underlying implementation of txid_status, which is mapped to an enum in\n+ * system_views.sql.\n+ */\n\nNot any more...\n\n+ errmsg(\"transaction ID \"UINT64_FORMAT\" is an invalid xid\",\n+ xid_with_epoch)));\n\nSpacing.\n\n+ if (TransactionIdIsCurrentTransactionId(xid) ||\nTransactionIdIsInProgress(xid))\n+ status = gettext_noop(\"in progress\");\n+ else if (TransactionIdDidCommit(xid))\n+ status = gettext_noop(\"committed\");\n+ else if (TransactionIdDidAbort(xid))\n+ status = gettext_noop(\"aborted\");\n+ else\n+ /* shouldn't happen */\n+ ereport(ERROR,\n+ (errmsg_internal(\"unable to determine commit status of\nxid \"UINT64_FORMAT, xid)));\n\nMaybe I'm all wet here, but it seems like there might be a problem\nhere if the XID is older than the CLOG truncation point but less than\none epoch old. get_xid_in_recent_past only guarantees that the\ntransaction is less than one epoch old, not that we still have CLOG\ndata for it. And there's nothing to keep NextXID from advancing under\nus, so if somebody asks about a transaction that's just under 2^32\ntransactions old, then get_xid_in_recent_past could say it's valid,\nthen NextXID advances and we look up the XID extracted from the txid\nand get the status of the new transaction instead of the old one!\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 Aug 2016 15:10:37 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 24 August 2016 at 03:10, Robert Haas <robertmhaas@gmail.com> wrote:\n>\n> On Tue, Aug 23, 2016 at 12:59 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> > Also fine by me. You're right, keep it simple. It means the potential set of\n> > values isn't discoverable the same way, but ... meh. Using it usefully means\n> > reading the docs anyway.\n> >\n> > The remaining 2 patches of interest are attached - txid_status() and\n> > txid_convert_if_recent(). Thanks for committing txid_current_if_assigned().\n> >\n> > Now I'd best stop pretending I'm in a sensible timezone.\n>\n> I reviewed this version some more and found some more problems.\n\n\nThanks. It took me a few days to prep a new patch as I found another\nissue in the process. Updated patch attached.\n\nThe updated series starts (0001) with a change to slru.c to release\nthe control lock when throwing an exception so that we don't deadlock\nwith ourselves when re-entering slru.c; explanation below.\n\nThen there's the txid_status (0002) patch with fixes, then\ntxid_convert_if_recent(0003).\n\nI omitted txid_incinerate() ; I have an updated version that sets xact\nstatus to aborted for burned xacts instead of leaving them at 0\n(in-progress), but haven't had time to finish it so it doesn't try to\nblindly set the status of xacts on pages where it didn't hold\nXidGenLock when the page was added to the clog.\n\n> + uint32 xid_epoch = (uint32)(xid_with_epoch >>32);\n> + TransactionId xid = (TransactionId)(xid_with_epoch);\n>\n> I think this is not project style. In particular, I think that the\n> first one needs a space after the cast and another space before the\n> 32; and I think the second one has an unnecessary set of parentheses\n> and needs a space added.\n\n\nOK, no problems. I didn't realise spacing around casts was specified.\n\n>\n> +/*\n> + * Underlying implementation of txid_status, which is mapped to an enum in\n> + * system_views.sql.\n> + */\n>\n> Not any more...\n\n\nThat's why I shouldn't revise a patch at 1am ;)\n\n>\n> + if (TransactionIdIsCurrentTransactionId(xid) ||\n> TransactionIdIsInProgress(xid))\n> + status = gettext_noop(\"in progress\");\n> + else if (TransactionIdDidCommit(xid))\n> + status = gettext_noop(\"committed\");\n> + else if (TransactionIdDidAbort(xid))\n> + status = gettext_noop(\"aborted\");\n> + else\n> + /* shouldn't happen */\n> + ereport(ERROR,\n> + (errmsg_internal(\"unable to determine commit status of\n> xid \"UINT64_FORMAT, xid)));\n>\n> Maybe I'm all wet here, but it seems like there might be a problem\n> here if the XID is older than the CLOG truncation point but less than\n> one epoch old. get_xid_in_recent_past only guarantees that the\n> transaction is less than one epoch old, not that we still have CLOG\n> data for it.\n\n\nGood point. The call would then fail with something like\n\n ERROR: could not access status of transaction 778793573\n DETAIL: could not open file \"pg_clog/02E6\": No such file or directory\n\nThis probably didn't come up in my wraparound testing because I'm\naggressively forcing wraparound by writing a lot of clog very quickly\nunder XidGenLock, and because I'm mostly looking at xacts that are\neither recent or past the xid boundary. I've added better add coverage\nfor xacts around 2^30 behind the nextXid to the wraparound tests;\ncan't add them to txid.sql since the xid never gets that far in normal\nregression testing.\n\nWhat I'd really like is to be able to ask transam.c to handle the\nxid_in_recent_past logic, treating an attempt to read an xid from\nbeyond the clog truncation threshold as a soft error indicating\nunknown xact state. But that involves delving into slru.c, and I\nreally, really don't want to touch that for what should be a simple\nand pretty noninvasive utility function.\n\nA PG_TRY to trap ERRCODE_UNDEFINED_FILE seems like it'd be sufficient,\nexcept for two issues:\n\n* I see no accepted way to access the errcode etc from within the\nPG_CATCH block, though. PG_RE_THROW() can do it because pg_re_throw()\nis in elog.c. I couldn't find any existing code that seems to check\ndetails about an error thrown in a PG_TRY block, only SPI calls. I\ndon't want to ignore all types of errors and potentially hide\nproblems, so I just used geterrcode() - which is meant for errcontext\ncallbacks - and changed the comment to say it can be used in PG_CATCH\ntoo. I don't see why it shouldn't be.\nWe should probably have some sort of PG_CATCH_INFO(varname) that\nexposes the top ErrorData, but that's not needed for this patch so I\nleft it alone.\n\n* TransactionIdGetStatus() releases the CLogControlLock taken by\nSimpleLruReadPage_ReadOnly() on normal exit but not after an exception\nthrown from SlruReportIOError(). It seems appropriate for\nSimpleLruReadPage() to release the LWLock before calling\nSlruReportIOError(), so I've done that as a separate patch (now 0001).\n\n\nI also removed the TransactionIdInProgress check in txid_status and\njust assume it's in progress if it isn't our xid, committed or\naborted. TransactionIdInProgress looks like it's potentially more\nexpensive, and most of the time we'll be looking at committed or\naborted xacts anyway. I can't sanity-check TransactionIdInProgress\nafter commited/aborted, because there's then a race where the xact can\ncommit or abort after we decide it's not committed/aborted so it's not\nin progress when we go to check that.\n\n\n\n> And there's nothing to keep NextXID from advancing under\n> us, so if somebody asks about a transaction that's just under 2^32\n> transactions old, then get_xid_in_recent_past could say it's valid,\n> then NextXID advances and we look up the XID extracted from the txid\n> and get the status of the new transaction instead of the old one!\n\nHm, yeah. Though due to the clog truncation issue you noticed it\nprobably won't happen.\n\nWe could require that XidGenLock be held at least as LW_SHARED when\nentering get_xid_in_recent_past(), but I'd rather not since that'd be\nan otherwise-unnecessary lwlock for txid_convert_ifrecent().\n\nInstead, I think I'll rename the wraparound flag to too_old and set it\nif the xact is more than say 2^30 from the epoch struct's last_xid,\nleaving a race window so ridiculously improbable that the nearly\nimpossible chance of failing with a clog access error is not a worry.\nIf the server's managing to have a proc stuck that long then it's\nalready on fire. We're only interested in reasonably recent xacts\nsince we can only work with xacts before wraparound / clog truncation.\nThis just moves the threshold for \"reasonably recent\" a bit closer.\n\nAll this certainly reinforces my view that users handling 'xid'\ndirectly or trying to extract it from a bigint epoch-extended xid is a\nbad idea that needs to go away soon.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\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","msg_date":"Mon, 29 Aug 2016 11:25:39 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"Hi,\n\nOn 2016-08-29 11:25:39 +0800, Craig Ringer wrote:\n> ERROR: could not access status of transaction 778793573\n> DETAIL: could not open file \"pg_clog/02E6\": No such file or directory\n> \n> What I'd really like is to be able to ask transam.c to handle the\n> xid_in_recent_past logic, treating an attempt to read an xid from\n> beyond the clog truncation threshold as a soft error indicating\n> unknown xact state. But that involves delving into slru.c, and I\n> really, really don't want to touch that for what should be a simple\n> and pretty noninvasive utility function.\n\nCan't you \"just\" check this against ShmemVariableCache->oldestXid while\nholding appropriate locks?\n\n\n> A PG_TRY to trap ERRCODE_UNDEFINED_FILE seems like it'd be sufficient,\n> except for two issues:\n\nIt seems like a bad idea to PG_CATCH and not re-throw an error. That\ngenerally is quite error prone. At the very least locking and such gets\na lot more complicated (as you noticed below).\n\n> * TransactionIdGetStatus() releases the CLogControlLock taken by\n> SimpleLruReadPage_ReadOnly() on normal exit but not after an exception\n> thrown from SlruReportIOError(). It seems appropriate for\n> SimpleLruReadPage() to release the LWLock before calling\n> SlruReportIOError(), so I've done that as a separate patch (now 0001).\n\nWe normally prefer to handle this via the \"bulk\" releases in the error\nhandlers. It's otherwise hard to write code that handles these\nsituations reliably. It's different for spinlocks, but those normally\nprotect far smaller regions of code.\n\n\nAndres\n\n\n-- \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, 28 Aug 2016 20:45:14 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Sun, Aug 28, 2016 at 11:25 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> What I'd really like is to be able to ask transam.c to handle the\n> xid_in_recent_past logic, treating an attempt to read an xid from\n> beyond the clog truncation threshold as a soft error indicating\n> unknown xact state. But that involves delving into slru.c, and I\n> really, really don't want to touch that for what should be a simple\n> and pretty noninvasive utility function.\n\nI think you're going to have to bite the bullet and do that, though, because ...\n\n> A PG_TRY to trap ERRCODE_UNDEFINED_FILE seems like it'd be sufficient,\n\n...I don't think this has any chance of being acceptable. You can't\ncatch errors and not re-throw them. That's bad news.\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, 29 Aug 2016 02:01:32 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 29 August 2016 at 11:45, Andres Freund <andres@anarazel.de> wrote:\n> Hi,\n>\n> On 2016-08-29 11:25:39 +0800, Craig Ringer wrote:\n>> ERROR: could not access status of transaction 778793573\n>> DETAIL: could not open file \"pg_clog/02E6\": No such file or directory\n>>\n>> What I'd really like is to be able to ask transam.c to handle the\n>> xid_in_recent_past logic, treating an attempt to read an xid from\n>> beyond the clog truncation threshold as a soft error indicating\n>> unknown xact state. But that involves delving into slru.c, and I\n>> really, really don't want to touch that for what should be a simple\n>> and pretty noninvasive utility function.\n>\n> Can't you \"just\" check this against ShmemVariableCache->oldestXid while\n> holding appropriate locks?\n\nHm. Yeah, I should've thought of that. Thank you.\n\n>> A PG_TRY to trap ERRCODE_UNDEFINED_FILE seems like it'd be sufficient,\n>> except for two issues:\n>\n> It seems like a bad idea to PG_CATCH and not re-throw an error. That\n> generally is quite error prone. At the very least locking and such gets\n> a lot more complicated (as you noticed below).\n\nYeah, and as I remember from the \"fun\" of trying to write apply errors\nto tables in BDR. It wasn't my first choice.\n\n>> * TransactionIdGetStatus() releases the CLogControlLock taken by\n>> SimpleLruReadPage_ReadOnly() on normal exit but not after an exception\n>> thrown from SlruReportIOError(). It seems appropriate for\n>> SimpleLruReadPage() to release the LWLock before calling\n>> SlruReportIOError(), so I've done that as a separate patch (now 0001).\n>\n> We normally prefer to handle this via the \"bulk\" releases in the error\n> handlers. It's otherwise hard to write code that handles these\n> situations reliably. It's different for spinlocks, but those normally\n> protect far smaller regions of code.\n\nFair enough. It's not a complex path, but there are a _lot_ of\ncallers, and while I can't really imagine any of them relying on the\nCLogControLock being held on error it's not something I was keen to\nchange. I thought complicating the clog with a soft-error interface\nwas worse and didn't come up with a better approach.\n\nSaid better approach attached in revised series. Thanks.\n\nMy only real complaint with doing this is that it's a bit more\nconservative. But in practice clog truncation probably won't follow\nthat far behind oldestXmin so except in fairly contrived circumstances\nit won't hurt. Apps that need guarantees about how old an xid they can\nget status on can hold down xmin with a replication slot, a dummy\nprepared xact, or whatever. If we find that becomes a common need that\nshould be made simpler then appropriate API to allow apps to hold down\nclog truncation w/o blocking vacuuming can be added down the track.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\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":"Mon, 29 Aug 2016 15:53:34 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 29 August 2016 at 15:53, Craig Ringer <craig@2ndquadrant.com> wrote:\n\n> Said better approach attached in revised series. Thanks.\n\nHere's another minor update to the txid_status() and\ntxid_convert_if_recent() patches. The only change is moving\nget_xid_in_recent_past from src/backend/utils/adt/txid.c to\nsrc/backend/access/transam/xlog.c to permit its use by other code.\nSpecifically, I think it'll be needed for logical decoding on standby.\n\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\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":"Thu, 1 Sep 2016 13:08:45 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 1 September 2016 at 13:08, Craig Ringer <craig@2ndquadrant.com> wrote:\n> On 29 August 2016 at 15:53, Craig Ringer <craig@2ndquadrant.com> wrote:\n>\n>> Said better approach attached in revised series. Thanks.\n>\n> Here's another minor update to the txid_status() and\n> txid_convert_if_recent() patches. The only change is moving\n> get_xid_in_recent_past from src/backend/utils/adt/txid.c to\n> src/backend/access/transam/xlog.c to permit its use by other code.\n> Specifically, I think it'll be needed for logical decoding on standby.\n\nAhem, wrong patches. Attached correctly now.\n\n\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\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":"Thu, 1 Sep 2016 13:29:50 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"Here's another update to these patches. While working on support for\nlogical decoding on standbys I noticed that I needed something like\nget_xid_in_recent_past(...) there too. So I've moved it to xlog.c as\nTransactionIdInRecentPast too and flipped its arguments to be more\nconvenient. No other changes.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\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, 2 Sep 2016 20:16:53 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 2 September 2016 at 13:16, Craig Ringer <craig@2ndquadrant.com> wrote:\n\n> So I've moved it to xlog.c...\n\nI'm pretty sure it shouldn't live in xlog.c, but there may be some\ngood reason I can't see yet.\n\n-- \nSimon Riggs 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":"Fri, 2 Sep 2016 13:30:42 +0100","msg_from":"Simon Riggs <simon@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 2 Sep. 2016 8:30 pm, \"Simon Riggs\" <simon@2ndquadrant.com> wrote:\n>\n> On 2 September 2016 at 13:16, Craig Ringer <craig@2ndquadrant.com> wrote:\n>\n> > So I've moved it to xlog.c...\n>\n> I'm pretty sure it shouldn't live in xlog.c, but there may be some\n> good reason I can't see yet.\n\nUgh. Yes. transam.c would be rather saner.\n\nOnly for the helper to determine if an xid is recent though; txid_ status\nstays in adt/xact.c where it belongs along with txid_current() etc.\n\n> --\n> Simon Riggs http://www.2ndQuadrant.com/\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\nOn 2 Sep. 2016 8:30 pm, \"Simon Riggs\" <simon@2ndquadrant.com> wrote:\n>\n> On 2 September 2016 at 13:16, Craig Ringer <craig@2ndquadrant.com> wrote:\n>\n> > So I've moved it to xlog.c...\n>\n> I'm pretty sure it shouldn't live in xlog.c, but there may be some\n> good reason I can't see yet.\n Ugh. Yes. transam.c would be rather saner.\nOnly for the helper to determine if an xid is recent though; txid_ status stays in adt/xact.c where it belongs along with txid_current() etc.\n> --\n> Simon Riggs http://www.2ndQuadrant.com/\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Fri, 2 Sep 2016 20:38:39 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 2 September 2016 at 20:38, Craig Ringer <craig@2ndquadrant.com> wrote:\n> On 2 Sep. 2016 8:30 pm, \"Simon Riggs\" <simon@2ndquadrant.com> wrote:\n>>\n>> On 2 September 2016 at 13:16, Craig Ringer <craig@2ndquadrant.com> wrote:\n>>\n>> > So I've moved it to xlog.c...\n>>\n>> I'm pretty sure it shouldn't live in xlog.c, but there may be some\n>> good reason I can't see yet.\n>\n> Ugh. Yes. transam.c would be rather saner.\n>\n> Only for the helper to determine if an xid is recent though; txid_ status\n> stays in adt/xact.c where it belongs along with txid_current() etc.\n\nFixed, moved to transam.c.\n\n\n\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\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, 2 Sep 2016 21:01:58 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 2 September 2016 at 21:01, Craig Ringer <craig@2ndquadrant.com> wrote:\n> On 2 September 2016 at 20:38, Craig Ringer <craig@2ndquadrant.com> wrote:\n>> On 2 Sep. 2016 8:30 pm, \"Simon Riggs\" <simon@2ndquadrant.com> wrote:\n>>>\n>>> On 2 September 2016 at 13:16, Craig Ringer <craig@2ndquadrant.com> wrote:\n>>>\n>>> > So I've moved it to xlog.c...\n>>>\n>>> I'm pretty sure it shouldn't live in xlog.c, but there may be some\n>>> good reason I can't see yet.\n>>\n>> Ugh. Yes. transam.c would be rather saner.\n>>\n>> Only for the helper to determine if an xid is recent though; txid_ status\n>> stays in adt/xact.c where it belongs along with txid_current() etc.\n>\n> Fixed, moved to transam.c.\n\nMissed that this causes an undefined reference to GetNextXidAndEpoch()\nwhich is in xlog.c. I knew there was a reason I put it there. So most\nrecent patch is wrong, use the prior one.\n\nGetNextXidAndEpoch() needs to be in xlog.c because it uses XLogCtl's\nshmem copy of checkPoint.nextXidEpoch. So either transam.c needs to\n#include xlog.h (which seems a bit backwards) or\nTransactionIdInRecentPast() should go in xlog.c.\n\nI don't like either really. Opinion? I'm sure we'll want this\nfunctionality in other places as part of dealing with the problems\ndiscussed upthread with 'xid' exposed to users.\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":"Fri, 2 Sep 2016 21:46:19 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 02/09/16 15:46, Craig Ringer wrote:\n> On 2 September 2016 at 21:01, Craig Ringer <craig@2ndquadrant.com> wrote:\n>> On 2 September 2016 at 20:38, Craig Ringer <craig@2ndquadrant.com> wrote:\n>>> On 2 Sep. 2016 8:30 pm, \"Simon Riggs\" <simon@2ndquadrant.com> wrote:\n>>>>\n>>>> On 2 September 2016 at 13:16, Craig Ringer <craig@2ndquadrant.com> wrote:\n>>>>\n>>>>> So I've moved it to xlog.c...\n>>>>\n>>>> I'm pretty sure it shouldn't live in xlog.c, but there may be some\n>>>> good reason I can't see yet.\n>>>\n>>> Ugh. Yes. transam.c would be rather saner.\n>>>\n>>> Only for the helper to determine if an xid is recent though; txid_ status\n>>> stays in adt/xact.c where it belongs along with txid_current() etc.\n>>\n>> Fixed, moved to transam.c.\n>\n> Missed that this causes an undefined reference to GetNextXidAndEpoch()\n> which is in xlog.c. I knew there was a reason I put it there. So most\n> recent patch is wrong, use the prior one.\n>\n> GetNextXidAndEpoch() needs to be in xlog.c because it uses XLogCtl's\n> shmem copy of checkPoint.nextXidEpoch. So either transam.c needs to\n> #include xlog.h (which seems a bit backwards) or\n> TransactionIdInRecentPast() should go in xlog.c.\n>\n> I don't like either really. Opinion? I'm sure we'll want this\n> functionality in other places as part of dealing with the problems\n> discussed upthread with 'xid' exposed to users.\n>\n\nYou could put it to txid.c where all the other txid stuff is in?\n\n-- \n Petr Jelinek 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, 2 Sep 2016 17:29:34 +0200","msg_from":"Petr Jelinek <petr@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 2 September 2016 at 23:29, Petr Jelinek <petr@2ndquadrant.com> wrote:\n\n> You could put it to txid.c where all the other txid stuff is in?\n\nYeah, even though it's in adt/ I think it'll do.\n\nI thought I'd need get_xid_in_recent_past() for catalog_xmin hot\nstandby feedback, but upon closer examination the needed logic isn't\nthe same anymore. txid_status() wants to ensure clog lookups are safe\nand limit by oldest xid, wheras the walsender doesn't actually care\nabout that and is just avoiding wrapped xids.\n\nI'm just going back to how it was, all in adt/txid.c, and making it\nstatic again. We can move it and make it non-static if a need to do so\ncomes up.\n\nAttached rebased patch updated and vs master.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\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, 16 Sep 2016 08:52:17 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Thu, Sep 15, 2016 at 8:52 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> On 2 September 2016 at 23:29, Petr Jelinek <petr@2ndquadrant.com> wrote:\n>\n>> You could put it to txid.c where all the other txid stuff is in?\n>\n> Yeah, even though it's in adt/ I think it'll do.\n>\n> I thought I'd need get_xid_in_recent_past() for catalog_xmin hot\n> standby feedback, but upon closer examination the needed logic isn't\n> the same anymore. txid_status() wants to ensure clog lookups are safe\n> and limit by oldest xid, wheras the walsender doesn't actually care\n> about that and is just avoiding wrapped xids.\n>\n> I'm just going back to how it was, all in adt/txid.c, and making it\n> static again. We can move it and make it non-static if a need to do so\n> comes up.\n>\n> Attached rebased patch updated and vs master.\n\nYou appear to have attached the wrong patch set.\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, 16 Sep 2016 09:28:54 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 16 September 2016 at 21:28, Robert Haas <robertmhaas@gmail.com> wrote:\n> On Thu, Sep 15, 2016 at 8:52 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n>> On 2 September 2016 at 23:29, Petr Jelinek <petr@2ndquadrant.com> wrote:\n>>\n>>> You could put it to txid.c where all the other txid stuff is in?\n>>\n>> Yeah, even though it's in adt/ I think it'll do.\n>>\n>> I thought I'd need get_xid_in_recent_past() for catalog_xmin hot\n>> standby feedback, but upon closer examination the needed logic isn't\n>> the same anymore. txid_status() wants to ensure clog lookups are safe\n>> and limit by oldest xid, wheras the walsender doesn't actually care\n>> about that and is just avoiding wrapped xids.\n>>\n>> I'm just going back to how it was, all in adt/txid.c, and making it\n>> static again. We can move it and make it non-static if a need to do so\n>> comes up.\n>>\n>> Attached rebased patch updated and vs master.\n>\n> You appear to have attached the wrong patch set.\n\nWhoops, multitasking fail.\n\nSorry for the late response, hospitals are \"fun\".\n\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\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, 20 Sep 2016 09:54:28 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Mon, Sep 19, 2016 at 9:54 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n>> You appear to have attached the wrong patch set.\n>\n> Whoops, multitasking fail.\n>\n> Sorry for the late response, hospitals are \"fun\".\n\nI did some cleanup of 0001 (see attached) and was all set to commit it\nwhen I realized what I think is a problem: holding XidGenLock doesn't\nseem to help with the race condition between this function and CLOG\ntruncation, because vac_truncate_clog() updates the shared memory\nlimits AFTER performing the truncation. If the order of those\noperations were reversed, we'd be fine, because it would get stuck\ntrying to update the shared memory limits and wouldn't be able to\ntruncate until it did - and conversely, if it updated the shared\nmemory limits before we examined them, that would be OK, too, because\nwe'd be sure not to consult the pages that are about to be truncated.\nAs it is, though, I don't see that there's any real interlock here.\n\n(BTW, the stuff you moved from clog.c to clog.h doesn't actually need\nto be moved; one of the changes I made here was to undo that.)\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\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, 20 Sep 2016 10:46:39 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 20 September 2016 at 22:46, Robert Haas <robertmhaas@gmail.com> wrote:\n\n> I did some cleanup of 0001 (see attached) and was all set to commit it\n> when I realized what I think is a problem: holding XidGenLock doesn't\n> seem to help with the race condition between this function and CLOG\n> truncation, because vac_truncate_clog() updates the shared memory\n> limits AFTER performing the truncation.\n\nThanks ... and drat.\n\n> If the order of those\n> operations were reversed, we'd be fine, because it would get stuck\n> trying to update the shared memory limits and wouldn't be able to\n> truncate until it did - and conversely, if it updated the shared\n> memory limits before we examined them, that would be OK, too, because\n> we'd be sure not to consult the pages that are about to be truncated.\n\nI'm hesitant to mess with something that fundamental for what I was\nhoping was a low-impact feature, albeit one that seems to be trying\nhard not to be at every turn. It looks pretty reasonable to update\noldestXid before clog truncation but I don't want to be wrong and\ncreate some obscure race or crash recovery issue related to wraparound\nand clog truncation. It could well be a problem if we're very close to\nwraparound.\n\nSo far nothing has had any reason to care about this, since there's no\nway to attempt to access an older xid in a normally functioning\nsystem. The commit timestamp lookup code doesn't care whether the xid\nis still in clog or not and nothing else does lookups based on xids\nsupplied by the user. If anything else did or does in future it will\nhave the same problem as txid_status.\n\nWe've already ruled out releasing the slru LWLock in SlruReportIOError\nthen PG_TRY / PG_CATCH()ing clog access errors in txid_status() per my\noriginal approach to this issue.\n\nSo I see a few options now:\n\n* Do nothing. Permit this race to exist, document the error, and\nexpect apps to cope. I'm pretty tempted to go for exactly this since\nit pushes the cost onto users of the feature and doesn't make a mess\nelsewhere. People who use this will typically be invoking it as a\nstandalone toplevel function anyway, so it's mostly just a bit of\nnoise in the error logs if you lose the race - and we have plenty of\nother sources of that already.\n\n* Rather than calling TransactionIdDidCommit / TransactionIdDidAbort,\ncall clog.c's TransactionIdGetStatus with a new missing_ok flag. That\nsets a bool* missing param added to SimpleLruReadPage_ReadOnly(...)\nand in turn to SimpleLruReadPage(...) that's set instead of calling\nSlruReportIOError(...). This seems rather intrusive and will add\nlittle-used params and paths to fairly hot slru.c code so I'm not\nkeen.\n\n* Take CLogControlLock LW_SHARED in txid_status() to prevent\ntruncation before reading oldestXid. We'd need a way to pass an\n\"already locked\" state through TransactionIdGetStatus(...) to\nSimpleLruReadPage_ReadOnly(...), which isn't great since again it's\npretty hot code.\n\n* Don't release the slru LWLock in SlruReportIOError; instead release\nCLogControlLock from txid_status on clog access failure. As before\nmeans PG_TRY / PG_CATCH without PG_RE_THROW, but it means the only\nplace it affects is callers of txid_status(...). For added safety,\nrestrict txid_status() to being called in a toplevel virtual xact so\nwe know we'll finish up promptly. It's a horrible layering violation\nhaving txid_status(...) release the clog control lock though, and\nseems risky.\n\nThe only non-horrible one of those is IMO to just let the caller see\nan error if they lose the race. It's a function that's intended for\nuse when you're dealing with indeterminate transaction state after a\nserver or application error anyway, so it's part of an error path\nalready. So I say we just document the behaviour. Because slru.c\ndoesn't release its LWLock on error we also need to ensure\ntxid_status(...) is also only called from a toplevel xact so the user\ndoesn't attempt to wrap it in plpgsql BEGIN ... EXCEPTION block and it\ncauses the xact to abort.\n\nWill follow up with just that.\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, 21 Sep 2016 15:40:34 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Wed, Sep 21, 2016 at 3:40 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> The only non-horrible one of those is IMO to just let the caller see\n> an error if they lose the race. It's a function that's intended for\n> use when you're dealing with indeterminate transaction state after a\n> server or application error anyway, so it's part of an error path\n> already. So I say we just document the behaviour.\n\nI am not keen on that idea. The errors we're likely to be exposing\nare going to look like low-level internal failures, which might scare\nsome DBA. Even if they don't, I think it's playing with fire. The\nsystem is designed on the assumption that nobody will try to look up\nan XID that's too old, and if we start to violate that assumption I\nthink we're undermining the design integrity of the system in a way\nwe'll likely come to regret. To put that more plainly, when the code\nis written with the assumption that X will never happen, it's usually\na bad idea to casually add code that does X.\n\n> Because slru.c\n> doesn't release its LWLock on error we also need to ensure\n> txid_status(...) is also only called from a toplevel xact so the user\n> doesn't attempt to wrap it in plpgsql BEGIN ... EXCEPTION block and it\n> causes the xact to abort.\n\nI think this is muddled, because an error aborts the transaction, and\nAbortTransaction() and AbortSubTransaction() start with\nLWLockReleaseAll().\n\nIt might not be too hard to add a second copy of oldestXid in shared\nmemory that is updated before truncation rather than afterward... but\nyeah, like you, I'm not finding this nearly as straightforward as\nmight have been hoped.\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, 21 Sep 2016 10:16:49 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 21 September 2016 at 22:16, Robert Haas <robertmhaas@gmail.com> wrote:\n> On Wed, Sep 21, 2016 at 3:40 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n>> The only non-horrible one of those is IMO to just let the caller see\n>> an error if they lose the race. It's a function that's intended for\n>> use when you're dealing with indeterminate transaction state after a\n>> server or application error anyway, so it's part of an error path\n>> already. So I say we just document the behaviour.\n>\n> I am not keen on that idea. The errors we're likely to be exposing\n> are going to look like low-level internal failures, which might scare\n> some DBA. Even if they don't, I think it's playing with fire. The\n> system is designed on the assumption that nobody will try to look up\n> an XID that's too old, and if we start to violate that assumption I\n> think we're undermining the design integrity of the system in a way\n> we'll likely come to regret. To put that more plainly, when the code\n> is written with the assumption that X will never happen, it's usually\n> a bad idea to casually add code that does X.\n\nFair point.\n\n> [snip]\n>\n> It might not be too hard to add a second copy of oldestXid in shared\n> memory that is updated before truncation rather than afterward... but\n> yeah, like you, I'm not finding this nearly as straightforward as\n> might have been hoped.\n\nYeah.\n\nI suspect that'll be the way to go, to add another copy that's updated\nbefore clog truncation. It just seems ... unclean. Like it shouldn't\nbe necessary, something else isn't right. But it's probably the lowest\npain option.\n\nI'm going to take a step back on this and see if I can spot an alternative.\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":"Tue, 27 Sep 2016 09:23:21 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 27 September 2016 at 09:23, Craig Ringer <craig@2ndquadrant.com> wrote:\n> On 21 September 2016 at 22:16, Robert Haas <robertmhaas@gmail.com> wrote:\n>>\n>> It might not be too hard to add a second copy of oldestXid in shared\n>> memory that is updated before truncation rather than afterward... but\n>> yeah, like you, I'm not finding this nearly as straightforward as\n>> might have been hoped.\n>\n> Yeah.\n>\n> I suspect that'll be the way to go, to add another copy that's updated\n> before clog truncation. It just seems ... unclean. Like it shouldn't\n> be necessary, something else isn't right. But it's probably the lowest\n> pain option.\n\nOK, so summarizing the problem:\n\nslru.c and clog.c have no soft-failure entrypoints where we can look\nit up and fail gracefully if the xid isn't in the clog.\nvac_truncate_clog() calls TruncateCLOG to chop SLRU pages from the\nclog before it takes XidGenLock to advance oldestXid. So we cannot use\noldestXid protected by XidGenLock as an interlock against concurrent\nclog truncation to prevent SLRU access errors looking up an xid, and\nthere's no other candidate lock.\n\nThis hasn't been an issue before because nobody looks up arbitrary\nuser-supplied XIDs in clog, we only look up things that're protected\nby datfrozenxid etc. But the whole point of txid_status is to look up\nuser-supplied xids.\n\nWe can't just take ClogControlLock from txid_status() to block\ntruncation because clog.c expects to own that lock, and takes it (via\nslru.c) in TransactionIdGetStatus, with no way to pass an\nalready_locked state. We'd self-deadlock.\n\nAdding a second copy of oldestXid - say pendingOldestXid - won't\nactually help us unless we also take some suitable LWLock before\nupdating it, otherwise truncation can continue after we look at\npendingOldestXid but before we do the clog lookup. That means an extra\nLWLock for each clog truncation, but compared to the I/O done during\nclog truncation and the cost of the SlruScanDirectory() pre-check done\nby TruncateCLOG it's nothing.\n\nWe could take XidGenLock twice, once to update this new\npendingOldestXid field and once to update oldestXid, but that's a\nhighly contested lock I'd rather not mess with even on a path that's\nnot hit much.\n\nInstead, I've added a new LWLock, ClogTruncationLock, for that\npurpose. vac_truncate_clog() takes it if it decides to attempt clog\ntruncation. This lock is held throughout the whole process of clog\ntruncation and oldestXid advance, so there's no need for a new\npendingOldestXid field in ShmemVariableCache. We just take the lock\nthen look at oldestXid, knowing that it's guaranteed to correspond to\nan existing clog page that won't go away while we're looking.\nClogTruncationLock is utterly uncontested so it's going to have no\nmeaningful impact compared to all the work we do scanning the clog to\ndecide whether we're even going to try truncating it, etc.\n\n\n(BTW, it seems like a pity that lwlocknames.txt doesn't have comments\non each lock. We could have\nsrc/backend/storage/lmgr/generate-lwlocknames.pl transform the #\ncomments into /* comments */ on the generated header. Thoughts?)\n\n\n\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, 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":"Wed, 21 Dec 2016 16:02:15 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Wed, Dec 21, 2016 at 3:02 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> Instead, I've added a new LWLock, ClogTruncationLock, for that\n> purpose. vac_truncate_clog() takes it if it decides to attempt clog\n> truncation. This lock is held throughout the whole process of clog\n> truncation and oldestXid advance, so there's no need for a new\n> pendingOldestXid field in ShmemVariableCache. We just take the lock\n> then look at oldestXid, knowing that it's guaranteed to correspond to\n> an existing clog page that won't go away while we're looking.\n> ClogTruncationLock is utterly uncontested so it's going to have no\n> meaningful impact compared to all the work we do scanning the clog to\n> decide whether we're even going to try truncating it, etc.\n\nThat makes everything that happens between when we acquire that lock\nand when we release it non-interruptible, which seems undesirable. I\nthink that extra copy of oldestXid is a nicer approach.\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, 21 Dec 2016 11:30:44 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 22 December 2016 at 00:30, Robert Haas <robertmhaas@gmail.com> wrote:\n\n> That makes everything that happens between when we acquire that lock\n> and when we release it non-interruptible, which seems undesirable. I\n> think that extra copy of oldestXid is a nicer approach.\n\nThat's a side-effect I didn't realise. Given that, yes, I agree.\n\nSince we don't truncate clog much, do you think it's reasonable to\njust take XidGenLock again before we proceed? I'm reluctant to add\nanother acquisition of a frequently contested lock for something 99.9%\nof the codebase won't care about, so I think it's probably better to\nadd a new LWLock, and I'll resubmit on that basis, but figure it's\nworth asking.\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":"Thu, 22 Dec 2016 07:49:40 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 22 December 2016 at 07:49, Craig Ringer <craig@2ndquadrant.com> wrote:\n> On 22 December 2016 at 00:30, Robert Haas <robertmhaas@gmail.com> wrote:\n>\n>> That makes everything that happens between when we acquire that lock\n>> and when we release it non-interruptible, which seems undesirable. I\n>> think that extra copy of oldestXid is a nicer approach.\n>\n> That's a side-effect I didn't realise. Given that, yes, I agree.\n>\n> Since we don't truncate clog much, do you think it's reasonable to\n> just take XidGenLock again before we proceed? I'm reluctant to add\n> another acquisition of a frequently contested lock for something 99.9%\n> of the codebase won't care about, so I think it's probably better to\n> add a new LWLock, and I'll resubmit on that basis, but figure it's\n> worth asking.\n\nUpdated.\n\nIf you think it's better to just take XidGenLock again, let me know.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, 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, 22 Dec 2016 09:55:06 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 22 December 2016 at 09:55, Craig Ringer <craig@2ndquadrant.com> wrote:\n\n> Updated.\n>\n> If you think it's better to just take XidGenLock again, let me know.\n\nHere's a further update that merges in Robert's changes from the patch\nposted upthread.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, 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, 22 Dec 2016 13:12:10 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Thu, Dec 22, 2016 at 12:12 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> On 22 December 2016 at 09:55, Craig Ringer <craig@2ndquadrant.com> wrote:\n>\n>> Updated.\n>>\n>> If you think it's better to just take XidGenLock again, let me know.\n>\n> Here's a further update that merges in Robert's changes from the patch\n> posted upthread.\n\nThis patch contains unresolved merge conflicts. Maybe\nSetPendingTransactionIdLimit could happen in TruncateCLOG rather than\nthe caller. Instead of using an LWLock, how about a spin lock?\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, 22 Dec 2016 12:26:22 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 23 December 2016 at 01:26, Robert Haas <robertmhaas@gmail.com> wrote:\n\n> This patch contains unresolved merge conflicts.\n\nAh, it conflicts with fe0a0b59, the PostmasterRandom changes. I\nthought I'd rebased more recently than that.\n\n> Maybe SetPendingTransactionIdLimit could happen in TruncateCLOG rather than\n> the caller.\n\nI've spent a while looking at how committs and multixact handle the\nrace hazard of the gap between clog truncation and shmem state change\n- and the related race on standby, where the hazard is the gap between\nreplay of CLOG_TRUNCATE records and replay of the next checkpoint\ncontaining updated limits.\n\npg_xact_commit_timestamp() and the underlying\nTransactionIdGetCommitTsData are supposed to accept arbitrary xids,\nlike txid_status(), so they should handle this.\n\nHowever, as far as I can tell, committs has the same issues presented\nby txid_status(). We only update ShmemVariableCache->oldestCommitTsXid\n_after_ we truncate the committs SLRU, no lock is held over the\nduration, and no other lock-protected var is set before truncation. We\ndon't record the new limit values in COMMIT_TS_TRUNCATE records so\nthey only becomes known to the standby at replay of the next\ncheckpoint.\n\nmultixact instead runs a single critical section to write xlog, set\nshmem state, and _then_ truncate the SLRUs. It's more complex than\nclog or commit_ts since it has two inter-related SLRUs, though.\n\nSo: I think we should be setting ShmemVariableCache->oldestXid (under\nXidGenLock) from TruncateCLOG, after we write xlog but _before_\ntruncation. Then the rest of SetTransactionIdLimit(...) proceeds as\nbefore. I don't _think_ we need the critical section since we only\nhave the one SLRU to worry about. We should also add oldestXid to\nCLOG_TRUNCATE xlog records so it is up to date on standbys.\n\nIt's a little annoying to take XidGenLock twice - once for oldestXid,\nonce for the other xid limits - but we can just delay advancing\noldestXid entirely until we've got a clog page to truncate away, in\nwhich case there's a big enough cost that it doesn't matter. We have\nto wait to advance the other limits until after we truncate clog to\nprevent new (wrapped) xids trying to set values in clog for the\nbefore-wrap xids we're about to truncate away.\n\nOn the upside, the need for pendingOldestXid, which always felt hacky,\ngoes away.\n\nWhile we're at it, lets do the same thing for commit_ts. Advance its\nlimit before truncation, not after, and record that limit in its xlog\nrecords for redo. There's no need for any special dancing around\nthere.\n\n\n\n> Instead of using an LWLock, how about a spin lock?\n\nI wrote an explanation of how that wouldn't work, but then I found the\nproblem with standby, so it no longer matters.\n\nI'll have to follow up with a patch soon, as it's Toddler o'Clock.\n\n(It's interesting that most of what I'm doing at the moment is\nwrestling with resource retention limits and how they're often quite\nimplicit. All the catalog_xmin stuff for logical decoding on standby\nhas parallels to this, for example, though not enough to create useful\noverlap of functionality.)\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":"Fri, 23 Dec 2016 18:00:43 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 23 December 2016 at 18:00, Craig Ringer <craig@2ndquadrant.com> wrote:\n\n> I'll have to follow up with a patch soon, as it's Toddler o'Clock.\n\nHere we go.\n\nThis patch advances oldestXid, under XidGenLock, _before_ truncating clog.\n\ntxid_status() holds XidGenLock from when it tests oldestXid until it's\ndone looking up clog, thus eliminating the race.\n\nCLOG_TRUNCATE records now contain the oldestXid, so they can advance\noldestXid on a standby, or when we've truncated clog since the most\nrecent checkpoint on the master during recovery. It's advanced under\nXidGenLock during redo to protect against this race on standby.\n\nAs outlined in my prior mail I think this is the right approach. I\ndon't like taking XidGenLock twice, but we don't advance datfrozenxid\nmuch so it's not a big concern. While a separate ClogTruncationLock\ncould be added like in my earlier patch, oldestXid is currently under\nXidGenLock and I'd rather not change that.\n\nThe biggest change here is that oldestXid is advanced separately to\nthe vac limits in the rest of ShmemVariableCache. As far as I can tell\nwe don't prevent two manual VACUUMs on different DBs from trying to\nconcurrently run vac_truncate_clog, so this has to be safe against two\ninvocations racing each other. Rather than try to lock out such\nconcurrency, the patch ensures that oldestXid can never go backwards.\nIt doesn't really matter if the vac limits go backwards, it's no worse\nthan what can already happen in the current code.\n\nWe cannot advance the vacuum limits before we truncate the clog away,\nin case someone tries to access a very new xid (if we're near\nwraparound)\n\nI'm pretty sure that commit timestamps suffer from the same flaw as\nRobert identified upthread with clog. This patch fixes the clog race,\nbut not the similar one in commit timestamps. Unlike the clog race\nwith txid_status(), the commit timestamps one is already potentially\nuser-visible since we allow arbitrary xids to be looked up for commit\ntimestamps. I'll address that separately.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, 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":"Wed, 28 Dec 2016 18:00:37 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 28 December 2016 at 18:00, Craig Ringer <craig@2ndquadrant.com> wrote:\n> On 23 December 2016 at 18:00, Craig Ringer <craig@2ndquadrant.com> wrote:\n>\n>> I'll have to follow up with a patch soon, as it's Toddler o'Clock.\n>\n> Here we go.\n>\n> This patch advances oldestXid, under XidGenLock, _before_ truncating clog.\n>\n> txid_status() holds XidGenLock from when it tests oldestXid until it's\n> done looking up clog, thus eliminating the race.\n>\n> CLOG_TRUNCATE records now contain the oldestXid, so they can advance\n> oldestXid on a standby, or when we've truncated clog since the most\n> recent checkpoint on the master during recovery. It's advanced under\n> XidGenLock during redo to protect against this race on standby.\n>\n> As outlined in my prior mail I think this is the right approach. I\n> don't like taking XidGenLock twice, but we don't advance datfrozenxid\n> much so it's not a big concern. While a separate ClogTruncationLock\n> could be added like in my earlier patch, oldestXid is currently under\n> XidGenLock and I'd rather not change that.\n>\n> The biggest change here is that oldestXid is advanced separately to\n> the vac limits in the rest of ShmemVariableCache. As far as I can tell\n> we don't prevent two manual VACUUMs on different DBs from trying to\n> concurrently run vac_truncate_clog, so this has to be safe against two\n> invocations racing each other. Rather than try to lock out such\n> concurrency, the patch ensures that oldestXid can never go backwards.\n> It doesn't really matter if the vac limits go backwards, it's no worse\n> than what can already happen in the current code.\n>\n> We cannot advance the vacuum limits before we truncate the clog away,\n> in case someone tries to access a very new xid (if we're near\n> wraparound)\n>\n> I'm pretty sure that commit timestamps suffer from the same flaw as\n> Robert identified upthread with clog. This patch fixes the clog race,\n> but not the similar one in commit timestamps. Unlike the clog race\n> with txid_status(), the commit timestamps one is already potentially\n> user-visible since we allow arbitrary xids to be looked up for commit\n> timestamps. I'll address that separately.\n\nRebased patch attached. I've split the clog changes out from\ntxid_status() its self.\n\nThere is relevant discussion on the commit timestamp truncation fix\nthread where the similar fix for commit_ts got committed.\n\nhttps://www.postgresql.org/message-id/flat/979ff13d-0b8e-4937-01e8-2925c0adc306%402ndquadrant.com#979ff13d-0b8e-4937-01e8-2925c0adc306@2ndquadrant.com\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\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":"Mon, 23 Jan 2017 14:32:13 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Mon, Jan 23, 2017 at 1:32 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> Rebased patch attached. I've split the clog changes out from\n> txid_status() its self.\n\nI think it's fairly surprising that TruncateCLOG() has responsibility\nfor writing the xlog record that protects advancing\nShmemVariableCache->oldestXid, but not the responsibility for actually\nadvancing that value. In other words, I think the AdvanceOldestXid()\ncall in vac_truncate_clog() should be moved into TruncateClog().\n(Similarly, one wonders why AdvanceOldestCommitTsXid() isn't the\nresponsibility of TruncateCommitTs().)\n\nI think it is not correct to advance oldestXid but not oldestXidDB.\nOtherwise, GetNewTransactionId() might complain about the wrong\ndatabase.\n\nThe way that SetTransactionIdLimit() now works looks a bit dangerous.\nxidWrapLimit, xidStopLimit, and xidWarnLimit are computed based on the\npassed-in oldestXid value and written straight into shared memory.\nBut the shared memory copy of oldestXid could have a different value.\nI'm not sure if that breaks anything, but it certainly weakens any\nconfidence callers might have had that all those values are consistent\nwith each other.\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, 24 Jan 2017 10:49:57 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 24 January 2017 at 23:49, Robert Haas <robertmhaas@gmail.com> wrote:\n\n> I think it's fairly surprising that TruncateCLOG() has responsibility\n> for writing the xlog record that protects advancing\n> ShmemVariableCache->oldestXid, but not the responsibility for actually\n> advancing that value. In other words, I think the AdvanceOldestXid()\n> call in vac_truncate_clog() should be moved into TruncateClog().\n> (Similarly, one wonders why AdvanceOldestCommitTsXid() isn't the\n> responsibility of TruncateCommitTs().)\n\nAgreed, and done in attached.\n\nI haven't done the same for commit_ts but will do so on the thread for\nthe commit_ts race fix if we go ahead with this change here.\n\n> I think it is not correct to advance oldestXid but not oldestXidDB.\n> Otherwise, GetNewTransactionId() might complain about the wrong\n> database.\n\nThat's a clear oversight. Fixed in attached.\n\n> The way that SetTransactionIdLimit() now works looks a bit dangerous.\n> xidWrapLimit, xidStopLimit, and xidWarnLimit are computed based on the\n> passed-in oldestXid value and written straight into shared memory.\n> But the shared memory copy of oldestXid could have a different value.\n> I'm not sure if that breaks anything, but it certainly weakens any\n> confidence callers might have had that all those values are consistent\n> with each other.\n\nThis was my main hesitation with the whole thing too.\n\nIt's necessary to advance oldestXmin before we xlog the advance and\ntruncate clog, and necessary to advance the vacuum limits only\nafterwards.\n\nI thought it sensible to be conservative about the xid limit to\nminimise the change from current behaviour, i.e. advance only up to\nxid limits calculated from the oldestXid determined by the currently\nvac_truncate_clog. But the current one is actually wrong; in the\n(unlikely if it's possible at all) case where two vac_truncate_clog()s\nA and B run with ordering A(advanceOldestXmin) B(advanceOldestXmin)\nA(truncate) B(truncate) B(advanceLimits) A(advanceLimits), A's advance\nof the limits would clobber b's. Not too bad, but it'd delay\nadvancing the limits until the next vacuum, and some xid could get\nallocated before the limits go backwards.\n\nIt's safer to take XidGenLock for slightly longer in order to capture\nthe oldestXid from shmem and calculate using it. That ensures we never\nhave any risk of going backwards. The attached updated patch does so.\n\nBTW, I find it quite amusing that this was intended to be a small,\nunintrusive patch, and now it's messing with xid limits and clog\ntruncation. I'm almost tempted to say we should commit it with the\n(tiny) race with clog truncation in place. We certainly haven't had\nany complaints about the same race from people using commit_ts. But\nthe race window on standby is quite large and I'd rather not knowingly\nintroduce new bugs.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\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":"Wed, 25 Jan 2017 13:44:04 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 25 January 2017 at 13:44, Craig Ringer <craig@2ndquadrant.com> wrote:\n> On 24 January 2017 at 23:49, Robert Haas <robertmhaas@gmail.com> wrote:\n>\n>> I think it's fairly surprising that TruncateCLOG() has responsibility\n>> for writing the xlog record that protects advancing\n>> ShmemVariableCache->oldestXid, but not the responsibility for actually\n>> advancing that value. In other words, I think the AdvanceOldestXid()\n>> call in vac_truncate_clog() should be moved into TruncateClog().\n>> (Similarly, one wonders why AdvanceOldestCommitTsXid() isn't the\n>> responsibility of TruncateCommitTs().)\n>\n> Agreed, and done in attached.\n>\n> I haven't done the same for commit_ts but will do so on the thread for\n> the commit_ts race fix if we go ahead with this change here.\n>\n>> I think it is not correct to advance oldestXid but not oldestXidDB.\n>> Otherwise, GetNewTransactionId() might complain about the wrong\n>> database.\n>\n> That's a clear oversight. Fixed in attached.\n\nNew attached also records it in xlog and applies it to the standby.\n\nIf we want to save the 4 bytes per xmin advance (probably not worth\ncaring) we can instead skip setting it on the standby, in which case\nit'll be potentially wrong until the next checkpoint. I'd rather make\nsure it stays correct.\n\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\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":"Wed, 25 Jan 2017 15:02:26 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Wed, Jan 25, 2017 at 4:02 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> If we want to save the 4 bytes per xmin advance (probably not worth\n> caring) we can instead skip setting it on the standby, in which case\n> it'll be potentially wrong until the next checkpoint. I'd rather make\n> sure it stays correct.\n\nThose patches still apply and no reviews have come in yet, so moved to\nCF 2017-03.\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, 1 Feb 2017 15:10:21 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Wed, Jan 25, 2017 at 12:44 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n>> The way that SetTransactionIdLimit() now works looks a bit dangerous.\n>> xidWrapLimit, xidStopLimit, and xidWarnLimit are computed based on the\n>> passed-in oldestXid value and written straight into shared memory.\n>> But the shared memory copy of oldestXid could have a different value.\n>> I'm not sure if that breaks anything, but it certainly weakens any\n>> confidence callers might have had that all those values are consistent\n>> with each other.\n>\n> This was my main hesitation with the whole thing too.\n>\n> It's necessary to advance oldestXmin before we xlog the advance and\n> truncate clog, and necessary to advance the vacuum limits only\n> afterwards.\n\nWell, that's why I tried to advocate that their should be two separate\nXID limits in shared memory: leave what's there now the way it is, and\nthen add a new limit for \"don't try to look up XIDs before this point:\nsplat\". I still think that'd be less risky.\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, 9 Mar 2017 13:55:25 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 10 March 2017 at 02:55, Robert Haas <robertmhaas@gmail.com> wrote:\n\n> Well, that's why I tried to advocate that their should be two separate\n> XID limits in shared memory: leave what's there now the way it is, and\n> then add a new limit for \"don't try to look up XIDs before this point:\n> splat\". I still think that'd be less risky.\n\nI'm coming back to this \"cold\" after an extended break, so I hope I\nget the details right.\n\nTL;DR: doing it that way duplicates a bunch of stuff and is ugly\nwithout offering significant benefit over just fixing the original.\n\n\nI started out with the approach you suggested, but it turns out to be\nless helpful than you'd think. Simply advancing a new lower limit\nfield before beginning truncation is no help; there's then a race\nwhere the lower-limit field can be advanced after we test it but\nbefore we actually do the SLRU read from clog. To guard against this\nit's necessary for SLRU truncation to take an exclusive lock during\nwhich it advances the lower bound. That way a concurrent reader can\ntake the lock in shared mode before reading the lower bound and hold\nit until it finishes the clog read, knowing that vacuum cannot then\ntruncate the data out from under it because it'll block trying to\nadvance the lower limit.\n\nA spinlock isn't suitable for this. While we can take the lock only\nbriefly to update the limit field in vac_truncate_clog, in\ntxid_status() we have to hold it from when we test the boundary\nthrough to when we finish the SLRU clog lookup, and that lookup does\nI/O and might sleep. If we release it after testing the lower bound\nbut before the SLRU lookup our race comes back since vacuum can jump\nin and truncate it out from under us. So now we need a new LWLock used\nonly for vac_truncate_clog before advancing the clog truncation.\n\nI implemented just that - a new ClogTruncateLog in the lwlocks array\nand a new field in ShmemVariableCache for the lower xid bound, IIRC.\nOther than requiring an extra lwlock acquisition for vac_truncate_clog\nit works fine ... for the master.\n\nBut it doesn't fix the much bigger race on the standby. We only emit\nWAL for xid limits after we truncate clog, and the clog truncation\nrecord doesn't record the new limit.\n\nSo now we need a new, somewhat redundant, xlog record and redo\nfunction for this lower clog bound pointer. Which, really, is only\nactually tracking a slightly more up to date version of oldestXid.\n\nAt that point I was just papering around a race that should just be\nfixed at its source instead. Advance oldestXid before truncating clog,\nand write a clog truncation record that includes the new oldestXid.\n\nSo... I can go back to the old approach and just add the new xlog\nrecord and redo method, new lwlock, new shmemvariablecache field, etc,\nif you're really concerned this approach is risky. But I'd rather fix\nthe original problem instead.\n\nIt might be helpful if I separate out the patch that touches oldestXid\nfrom the rest for separate review, so I'll update here with a 2-patch\nseries instead of a squashed single patch soon.\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":"Fri, 10 Mar 2017 15:00:16 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Fri, Mar 10, 2017 at 2:00 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> On 10 March 2017 at 02:55, Robert Haas <robertmhaas@gmail.com> wrote:\n>> Well, that's why I tried to advocate that their should be two separate\n>> XID limits in shared memory: leave what's there now the way it is, and\n>> then add a new limit for \"don't try to look up XIDs before this point:\n>> splat\". I still think that'd be less risky.\n>\n> I'm coming back to this \"cold\" after an extended break, so I hope I\n> get the details right.\n\nYeah, sorry I've been away from this for a while.\n\n> TL;DR: doing it that way duplicates a bunch of stuff and is ugly\n> without offering significant benefit over just fixing the original.\n>\n> I started out with the approach you suggested, but it turns out to be\n> less helpful than you'd think. Simply advancing a new lower limit\n> field before beginning truncation is no help; there's then a race\n> where the lower-limit field can be advanced after we test it but\n> before we actually do the SLRU read from clog. To guard against this\n> it's necessary for SLRU truncation to take an exclusive lock during\n> which it advances the lower bound. That way a concurrent reader can\n> take the lock in shared mode before reading the lower bound and hold\n> it until it finishes the clog read, knowing that vacuum cannot then\n> truncate the data out from under it because it'll block trying to\n> advance the lower limit.\n\nThat's a good point which I hadn't fully considered. On the other\nhand, there really are two separate notions of the \"oldest\" XID.\nThere's the oldest XID that we can safely look up, and then there's\nthe oldest XID that we can't reuse. These two are the same when no\ntruncation is in progress, but when a truncation is in progress then\nthey're different: the oldest XID that's safe to look up is the first\none after whatever we're truncating away, but the oldest XID that we\ncan't reuse is the newest one preceding the stuff that we're busy\ntruncating. IOW, when truncation is happening, there's a portion of\nthe XID space whose clog files are being removed - and the XIDs that\nare in that range aren't safe to look up any more, but are also not\navailable for reuse to prevent wraparound. Right now, all of the\nrelevant fields in VariableCacheData are based on the ready-for-reuse\nconcept, and I don't think that switching some but not all of them to\nbe based on the safe-to-look-up concept necessarily qualifies as an\nimprovement. It's different, but I'm not sure it's better.\n\nWhat if we approached this problem from the other end? Suppose we use\na heavyweight lock on, say, transaction ID 1 to represent the right to\ntruncate CLOG. We grab this lock in exclusive mode before beginning\nto truncate, and in shared mode while looking up XIDs.\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, 10 Mar 2017 16:09:52 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 11 March 2017 at 05:09, Robert Haas <robertmhaas@gmail.com> wrote:\n\n> On the other\n> hand, there really are two separate notions of the \"oldest\" XID.\n> There's the oldest XID that we can safely look up, and then there's\n> the oldest XID that we can't reuse. These two are the same when no\n> truncation is in progress, but when a truncation is in progress then\n> they're different: the oldest XID that's safe to look up is the first\n> one after whatever we're truncating away, but the oldest XID that we\n> can't reuse is the newest one preceding the stuff that we're busy\n> truncating.\n\nRight.\n\nMy view here is that the oldest xid we cannot reuse is already guarded\nby xidWrapLimit, which we advance after clog truncation. Whether as\nthis advances at the same time as or after we advance oldestXid and\ntruncate clog doesn't actually matter, we must just ensure that it\nnever advances _before_.\n\nSo tracking a second copy of oldestXid whose only purpose is to\nrecalculate xidWrapLimit serves no real purpose. It's redundant except\nduring vac_truncate_clog, during which time local state is sufficient\n*if* we add oldestXid to the clog truncation xlog record, which we\nmust do anyway because:\n\nAny number of locking hoop-jumping schemes fail to solve the problem\nof outdated oldestXid information on standbys. Right now we truncate\nclog and xlog the truncation before we write the new oldestXid limit\nto xlog. In fact, we don't write the new xid limit to xlog until the\nnext checkpoint. So the standby has a huge window where its idea of\noldestXid is completely wrong, and unless we at least add the new\noldestXid to the clog truncation xlog record we can't fix that.\n\nWe only get away with this now because there's no way to look up an\narbitrary xid's status.\n\nNo locking scheme on the master can solve this, because the locks on\nthe master do not affect the standby or vice versa.\n\nTherefore, we _must_ advance oldestXid (or a copy of it used only for\n\"oldest xid still in clog) before truncating clog.\n\nIf we're going to do that we might as well just make sure the\nstandby's xid limits are updated correctly when we truncate clog\nrather than doing it lazily at checkpoints. Advance oldestXid before\ntruncating clog away, and record the new xid in the clog truncation\nxlog record. On redo after master crash, and on standbys, we're\nguaranteed to re-do the whole clog truncation operation - advance\noldestXid, truncate clog, advance xidWrapLimit etc - and everything\nstays consistent.\n\nI'll extract this part of the patch so it can be looked at separately,\nit'll be clearer that way.\n\nI think of it as slightly contracting then slightly expanding the xid\nrange window during clog truncation. Advance the oldest xid slightly\nbefore the xidWrapLimit, so temporarily the range of xids is narrower\nthan 2^31. xlog it first so we ensure it's all redone on crash and on\nstandby. Because no lock is held throughout all of vac_truncate_clog,\nmake sure the ordering of the different phases between concurrent\nvac_truncate_xlog runs doesn't matter.\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":"Sat, 11 Mar 2017 14:32:08 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 11 March 2017 at 14:32, Craig Ringer <craig@2ndquadrant.com> wrote:\n\n> I'll extract this part of the patch so it can be looked at separately,\n> it'll be clearer that way.\n\nApparently I thought that last time too since I already posted it\nsplit up. Ahem. Working on too many different things at once.\n\nLast-posted patches apply fine, no need for an update.\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":"Mon, 13 Mar 2017 11:22:48 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Sat, Mar 11, 2017 at 1:32 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> On 11 March 2017 at 05:09, Robert Haas <robertmhaas@gmail.com> wrote:\n>> On the other\n>> hand, there really are two separate notions of the \"oldest\" XID.\n>> There's the oldest XID that we can safely look up, and then there's\n>> the oldest XID that we can't reuse. These two are the same when no\n>> truncation is in progress, but when a truncation is in progress then\n>> they're different: the oldest XID that's safe to look up is the first\n>> one after whatever we're truncating away, but the oldest XID that we\n>> can't reuse is the newest one preceding the stuff that we're busy\n>> truncating.\n>\n> Right.\n>\n> My view here is that the oldest xid we cannot reuse is already guarded\n> by xidWrapLimit, which we advance after clog truncation. Whether as\n> this advances at the same time as or after we advance oldestXid and\n> truncate clog doesn't actually matter, we must just ensure that it\n> never advances _before_.\n>\n> So tracking a second copy of oldestXid whose only purpose is to\n> recalculate xidWrapLimit serves no real purpose.\n\nHmm, so what this patch is doing changed quite a bit between January\n23rd and January 25th. In the January 23rd version, oldestXid and\noldestXidDB are changed to track the oldest XID that we can safely\nlook up, and the remaining related fields are still relative to the\noldest XID that can be reused. That seems, as I said before, scary.\nBut in the January 25th version, *all* of the related fields have been\nchanged to track the oldest XID that we can safely look up, because\nSetTransactionIdLimit() now uses the values set by AdvanceOldestXid()\nto compute all of the other values, which seems flat-out incorrect.\nAdvanceOldestXid() is called to advance that limit before clog\ntruncation happens, and if somebody then calls SetTransactionIdLimit()\nbefore clog truncation is complete, we'll advanced those derived\nlimits prematurely.\n\nFor example, suppose vacuum #1 comes along, advances the limits,\ntruncates clog, and then gets descheduled. Now vacuum #2 comes along,\nadvances the limits further, and then gets descheduled. Now vacuum #1\nwakes up and calls SetTransactionIdLimit() and prematurely advances\nxidWrapLimit. Oops.\n\nThe way you put it is that having a second copy of oldestXid whose\nonly purpose is to recompute xidWrapLimit is pointless, but the way\nI'd say is that you're trying to make one variable do two jobs.\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, 13 Mar 2017 17:43:03 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 14 March 2017 at 05:43, Robert Haas <robertmhaas@gmail.com> wrote:\n\n> For example, suppose vacuum #1 comes along, advances the limits,\n> truncates clog, and then gets descheduled. Now vacuum #2 comes along,\n> advances the limits further, and then gets descheduled. Now vacuum #1\n> wakes up and calls SetTransactionIdLimit() and prematurely advances\n> xidWrapLimit. Oops.\n\nMm, right. And without a lock held from when oldestXid advances\nthrough to completion of clog truncation, then taking the same lock in\nSetTransactionIdLimit, there's not really a way around it.\n\nI'm embarrassed not to have seen that.\n\nDoing things the other way around, per the earlier patch, can cause\nSetTransactionIdLimit to not to advance as far as it should.\n\nOK, I'm convinced, a new field is safer, even if it's redundant most\nof the time.\n\nI'll introduce a new LWLock, ClogTruncationLock, which will be held\nfrom when we advance the new clogOldestXid field through to when clog\ntruncation completes.\n\nMost of the rest can stay the same. In particular, the expanded xlog\nrecord for clog truncation will still be required.\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":"Tue, 14 Mar 2017 10:22:59 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Mon, Mar 13, 2017 at 10:22 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> I'll introduce a new LWLock, ClogTruncationLock, which will be held\n> from when we advance the new clogOldestXid field through to when clog\n> truncation completes.\n\n+1.\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, 14 Mar 2017 07:57:08 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 14 March 2017 at 19:57, Robert Haas <robertmhaas@gmail.com> wrote:\n> On Mon, Mar 13, 2017 at 10:22 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n>> I'll introduce a new LWLock, ClogTruncationLock, which will be held\n>> from when we advance the new clogOldestXid field through to when clog\n>> truncation completes.\n>\n> +1.\n\nOK, here's the revised patch. Relevant comments added where needed.\n\nIt still changes the xlog record for clog truncation to carry the new\nxid, but I've left advancing oldestXid until the checkpoint as is\ncurrently the case, and only advanced oldestClogXid when we replay\nclog truncation.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\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":"Mon, 20 Mar 2017 13:38:18 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Mon, Mar 20, 2017 at 1:38 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> On 14 March 2017 at 19:57, Robert Haas <robertmhaas@gmail.com> wrote:\n>> On Mon, Mar 13, 2017 at 10:22 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n>>> I'll introduce a new LWLock, ClogTruncationLock, which will be held\n>>> from when we advance the new clogOldestXid field through to when clog\n>>> truncation completes.\n>>\n>> +1.\n>\n> OK, here's the revised patch. Relevant comments added where needed.\n>\n> It still changes the xlog record for clog truncation to carry the new\n> xid, but I've left advancing oldestXid until the checkpoint as is\n> currently the case, and only advanced oldestClogXid when we replay\n> clog truncation.\n\n/me smacks forehead. Actually, it should be CLogTruncationLock, with\na capital L, for consistency with CLogControlLock.\n\nThe new lock needs to be added to the table in monitoring.sgml.\n\nI don't think the new header comments in TransactionIdDidCommit and\nTransactionIdDidAbort are super-clear. I'm not sure you're going to\nbe able to explain it there in a reasonable number of words, but I\nthink that speaking of \"testing against oldestClogXid\" will leave\npeople wondering what exactly that means. Maybe just write \"caller is\nresponsible for ensuring that the clog records covering XID being\nlooked up can't be truncated away while the lookup is in progress\",\nand then leave the bit about CLogTruncationLock to be explained by the\ncallers that do that. Or you could drop these comments entirely.\n\nOverall, though, I think that 0001 looks far better than any previous\niteration. It's simple. It looks safe. It seems unlikely to break\nanything that works now. Woo hoo!\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, 21 Mar 2017 13:49:49 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 22 March 2017 at 01:49, Robert Haas <robertmhaas@gmail.com> wrote:\n\n> /me smacks forehead. Actually, it should be CLogTruncationLock, with\n> a capital L, for consistency with CLogControlLock.\n\nWill do.\n\n> The new lock needs to be added to the table in monitoring.sgml.\n\nSame.\n\n> I don't think the new header comments in TransactionIdDidCommit and\n> TransactionIdDidAbort are super-clear. I'm not sure you're going to\n> be able to explain it there in a reasonable number of words, but I\n> think that speaking of \"testing against oldestClogXid\" will leave\n> people wondering what exactly that means. Maybe just write \"caller is\n> responsible for ensuring that the clog records covering XID being\n> looked up can't be truncated away while the lookup is in progress\",\n> and then leave the bit about CLogTruncationLock to be explained by the\n> callers that do that. Or you could drop these comments entirely.\n\nOK. I'll revisit and see if I can clean it up, otherwise remove it.\n\n> Overall, though, I think that 0001 looks far better than any previous\n> iteration. It's simple. It looks safe. It seems unlikely to break\n> anything that works now. Woo hoo!\n\nFunny that this started with \"hey, here's a simple, non-invasive\nfunction for looking up the status of an arbitrary xid\".\n\nMature, complex systems eh?\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 Mar 2017 09:49:55 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 22 March 2017 at 09:49, Craig Ringer <craig@2ndquadrant.com> wrote:\n\n>> Overall, though, I think that 0001 looks far better than any previous\n>> iteration. It's simple. It looks safe. It seems unlikely to break\n>> anything that works now. Woo hoo!\n>\n> Funny that this started with \"hey, here's a simple, non-invasive\n> function for looking up the status of an arbitrary xid\".\n\nChanges made per discussion.\n\nRemoved the comments on TransactionIdDidCommit and\nTransactionIdDidAbort . It's not going to be relevant for the immense\nmajority of callers anyway, and callers that are looking up arbitrary\nuser supplied XIDs will (hopefully) be looking at\nTransactionIdInRecentPast anyway.\n\nI'll be leaving the 'xid' vs 'bigint' issues elsewhere in Pg for next\nrelease, nowhere near time for that now.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\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":"Wed, 22 Mar 2017 11:35:12 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 22 March 2017 at 03:35, Craig Ringer <craig@2ndquadrant.com> wrote:\n> On 22 March 2017 at 09:49, Craig Ringer <craig@2ndquadrant.com> wrote:\n>\n>>> Overall, though, I think that 0001 looks far better than any previous\n>>> iteration. It's simple. It looks safe. It seems unlikely to break\n>>> anything that works now. Woo hoo!\n>>\n>> Funny that this started with \"hey, here's a simple, non-invasive\n>> function for looking up the status of an arbitrary xid\".\n>\n> Changes made per discussion.\n\nThis looks good to me also. Does what we need it to do.\n\nI was a little worried by possible performance of new lock, but its\nnot intended to be run frequently, so its OK.\n\n-- \nSimon Riggs 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, 22 Mar 2017 07:13:32 +0000","msg_from":"Simon Riggs <simon@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Wed, Mar 22, 2017 at 3:13 AM, Simon Riggs <simon@2ndquadrant.com> wrote:\n>> Changes made per discussion.\n>\n> This looks good to me also. Does what we need it to do.\n>\n> I was a little worried by possible performance of new lock, but its\n> not intended to be run frequently, so its OK.\n\nAgreed.\n\nReviewing 0002:\n\n+ if (!TransactionIdIsValid(xid))\n+ {\n+ LWLockRelease(XidGenLock);\n+ ereport(ERROR,\n+ (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\n+ errmsg(\"transaction ID \" UINT64_FORMAT \" is an invalid xid\",\n+ xid_with_epoch)));\n+ }\n\nIt's unnecessary to release LWLockRelease() before throwing an error,\nand it's also wrong because we haven't acquired XidGenLock in this\ncode path. But maybe it would be better to just remove this entirely\nand instead have TransactionIdInRecentPast return false for\nInvalidTransactionId. Then we'd avoid adding a translatable message\nfor a case that is basically harmless to allow.\n\n+ if (TransactionIdIsCurrentTransactionId(xid))\n+ status = gettext_noop(\"in progress\");\n+ else if (TransactionIdDidCommit(xid))\n+ status = gettext_noop(\"committed\");\n+ else if (TransactionIdDidAbort(xid))\n+ status = gettext_noop(\"aborted\");\n+ else\n+\n+ /*\n+ * can't test TransactionIdIsInProgress here or we race with\n+ * concurrent commit/abort. There's no point anyway, since it\n+ * might then commit/abort just after we check.\n+ */\n+ status = gettext_noop(\"in progress\");\n\nI am not sure this is going to do the right thing for transactions\nthat are aborted by a crash without managing to write an abort record.\nIt seems that the first check will say the transaction isn't in\nprogress, and the second and third checks will say it isn't either\ncommitted or aborted since, if I am reading this correctly, they just\nread clog directly. Compare HeapTupleSatisfiesMVCC, which assumes\nthat a not-in-progress, not-committed transaction must necessarily\nhave aborted. I think your comment is pointing to a real issue,\nthough. It seems like what might be needed is to add one more check.\nBefore where you have the \"else\" clause, check if the XID is old, e.g.\nprecedes our snapshot's xmin. If so, it must be committed or aborted\nand, since it didn't commit, it aborted. If not, it must've changed\nfrom in progress to not-in-progress just as we were in the midst\nchecking, so labeling it as in progress is fine.\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 Mar 2017 13:41:47 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 22 March 2017 at 17:41, Robert Haas <robertmhaas@gmail.com> wrote:\n\n> + if (TransactionIdIsCurrentTransactionId(xid))\n> + status = gettext_noop(\"in progress\");\n> + else if (TransactionIdDidCommit(xid))\n> + status = gettext_noop(\"committed\");\n> + else if (TransactionIdDidAbort(xid))\n> + status = gettext_noop(\"aborted\");\n> + else\n> +\n> + /*\n> + * can't test TransactionIdIsInProgress here or we race with\n> + * concurrent commit/abort. There's no point anyway, since it\n> + * might then commit/abort just after we check.\n> + */\n> + status = gettext_noop(\"in progress\");\n>\n> I am not sure this is going to do the right thing for transactions\n> that are aborted by a crash without managing to write an abort record.\n\nYes, perhaps we should report that state as \"aborted - incomplete\".\n\nAnd of course, we might return \"subcommitted\" also, which could\ntechnically also be an abort in some cases, so we'd need to do a wait\nloop on that.\n\nWhich makes me think it would be confusing to say \"in progress\" for\nwhen it is our current xid, since the user might wait until it is\ncomplete and then wait forever. Prefer it if it said \"in progress -\ncurrent transaction\"\n\n-- \nSimon Riggs 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, 22 Mar 2017 18:08:55 +0000","msg_from":"Simon Riggs <simon@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Wed, Mar 22, 2017 at 2:08 PM, Simon Riggs <simon@2ndquadrant.com> wrote:\n> On 22 March 2017 at 17:41, Robert Haas <robertmhaas@gmail.com> wrote:\n>> + if (TransactionIdIsCurrentTransactionId(xid))\n>> + status = gettext_noop(\"in progress\");\n>> + else if (TransactionIdDidCommit(xid))\n>> + status = gettext_noop(\"committed\");\n>> + else if (TransactionIdDidAbort(xid))\n>> + status = gettext_noop(\"aborted\");\n>> + else\n>> +\n>> + /*\n>> + * can't test TransactionIdIsInProgress here or we race with\n>> + * concurrent commit/abort. There's no point anyway, since it\n>> + * might then commit/abort just after we check.\n>> + */\n>> + status = gettext_noop(\"in progress\");\n>>\n>> I am not sure this is going to do the right thing for transactions\n>> that are aborted by a crash without managing to write an abort record.\n>\n> Yes, perhaps we should report that state as \"aborted - incomplete\".\n>\n> And of course, we might return \"subcommitted\" also, which could\n> technically also be an abort in some cases, so we'd need to do a wait\n> loop on that.\n\nI actually don't think those are things we should expose to users.\nThey're internal implementation details. The user had better not care\nwhether an abort was the type of abort that wrote an abort record or\nthe type that didn't.\n\n> Which makes me think it would be confusing to say \"in progress\" for\n> when it is our current xid, since the user might wait until it is\n> complete and then wait forever. Prefer it if it said \"in progress -\n> current transaction\"\n\nHmm, or just \"current transaction\", maybe?\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 Mar 2017 14:10:28 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 23 March 2017 at 02:08, Simon Riggs <simon@2ndquadrant.com> wrote:\n\n> And of course, we might return \"subcommitted\" also, which could\n> technically also be an abort in some cases, so we'd need to do a wait\n> loop on that.\n\nUsers generally don't see subxact IDs, so it wasn't something I was\noverly concerned by. Most notably, txid_current() doesn't report\nthem.\n\nUsers who want to know the commit status of an xact that had a commit\nin-flight when they lost access to it due to server crash, network\nloss, etc aren't going to care about subxacts. If you lose your\nconnection after a RELEASE SAVEPOINT you know the outer xact will get\naborted or the state of the individual subxacts.\n\nThey're visible in heap tuples, but you can only see uncommitted heap\ntuples from your own top-level xid. For anything else, if you can see\nthe xid in xmin you know it committed. There isn't really any reason\nyou'd be looking up tuple xids with txid_status anyway, and if you did\nyou'd have to pay attention to mess like multixacts in xmax ... but\nyou can't tell from the xid alone if it's a multixact or not, so this\ndoesn't make sense with xids that could be multis anyway.\n\nIf we're going to handle subxacts specially, we should probably report\nthem as \"sub-committed\" if we find that the current xid is a committed\nsubxact member of an outer xact that is still in-progress. But IMO\nit's pretty pointless since you won't be dealing with subxact IDs in\nthe application anyway.\n\n> Which makes me think it would be confusing to say \"in progress\" for\n> when it is our current xid, since the user might wait until it is\n> complete and then wait forever. Prefer it if it said \"in progress -\n> current transaction\"\n\nI'm fine with \"current transaction\". Though I think it's kind of a\nmoot point as I don't see any reason for an application to ever be\ndoing the equivalent of\n\n txid_status(txid_current())\n\nin the first place. But it's harmless enough.\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":"Thu, 23 Mar 2017 09:40:49 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 23 March 2017 at 01:41, Robert Haas <robertmhaas@gmail.com> wrote:\n> On Wed, Mar 22, 2017 at 3:13 AM, Simon Riggs <simon@2ndquadrant.com> wrote:\n>>> Changes made per discussion.\n>>\n>> This looks good to me also. Does what we need it to do.\n>>\n>> I was a little worried by possible performance of new lock, but its\n>> not intended to be run frequently, so its OK.\n>\n> Agreed.\n>\n> Reviewing 0002:\n>\n> + if (!TransactionIdIsValid(xid))\n> + {\n> + LWLockRelease(XidGenLock);\n> + ereport(ERROR,\n> + (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\n> + errmsg(\"transaction ID \" UINT64_FORMAT \" is an invalid xid\",\n> + xid_with_epoch)));\n> + }\n>\n> It's unnecessary to release LWLockRelease() before throwing an error,\n> and it's also wrong because we haven't acquired XidGenLock in this\n> code path. But maybe it would be better to just remove this entirely\n> and instead have TransactionIdInRecentPast return false for\n> InvalidTransactionId. Then we'd avoid adding a translatable message\n> for a case that is basically harmless to allow.\n\nAgreed, that's better.\n\n> + if (TransactionIdIsCurrentTransactionId(xid))\n> + status = gettext_noop(\"in progress\");\n> + else if (TransactionIdDidCommit(xid))\n> + status = gettext_noop(\"committed\");\n> + else if (TransactionIdDidAbort(xid))\n> + status = gettext_noop(\"aborted\");\n> + else\n> +\n> + /*\n> + * can't test TransactionIdIsInProgress here or we race with\n> + * concurrent commit/abort. There's no point anyway, since it\n> + * might then commit/abort just after we check.\n> + */\n> + status = gettext_noop(\"in progress\");\n>\n> I am not sure this is going to do the right thing for transactions\n> that are aborted by a crash without managing to write an abort record.\n\nYou're right. It'll report in-progress, since the clog entry will\nstill be 0. We don't appear to explicitly update the clog during\nrecovery.\n\nFunny, I got so focused on the clog access safety stuff in the end\nthat I missed the bigger picture.\n\nGiven that part of the point of this is xact status after crash,\nthat's kind of important. I've added a TAP test for this, as part of a\nnew test set, \"011_crash_recovery.pl\". The recovery tests don't really\nhave much on crash behaviour at the moment so might as well start it\nand add more as we go. It doesn't make sense to add a whole new top\nlevel TAP test just for txid_status.\n\n(I *love* the TAP framework now, it took 10 mins to write a test\nthat's trivial to repeat in 5 seconds per run for this).\n\n> It seems that the first check will say the transaction isn't in\n> progress, and the second and third checks will say it isn't either\n> committed or aborted since, if I am reading this correctly, they just\n> read clog directly.\n\nRight. They're not concerned with subxacts or multixacts.\n\n> Compare HeapTupleSatisfiesMVCC, which assumes\n> that a not-in-progress, not-committed transaction must necessarily\n> have aborted.\n\nRight.\n\nWe don't have a HeapTuple here, of course, so we can't test flags.\nMost importantly this means we can't handle multixacts. If we ever did\ndecide to expose multixacts as bigint epoch-qualified xids for general\nusers, we'd probably reserve the high bit of the epoch the bigint xid\nrepresentation for the equivalent of HEAP_XMAX_IS_MULTI, and maybe\nit's worth reserving that bit now and ERROR'ing if we find it set. But\nright now we never produce a bigint xid with a multixact.\n\nI wonder if a note in the docs warning not to cast xid to bigint is\nworth it. Probably not. You have to cast xid::text::bigint which is\nkind of a hint, plus it doesn't make sense to test txid_status() for\ntuples' xmin/xmax . I guess you might use it with xids from\npg_stat_activity, but there's not much point when you can just look at\npg_stat_activity again to see if the proc of interest is still there\nand has the same xid.\n\nAnyway...\n\n> I think your comment is pointing to a real issue,\n> though. It seems like what might be needed is to add one more check.\n> Before where you have the \"else\" clause, check if the XID is old, e.g.\n> precedes our snapshot's xmin. If so, it must be committed or aborted\n> and, since it didn't commit, it aborted. If not, it must've changed\n> from in progress to not-in-progress just as we were in the midst\n> checking, so labeling it as in progress is fine.\n\nThat seems clear and sensible.\n\nXidInMVCCSnapshot simply tests TransactionIdPrecedes(xid,\nsnapshot->xmin), as does HeapTupleSatisfiesHistoricMVCC . It seems\nreasonable enough to just examine the snapshot xmin directly in\ntxid_status too; it's already done in replication/logical/snapbuild.c\nand lmgr/predicate.c.\n\nWe're running in an SQL-called function so we'll have a good snapshot\nand GetSnapshotData will have run, so GetActiveSnapshot()->xmin should\nbe sufficient.\n\nAmended patch attached, with added TAP test included.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, 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, 23 Mar 2017 11:25:07 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 23 March 2017 at 11:25, Craig Ringer <craig@2ndquadrant.com> wrote:\n\n> Amended patch attached, with added TAP test included.\n\nManaged to omit it, sigh.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, 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, 23 Mar 2017 12:22:23 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Tue, Mar 21, 2017 at 11:35 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> Changes made per discussion.\n\nCommitted 0001.\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, 23 Mar 2017 14:29:39 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 24 March 2017 at 02:29, Robert Haas <robertmhaas@gmail.com> wrote:\n> On Tue, Mar 21, 2017 at 11:35 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n>> Changes made per discussion.\n>\n> Committed 0001.\n\nMuch appreciated.\n\nHere's the 2nd patch rebased on top of master, with the TAP test\nincluded this time. Looks ready to go.\n\nI really appreciate the time you've taken to look at this. Point me at\nanything from your team you want some outside review on.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, 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":"Fri, 24 Mar 2017 14:27:44 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Fri, Mar 24, 2017 at 2:27 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> On 24 March 2017 at 02:29, Robert Haas <robertmhaas@gmail.com> wrote:\n>> On Tue, Mar 21, 2017 at 11:35 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n>>> Changes made per discussion.\n>>\n>> Committed 0001.\n>\n> Much appreciated.\n>\n> Here's the 2nd patch rebased on top of master, with the TAP test\n> included this time. Looks ready to go.\n\nCommitted.\n\n> I really appreciate the time you've taken to look at this. Point me at\n> anything from your team you want some outside review on.\n\nThanks, that is a valuable offer which I will be pleased to accept!\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 Mar 2017 12:28:24 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 3/24/17 02:27, Craig Ringer wrote:\n> On 24 March 2017 at 02:29, Robert Haas <robertmhaas@gmail.com> wrote:\n>> On Tue, Mar 21, 2017 at 11:35 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n>>> Changes made per discussion.\n>>\n>> Committed 0001.\n> \n> Much appreciated.\n> \n> Here's the 2nd patch rebased on top of master, with the TAP test\n> included this time. Looks ready to go.\n\nI'm experiencing hangs in the new t/011_crash_recovery.pl test. It\nseems to hang after the first call to SELECT txid_current();.\n\n-- \nPeter Eisentraut 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":"Fri, 24 Mar 2017 19:31:37 -0400","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 25 March 2017 at 07:31, Peter Eisentraut\n<peter.eisentraut@2ndquadrant.com> wrote:\n> On 3/24/17 02:27, Craig Ringer wrote:\n>> On 24 March 2017 at 02:29, Robert Haas <robertmhaas@gmail.com> wrote:\n>>> On Tue, Mar 21, 2017 at 11:35 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n>>>> Changes made per discussion.\n>>>\n>>> Committed 0001.\n>>\n>> Much appreciated.\n>>\n>> Here's the 2nd patch rebased on top of master, with the TAP test\n>> included this time. Looks ready to go.\n>\n> I'm experiencing hangs in the new t/011_crash_recovery.pl test. It\n> seems to hang after the first call to SELECT txid_current();.\n\nif you add\n\nnote \"txid is $xid\";\n\nafter\n\n+chomp($xid);\n\ndoes it report the xid?\n\nAlternately, can you see a 'psql' process and a backend doing a SELECT\nwhen it stops progressing?\n\nI'm wondering if this is a perl version/platform issue around\n\n $tx->pump until $stdout =~ /[[:digit:]]+[\\r\\n]$/;\n\nwhere we're not recognising the required output from psql when we get it.\n\nWhat's in src/test/recovery/tmp_check/log/regress_log_011* ?\n\nI couldn't use PostgresNode::psql or PostgresNode::safe_psql here\nbecause the session must stay open.\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":"Sat, 25 Mar 2017 08:08:01 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"> I'm wondering if this is a perl version/platform issue around\n> \n> $tx->pump until $stdout =~ /[[:digit:]]+[\\r\\n]$/;\n> \n> where we're not recognising the required output from psql when we get it.\n> \n> What's in src/test/recovery/tmp_check/log/regress_log_011* ?\n> \n> I couldn't use PostgresNode::psql or PostgresNode::safe_psql here\n> because the session must stay open.\n\nThe problem was that psql needs to be called with -X. Fix committed.\n\n-- \nPeter Eisentraut 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":"Sat, 25 Mar 2017 00:12:48 -0400","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 3/25/17 12:12 AM, Peter Eisentraut wrote:\n>> I'm wondering if this is a perl version/platform issue around\n>>\n>> $tx->pump until $stdout =~ /[[:digit:]]+[\\r\\n]$/;\n>>\n>> where we're not recognising the required output from psql when we get it.\n>>\n>> What's in src/test/recovery/tmp_check/log/regress_log_011* ?\n>>\n>> I couldn't use PostgresNode::psql or PostgresNode::safe_psql here\n>> because the session must stay open.\n>\n> The problem was that psql needs to be called with -X. Fix committed.\n\nIt's not clear to me what remains to be done on this patch. I feel, at \nthe least, that patches 3 and 4 need to be rebased and the status set \nback to \"Needs Review\".\n\nThis thread has been idle for six days. Please respond with a new patch \nby 2017-04-04 00:00 AoE (UTC-12) or this submission will be marked \n\"Returned with Feedback\".\n\n-- \n-David\ndavid@pgmasters.net\n\n\n-- \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, 31 Mar 2017 10:31:20 -0400","msg_from":"David Steele <david@pgmasters.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 31 Mar. 2017 22:31, \"David Steele\" <david@pgmasters.net> wrote:\n\nOn 3/25/17 12:12 AM, Peter Eisentraut wrote:\n\n> I'm wondering if this is a perl version/platform issue around\n>>\n>> $tx->pump until $stdout =~ /[[:digit:]]+[\\r\\n]$/;\n>>\n>> where we're not recognising the required output from psql when we get it.\n>>\n>> What's in src/test/recovery/tmp_check/log/regress_log_011* ?\n>>\n>> I couldn't use PostgresNode::psql or PostgresNode::safe_psql here\n>> because the session must stay open.\n>>\n>\n> The problem was that psql needs to be called with -X. Fix committed.\n>\n\nIt's not clear to me what remains to be done on this patch. I feel, at the\nleast, that patches 3 and 4 need to be rebased and the status set back to\n\"Needs Review\".\n\nThis thread has been idle for six days. Please respond with a new patch by\n2017-04-04 00:00 AoE (UTC-12) or this submission will be marked \"Returned\nwith Feedback\".\n\n\nI consider the feature complete and committed.\n\nPatch 3 is optional and per discussion buried upthread we generally agreed\nthat it'd be better to replace most user visible uses of the 'xid' data\ntype with a new epoch extended 'xid64' or similar.\n\nPatch 4, txid_incinerate, has never been intended for commit. It's a\ntesting tool.\n\nPatches 1 and 2 were the key parts and thanks to Robert's helpful review,\nadvice and edits they're committed now.\n\nCommitted, done. Yay.\n\nOn 31 Mar. 2017 22:31, \"David Steele\" <david@pgmasters.net> wrote:On 3/25/17 12:12 AM, Peter Eisentraut wrote:\n\nI'm wondering if this is a perl version/platform issue around\n\n $tx->pump until $stdout =~ /[[:digit:]]+[\\r\\n]$/;\n\nwhere we're not recognising the required output from psql when we get it.\n\nWhat's in src/test/recovery/tmp_check/log/regress_log_011* ?\n\nI couldn't use PostgresNode::psql or PostgresNode::safe_psql here\nbecause the session must stay open.\n\n\nThe problem was that psql needs to be called with -X. Fix committed.\n\n\nIt's not clear to me what remains to be done on this patch. I feel, at the least, that patches 3 and 4 need to be rebased and the status set back to \"Needs Review\".\n\nThis thread has been idle for six days. Please respond with a new patch by 2017-04-04 00:00 AoE (UTC-12) or this submission will be marked \"Returned with Feedback\".I consider the feature complete and committed.Patch 3 is optional and per discussion buried upthread we generally agreed that it'd be better to replace most user visible uses of the 'xid' data type with a new epoch extended 'xid64' or similar.Patch 4, txid_incinerate, has never been intended for commit. It's a testing tool.Patches 1 and 2 were the key parts and thanks to Robert's helpful review, advice and edits they're committed now.Committed, done. Yay.","msg_date":"Fri, 31 Mar 2017 22:46:33 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On 3/31/17 10:46 AM, Craig Ringer wrote:\n>\n> Patches 1 and 2 were the key parts and thanks to Robert's helpful\n> review, advice and edits they're committed now.\n>\n> Committed, done. Yay.\n\nExcellent. I have marked this a \"Committed\" by Robert.\n\nOne down...\n\n-- \n-David\ndavid@pgmasters.net\n\n\n-- \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, 31 Mar 2017 11:08:05 -0400","msg_from":"David Steele <david@pgmasters.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"},{"msg_contents":"On Fri, Mar 31, 2017 at 11:08 AM, David Steele <david@pgmasters.net> wrote:\n> On 3/31/17 10:46 AM, Craig Ringer wrote:\n>> Patches 1 and 2 were the key parts and thanks to Robert's helpful\n>> review, advice and edits they're committed now.\n>>\n>> Committed, done. Yay.\n>\n> Excellent. I have marked this a \"Committed\" by Robert.\n>\n> One down...\n\n...71 to go?\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, 31 Mar 2017 13:28:15 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Transaction traceability - txid_status(bigint)"}],"string":"[\n {\n \"msg_contents\": \"Hi all\\n\\nFollowing on from\\n\\nbigint txids vs 'xid' type, new txid_recent(bigint) => xid\\n\\nhttps://www.postgresql.org/message-id/CAMsr+YFDZMN_iZ7KrRoe+j0KVLQvFVgvZxbcVxR-MLjgtoZugA@mail.gmail.com\\n\\n\\nhere's a patch that implements a txid_status(bigint) function to report the\\ncommit status of a function.\\n\\nIf an application is disconnected while a COMMIT request is in flight,\\nthe backend crashes mid-commit, etc, then an application may not be\\nsure whether or not a commit completed successfully or was rolled\\nback. While two-phase commit solves this it does so at a considerable\\noverhead, so introduce a lighter alternative.\\n\\ntxid_status(bigint) lets an application determine the status of a\\na commit based on an xid-with-epoch as returned by txid_current()\\nor similar. Status may be committed, aborted, in-progress (including\\nprepared xacts) or null if the xact is too old for its commit status\\nto still be retained because it has passed the wrap-around epoch\\nboundary.\\n\\nApplications must call txid_current() in their transactions to make\\nmuch use of this since PostgreSQL does not automatically report an xid\\nto the client when one is assigned.\\n\\nA future protocol enhancement to report txid assignment would be very\\nuseful, but quite separate to this.\\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\": \"Sat, 20 Aug 2016 21:24:47 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"[PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 20 August 2016 at 21:24, Craig Ringer <craig@2ndquadrant.com> wrote:\\n\\n> Hi all\\n>\\n> Following on from\\n>\\n> bigint txids vs 'xid' type, new txid_recent(bigint) => xid\\n>\\n>\\nAhem. Forgot to squash in a fixup commit. Updated patch of\\ntxid_status(bigint) attachd.\\n\\nA related patch follows, adding a new txid_current_ifassigned(bigint)\\nfunction as suggested by Jim Nasby. It's usefully connected to\\ntxid_status() and might as well be added at the same time.\\n\\nSince it builds on the same history I've also attached an updated version\\nof txid_recent(bigint) now called txid_convert_ifrecent(bigint), per the\\nabove-linked thread.\\n\\nFinally, and not intended for commit, is a useful test function I wrote to\\ncause extremely rapid xid wraparound, bundled up into a src/test/regress\\ntest case. txid_incinerate() can jump the server about UINT32/2 xids in ~2\\nseconds if fsync is off, making it handy for testing. Posting so others\\ncan use it for their own test needs later and because it's useful for\\ntesting these patches that touch on the xid epoch.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\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\": \"Sat, 20 Aug 2016 21:46:55 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Sat, Aug 20, 2016 at 9:46 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> Ahem. Forgot to squash in a fixup commit. Updated patch of\\n> txid_status(bigint) attachd.\\n>\\n> A related patch follows, adding a new txid_current_ifassigned(bigint)\\n> function as suggested by Jim Nasby. It's usefully connected to txid_status()\\n> and might as well be added at the same time.\\n>\\n> Since it builds on the same history I've also attached an updated version of\\n> txid_recent(bigint) now called txid_convert_ifrecent(bigint), per the\\n> above-linked thread.\\n>\\n> Finally, and not intended for commit, is a useful test function I wrote to\\n> cause extremely rapid xid wraparound, bundled up into a src/test/regress\\n> test case. txid_incinerate() can jump the server about UINT32/2 xids in ~2\\n> seconds if fsync is off, making it handy for testing. Posting so others can\\n> use it for their own test needs later and because it's useful for testing\\n> these patches that touch on the xid epoch.\\n\\nI think you should use underscores to separate all of the words\\ninstead of only some of them.\\n\\nAlso, note that the corresponding internal function is\\nGetTopTransactionIdIfAny(), which might suggest txid_current_if_any()\\nrather than txid_current_if_assigned(), but you could argue that your\\nnaming is better...\\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, 22 Aug 2016 13:03:25 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 23 August 2016 at 01:03, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n\\n> I think you should use underscores to separate all of the words\\n> instead of only some of them.\\n>\\n\\nRight. Will fix.\\n\\nThanks for taking a look.\\n\\n\\n> Also, note that the corresponding internal function is\\n> GetTopTransactionIdIfAny(), which might suggest txid_current_if_any()\\n> rather than txid_current_if_assigned(), but you could argue that your\\n> naming is better.\\n\\n\\nYeah, I do argue that in this case. Not a hugely strong opinion, but we\\nrefer to txid_current() in the docs as:\\n\\n\\\"get current transaction ID, assigning a new one if the current transaction\\ndoes not have one\\\"\\n\\nso I thought it'd be worth being consistent with that. It's not like\\ntxid_current() mirrors the name of GetTopTransactionId() after all ;) and I\\nthink it's more important to be consistent with what the user sees than\\nwith the code.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 23 August 2016 at 01:03, Robert Haas <robertmhaas@gmail.com> wrote: I think you should use underscores to separate all of the words\\ninstead of only some of them.Right. Will fix.Thanks for taking a look. Also, note that the corresponding internal function is\\nGetTopTransactionIdIfAny(), which might suggest txid_current_if_any()\\nrather than txid_current_if_assigned(), but you could argue that your\\nnaming is better.Yeah, I do argue that in this case. Not a hugely strong opinion, but we refer to txid_current() in the docs as:\\\"get current transaction ID, assigning a new one if the current transaction does not have one\\\"so I thought it'd be worth being consistent with that. It's not like txid_current() mirrors the name of GetTopTransactionId() after all ;) and I think it's more important to be consistent with what the user sees than with the code.-- Craig Ringer http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Tue, 23 Aug 2016 08:31:26 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 23 August 2016 at 01:03, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n\\n>\\n> I think you should use underscores to separate all of the words\\n> instead of only some of them.\\n>\\n>\\nifassigned => if_assigned\\n\\nifrecent=> if_recent\\n\\nUpdated patch series attached. As before, 0-4 intended for commit, 5 just\\nbecause it'll be handy to have around for people doing wraparound related\\ntesting.\\n\\nAgain, thanks for taking a look.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\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\": \"Tue, 23 Aug 2016 08:55:00 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 23/08/16 02:55, Craig Ringer wrote:\\n> On 23 August 2016 at 01:03, Robert Haas <robertmhaas@gmail.com\\n> <mailto:robertmhaas@gmail.com>> wrote:\\n>\\n>\\n>\\n> I think you should use underscores to separate all of the words\\n> instead of only some of them.\\n>\\n>\\n> ifassigned => if_assigned\\n>\\n> ifrecent=> if_recent\\n>\\n> Updated patch series attached. As before, 0-4 intended for commit, 5\\n> just because it'll be handy to have around for people doing wraparound\\n> related testing.\\n\\nI guess you mean 0-3 for commit and 4 is just handy?\\n\\n From the point of code this patch seems good to me.\\n\\nI do wonder about the 3rd patch though. I wonder if it would not be \\nbetter to have the opposite function instead - converting xid to txid as \\nthat will always work and does not have to have the NULL case and would \\nbe simpler in terms of code.\\n\\n-- \\n Petr Jelinek 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, 23 Aug 2016 10:02:16 +0200\",\n \"msg_from\": \"Petr Jelinek <petr@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 23 Aug 2016 16:02, \\\"Petr Jelinek\\\" <petr@2ndquadrant.com> wrote:\\n>\\n> On 23/08/16 02:55, Craig Ringer wrote:\\n>>\\n>> On 23 August 2016 at 01:03, Robert Haas <robertmhaas@gmail.com\\n>> <mailto:robertmhaas@gmail.com>> wrote:\\n>>\\n>>\\n>>\\n>> I think you should use underscores to separate all of the words\\n>> instead of only some of them.\\n>>\\n>>\\n>> ifassigned => if_assigned\\n>>\\n>> ifrecent=> if_recent\\n>>\\n>> Updated patch series attached. As before, 0-4 intended for commit, 5\\n>> just because it'll be handy to have around for people doing wraparound\\n>> related testing.\\n>\\n>\\n> I guess you mean 0-3 for commit and 4 is just handy?\\n\\nEr. Right. 1-3. 4 just as handy test/tool.\\n\\n1 most important and useful. Then 2. Then 3.\\n\\n> From the point of code this patch seems good to me.\\n\\nThanks.\\n\\n> I do wonder about the 3rd patch though. I wonder if it would not be\\nbetter to have the opposite function instead - converting xid to txid as\\nthat will always work and does not have to have the NULL case and would be\\nsimpler in terms of code.\\n\\nYeah, but it wouldn't solve the need to take txid_current() output and do\\nstuff with it other than ordinal comparison. Like pass to commit ts\\nfunctions and others that take xid. If we extend all funcs that take xid to\\ntake bigint instead, they just get to use the same epoch logic in them,\\ncomplete with some way to deal with wrapped xids sensibly. It has to be\\ndone somewhere. Though it's prettier if hidden from the user.\\n\\nMore importantly imo, txid => bigint has to assume the current epoch. We\\nhave no way to make sure the user doesn't try to use something already\\nwrapped.\\n\\nI don't mind if everyone decides it's better to make xid go away and use\\nbigint everywhere user facing. Or even a new bigxid type. More work than I\\ncan really afford but can manage; shouldn't block #1 and #2 though as they\\nalready use bigint.\\n\\nOn 23 Aug 2016 16:02, \\\"Petr Jelinek\\\" <petr@2ndquadrant.com> wrote:\\n>\\n> On 23/08/16 02:55, Craig Ringer wrote:\\n>>\\n>> On 23 August 2016 at 01:03, Robert Haas <robertmhaas@gmail.com\\n>> <mailto:robertmhaas@gmail.com>> wrote:\\n>>\\n>>\\n>>\\n>> I think you should use underscores to separate all of the words\\n>> instead of only some of them.\\n>>\\n>>\\n>> ifassigned => if_assigned\\n>>\\n>> ifrecent=> if_recent\\n>>\\n>> Updated patch series attached. As before, 0-4 intended for commit, 5\\n>> just because it'll be handy to have around for people doing wraparound\\n>> related testing.\\n>\\n>\\n> I guess you mean 0-3 for commit and 4 is just handy?\\nEr. Right. 1-3. 4 just as handy test/tool.\\n1 most important and useful. Then 2. Then 3.\\n> From the point of code this patch seems good to me.\\nThanks.\\n> I do wonder about the 3rd patch though. I wonder if it would not be better to have the opposite function instead - converting xid to txid as that will always work and does not have to have the NULL case and would be simpler in terms of code.\\nYeah, but it wouldn't solve the need to take txid_current() output and do stuff with it other than ordinal comparison. Like pass to commit ts functions and others that take xid. If we extend all funcs that take xid to take bigint instead, they just get to use the same epoch logic in them, complete with some way to deal with wrapped xids sensibly. It has to be done somewhere. Though it's prettier if hidden from the user.\\nMore importantly imo, txid => bigint has to assume the current epoch. We have no way to make sure the user doesn't try to use something already wrapped.\\nI don't mind if everyone decides it's better to make xid go away and use bigint everywhere user facing. Or even a new bigxid type. More work than I can really afford but can manage; shouldn't block #1 and #2 though as they already use bigint.\",\n \"msg_date\": \"Tue, 23 Aug 2016 17:27:02 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 23/08/16 11:27, Craig Ringer wrote:\\n> On 23 Aug 2016 16:02, \\\"Petr Jelinek\\\" <petr@2ndquadrant.com\\n> <mailto:petr@2ndquadrant.com>> wrote:\\n>>\\n>> On 23/08/16 02:55, Craig Ringer wrote:\\n>>>\\n>>> On 23 August 2016 at 01:03, Robert Haas <robertmhaas@gmail.com\\n> <mailto:robertmhaas@gmail.com>\\n>>> <mailto:robertmhaas@gmail.com <mailto:robertmhaas@gmail.com>>> wrote:\\n>>>\\n>>>\\n>>>\\n>>> I think you should use underscores to separate all of the words\\n>>> instead of only some of them.\\n>>>\\n>>>\\n>>> ifassigned => if_assigned\\n>>>\\n>>> ifrecent=> if_recent\\n>>>\\n>>> Updated patch series attached. As before, 0-4 intended for commit, 5\\n>>> just because it'll be handy to have around for people doing wraparound\\n>>> related testing.\\n>>\\n>>\\n>> I guess you mean 0-3 for commit and 4 is just handy?\\n>\\n> Er. Right. 1-3. 4 just as handy test/tool.\\n>\\n> 1 most important and useful. Then 2. Then 3.\\n>\\n>> From the point of code this patch seems good to me.\\n>\\n> Thanks.\\n>\\n>> I do wonder about the 3rd patch though. I wonder if it would not be\\n> better to have the opposite function instead - converting xid to txid as\\n> that will always work and does not have to have the NULL case and would\\n> be simpler in terms of code.\\n>\\n> Yeah, but it wouldn't solve the need to take txid_current() output and\\n> do stuff with it other than ordinal comparison. Like pass to commit ts\\n> functions and others that take xid. If we extend all funcs that take xid\\n> to take bigint instead, they just get to use the same epoch logic in\\n> them, complete with some way to deal with wrapped xids sensibly. It has\\n> to be done somewhere. Though it's prettier if hidden from the user.\\n>\\n> More importantly imo, txid => bigint has to assume the current epoch. We\\n> have no way to make sure the user doesn't try to use something already\\n> wrapped.\\n>\\n\\nOkay, fair points.\\n\\n> I don't mind if everyone decides it's better to make xid go away and use\\n> bigint everywhere user facing. Or even a new bigxid type. More work than\\n> I can really afford but can manage; shouldn't block #1 and #2 though as\\n> they already use bigint.\\n>\\n\\nI don't think that would be very straightforward to be honest. I guess \\nfor what you want to achieve the approach chosen is the best one then.\\n\\n-- \\n Petr Jelinek 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, 23 Aug 2016 12:04:21 +0200\",\n \"msg_from\": \"Petr Jelinek <petr@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 22, 2016 at 8:55 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> Updated patch series attached. As before, 0-4 intended for commit, 5 just\\n> because it'll be handy to have around for people doing wraparound related\\n> testing.\\n>\\n> Again, thanks for taking a look.\\n\\n/me reviews a bit more deeply.\\n\\nIn 0001, it seems to me that \\\"in-progress\\\" should be \\\"in progress\\\". I\\ndon't think it's normal to hyphenate that. We have admittedly\\nsometimes done so, but:\\n\\n[rhaas pgsql]$ git grep 'in-progress' | wc -l\\n 63\\n[rhaas pgsql]$ git grep 'in progress' | wc -l\\n 346\\n\\nIt may make sense to speak of an \\\"in-progress transaction\\\" but I would\\nsay \\\"the transaction is in progress\\\" not \\\"the transaction is\\nin-progress\\\", which seems to me to argue for a space as the proper\\nseparator here.\\n\\nAlso:\\n\\n+CREATE TYPE txid_status AS ENUM ('committed', 'in-progress', 'aborted');\\n+\\n+CREATE FUNCTION\\n+ txid_status(txid bigint)\\n+RETURNS txid_status\\n+LANGUAGE sql\\n+VOLATILE PARALLEL SAFE\\n+AS $$\\n+SELECT CASE\\n+ WHEN s IS NULL THEN NULL::txid_status\\n+ WHEN s = -1 THEN 'aborted'::txid_status\\n+ WHEN s = 0 THEN 'in-progress'::txid_status\\n+ WHEN s = 1 THEN 'committed'::txid_status\\n+END\\n+FROM pg_catalog.txid_status_internal($1) s;\\n+$$;\\n+\\n+COMMENT ON FUNCTION txid_status(bigint)\\n+IS 'get commit status of given recent xid or null if too old';\\n\\nI'm not really that keen on this approach. I don't think we need to\\nintroduce a new data type for this, and I would rather not use SQL,\\neither. It would be faster and simpler to just return the appropriate\\nstring from a C function defined directly.\\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 Aug 2016 10:18:47 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 23, 2016 at 10:18 AM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Mon, Aug 22, 2016 at 8:55 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>> Updated patch series attached. As before, 0-4 intended for commit, 5 just\\n>> because it'll be handy to have around for people doing wraparound related\\n>> testing.\\n>>\\n>> Again, thanks for taking a look.\\n>\\n> /me reviews a bit more deeply.\\n>\\n> In 0001, ...\\n\\n0002 looks good, so I committed it. You forgot a function prototype\\nfor the new SQL-callable function, though, so I added that. For me,\\nit generates a compiler warning if that's missing; you might want to\\ntry to achieve a similar setup.\\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 Aug 2016 10:34:17 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 23 August 2016 at 22:18, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On Mon, Aug 22, 2016 at 8:55 PM, Craig Ringer <craig@2ndquadrant.com>\\n> wrote:\\n> > Updated patch series attached. As before, 0-4 intended for commit, 5 just\\n> > because it'll be handy to have around for people doing wraparound related\\n> > testing.\\n> >\\n> > Again, thanks for taking a look.\\n>\\n> /me reviews a bit more deeply.\\n>\\n> In 0001, it seems to me that \\\"in-progress\\\" should be \\\"in progress\\\".\\n>\\n\\nFine by me. I was on the fence about it anyway.\\n\\n+CREATE TYPE txid_status AS ENUM ('committed', 'in-progress', 'aborted');\\n>\\n\\n\\n> I'm not really that keen on this approach. I don't think we need to\\n> introduce a new data type for this, and I would rather not use SQL,\\n> either. It would be faster and simpler to just return the appropriate\\n> string from a C function defined directly.\\n>\\n\\nAlso fine by me. You're right, keep it simple. It means the potential set\\nof values isn't discoverable the same way, but ... meh. Using it usefully\\nmeans reading the docs anyway.\\n\\nThe remaining 2 patches of interest are attached - txid_status() and\\ntxid_convert_if_recent(). Thanks for committing txid_current_if_assigned().\\n\\nNow I'd best stop pretending I'm in a sensible timezone.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, 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\": \"Wed, 24 Aug 2016 00:59:35 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 23, 2016 at 12:59 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> Also fine by me. You're right, keep it simple. It means the potential set of\\n> values isn't discoverable the same way, but ... meh. Using it usefully means\\n> reading the docs anyway.\\n>\\n> The remaining 2 patches of interest are attached - txid_status() and\\n> txid_convert_if_recent(). Thanks for committing txid_current_if_assigned().\\n>\\n> Now I'd best stop pretending I'm in a sensible timezone.\\n\\nI reviewed this version some more and found some more problems.\\n\\n+ uint32 xid_epoch = (uint32)(xid_with_epoch >>32);\\n+ TransactionId xid = (TransactionId)(xid_with_epoch);\\n\\nI think this is not project style. In particular, I think that the\\nfirst one needs a space after the cast and another space before the\\n32; and I think the second one has an unnecessary set of parentheses\\nand needs a space added.\\n\\n+/*\\n+ * Underlying implementation of txid_status, which is mapped to an enum in\\n+ * system_views.sql.\\n+ */\\n\\nNot any more...\\n\\n+ errmsg(\\\"transaction ID \\\"UINT64_FORMAT\\\" is an invalid xid\\\",\\n+ xid_with_epoch)));\\n\\nSpacing.\\n\\n+ if (TransactionIdIsCurrentTransactionId(xid) ||\\nTransactionIdIsInProgress(xid))\\n+ status = gettext_noop(\\\"in progress\\\");\\n+ else if (TransactionIdDidCommit(xid))\\n+ status = gettext_noop(\\\"committed\\\");\\n+ else if (TransactionIdDidAbort(xid))\\n+ status = gettext_noop(\\\"aborted\\\");\\n+ else\\n+ /* shouldn't happen */\\n+ ereport(ERROR,\\n+ (errmsg_internal(\\\"unable to determine commit status of\\nxid \\\"UINT64_FORMAT, xid)));\\n\\nMaybe I'm all wet here, but it seems like there might be a problem\\nhere if the XID is older than the CLOG truncation point but less than\\none epoch old. get_xid_in_recent_past only guarantees that the\\ntransaction is less than one epoch old, not that we still have CLOG\\ndata for it. And there's nothing to keep NextXID from advancing under\\nus, so if somebody asks about a transaction that's just under 2^32\\ntransactions old, then get_xid_in_recent_past could say it's valid,\\nthen NextXID advances and we look up the XID extracted from the txid\\nand get the status of the new transaction instead of the old one!\\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 Aug 2016 15:10:37 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 24 August 2016 at 03:10, Robert Haas <robertmhaas@gmail.com> wrote:\\n>\\n> On Tue, Aug 23, 2016 at 12:59 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> > Also fine by me. You're right, keep it simple. It means the potential set of\\n> > values isn't discoverable the same way, but ... meh. Using it usefully means\\n> > reading the docs anyway.\\n> >\\n> > The remaining 2 patches of interest are attached - txid_status() and\\n> > txid_convert_if_recent(). Thanks for committing txid_current_if_assigned().\\n> >\\n> > Now I'd best stop pretending I'm in a sensible timezone.\\n>\\n> I reviewed this version some more and found some more problems.\\n\\n\\nThanks. It took me a few days to prep a new patch as I found another\\nissue in the process. Updated patch attached.\\n\\nThe updated series starts (0001) with a change to slru.c to release\\nthe control lock when throwing an exception so that we don't deadlock\\nwith ourselves when re-entering slru.c; explanation below.\\n\\nThen there's the txid_status (0002) patch with fixes, then\\ntxid_convert_if_recent(0003).\\n\\nI omitted txid_incinerate() ; I have an updated version that sets xact\\nstatus to aborted for burned xacts instead of leaving them at 0\\n(in-progress), but haven't had time to finish it so it doesn't try to\\nblindly set the status of xacts on pages where it didn't hold\\nXidGenLock when the page was added to the clog.\\n\\n> + uint32 xid_epoch = (uint32)(xid_with_epoch >>32);\\n> + TransactionId xid = (TransactionId)(xid_with_epoch);\\n>\\n> I think this is not project style. In particular, I think that the\\n> first one needs a space after the cast and another space before the\\n> 32; and I think the second one has an unnecessary set of parentheses\\n> and needs a space added.\\n\\n\\nOK, no problems. I didn't realise spacing around casts was specified.\\n\\n>\\n> +/*\\n> + * Underlying implementation of txid_status, which is mapped to an enum in\\n> + * system_views.sql.\\n> + */\\n>\\n> Not any more...\\n\\n\\nThat's why I shouldn't revise a patch at 1am ;)\\n\\n>\\n> + if (TransactionIdIsCurrentTransactionId(xid) ||\\n> TransactionIdIsInProgress(xid))\\n> + status = gettext_noop(\\\"in progress\\\");\\n> + else if (TransactionIdDidCommit(xid))\\n> + status = gettext_noop(\\\"committed\\\");\\n> + else if (TransactionIdDidAbort(xid))\\n> + status = gettext_noop(\\\"aborted\\\");\\n> + else\\n> + /* shouldn't happen */\\n> + ereport(ERROR,\\n> + (errmsg_internal(\\\"unable to determine commit status of\\n> xid \\\"UINT64_FORMAT, xid)));\\n>\\n> Maybe I'm all wet here, but it seems like there might be a problem\\n> here if the XID is older than the CLOG truncation point but less than\\n> one epoch old. get_xid_in_recent_past only guarantees that the\\n> transaction is less than one epoch old, not that we still have CLOG\\n> data for it.\\n\\n\\nGood point. The call would then fail with something like\\n\\n ERROR: could not access status of transaction 778793573\\n DETAIL: could not open file \\\"pg_clog/02E6\\\": No such file or directory\\n\\nThis probably didn't come up in my wraparound testing because I'm\\naggressively forcing wraparound by writing a lot of clog very quickly\\nunder XidGenLock, and because I'm mostly looking at xacts that are\\neither recent or past the xid boundary. I've added better add coverage\\nfor xacts around 2^30 behind the nextXid to the wraparound tests;\\ncan't add them to txid.sql since the xid never gets that far in normal\\nregression testing.\\n\\nWhat I'd really like is to be able to ask transam.c to handle the\\nxid_in_recent_past logic, treating an attempt to read an xid from\\nbeyond the clog truncation threshold as a soft error indicating\\nunknown xact state. But that involves delving into slru.c, and I\\nreally, really don't want to touch that for what should be a simple\\nand pretty noninvasive utility function.\\n\\nA PG_TRY to trap ERRCODE_UNDEFINED_FILE seems like it'd be sufficient,\\nexcept for two issues:\\n\\n* I see no accepted way to access the errcode etc from within the\\nPG_CATCH block, though. PG_RE_THROW() can do it because pg_re_throw()\\nis in elog.c. I couldn't find any existing code that seems to check\\ndetails about an error thrown in a PG_TRY block, only SPI calls. I\\ndon't want to ignore all types of errors and potentially hide\\nproblems, so I just used geterrcode() - which is meant for errcontext\\ncallbacks - and changed the comment to say it can be used in PG_CATCH\\ntoo. I don't see why it shouldn't be.\\nWe should probably have some sort of PG_CATCH_INFO(varname) that\\nexposes the top ErrorData, but that's not needed for this patch so I\\nleft it alone.\\n\\n* TransactionIdGetStatus() releases the CLogControlLock taken by\\nSimpleLruReadPage_ReadOnly() on normal exit but not after an exception\\nthrown from SlruReportIOError(). It seems appropriate for\\nSimpleLruReadPage() to release the LWLock before calling\\nSlruReportIOError(), so I've done that as a separate patch (now 0001).\\n\\n\\nI also removed the TransactionIdInProgress check in txid_status and\\njust assume it's in progress if it isn't our xid, committed or\\naborted. TransactionIdInProgress looks like it's potentially more\\nexpensive, and most of the time we'll be looking at committed or\\naborted xacts anyway. I can't sanity-check TransactionIdInProgress\\nafter commited/aborted, because there's then a race where the xact can\\ncommit or abort after we decide it's not committed/aborted so it's not\\nin progress when we go to check that.\\n\\n\\n\\n> And there's nothing to keep NextXID from advancing under\\n> us, so if somebody asks about a transaction that's just under 2^32\\n> transactions old, then get_xid_in_recent_past could say it's valid,\\n> then NextXID advances and we look up the XID extracted from the txid\\n> and get the status of the new transaction instead of the old one!\\n\\nHm, yeah. Though due to the clog truncation issue you noticed it\\nprobably won't happen.\\n\\nWe could require that XidGenLock be held at least as LW_SHARED when\\nentering get_xid_in_recent_past(), but I'd rather not since that'd be\\nan otherwise-unnecessary lwlock for txid_convert_ifrecent().\\n\\nInstead, I think I'll rename the wraparound flag to too_old and set it\\nif the xact is more than say 2^30 from the epoch struct's last_xid,\\nleaving a race window so ridiculously improbable that the nearly\\nimpossible chance of failing with a clog access error is not a worry.\\nIf the server's managing to have a proc stuck that long then it's\\nalready on fire. We're only interested in reasonably recent xacts\\nsince we can only work with xacts before wraparound / clog truncation.\\nThis just moves the threshold for \\\"reasonably recent\\\" a bit closer.\\n\\nAll this certainly reinforces my view that users handling 'xid'\\ndirectly or trying to extract it from a bigint epoch-extended xid is a\\nbad idea that needs to go away soon.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\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, 29 Aug 2016 11:25:39 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2016-08-29 11:25:39 +0800, Craig Ringer wrote:\\n> ERROR: could not access status of transaction 778793573\\n> DETAIL: could not open file \\\"pg_clog/02E6\\\": No such file or directory\\n> \\n> What I'd really like is to be able to ask transam.c to handle the\\n> xid_in_recent_past logic, treating an attempt to read an xid from\\n> beyond the clog truncation threshold as a soft error indicating\\n> unknown xact state. But that involves delving into slru.c, and I\\n> really, really don't want to touch that for what should be a simple\\n> and pretty noninvasive utility function.\\n\\nCan't you \\\"just\\\" check this against ShmemVariableCache->oldestXid while\\nholding appropriate locks?\\n\\n\\n> A PG_TRY to trap ERRCODE_UNDEFINED_FILE seems like it'd be sufficient,\\n> except for two issues:\\n\\nIt seems like a bad idea to PG_CATCH and not re-throw an error. That\\ngenerally is quite error prone. At the very least locking and such gets\\na lot more complicated (as you noticed below).\\n\\n> * TransactionIdGetStatus() releases the CLogControlLock taken by\\n> SimpleLruReadPage_ReadOnly() on normal exit but not after an exception\\n> thrown from SlruReportIOError(). It seems appropriate for\\n> SimpleLruReadPage() to release the LWLock before calling\\n> SlruReportIOError(), so I've done that as a separate patch (now 0001).\\n\\nWe normally prefer to handle this via the \\\"bulk\\\" releases in the error\\nhandlers. It's otherwise hard to write code that handles these\\nsituations reliably. It's different for spinlocks, but those normally\\nprotect far smaller regions of code.\\n\\n\\nAndres\\n\\n\\n-- \\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, 28 Aug 2016 20:45:14 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Sun, Aug 28, 2016 at 11:25 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> What I'd really like is to be able to ask transam.c to handle the\\n> xid_in_recent_past logic, treating an attempt to read an xid from\\n> beyond the clog truncation threshold as a soft error indicating\\n> unknown xact state. But that involves delving into slru.c, and I\\n> really, really don't want to touch that for what should be a simple\\n> and pretty noninvasive utility function.\\n\\nI think you're going to have to bite the bullet and do that, though, because ...\\n\\n> A PG_TRY to trap ERRCODE_UNDEFINED_FILE seems like it'd be sufficient,\\n\\n...I don't think this has any chance of being acceptable. You can't\\ncatch errors and not re-throw them. That's bad news.\\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, 29 Aug 2016 02:01:32 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 29 August 2016 at 11:45, Andres Freund <andres@anarazel.de> wrote:\\n> Hi,\\n>\\n> On 2016-08-29 11:25:39 +0800, Craig Ringer wrote:\\n>> ERROR: could not access status of transaction 778793573\\n>> DETAIL: could not open file \\\"pg_clog/02E6\\\": No such file or directory\\n>>\\n>> What I'd really like is to be able to ask transam.c to handle the\\n>> xid_in_recent_past logic, treating an attempt to read an xid from\\n>> beyond the clog truncation threshold as a soft error indicating\\n>> unknown xact state. But that involves delving into slru.c, and I\\n>> really, really don't want to touch that for what should be a simple\\n>> and pretty noninvasive utility function.\\n>\\n> Can't you \\\"just\\\" check this against ShmemVariableCache->oldestXid while\\n> holding appropriate locks?\\n\\nHm. Yeah, I should've thought of that. Thank you.\\n\\n>> A PG_TRY to trap ERRCODE_UNDEFINED_FILE seems like it'd be sufficient,\\n>> except for two issues:\\n>\\n> It seems like a bad idea to PG_CATCH and not re-throw an error. That\\n> generally is quite error prone. At the very least locking and such gets\\n> a lot more complicated (as you noticed below).\\n\\nYeah, and as I remember from the \\\"fun\\\" of trying to write apply errors\\nto tables in BDR. It wasn't my first choice.\\n\\n>> * TransactionIdGetStatus() releases the CLogControlLock taken by\\n>> SimpleLruReadPage_ReadOnly() on normal exit but not after an exception\\n>> thrown from SlruReportIOError(). It seems appropriate for\\n>> SimpleLruReadPage() to release the LWLock before calling\\n>> SlruReportIOError(), so I've done that as a separate patch (now 0001).\\n>\\n> We normally prefer to handle this via the \\\"bulk\\\" releases in the error\\n> handlers. It's otherwise hard to write code that handles these\\n> situations reliably. It's different for spinlocks, but those normally\\n> protect far smaller regions of code.\\n\\nFair enough. It's not a complex path, but there are a _lot_ of\\ncallers, and while I can't really imagine any of them relying on the\\nCLogControLock being held on error it's not something I was keen to\\nchange. I thought complicating the clog with a soft-error interface\\nwas worse and didn't come up with a better approach.\\n\\nSaid better approach attached in revised series. Thanks.\\n\\nMy only real complaint with doing this is that it's a bit more\\nconservative. But in practice clog truncation probably won't follow\\nthat far behind oldestXmin so except in fairly contrived circumstances\\nit won't hurt. Apps that need guarantees about how old an xid they can\\nget status on can hold down xmin with a replication slot, a dummy\\nprepared xact, or whatever. If we find that becomes a common need that\\nshould be made simpler then appropriate API to allow apps to hold down\\nclog truncation w/o blocking vacuuming can be added down the track.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\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, 29 Aug 2016 15:53:34 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 29 August 2016 at 15:53, Craig Ringer <craig@2ndquadrant.com> wrote:\\n\\n> Said better approach attached in revised series. Thanks.\\n\\nHere's another minor update to the txid_status() and\\ntxid_convert_if_recent() patches. The only change is moving\\nget_xid_in_recent_past from src/backend/utils/adt/txid.c to\\nsrc/backend/access/transam/xlog.c to permit its use by other code.\\nSpecifically, I think it'll be needed for logical decoding on standby.\\n\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\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, 1 Sep 2016 13:08:45 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 1 September 2016 at 13:08, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> On 29 August 2016 at 15:53, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>\\n>> Said better approach attached in revised series. Thanks.\\n>\\n> Here's another minor update to the txid_status() and\\n> txid_convert_if_recent() patches. The only change is moving\\n> get_xid_in_recent_past from src/backend/utils/adt/txid.c to\\n> src/backend/access/transam/xlog.c to permit its use by other code.\\n> Specifically, I think it'll be needed for logical decoding on standby.\\n\\nAhem, wrong patches. Attached correctly now.\\n\\n\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\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, 1 Sep 2016 13:29:50 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"Here's another update to these patches. While working on support for\\nlogical decoding on standbys I noticed that I needed something like\\nget_xid_in_recent_past(...) there too. So I've moved it to xlog.c as\\nTransactionIdInRecentPast too and flipped its arguments to be more\\nconvenient. No other changes.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\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, 2 Sep 2016 20:16:53 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 2 September 2016 at 13:16, Craig Ringer <craig@2ndquadrant.com> wrote:\\n\\n> So I've moved it to xlog.c...\\n\\nI'm pretty sure it shouldn't live in xlog.c, but there may be some\\ngood reason I can't see yet.\\n\\n-- \\nSimon Riggs 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\": \"Fri, 2 Sep 2016 13:30:42 +0100\",\n \"msg_from\": \"Simon Riggs <simon@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 2 Sep. 2016 8:30 pm, \\\"Simon Riggs\\\" <simon@2ndquadrant.com> wrote:\\n>\\n> On 2 September 2016 at 13:16, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>\\n> > So I've moved it to xlog.c...\\n>\\n> I'm pretty sure it shouldn't live in xlog.c, but there may be some\\n> good reason I can't see yet.\\n\\nUgh. Yes. transam.c would be rather saner.\\n\\nOnly for the helper to determine if an xid is recent though; txid_ status\\nstays in adt/xact.c where it belongs along with txid_current() etc.\\n\\n> --\\n> Simon Riggs http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\nOn 2 Sep. 2016 8:30 pm, \\\"Simon Riggs\\\" <simon@2ndquadrant.com> wrote:\\n>\\n> On 2 September 2016 at 13:16, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>\\n> > So I've moved it to xlog.c...\\n>\\n> I'm pretty sure it shouldn't live in xlog.c, but there may be some\\n> good reason I can't see yet.\\n Ugh. Yes. transam.c would be rather saner.\\nOnly for the helper to determine if an xid is recent though; txid_ status stays in adt/xact.c where it belongs along with txid_current() etc.\\n> --\\n> Simon Riggs http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Fri, 2 Sep 2016 20:38:39 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 2 September 2016 at 20:38, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> On 2 Sep. 2016 8:30 pm, \\\"Simon Riggs\\\" <simon@2ndquadrant.com> wrote:\\n>>\\n>> On 2 September 2016 at 13:16, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>>\\n>> > So I've moved it to xlog.c...\\n>>\\n>> I'm pretty sure it shouldn't live in xlog.c, but there may be some\\n>> good reason I can't see yet.\\n>\\n> Ugh. Yes. transam.c would be rather saner.\\n>\\n> Only for the helper to determine if an xid is recent though; txid_ status\\n> stays in adt/xact.c where it belongs along with txid_current() etc.\\n\\nFixed, moved to transam.c.\\n\\n\\n\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\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, 2 Sep 2016 21:01:58 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 2 September 2016 at 21:01, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> On 2 September 2016 at 20:38, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>> On 2 Sep. 2016 8:30 pm, \\\"Simon Riggs\\\" <simon@2ndquadrant.com> wrote:\\n>>>\\n>>> On 2 September 2016 at 13:16, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>>>\\n>>> > So I've moved it to xlog.c...\\n>>>\\n>>> I'm pretty sure it shouldn't live in xlog.c, but there may be some\\n>>> good reason I can't see yet.\\n>>\\n>> Ugh. Yes. transam.c would be rather saner.\\n>>\\n>> Only for the helper to determine if an xid is recent though; txid_ status\\n>> stays in adt/xact.c where it belongs along with txid_current() etc.\\n>\\n> Fixed, moved to transam.c.\\n\\nMissed that this causes an undefined reference to GetNextXidAndEpoch()\\nwhich is in xlog.c. I knew there was a reason I put it there. So most\\nrecent patch is wrong, use the prior one.\\n\\nGetNextXidAndEpoch() needs to be in xlog.c because it uses XLogCtl's\\nshmem copy of checkPoint.nextXidEpoch. So either transam.c needs to\\n#include xlog.h (which seems a bit backwards) or\\nTransactionIdInRecentPast() should go in xlog.c.\\n\\nI don't like either really. Opinion? I'm sure we'll want this\\nfunctionality in other places as part of dealing with the problems\\ndiscussed upthread with 'xid' exposed to users.\\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\": \"Fri, 2 Sep 2016 21:46:19 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 02/09/16 15:46, Craig Ringer wrote:\\n> On 2 September 2016 at 21:01, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>> On 2 September 2016 at 20:38, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>>> On 2 Sep. 2016 8:30 pm, \\\"Simon Riggs\\\" <simon@2ndquadrant.com> wrote:\\n>>>>\\n>>>> On 2 September 2016 at 13:16, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>>>>\\n>>>>> So I've moved it to xlog.c...\\n>>>>\\n>>>> I'm pretty sure it shouldn't live in xlog.c, but there may be some\\n>>>> good reason I can't see yet.\\n>>>\\n>>> Ugh. Yes. transam.c would be rather saner.\\n>>>\\n>>> Only for the helper to determine if an xid is recent though; txid_ status\\n>>> stays in adt/xact.c where it belongs along with txid_current() etc.\\n>>\\n>> Fixed, moved to transam.c.\\n>\\n> Missed that this causes an undefined reference to GetNextXidAndEpoch()\\n> which is in xlog.c. I knew there was a reason I put it there. So most\\n> recent patch is wrong, use the prior one.\\n>\\n> GetNextXidAndEpoch() needs to be in xlog.c because it uses XLogCtl's\\n> shmem copy of checkPoint.nextXidEpoch. So either transam.c needs to\\n> #include xlog.h (which seems a bit backwards) or\\n> TransactionIdInRecentPast() should go in xlog.c.\\n>\\n> I don't like either really. Opinion? I'm sure we'll want this\\n> functionality in other places as part of dealing with the problems\\n> discussed upthread with 'xid' exposed to users.\\n>\\n\\nYou could put it to txid.c where all the other txid stuff is in?\\n\\n-- \\n Petr Jelinek 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, 2 Sep 2016 17:29:34 +0200\",\n \"msg_from\": \"Petr Jelinek <petr@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 2 September 2016 at 23:29, Petr Jelinek <petr@2ndquadrant.com> wrote:\\n\\n> You could put it to txid.c where all the other txid stuff is in?\\n\\nYeah, even though it's in adt/ I think it'll do.\\n\\nI thought I'd need get_xid_in_recent_past() for catalog_xmin hot\\nstandby feedback, but upon closer examination the needed logic isn't\\nthe same anymore. txid_status() wants to ensure clog lookups are safe\\nand limit by oldest xid, wheras the walsender doesn't actually care\\nabout that and is just avoiding wrapped xids.\\n\\nI'm just going back to how it was, all in adt/txid.c, and making it\\nstatic again. We can move it and make it non-static if a need to do so\\ncomes up.\\n\\nAttached rebased patch updated and vs master.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\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 Sep 2016 08:52:17 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Thu, Sep 15, 2016 at 8:52 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> On 2 September 2016 at 23:29, Petr Jelinek <petr@2ndquadrant.com> wrote:\\n>\\n>> You could put it to txid.c where all the other txid stuff is in?\\n>\\n> Yeah, even though it's in adt/ I think it'll do.\\n>\\n> I thought I'd need get_xid_in_recent_past() for catalog_xmin hot\\n> standby feedback, but upon closer examination the needed logic isn't\\n> the same anymore. txid_status() wants to ensure clog lookups are safe\\n> and limit by oldest xid, wheras the walsender doesn't actually care\\n> about that and is just avoiding wrapped xids.\\n>\\n> I'm just going back to how it was, all in adt/txid.c, and making it\\n> static again. We can move it and make it non-static if a need to do so\\n> comes up.\\n>\\n> Attached rebased patch updated and vs master.\\n\\nYou appear to have attached the wrong patch set.\\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, 16 Sep 2016 09:28:54 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 16 September 2016 at 21:28, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Thu, Sep 15, 2016 at 8:52 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>> On 2 September 2016 at 23:29, Petr Jelinek <petr@2ndquadrant.com> wrote:\\n>>\\n>>> You could put it to txid.c where all the other txid stuff is in?\\n>>\\n>> Yeah, even though it's in adt/ I think it'll do.\\n>>\\n>> I thought I'd need get_xid_in_recent_past() for catalog_xmin hot\\n>> standby feedback, but upon closer examination the needed logic isn't\\n>> the same anymore. txid_status() wants to ensure clog lookups are safe\\n>> and limit by oldest xid, wheras the walsender doesn't actually care\\n>> about that and is just avoiding wrapped xids.\\n>>\\n>> I'm just going back to how it was, all in adt/txid.c, and making it\\n>> static again. We can move it and make it non-static if a need to do so\\n>> comes up.\\n>>\\n>> Attached rebased patch updated and vs master.\\n>\\n> You appear to have attached the wrong patch set.\\n\\nWhoops, multitasking fail.\\n\\nSorry for the late response, hospitals are \\\"fun\\\".\\n\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\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, 20 Sep 2016 09:54:28 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Mon, Sep 19, 2016 at 9:54 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>> You appear to have attached the wrong patch set.\\n>\\n> Whoops, multitasking fail.\\n>\\n> Sorry for the late response, hospitals are \\\"fun\\\".\\n\\nI did some cleanup of 0001 (see attached) and was all set to commit it\\nwhen I realized what I think is a problem: holding XidGenLock doesn't\\nseem to help with the race condition between this function and CLOG\\ntruncation, because vac_truncate_clog() updates the shared memory\\nlimits AFTER performing the truncation. If the order of those\\noperations were reversed, we'd be fine, because it would get stuck\\ntrying to update the shared memory limits and wouldn't be able to\\ntruncate until it did - and conversely, if it updated the shared\\nmemory limits before we examined them, that would be OK, too, because\\nwe'd be sure not to consult the pages that are about to be truncated.\\nAs it is, though, I don't see that there's any real interlock here.\\n\\n(BTW, the stuff you moved from clog.c to clog.h doesn't actually need\\nto be moved; one of the changes I made here was to undo that.)\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\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, 20 Sep 2016 10:46:39 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 20 September 2016 at 22:46, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> I did some cleanup of 0001 (see attached) and was all set to commit it\\n> when I realized what I think is a problem: holding XidGenLock doesn't\\n> seem to help with the race condition between this function and CLOG\\n> truncation, because vac_truncate_clog() updates the shared memory\\n> limits AFTER performing the truncation.\\n\\nThanks ... and drat.\\n\\n> If the order of those\\n> operations were reversed, we'd be fine, because it would get stuck\\n> trying to update the shared memory limits and wouldn't be able to\\n> truncate until it did - and conversely, if it updated the shared\\n> memory limits before we examined them, that would be OK, too, because\\n> we'd be sure not to consult the pages that are about to be truncated.\\n\\nI'm hesitant to mess with something that fundamental for what I was\\nhoping was a low-impact feature, albeit one that seems to be trying\\nhard not to be at every turn. It looks pretty reasonable to update\\noldestXid before clog truncation but I don't want to be wrong and\\ncreate some obscure race or crash recovery issue related to wraparound\\nand clog truncation. It could well be a problem if we're very close to\\nwraparound.\\n\\nSo far nothing has had any reason to care about this, since there's no\\nway to attempt to access an older xid in a normally functioning\\nsystem. The commit timestamp lookup code doesn't care whether the xid\\nis still in clog or not and nothing else does lookups based on xids\\nsupplied by the user. If anything else did or does in future it will\\nhave the same problem as txid_status.\\n\\nWe've already ruled out releasing the slru LWLock in SlruReportIOError\\nthen PG_TRY / PG_CATCH()ing clog access errors in txid_status() per my\\noriginal approach to this issue.\\n\\nSo I see a few options now:\\n\\n* Do nothing. Permit this race to exist, document the error, and\\nexpect apps to cope. I'm pretty tempted to go for exactly this since\\nit pushes the cost onto users of the feature and doesn't make a mess\\nelsewhere. People who use this will typically be invoking it as a\\nstandalone toplevel function anyway, so it's mostly just a bit of\\nnoise in the error logs if you lose the race - and we have plenty of\\nother sources of that already.\\n\\n* Rather than calling TransactionIdDidCommit / TransactionIdDidAbort,\\ncall clog.c's TransactionIdGetStatus with a new missing_ok flag. That\\nsets a bool* missing param added to SimpleLruReadPage_ReadOnly(...)\\nand in turn to SimpleLruReadPage(...) that's set instead of calling\\nSlruReportIOError(...). This seems rather intrusive and will add\\nlittle-used params and paths to fairly hot slru.c code so I'm not\\nkeen.\\n\\n* Take CLogControlLock LW_SHARED in txid_status() to prevent\\ntruncation before reading oldestXid. We'd need a way to pass an\\n\\\"already locked\\\" state through TransactionIdGetStatus(...) to\\nSimpleLruReadPage_ReadOnly(...), which isn't great since again it's\\npretty hot code.\\n\\n* Don't release the slru LWLock in SlruReportIOError; instead release\\nCLogControlLock from txid_status on clog access failure. As before\\nmeans PG_TRY / PG_CATCH without PG_RE_THROW, but it means the only\\nplace it affects is callers of txid_status(...). For added safety,\\nrestrict txid_status() to being called in a toplevel virtual xact so\\nwe know we'll finish up promptly. It's a horrible layering violation\\nhaving txid_status(...) release the clog control lock though, and\\nseems risky.\\n\\nThe only non-horrible one of those is IMO to just let the caller see\\nan error if they lose the race. It's a function that's intended for\\nuse when you're dealing with indeterminate transaction state after a\\nserver or application error anyway, so it's part of an error path\\nalready. So I say we just document the behaviour. Because slru.c\\ndoesn't release its LWLock on error we also need to ensure\\ntxid_status(...) is also only called from a toplevel xact so the user\\ndoesn't attempt to wrap it in plpgsql BEGIN ... EXCEPTION block and it\\ncauses the xact to abort.\\n\\nWill follow up with just that.\\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, 21 Sep 2016 15:40:34 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Wed, Sep 21, 2016 at 3:40 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> The only non-horrible one of those is IMO to just let the caller see\\n> an error if they lose the race. It's a function that's intended for\\n> use when you're dealing with indeterminate transaction state after a\\n> server or application error anyway, so it's part of an error path\\n> already. So I say we just document the behaviour.\\n\\nI am not keen on that idea. The errors we're likely to be exposing\\nare going to look like low-level internal failures, which might scare\\nsome DBA. Even if they don't, I think it's playing with fire. The\\nsystem is designed on the assumption that nobody will try to look up\\nan XID that's too old, and if we start to violate that assumption I\\nthink we're undermining the design integrity of the system in a way\\nwe'll likely come to regret. To put that more plainly, when the code\\nis written with the assumption that X will never happen, it's usually\\na bad idea to casually add code that does X.\\n\\n> Because slru.c\\n> doesn't release its LWLock on error we also need to ensure\\n> txid_status(...) is also only called from a toplevel xact so the user\\n> doesn't attempt to wrap it in plpgsql BEGIN ... EXCEPTION block and it\\n> causes the xact to abort.\\n\\nI think this is muddled, because an error aborts the transaction, and\\nAbortTransaction() and AbortSubTransaction() start with\\nLWLockReleaseAll().\\n\\nIt might not be too hard to add a second copy of oldestXid in shared\\nmemory that is updated before truncation rather than afterward... but\\nyeah, like you, I'm not finding this nearly as straightforward as\\nmight have been hoped.\\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, 21 Sep 2016 10:16:49 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 21 September 2016 at 22:16, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Wed, Sep 21, 2016 at 3:40 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>> The only non-horrible one of those is IMO to just let the caller see\\n>> an error if they lose the race. It's a function that's intended for\\n>> use when you're dealing with indeterminate transaction state after a\\n>> server or application error anyway, so it's part of an error path\\n>> already. So I say we just document the behaviour.\\n>\\n> I am not keen on that idea. The errors we're likely to be exposing\\n> are going to look like low-level internal failures, which might scare\\n> some DBA. Even if they don't, I think it's playing with fire. The\\n> system is designed on the assumption that nobody will try to look up\\n> an XID that's too old, and if we start to violate that assumption I\\n> think we're undermining the design integrity of the system in a way\\n> we'll likely come to regret. To put that more plainly, when the code\\n> is written with the assumption that X will never happen, it's usually\\n> a bad idea to casually add code that does X.\\n\\nFair point.\\n\\n> [snip]\\n>\\n> It might not be too hard to add a second copy of oldestXid in shared\\n> memory that is updated before truncation rather than afterward... but\\n> yeah, like you, I'm not finding this nearly as straightforward as\\n> might have been hoped.\\n\\nYeah.\\n\\nI suspect that'll be the way to go, to add another copy that's updated\\nbefore clog truncation. It just seems ... unclean. Like it shouldn't\\nbe necessary, something else isn't right. But it's probably the lowest\\npain option.\\n\\nI'm going to take a step back on this and see if I can spot an alternative.\\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\": \"Tue, 27 Sep 2016 09:23:21 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 27 September 2016 at 09:23, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> On 21 September 2016 at 22:16, Robert Haas <robertmhaas@gmail.com> wrote:\\n>>\\n>> It might not be too hard to add a second copy of oldestXid in shared\\n>> memory that is updated before truncation rather than afterward... but\\n>> yeah, like you, I'm not finding this nearly as straightforward as\\n>> might have been hoped.\\n>\\n> Yeah.\\n>\\n> I suspect that'll be the way to go, to add another copy that's updated\\n> before clog truncation. It just seems ... unclean. Like it shouldn't\\n> be necessary, something else isn't right. But it's probably the lowest\\n> pain option.\\n\\nOK, so summarizing the problem:\\n\\nslru.c and clog.c have no soft-failure entrypoints where we can look\\nit up and fail gracefully if the xid isn't in the clog.\\nvac_truncate_clog() calls TruncateCLOG to chop SLRU pages from the\\nclog before it takes XidGenLock to advance oldestXid. So we cannot use\\noldestXid protected by XidGenLock as an interlock against concurrent\\nclog truncation to prevent SLRU access errors looking up an xid, and\\nthere's no other candidate lock.\\n\\nThis hasn't been an issue before because nobody looks up arbitrary\\nuser-supplied XIDs in clog, we only look up things that're protected\\nby datfrozenxid etc. But the whole point of txid_status is to look up\\nuser-supplied xids.\\n\\nWe can't just take ClogControlLock from txid_status() to block\\ntruncation because clog.c expects to own that lock, and takes it (via\\nslru.c) in TransactionIdGetStatus, with no way to pass an\\nalready_locked state. We'd self-deadlock.\\n\\nAdding a second copy of oldestXid - say pendingOldestXid - won't\\nactually help us unless we also take some suitable LWLock before\\nupdating it, otherwise truncation can continue after we look at\\npendingOldestXid but before we do the clog lookup. That means an extra\\nLWLock for each clog truncation, but compared to the I/O done during\\nclog truncation and the cost of the SlruScanDirectory() pre-check done\\nby TruncateCLOG it's nothing.\\n\\nWe could take XidGenLock twice, once to update this new\\npendingOldestXid field and once to update oldestXid, but that's a\\nhighly contested lock I'd rather not mess with even on a path that's\\nnot hit much.\\n\\nInstead, I've added a new LWLock, ClogTruncationLock, for that\\npurpose. vac_truncate_clog() takes it if it decides to attempt clog\\ntruncation. This lock is held throughout the whole process of clog\\ntruncation and oldestXid advance, so there's no need for a new\\npendingOldestXid field in ShmemVariableCache. We just take the lock\\nthen look at oldestXid, knowing that it's guaranteed to correspond to\\nan existing clog page that won't go away while we're looking.\\nClogTruncationLock is utterly uncontested so it's going to have no\\nmeaningful impact compared to all the work we do scanning the clog to\\ndecide whether we're even going to try truncating it, etc.\\n\\n\\n(BTW, it seems like a pity that lwlocknames.txt doesn't have comments\\non each lock. We could have\\nsrc/backend/storage/lmgr/generate-lwlocknames.pl transform the #\\ncomments into /* comments */ on the generated header. Thoughts?)\\n\\n\\n\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, 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\": \"Wed, 21 Dec 2016 16:02:15 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 21, 2016 at 3:02 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> Instead, I've added a new LWLock, ClogTruncationLock, for that\\n> purpose. vac_truncate_clog() takes it if it decides to attempt clog\\n> truncation. This lock is held throughout the whole process of clog\\n> truncation and oldestXid advance, so there's no need for a new\\n> pendingOldestXid field in ShmemVariableCache. We just take the lock\\n> then look at oldestXid, knowing that it's guaranteed to correspond to\\n> an existing clog page that won't go away while we're looking.\\n> ClogTruncationLock is utterly uncontested so it's going to have no\\n> meaningful impact compared to all the work we do scanning the clog to\\n> decide whether we're even going to try truncating it, etc.\\n\\nThat makes everything that happens between when we acquire that lock\\nand when we release it non-interruptible, which seems undesirable. I\\nthink that extra copy of oldestXid is a nicer approach.\\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, 21 Dec 2016 11:30:44 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 22 December 2016 at 00:30, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> That makes everything that happens between when we acquire that lock\\n> and when we release it non-interruptible, which seems undesirable. I\\n> think that extra copy of oldestXid is a nicer approach.\\n\\nThat's a side-effect I didn't realise. Given that, yes, I agree.\\n\\nSince we don't truncate clog much, do you think it's reasonable to\\njust take XidGenLock again before we proceed? I'm reluctant to add\\nanother acquisition of a frequently contested lock for something 99.9%\\nof the codebase won't care about, so I think it's probably better to\\nadd a new LWLock, and I'll resubmit on that basis, but figure it's\\nworth asking.\\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\": \"Thu, 22 Dec 2016 07:49:40 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 22 December 2016 at 07:49, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> On 22 December 2016 at 00:30, Robert Haas <robertmhaas@gmail.com> wrote:\\n>\\n>> That makes everything that happens between when we acquire that lock\\n>> and when we release it non-interruptible, which seems undesirable. I\\n>> think that extra copy of oldestXid is a nicer approach.\\n>\\n> That's a side-effect I didn't realise. Given that, yes, I agree.\\n>\\n> Since we don't truncate clog much, do you think it's reasonable to\\n> just take XidGenLock again before we proceed? I'm reluctant to add\\n> another acquisition of a frequently contested lock for something 99.9%\\n> of the codebase won't care about, so I think it's probably better to\\n> add a new LWLock, and I'll resubmit on that basis, but figure it's\\n> worth asking.\\n\\nUpdated.\\n\\nIf you think it's better to just take XidGenLock again, let me know.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, 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, 22 Dec 2016 09:55:06 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 22 December 2016 at 09:55, Craig Ringer <craig@2ndquadrant.com> wrote:\\n\\n> Updated.\\n>\\n> If you think it's better to just take XidGenLock again, let me know.\\n\\nHere's a further update that merges in Robert's changes from the patch\\nposted upthread.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, 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, 22 Dec 2016 13:12:10 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 22, 2016 at 12:12 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> On 22 December 2016 at 09:55, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>\\n>> Updated.\\n>>\\n>> If you think it's better to just take XidGenLock again, let me know.\\n>\\n> Here's a further update that merges in Robert's changes from the patch\\n> posted upthread.\\n\\nThis patch contains unresolved merge conflicts. Maybe\\nSetPendingTransactionIdLimit could happen in TruncateCLOG rather than\\nthe caller. Instead of using an LWLock, how about a spin lock?\\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, 22 Dec 2016 12:26:22 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 23 December 2016 at 01:26, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> This patch contains unresolved merge conflicts.\\n\\nAh, it conflicts with fe0a0b59, the PostmasterRandom changes. I\\nthought I'd rebased more recently than that.\\n\\n> Maybe SetPendingTransactionIdLimit could happen in TruncateCLOG rather than\\n> the caller.\\n\\nI've spent a while looking at how committs and multixact handle the\\nrace hazard of the gap between clog truncation and shmem state change\\n- and the related race on standby, where the hazard is the gap between\\nreplay of CLOG_TRUNCATE records and replay of the next checkpoint\\ncontaining updated limits.\\n\\npg_xact_commit_timestamp() and the underlying\\nTransactionIdGetCommitTsData are supposed to accept arbitrary xids,\\nlike txid_status(), so they should handle this.\\n\\nHowever, as far as I can tell, committs has the same issues presented\\nby txid_status(). We only update ShmemVariableCache->oldestCommitTsXid\\n_after_ we truncate the committs SLRU, no lock is held over the\\nduration, and no other lock-protected var is set before truncation. We\\ndon't record the new limit values in COMMIT_TS_TRUNCATE records so\\nthey only becomes known to the standby at replay of the next\\ncheckpoint.\\n\\nmultixact instead runs a single critical section to write xlog, set\\nshmem state, and _then_ truncate the SLRUs. It's more complex than\\nclog or commit_ts since it has two inter-related SLRUs, though.\\n\\nSo: I think we should be setting ShmemVariableCache->oldestXid (under\\nXidGenLock) from TruncateCLOG, after we write xlog but _before_\\ntruncation. Then the rest of SetTransactionIdLimit(...) proceeds as\\nbefore. I don't _think_ we need the critical section since we only\\nhave the one SLRU to worry about. We should also add oldestXid to\\nCLOG_TRUNCATE xlog records so it is up to date on standbys.\\n\\nIt's a little annoying to take XidGenLock twice - once for oldestXid,\\nonce for the other xid limits - but we can just delay advancing\\noldestXid entirely until we've got a clog page to truncate away, in\\nwhich case there's a big enough cost that it doesn't matter. We have\\nto wait to advance the other limits until after we truncate clog to\\nprevent new (wrapped) xids trying to set values in clog for the\\nbefore-wrap xids we're about to truncate away.\\n\\nOn the upside, the need for pendingOldestXid, which always felt hacky,\\ngoes away.\\n\\nWhile we're at it, lets do the same thing for commit_ts. Advance its\\nlimit before truncation, not after, and record that limit in its xlog\\nrecords for redo. There's no need for any special dancing around\\nthere.\\n\\n\\n\\n> Instead of using an LWLock, how about a spin lock?\\n\\nI wrote an explanation of how that wouldn't work, but then I found the\\nproblem with standby, so it no longer matters.\\n\\nI'll have to follow up with a patch soon, as it's Toddler o'Clock.\\n\\n(It's interesting that most of what I'm doing at the moment is\\nwrestling with resource retention limits and how they're often quite\\nimplicit. All the catalog_xmin stuff for logical decoding on standby\\nhas parallels to this, for example, though not enough to create useful\\noverlap of functionality.)\\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\": \"Fri, 23 Dec 2016 18:00:43 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 23 December 2016 at 18:00, Craig Ringer <craig@2ndquadrant.com> wrote:\\n\\n> I'll have to follow up with a patch soon, as it's Toddler o'Clock.\\n\\nHere we go.\\n\\nThis patch advances oldestXid, under XidGenLock, _before_ truncating clog.\\n\\ntxid_status() holds XidGenLock from when it tests oldestXid until it's\\ndone looking up clog, thus eliminating the race.\\n\\nCLOG_TRUNCATE records now contain the oldestXid, so they can advance\\noldestXid on a standby, or when we've truncated clog since the most\\nrecent checkpoint on the master during recovery. It's advanced under\\nXidGenLock during redo to protect against this race on standby.\\n\\nAs outlined in my prior mail I think this is the right approach. I\\ndon't like taking XidGenLock twice, but we don't advance datfrozenxid\\nmuch so it's not a big concern. While a separate ClogTruncationLock\\ncould be added like in my earlier patch, oldestXid is currently under\\nXidGenLock and I'd rather not change that.\\n\\nThe biggest change here is that oldestXid is advanced separately to\\nthe vac limits in the rest of ShmemVariableCache. As far as I can tell\\nwe don't prevent two manual VACUUMs on different DBs from trying to\\nconcurrently run vac_truncate_clog, so this has to be safe against two\\ninvocations racing each other. Rather than try to lock out such\\nconcurrency, the patch ensures that oldestXid can never go backwards.\\nIt doesn't really matter if the vac limits go backwards, it's no worse\\nthan what can already happen in the current code.\\n\\nWe cannot advance the vacuum limits before we truncate the clog away,\\nin case someone tries to access a very new xid (if we're near\\nwraparound)\\n\\nI'm pretty sure that commit timestamps suffer from the same flaw as\\nRobert identified upthread with clog. This patch fixes the clog race,\\nbut not the similar one in commit timestamps. Unlike the clog race\\nwith txid_status(), the commit timestamps one is already potentially\\nuser-visible since we allow arbitrary xids to be looked up for commit\\ntimestamps. I'll address that separately.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, 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\": \"Wed, 28 Dec 2016 18:00:37 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 28 December 2016 at 18:00, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> On 23 December 2016 at 18:00, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>\\n>> I'll have to follow up with a patch soon, as it's Toddler o'Clock.\\n>\\n> Here we go.\\n>\\n> This patch advances oldestXid, under XidGenLock, _before_ truncating clog.\\n>\\n> txid_status() holds XidGenLock from when it tests oldestXid until it's\\n> done looking up clog, thus eliminating the race.\\n>\\n> CLOG_TRUNCATE records now contain the oldestXid, so they can advance\\n> oldestXid on a standby, or when we've truncated clog since the most\\n> recent checkpoint on the master during recovery. It's advanced under\\n> XidGenLock during redo to protect against this race on standby.\\n>\\n> As outlined in my prior mail I think this is the right approach. I\\n> don't like taking XidGenLock twice, but we don't advance datfrozenxid\\n> much so it's not a big concern. While a separate ClogTruncationLock\\n> could be added like in my earlier patch, oldestXid is currently under\\n> XidGenLock and I'd rather not change that.\\n>\\n> The biggest change here is that oldestXid is advanced separately to\\n> the vac limits in the rest of ShmemVariableCache. As far as I can tell\\n> we don't prevent two manual VACUUMs on different DBs from trying to\\n> concurrently run vac_truncate_clog, so this has to be safe against two\\n> invocations racing each other. Rather than try to lock out such\\n> concurrency, the patch ensures that oldestXid can never go backwards.\\n> It doesn't really matter if the vac limits go backwards, it's no worse\\n> than what can already happen in the current code.\\n>\\n> We cannot advance the vacuum limits before we truncate the clog away,\\n> in case someone tries to access a very new xid (if we're near\\n> wraparound)\\n>\\n> I'm pretty sure that commit timestamps suffer from the same flaw as\\n> Robert identified upthread with clog. This patch fixes the clog race,\\n> but not the similar one in commit timestamps. Unlike the clog race\\n> with txid_status(), the commit timestamps one is already potentially\\n> user-visible since we allow arbitrary xids to be looked up for commit\\n> timestamps. I'll address that separately.\\n\\nRebased patch attached. I've split the clog changes out from\\ntxid_status() its self.\\n\\nThere is relevant discussion on the commit timestamp truncation fix\\nthread where the similar fix for commit_ts got committed.\\n\\nhttps://www.postgresql.org/message-id/flat/979ff13d-0b8e-4937-01e8-2925c0adc306%402ndquadrant.com#979ff13d-0b8e-4937-01e8-2925c0adc306@2ndquadrant.com\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\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, 23 Jan 2017 14:32:13 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 23, 2017 at 1:32 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> Rebased patch attached. I've split the clog changes out from\\n> txid_status() its self.\\n\\nI think it's fairly surprising that TruncateCLOG() has responsibility\\nfor writing the xlog record that protects advancing\\nShmemVariableCache->oldestXid, but not the responsibility for actually\\nadvancing that value. In other words, I think the AdvanceOldestXid()\\ncall in vac_truncate_clog() should be moved into TruncateClog().\\n(Similarly, one wonders why AdvanceOldestCommitTsXid() isn't the\\nresponsibility of TruncateCommitTs().)\\n\\nI think it is not correct to advance oldestXid but not oldestXidDB.\\nOtherwise, GetNewTransactionId() might complain about the wrong\\ndatabase.\\n\\nThe way that SetTransactionIdLimit() now works looks a bit dangerous.\\nxidWrapLimit, xidStopLimit, and xidWarnLimit are computed based on the\\npassed-in oldestXid value and written straight into shared memory.\\nBut the shared memory copy of oldestXid could have a different value.\\nI'm not sure if that breaks anything, but it certainly weakens any\\nconfidence callers might have had that all those values are consistent\\nwith each other.\\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, 24 Jan 2017 10:49:57 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 24 January 2017 at 23:49, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> I think it's fairly surprising that TruncateCLOG() has responsibility\\n> for writing the xlog record that protects advancing\\n> ShmemVariableCache->oldestXid, but not the responsibility for actually\\n> advancing that value. In other words, I think the AdvanceOldestXid()\\n> call in vac_truncate_clog() should be moved into TruncateClog().\\n> (Similarly, one wonders why AdvanceOldestCommitTsXid() isn't the\\n> responsibility of TruncateCommitTs().)\\n\\nAgreed, and done in attached.\\n\\nI haven't done the same for commit_ts but will do so on the thread for\\nthe commit_ts race fix if we go ahead with this change here.\\n\\n> I think it is not correct to advance oldestXid but not oldestXidDB.\\n> Otherwise, GetNewTransactionId() might complain about the wrong\\n> database.\\n\\nThat's a clear oversight. Fixed in attached.\\n\\n> The way that SetTransactionIdLimit() now works looks a bit dangerous.\\n> xidWrapLimit, xidStopLimit, and xidWarnLimit are computed based on the\\n> passed-in oldestXid value and written straight into shared memory.\\n> But the shared memory copy of oldestXid could have a different value.\\n> I'm not sure if that breaks anything, but it certainly weakens any\\n> confidence callers might have had that all those values are consistent\\n> with each other.\\n\\nThis was my main hesitation with the whole thing too.\\n\\nIt's necessary to advance oldestXmin before we xlog the advance and\\ntruncate clog, and necessary to advance the vacuum limits only\\nafterwards.\\n\\nI thought it sensible to be conservative about the xid limit to\\nminimise the change from current behaviour, i.e. advance only up to\\nxid limits calculated from the oldestXid determined by the currently\\nvac_truncate_clog. But the current one is actually wrong; in the\\n(unlikely if it's possible at all) case where two vac_truncate_clog()s\\nA and B run with ordering A(advanceOldestXmin) B(advanceOldestXmin)\\nA(truncate) B(truncate) B(advanceLimits) A(advanceLimits), A's advance\\nof the limits would clobber b's. Not too bad, but it'd delay\\nadvancing the limits until the next vacuum, and some xid could get\\nallocated before the limits go backwards.\\n\\nIt's safer to take XidGenLock for slightly longer in order to capture\\nthe oldestXid from shmem and calculate using it. That ensures we never\\nhave any risk of going backwards. The attached updated patch does so.\\n\\nBTW, I find it quite amusing that this was intended to be a small,\\nunintrusive patch, and now it's messing with xid limits and clog\\ntruncation. I'm almost tempted to say we should commit it with the\\n(tiny) race with clog truncation in place. We certainly haven't had\\nany complaints about the same race from people using commit_ts. But\\nthe race window on standby is quite large and I'd rather not knowingly\\nintroduce new bugs.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\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, 25 Jan 2017 13:44:04 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 25 January 2017 at 13:44, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> On 24 January 2017 at 23:49, Robert Haas <robertmhaas@gmail.com> wrote:\\n>\\n>> I think it's fairly surprising that TruncateCLOG() has responsibility\\n>> for writing the xlog record that protects advancing\\n>> ShmemVariableCache->oldestXid, but not the responsibility for actually\\n>> advancing that value. In other words, I think the AdvanceOldestXid()\\n>> call in vac_truncate_clog() should be moved into TruncateClog().\\n>> (Similarly, one wonders why AdvanceOldestCommitTsXid() isn't the\\n>> responsibility of TruncateCommitTs().)\\n>\\n> Agreed, and done in attached.\\n>\\n> I haven't done the same for commit_ts but will do so on the thread for\\n> the commit_ts race fix if we go ahead with this change here.\\n>\\n>> I think it is not correct to advance oldestXid but not oldestXidDB.\\n>> Otherwise, GetNewTransactionId() might complain about the wrong\\n>> database.\\n>\\n> That's a clear oversight. Fixed in attached.\\n\\nNew attached also records it in xlog and applies it to the standby.\\n\\nIf we want to save the 4 bytes per xmin advance (probably not worth\\ncaring) we can instead skip setting it on the standby, in which case\\nit'll be potentially wrong until the next checkpoint. I'd rather make\\nsure it stays correct.\\n\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\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, 25 Jan 2017 15:02:26 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 25, 2017 at 4:02 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> If we want to save the 4 bytes per xmin advance (probably not worth\\n> caring) we can instead skip setting it on the standby, in which case\\n> it'll be potentially wrong until the next checkpoint. I'd rather make\\n> sure it stays correct.\\n\\nThose patches still apply and no reviews have come in yet, so moved to\\nCF 2017-03.\\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, 1 Feb 2017 15:10:21 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 25, 2017 at 12:44 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>> The way that SetTransactionIdLimit() now works looks a bit dangerous.\\n>> xidWrapLimit, xidStopLimit, and xidWarnLimit are computed based on the\\n>> passed-in oldestXid value and written straight into shared memory.\\n>> But the shared memory copy of oldestXid could have a different value.\\n>> I'm not sure if that breaks anything, but it certainly weakens any\\n>> confidence callers might have had that all those values are consistent\\n>> with each other.\\n>\\n> This was my main hesitation with the whole thing too.\\n>\\n> It's necessary to advance oldestXmin before we xlog the advance and\\n> truncate clog, and necessary to advance the vacuum limits only\\n> afterwards.\\n\\nWell, that's why I tried to advocate that their should be two separate\\nXID limits in shared memory: leave what's there now the way it is, and\\nthen add a new limit for \\\"don't try to look up XIDs before this point:\\nsplat\\\". I still think that'd be less risky.\\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, 9 Mar 2017 13:55:25 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 10 March 2017 at 02:55, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> Well, that's why I tried to advocate that their should be two separate\\n> XID limits in shared memory: leave what's there now the way it is, and\\n> then add a new limit for \\\"don't try to look up XIDs before this point:\\n> splat\\\". I still think that'd be less risky.\\n\\nI'm coming back to this \\\"cold\\\" after an extended break, so I hope I\\nget the details right.\\n\\nTL;DR: doing it that way duplicates a bunch of stuff and is ugly\\nwithout offering significant benefit over just fixing the original.\\n\\n\\nI started out with the approach you suggested, but it turns out to be\\nless helpful than you'd think. Simply advancing a new lower limit\\nfield before beginning truncation is no help; there's then a race\\nwhere the lower-limit field can be advanced after we test it but\\nbefore we actually do the SLRU read from clog. To guard against this\\nit's necessary for SLRU truncation to take an exclusive lock during\\nwhich it advances the lower bound. That way a concurrent reader can\\ntake the lock in shared mode before reading the lower bound and hold\\nit until it finishes the clog read, knowing that vacuum cannot then\\ntruncate the data out from under it because it'll block trying to\\nadvance the lower limit.\\n\\nA spinlock isn't suitable for this. While we can take the lock only\\nbriefly to update the limit field in vac_truncate_clog, in\\ntxid_status() we have to hold it from when we test the boundary\\nthrough to when we finish the SLRU clog lookup, and that lookup does\\nI/O and might sleep. If we release it after testing the lower bound\\nbut before the SLRU lookup our race comes back since vacuum can jump\\nin and truncate it out from under us. So now we need a new LWLock used\\nonly for vac_truncate_clog before advancing the clog truncation.\\n\\nI implemented just that - a new ClogTruncateLog in the lwlocks array\\nand a new field in ShmemVariableCache for the lower xid bound, IIRC.\\nOther than requiring an extra lwlock acquisition for vac_truncate_clog\\nit works fine ... for the master.\\n\\nBut it doesn't fix the much bigger race on the standby. We only emit\\nWAL for xid limits after we truncate clog, and the clog truncation\\nrecord doesn't record the new limit.\\n\\nSo now we need a new, somewhat redundant, xlog record and redo\\nfunction for this lower clog bound pointer. Which, really, is only\\nactually tracking a slightly more up to date version of oldestXid.\\n\\nAt that point I was just papering around a race that should just be\\nfixed at its source instead. Advance oldestXid before truncating clog,\\nand write a clog truncation record that includes the new oldestXid.\\n\\nSo... I can go back to the old approach and just add the new xlog\\nrecord and redo method, new lwlock, new shmemvariablecache field, etc,\\nif you're really concerned this approach is risky. But I'd rather fix\\nthe original problem instead.\\n\\nIt might be helpful if I separate out the patch that touches oldestXid\\nfrom the rest for separate review, so I'll update here with a 2-patch\\nseries instead of a squashed single patch soon.\\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\": \"Fri, 10 Mar 2017 15:00:16 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 10, 2017 at 2:00 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> On 10 March 2017 at 02:55, Robert Haas <robertmhaas@gmail.com> wrote:\\n>> Well, that's why I tried to advocate that their should be two separate\\n>> XID limits in shared memory: leave what's there now the way it is, and\\n>> then add a new limit for \\\"don't try to look up XIDs before this point:\\n>> splat\\\". I still think that'd be less risky.\\n>\\n> I'm coming back to this \\\"cold\\\" after an extended break, so I hope I\\n> get the details right.\\n\\nYeah, sorry I've been away from this for a while.\\n\\n> TL;DR: doing it that way duplicates a bunch of stuff and is ugly\\n> without offering significant benefit over just fixing the original.\\n>\\n> I started out with the approach you suggested, but it turns out to be\\n> less helpful than you'd think. Simply advancing a new lower limit\\n> field before beginning truncation is no help; there's then a race\\n> where the lower-limit field can be advanced after we test it but\\n> before we actually do the SLRU read from clog. To guard against this\\n> it's necessary for SLRU truncation to take an exclusive lock during\\n> which it advances the lower bound. That way a concurrent reader can\\n> take the lock in shared mode before reading the lower bound and hold\\n> it until it finishes the clog read, knowing that vacuum cannot then\\n> truncate the data out from under it because it'll block trying to\\n> advance the lower limit.\\n\\nThat's a good point which I hadn't fully considered. On the other\\nhand, there really are two separate notions of the \\\"oldest\\\" XID.\\nThere's the oldest XID that we can safely look up, and then there's\\nthe oldest XID that we can't reuse. These two are the same when no\\ntruncation is in progress, but when a truncation is in progress then\\nthey're different: the oldest XID that's safe to look up is the first\\none after whatever we're truncating away, but the oldest XID that we\\ncan't reuse is the newest one preceding the stuff that we're busy\\ntruncating. IOW, when truncation is happening, there's a portion of\\nthe XID space whose clog files are being removed - and the XIDs that\\nare in that range aren't safe to look up any more, but are also not\\navailable for reuse to prevent wraparound. Right now, all of the\\nrelevant fields in VariableCacheData are based on the ready-for-reuse\\nconcept, and I don't think that switching some but not all of them to\\nbe based on the safe-to-look-up concept necessarily qualifies as an\\nimprovement. It's different, but I'm not sure it's better.\\n\\nWhat if we approached this problem from the other end? Suppose we use\\na heavyweight lock on, say, transaction ID 1 to represent the right to\\ntruncate CLOG. We grab this lock in exclusive mode before beginning\\nto truncate, and in shared mode while looking up XIDs.\\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, 10 Mar 2017 16:09:52 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 11 March 2017 at 05:09, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On the other\\n> hand, there really are two separate notions of the \\\"oldest\\\" XID.\\n> There's the oldest XID that we can safely look up, and then there's\\n> the oldest XID that we can't reuse. These two are the same when no\\n> truncation is in progress, but when a truncation is in progress then\\n> they're different: the oldest XID that's safe to look up is the first\\n> one after whatever we're truncating away, but the oldest XID that we\\n> can't reuse is the newest one preceding the stuff that we're busy\\n> truncating.\\n\\nRight.\\n\\nMy view here is that the oldest xid we cannot reuse is already guarded\\nby xidWrapLimit, which we advance after clog truncation. Whether as\\nthis advances at the same time as or after we advance oldestXid and\\ntruncate clog doesn't actually matter, we must just ensure that it\\nnever advances _before_.\\n\\nSo tracking a second copy of oldestXid whose only purpose is to\\nrecalculate xidWrapLimit serves no real purpose. It's redundant except\\nduring vac_truncate_clog, during which time local state is sufficient\\n*if* we add oldestXid to the clog truncation xlog record, which we\\nmust do anyway because:\\n\\nAny number of locking hoop-jumping schemes fail to solve the problem\\nof outdated oldestXid information on standbys. Right now we truncate\\nclog and xlog the truncation before we write the new oldestXid limit\\nto xlog. In fact, we don't write the new xid limit to xlog until the\\nnext checkpoint. So the standby has a huge window where its idea of\\noldestXid is completely wrong, and unless we at least add the new\\noldestXid to the clog truncation xlog record we can't fix that.\\n\\nWe only get away with this now because there's no way to look up an\\narbitrary xid's status.\\n\\nNo locking scheme on the master can solve this, because the locks on\\nthe master do not affect the standby or vice versa.\\n\\nTherefore, we _must_ advance oldestXid (or a copy of it used only for\\n\\\"oldest xid still in clog) before truncating clog.\\n\\nIf we're going to do that we might as well just make sure the\\nstandby's xid limits are updated correctly when we truncate clog\\nrather than doing it lazily at checkpoints. Advance oldestXid before\\ntruncating clog away, and record the new xid in the clog truncation\\nxlog record. On redo after master crash, and on standbys, we're\\nguaranteed to re-do the whole clog truncation operation - advance\\noldestXid, truncate clog, advance xidWrapLimit etc - and everything\\nstays consistent.\\n\\nI'll extract this part of the patch so it can be looked at separately,\\nit'll be clearer that way.\\n\\nI think of it as slightly contracting then slightly expanding the xid\\nrange window during clog truncation. Advance the oldest xid slightly\\nbefore the xidWrapLimit, so temporarily the range of xids is narrower\\nthan 2^31. xlog it first so we ensure it's all redone on crash and on\\nstandby. Because no lock is held throughout all of vac_truncate_clog,\\nmake sure the ordering of the different phases between concurrent\\nvac_truncate_xlog runs doesn't matter.\\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\": \"Sat, 11 Mar 2017 14:32:08 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 11 March 2017 at 14:32, Craig Ringer <craig@2ndquadrant.com> wrote:\\n\\n> I'll extract this part of the patch so it can be looked at separately,\\n> it'll be clearer that way.\\n\\nApparently I thought that last time too since I already posted it\\nsplit up. Ahem. Working on too many different things at once.\\n\\nLast-posted patches apply fine, no need for an update.\\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\": \"Mon, 13 Mar 2017 11:22:48 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Sat, Mar 11, 2017 at 1:32 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> On 11 March 2017 at 05:09, Robert Haas <robertmhaas@gmail.com> wrote:\\n>> On the other\\n>> hand, there really are two separate notions of the \\\"oldest\\\" XID.\\n>> There's the oldest XID that we can safely look up, and then there's\\n>> the oldest XID that we can't reuse. These two are the same when no\\n>> truncation is in progress, but when a truncation is in progress then\\n>> they're different: the oldest XID that's safe to look up is the first\\n>> one after whatever we're truncating away, but the oldest XID that we\\n>> can't reuse is the newest one preceding the stuff that we're busy\\n>> truncating.\\n>\\n> Right.\\n>\\n> My view here is that the oldest xid we cannot reuse is already guarded\\n> by xidWrapLimit, which we advance after clog truncation. Whether as\\n> this advances at the same time as or after we advance oldestXid and\\n> truncate clog doesn't actually matter, we must just ensure that it\\n> never advances _before_.\\n>\\n> So tracking a second copy of oldestXid whose only purpose is to\\n> recalculate xidWrapLimit serves no real purpose.\\n\\nHmm, so what this patch is doing changed quite a bit between January\\n23rd and January 25th. In the January 23rd version, oldestXid and\\noldestXidDB are changed to track the oldest XID that we can safely\\nlook up, and the remaining related fields are still relative to the\\noldest XID that can be reused. That seems, as I said before, scary.\\nBut in the January 25th version, *all* of the related fields have been\\nchanged to track the oldest XID that we can safely look up, because\\nSetTransactionIdLimit() now uses the values set by AdvanceOldestXid()\\nto compute all of the other values, which seems flat-out incorrect.\\nAdvanceOldestXid() is called to advance that limit before clog\\ntruncation happens, and if somebody then calls SetTransactionIdLimit()\\nbefore clog truncation is complete, we'll advanced those derived\\nlimits prematurely.\\n\\nFor example, suppose vacuum #1 comes along, advances the limits,\\ntruncates clog, and then gets descheduled. Now vacuum #2 comes along,\\nadvances the limits further, and then gets descheduled. Now vacuum #1\\nwakes up and calls SetTransactionIdLimit() and prematurely advances\\nxidWrapLimit. Oops.\\n\\nThe way you put it is that having a second copy of oldestXid whose\\nonly purpose is to recompute xidWrapLimit is pointless, but the way\\nI'd say is that you're trying to make one variable do two jobs.\\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, 13 Mar 2017 17:43:03 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 14 March 2017 at 05:43, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> For example, suppose vacuum #1 comes along, advances the limits,\\n> truncates clog, and then gets descheduled. Now vacuum #2 comes along,\\n> advances the limits further, and then gets descheduled. Now vacuum #1\\n> wakes up and calls SetTransactionIdLimit() and prematurely advances\\n> xidWrapLimit. Oops.\\n\\nMm, right. And without a lock held from when oldestXid advances\\nthrough to completion of clog truncation, then taking the same lock in\\nSetTransactionIdLimit, there's not really a way around it.\\n\\nI'm embarrassed not to have seen that.\\n\\nDoing things the other way around, per the earlier patch, can cause\\nSetTransactionIdLimit to not to advance as far as it should.\\n\\nOK, I'm convinced, a new field is safer, even if it's redundant most\\nof the time.\\n\\nI'll introduce a new LWLock, ClogTruncationLock, which will be held\\nfrom when we advance the new clogOldestXid field through to when clog\\ntruncation completes.\\n\\nMost of the rest can stay the same. In particular, the expanded xlog\\nrecord for clog truncation will still be required.\\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\": \"Tue, 14 Mar 2017 10:22:59 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 13, 2017 at 10:22 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> I'll introduce a new LWLock, ClogTruncationLock, which will be held\\n> from when we advance the new clogOldestXid field through to when clog\\n> truncation completes.\\n\\n+1.\\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, 14 Mar 2017 07:57:08 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 14 March 2017 at 19:57, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Mon, Mar 13, 2017 at 10:22 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>> I'll introduce a new LWLock, ClogTruncationLock, which will be held\\n>> from when we advance the new clogOldestXid field through to when clog\\n>> truncation completes.\\n>\\n> +1.\\n\\nOK, here's the revised patch. Relevant comments added where needed.\\n\\nIt still changes the xlog record for clog truncation to carry the new\\nxid, but I've left advancing oldestXid until the checkpoint as is\\ncurrently the case, and only advanced oldestClogXid when we replay\\nclog truncation.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\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 Mar 2017 13:38:18 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 20, 2017 at 1:38 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> On 14 March 2017 at 19:57, Robert Haas <robertmhaas@gmail.com> wrote:\\n>> On Mon, Mar 13, 2017 at 10:22 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>>> I'll introduce a new LWLock, ClogTruncationLock, which will be held\\n>>> from when we advance the new clogOldestXid field through to when clog\\n>>> truncation completes.\\n>>\\n>> +1.\\n>\\n> OK, here's the revised patch. Relevant comments added where needed.\\n>\\n> It still changes the xlog record for clog truncation to carry the new\\n> xid, but I've left advancing oldestXid until the checkpoint as is\\n> currently the case, and only advanced oldestClogXid when we replay\\n> clog truncation.\\n\\n/me smacks forehead. Actually, it should be CLogTruncationLock, with\\na capital L, for consistency with CLogControlLock.\\n\\nThe new lock needs to be added to the table in monitoring.sgml.\\n\\nI don't think the new header comments in TransactionIdDidCommit and\\nTransactionIdDidAbort are super-clear. I'm not sure you're going to\\nbe able to explain it there in a reasonable number of words, but I\\nthink that speaking of \\\"testing against oldestClogXid\\\" will leave\\npeople wondering what exactly that means. Maybe just write \\\"caller is\\nresponsible for ensuring that the clog records covering XID being\\nlooked up can't be truncated away while the lookup is in progress\\\",\\nand then leave the bit about CLogTruncationLock to be explained by the\\ncallers that do that. Or you could drop these comments entirely.\\n\\nOverall, though, I think that 0001 looks far better than any previous\\niteration. It's simple. It looks safe. It seems unlikely to break\\nanything that works now. Woo hoo!\\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, 21 Mar 2017 13:49:49 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 22 March 2017 at 01:49, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> /me smacks forehead. Actually, it should be CLogTruncationLock, with\\n> a capital L, for consistency with CLogControlLock.\\n\\nWill do.\\n\\n> The new lock needs to be added to the table in monitoring.sgml.\\n\\nSame.\\n\\n> I don't think the new header comments in TransactionIdDidCommit and\\n> TransactionIdDidAbort are super-clear. I'm not sure you're going to\\n> be able to explain it there in a reasonable number of words, but I\\n> think that speaking of \\\"testing against oldestClogXid\\\" will leave\\n> people wondering what exactly that means. Maybe just write \\\"caller is\\n> responsible for ensuring that the clog records covering XID being\\n> looked up can't be truncated away while the lookup is in progress\\\",\\n> and then leave the bit about CLogTruncationLock to be explained by the\\n> callers that do that. Or you could drop these comments entirely.\\n\\nOK. I'll revisit and see if I can clean it up, otherwise remove it.\\n\\n> Overall, though, I think that 0001 looks far better than any previous\\n> iteration. It's simple. It looks safe. It seems unlikely to break\\n> anything that works now. Woo hoo!\\n\\nFunny that this started with \\\"hey, here's a simple, non-invasive\\nfunction for looking up the status of an arbitrary xid\\\".\\n\\nMature, complex systems eh?\\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 Mar 2017 09:49:55 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 22 March 2017 at 09:49, Craig Ringer <craig@2ndquadrant.com> wrote:\\n\\n>> Overall, though, I think that 0001 looks far better than any previous\\n>> iteration. It's simple. It looks safe. It seems unlikely to break\\n>> anything that works now. Woo hoo!\\n>\\n> Funny that this started with \\\"hey, here's a simple, non-invasive\\n> function for looking up the status of an arbitrary xid\\\".\\n\\nChanges made per discussion.\\n\\nRemoved the comments on TransactionIdDidCommit and\\nTransactionIdDidAbort . It's not going to be relevant for the immense\\nmajority of callers anyway, and callers that are looking up arbitrary\\nuser supplied XIDs will (hopefully) be looking at\\nTransactionIdInRecentPast anyway.\\n\\nI'll be leaving the 'xid' vs 'bigint' issues elsewhere in Pg for next\\nrelease, nowhere near time for that now.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\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, 22 Mar 2017 11:35:12 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 22 March 2017 at 03:35, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> On 22 March 2017 at 09:49, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>\\n>>> Overall, though, I think that 0001 looks far better than any previous\\n>>> iteration. It's simple. It looks safe. It seems unlikely to break\\n>>> anything that works now. Woo hoo!\\n>>\\n>> Funny that this started with \\\"hey, here's a simple, non-invasive\\n>> function for looking up the status of an arbitrary xid\\\".\\n>\\n> Changes made per discussion.\\n\\nThis looks good to me also. Does what we need it to do.\\n\\nI was a little worried by possible performance of new lock, but its\\nnot intended to be run frequently, so its OK.\\n\\n-- \\nSimon Riggs 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, 22 Mar 2017 07:13:32 +0000\",\n \"msg_from\": \"Simon Riggs <simon@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 22, 2017 at 3:13 AM, Simon Riggs <simon@2ndquadrant.com> wrote:\\n>> Changes made per discussion.\\n>\\n> This looks good to me also. Does what we need it to do.\\n>\\n> I was a little worried by possible performance of new lock, but its\\n> not intended to be run frequently, so its OK.\\n\\nAgreed.\\n\\nReviewing 0002:\\n\\n+ if (!TransactionIdIsValid(xid))\\n+ {\\n+ LWLockRelease(XidGenLock);\\n+ ereport(ERROR,\\n+ (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\\n+ errmsg(\\\"transaction ID \\\" UINT64_FORMAT \\\" is an invalid xid\\\",\\n+ xid_with_epoch)));\\n+ }\\n\\nIt's unnecessary to release LWLockRelease() before throwing an error,\\nand it's also wrong because we haven't acquired XidGenLock in this\\ncode path. But maybe it would be better to just remove this entirely\\nand instead have TransactionIdInRecentPast return false for\\nInvalidTransactionId. Then we'd avoid adding a translatable message\\nfor a case that is basically harmless to allow.\\n\\n+ if (TransactionIdIsCurrentTransactionId(xid))\\n+ status = gettext_noop(\\\"in progress\\\");\\n+ else if (TransactionIdDidCommit(xid))\\n+ status = gettext_noop(\\\"committed\\\");\\n+ else if (TransactionIdDidAbort(xid))\\n+ status = gettext_noop(\\\"aborted\\\");\\n+ else\\n+\\n+ /*\\n+ * can't test TransactionIdIsInProgress here or we race with\\n+ * concurrent commit/abort. There's no point anyway, since it\\n+ * might then commit/abort just after we check.\\n+ */\\n+ status = gettext_noop(\\\"in progress\\\");\\n\\nI am not sure this is going to do the right thing for transactions\\nthat are aborted by a crash without managing to write an abort record.\\nIt seems that the first check will say the transaction isn't in\\nprogress, and the second and third checks will say it isn't either\\ncommitted or aborted since, if I am reading this correctly, they just\\nread clog directly. Compare HeapTupleSatisfiesMVCC, which assumes\\nthat a not-in-progress, not-committed transaction must necessarily\\nhave aborted. I think your comment is pointing to a real issue,\\nthough. It seems like what might be needed is to add one more check.\\nBefore where you have the \\\"else\\\" clause, check if the XID is old, e.g.\\nprecedes our snapshot's xmin. If so, it must be committed or aborted\\nand, since it didn't commit, it aborted. If not, it must've changed\\nfrom in progress to not-in-progress just as we were in the midst\\nchecking, so labeling it as in progress is fine.\\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 Mar 2017 13:41:47 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 22 March 2017 at 17:41, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> + if (TransactionIdIsCurrentTransactionId(xid))\\n> + status = gettext_noop(\\\"in progress\\\");\\n> + else if (TransactionIdDidCommit(xid))\\n> + status = gettext_noop(\\\"committed\\\");\\n> + else if (TransactionIdDidAbort(xid))\\n> + status = gettext_noop(\\\"aborted\\\");\\n> + else\\n> +\\n> + /*\\n> + * can't test TransactionIdIsInProgress here or we race with\\n> + * concurrent commit/abort. There's no point anyway, since it\\n> + * might then commit/abort just after we check.\\n> + */\\n> + status = gettext_noop(\\\"in progress\\\");\\n>\\n> I am not sure this is going to do the right thing for transactions\\n> that are aborted by a crash without managing to write an abort record.\\n\\nYes, perhaps we should report that state as \\\"aborted - incomplete\\\".\\n\\nAnd of course, we might return \\\"subcommitted\\\" also, which could\\ntechnically also be an abort in some cases, so we'd need to do a wait\\nloop on that.\\n\\nWhich makes me think it would be confusing to say \\\"in progress\\\" for\\nwhen it is our current xid, since the user might wait until it is\\ncomplete and then wait forever. Prefer it if it said \\\"in progress -\\ncurrent transaction\\\"\\n\\n-- \\nSimon Riggs 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, 22 Mar 2017 18:08:55 +0000\",\n \"msg_from\": \"Simon Riggs <simon@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 22, 2017 at 2:08 PM, Simon Riggs <simon@2ndquadrant.com> wrote:\\n> On 22 March 2017 at 17:41, Robert Haas <robertmhaas@gmail.com> wrote:\\n>> + if (TransactionIdIsCurrentTransactionId(xid))\\n>> + status = gettext_noop(\\\"in progress\\\");\\n>> + else if (TransactionIdDidCommit(xid))\\n>> + status = gettext_noop(\\\"committed\\\");\\n>> + else if (TransactionIdDidAbort(xid))\\n>> + status = gettext_noop(\\\"aborted\\\");\\n>> + else\\n>> +\\n>> + /*\\n>> + * can't test TransactionIdIsInProgress here or we race with\\n>> + * concurrent commit/abort. There's no point anyway, since it\\n>> + * might then commit/abort just after we check.\\n>> + */\\n>> + status = gettext_noop(\\\"in progress\\\");\\n>>\\n>> I am not sure this is going to do the right thing for transactions\\n>> that are aborted by a crash without managing to write an abort record.\\n>\\n> Yes, perhaps we should report that state as \\\"aborted - incomplete\\\".\\n>\\n> And of course, we might return \\\"subcommitted\\\" also, which could\\n> technically also be an abort in some cases, so we'd need to do a wait\\n> loop on that.\\n\\nI actually don't think those are things we should expose to users.\\nThey're internal implementation details. The user had better not care\\nwhether an abort was the type of abort that wrote an abort record or\\nthe type that didn't.\\n\\n> Which makes me think it would be confusing to say \\\"in progress\\\" for\\n> when it is our current xid, since the user might wait until it is\\n> complete and then wait forever. Prefer it if it said \\\"in progress -\\n> current transaction\\\"\\n\\nHmm, or just \\\"current transaction\\\", maybe?\\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 Mar 2017 14:10:28 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 23 March 2017 at 02:08, Simon Riggs <simon@2ndquadrant.com> wrote:\\n\\n> And of course, we might return \\\"subcommitted\\\" also, which could\\n> technically also be an abort in some cases, so we'd need to do a wait\\n> loop on that.\\n\\nUsers generally don't see subxact IDs, so it wasn't something I was\\noverly concerned by. Most notably, txid_current() doesn't report\\nthem.\\n\\nUsers who want to know the commit status of an xact that had a commit\\nin-flight when they lost access to it due to server crash, network\\nloss, etc aren't going to care about subxacts. If you lose your\\nconnection after a RELEASE SAVEPOINT you know the outer xact will get\\naborted or the state of the individual subxacts.\\n\\nThey're visible in heap tuples, but you can only see uncommitted heap\\ntuples from your own top-level xid. For anything else, if you can see\\nthe xid in xmin you know it committed. There isn't really any reason\\nyou'd be looking up tuple xids with txid_status anyway, and if you did\\nyou'd have to pay attention to mess like multixacts in xmax ... but\\nyou can't tell from the xid alone if it's a multixact or not, so this\\ndoesn't make sense with xids that could be multis anyway.\\n\\nIf we're going to handle subxacts specially, we should probably report\\nthem as \\\"sub-committed\\\" if we find that the current xid is a committed\\nsubxact member of an outer xact that is still in-progress. But IMO\\nit's pretty pointless since you won't be dealing with subxact IDs in\\nthe application anyway.\\n\\n> Which makes me think it would be confusing to say \\\"in progress\\\" for\\n> when it is our current xid, since the user might wait until it is\\n> complete and then wait forever. Prefer it if it said \\\"in progress -\\n> current transaction\\\"\\n\\nI'm fine with \\\"current transaction\\\". Though I think it's kind of a\\nmoot point as I don't see any reason for an application to ever be\\ndoing the equivalent of\\n\\n txid_status(txid_current())\\n\\nin the first place. But it's harmless enough.\\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\": \"Thu, 23 Mar 2017 09:40:49 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 23 March 2017 at 01:41, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Wed, Mar 22, 2017 at 3:13 AM, Simon Riggs <simon@2ndquadrant.com> wrote:\\n>>> Changes made per discussion.\\n>>\\n>> This looks good to me also. Does what we need it to do.\\n>>\\n>> I was a little worried by possible performance of new lock, but its\\n>> not intended to be run frequently, so its OK.\\n>\\n> Agreed.\\n>\\n> Reviewing 0002:\\n>\\n> + if (!TransactionIdIsValid(xid))\\n> + {\\n> + LWLockRelease(XidGenLock);\\n> + ereport(ERROR,\\n> + (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\\n> + errmsg(\\\"transaction ID \\\" UINT64_FORMAT \\\" is an invalid xid\\\",\\n> + xid_with_epoch)));\\n> + }\\n>\\n> It's unnecessary to release LWLockRelease() before throwing an error,\\n> and it's also wrong because we haven't acquired XidGenLock in this\\n> code path. But maybe it would be better to just remove this entirely\\n> and instead have TransactionIdInRecentPast return false for\\n> InvalidTransactionId. Then we'd avoid adding a translatable message\\n> for a case that is basically harmless to allow.\\n\\nAgreed, that's better.\\n\\n> + if (TransactionIdIsCurrentTransactionId(xid))\\n> + status = gettext_noop(\\\"in progress\\\");\\n> + else if (TransactionIdDidCommit(xid))\\n> + status = gettext_noop(\\\"committed\\\");\\n> + else if (TransactionIdDidAbort(xid))\\n> + status = gettext_noop(\\\"aborted\\\");\\n> + else\\n> +\\n> + /*\\n> + * can't test TransactionIdIsInProgress here or we race with\\n> + * concurrent commit/abort. There's no point anyway, since it\\n> + * might then commit/abort just after we check.\\n> + */\\n> + status = gettext_noop(\\\"in progress\\\");\\n>\\n> I am not sure this is going to do the right thing for transactions\\n> that are aborted by a crash without managing to write an abort record.\\n\\nYou're right. It'll report in-progress, since the clog entry will\\nstill be 0. We don't appear to explicitly update the clog during\\nrecovery.\\n\\nFunny, I got so focused on the clog access safety stuff in the end\\nthat I missed the bigger picture.\\n\\nGiven that part of the point of this is xact status after crash,\\nthat's kind of important. I've added a TAP test for this, as part of a\\nnew test set, \\\"011_crash_recovery.pl\\\". The recovery tests don't really\\nhave much on crash behaviour at the moment so might as well start it\\nand add more as we go. It doesn't make sense to add a whole new top\\nlevel TAP test just for txid_status.\\n\\n(I *love* the TAP framework now, it took 10 mins to write a test\\nthat's trivial to repeat in 5 seconds per run for this).\\n\\n> It seems that the first check will say the transaction isn't in\\n> progress, and the second and third checks will say it isn't either\\n> committed or aborted since, if I am reading this correctly, they just\\n> read clog directly.\\n\\nRight. They're not concerned with subxacts or multixacts.\\n\\n> Compare HeapTupleSatisfiesMVCC, which assumes\\n> that a not-in-progress, not-committed transaction must necessarily\\n> have aborted.\\n\\nRight.\\n\\nWe don't have a HeapTuple here, of course, so we can't test flags.\\nMost importantly this means we can't handle multixacts. If we ever did\\ndecide to expose multixacts as bigint epoch-qualified xids for general\\nusers, we'd probably reserve the high bit of the epoch the bigint xid\\nrepresentation for the equivalent of HEAP_XMAX_IS_MULTI, and maybe\\nit's worth reserving that bit now and ERROR'ing if we find it set. But\\nright now we never produce a bigint xid with a multixact.\\n\\nI wonder if a note in the docs warning not to cast xid to bigint is\\nworth it. Probably not. You have to cast xid::text::bigint which is\\nkind of a hint, plus it doesn't make sense to test txid_status() for\\ntuples' xmin/xmax . I guess you might use it with xids from\\npg_stat_activity, but there's not much point when you can just look at\\npg_stat_activity again to see if the proc of interest is still there\\nand has the same xid.\\n\\nAnyway...\\n\\n> I think your comment is pointing to a real issue,\\n> though. It seems like what might be needed is to add one more check.\\n> Before where you have the \\\"else\\\" clause, check if the XID is old, e.g.\\n> precedes our snapshot's xmin. If so, it must be committed or aborted\\n> and, since it didn't commit, it aborted. If not, it must've changed\\n> from in progress to not-in-progress just as we were in the midst\\n> checking, so labeling it as in progress is fine.\\n\\nThat seems clear and sensible.\\n\\nXidInMVCCSnapshot simply tests TransactionIdPrecedes(xid,\\nsnapshot->xmin), as does HeapTupleSatisfiesHistoricMVCC . It seems\\nreasonable enough to just examine the snapshot xmin directly in\\ntxid_status too; it's already done in replication/logical/snapbuild.c\\nand lmgr/predicate.c.\\n\\nWe're running in an SQL-called function so we'll have a good snapshot\\nand GetSnapshotData will have run, so GetActiveSnapshot()->xmin should\\nbe sufficient.\\n\\nAmended patch attached, with added TAP test included.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, 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, 23 Mar 2017 11:25:07 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 23 March 2017 at 11:25, Craig Ringer <craig@2ndquadrant.com> wrote:\\n\\n> Amended patch attached, with added TAP test included.\\n\\nManaged to omit it, sigh.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, 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, 23 Mar 2017 12:22:23 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 21, 2017 at 11:35 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> Changes made per discussion.\\n\\nCommitted 0001.\\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 Mar 2017 14:29:39 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 24 March 2017 at 02:29, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Tue, Mar 21, 2017 at 11:35 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>> Changes made per discussion.\\n>\\n> Committed 0001.\\n\\nMuch appreciated.\\n\\nHere's the 2nd patch rebased on top of master, with the TAP test\\nincluded this time. Looks ready to go.\\n\\nI really appreciate the time you've taken to look at this. Point me at\\nanything from your team you want some outside review on.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, 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\": \"Fri, 24 Mar 2017 14:27:44 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 24, 2017 at 2:27 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> On 24 March 2017 at 02:29, Robert Haas <robertmhaas@gmail.com> wrote:\\n>> On Tue, Mar 21, 2017 at 11:35 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>>> Changes made per discussion.\\n>>\\n>> Committed 0001.\\n>\\n> Much appreciated.\\n>\\n> Here's the 2nd patch rebased on top of master, with the TAP test\\n> included this time. Looks ready to go.\\n\\nCommitted.\\n\\n> I really appreciate the time you've taken to look at this. Point me at\\n> anything from your team you want some outside review on.\\n\\nThanks, that is a valuable offer which I will be pleased to accept!\\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 Mar 2017 12:28:24 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 3/24/17 02:27, Craig Ringer wrote:\\n> On 24 March 2017 at 02:29, Robert Haas <robertmhaas@gmail.com> wrote:\\n>> On Tue, Mar 21, 2017 at 11:35 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>>> Changes made per discussion.\\n>>\\n>> Committed 0001.\\n> \\n> Much appreciated.\\n> \\n> Here's the 2nd patch rebased on top of master, with the TAP test\\n> included this time. Looks ready to go.\\n\\nI'm experiencing hangs in the new t/011_crash_recovery.pl test. It\\nseems to hang after the first call to SELECT txid_current();.\\n\\n-- \\nPeter Eisentraut 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\": \"Fri, 24 Mar 2017 19:31:37 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 25 March 2017 at 07:31, Peter Eisentraut\\n<peter.eisentraut@2ndquadrant.com> wrote:\\n> On 3/24/17 02:27, Craig Ringer wrote:\\n>> On 24 March 2017 at 02:29, Robert Haas <robertmhaas@gmail.com> wrote:\\n>>> On Tue, Mar 21, 2017 at 11:35 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>>>> Changes made per discussion.\\n>>>\\n>>> Committed 0001.\\n>>\\n>> Much appreciated.\\n>>\\n>> Here's the 2nd patch rebased on top of master, with the TAP test\\n>> included this time. Looks ready to go.\\n>\\n> I'm experiencing hangs in the new t/011_crash_recovery.pl test. It\\n> seems to hang after the first call to SELECT txid_current();.\\n\\nif you add\\n\\nnote \\\"txid is $xid\\\";\\n\\nafter\\n\\n+chomp($xid);\\n\\ndoes it report the xid?\\n\\nAlternately, can you see a 'psql' process and a backend doing a SELECT\\nwhen it stops progressing?\\n\\nI'm wondering if this is a perl version/platform issue around\\n\\n $tx->pump until $stdout =~ /[[:digit:]]+[\\\\r\\\\n]$/;\\n\\nwhere we're not recognising the required output from psql when we get it.\\n\\nWhat's in src/test/recovery/tmp_check/log/regress_log_011* ?\\n\\nI couldn't use PostgresNode::psql or PostgresNode::safe_psql here\\nbecause the session must stay open.\\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\": \"Sat, 25 Mar 2017 08:08:01 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"> I'm wondering if this is a perl version/platform issue around\\n> \\n> $tx->pump until $stdout =~ /[[:digit:]]+[\\\\r\\\\n]$/;\\n> \\n> where we're not recognising the required output from psql when we get it.\\n> \\n> What's in src/test/recovery/tmp_check/log/regress_log_011* ?\\n> \\n> I couldn't use PostgresNode::psql or PostgresNode::safe_psql here\\n> because the session must stay open.\\n\\nThe problem was that psql needs to be called with -X. Fix committed.\\n\\n-- \\nPeter Eisentraut 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\": \"Sat, 25 Mar 2017 00:12:48 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 3/25/17 12:12 AM, Peter Eisentraut wrote:\\n>> I'm wondering if this is a perl version/platform issue around\\n>>\\n>> $tx->pump until $stdout =~ /[[:digit:]]+[\\\\r\\\\n]$/;\\n>>\\n>> where we're not recognising the required output from psql when we get it.\\n>>\\n>> What's in src/test/recovery/tmp_check/log/regress_log_011* ?\\n>>\\n>> I couldn't use PostgresNode::psql or PostgresNode::safe_psql here\\n>> because the session must stay open.\\n>\\n> The problem was that psql needs to be called with -X. Fix committed.\\n\\nIt's not clear to me what remains to be done on this patch. I feel, at \\nthe least, that patches 3 and 4 need to be rebased and the status set \\nback to \\\"Needs Review\\\".\\n\\nThis thread has been idle for six days. Please respond with a new patch \\nby 2017-04-04 00:00 AoE (UTC-12) or this submission will be marked \\n\\\"Returned with Feedback\\\".\\n\\n-- \\n-David\\ndavid@pgmasters.net\\n\\n\\n-- \\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, 31 Mar 2017 10:31:20 -0400\",\n \"msg_from\": \"David Steele <david@pgmasters.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 31 Mar. 2017 22:31, \\\"David Steele\\\" <david@pgmasters.net> wrote:\\n\\nOn 3/25/17 12:12 AM, Peter Eisentraut wrote:\\n\\n> I'm wondering if this is a perl version/platform issue around\\n>>\\n>> $tx->pump until $stdout =~ /[[:digit:]]+[\\\\r\\\\n]$/;\\n>>\\n>> where we're not recognising the required output from psql when we get it.\\n>>\\n>> What's in src/test/recovery/tmp_check/log/regress_log_011* ?\\n>>\\n>> I couldn't use PostgresNode::psql or PostgresNode::safe_psql here\\n>> because the session must stay open.\\n>>\\n>\\n> The problem was that psql needs to be called with -X. Fix committed.\\n>\\n\\nIt's not clear to me what remains to be done on this patch. I feel, at the\\nleast, that patches 3 and 4 need to be rebased and the status set back to\\n\\\"Needs Review\\\".\\n\\nThis thread has been idle for six days. Please respond with a new patch by\\n2017-04-04 00:00 AoE (UTC-12) or this submission will be marked \\\"Returned\\nwith Feedback\\\".\\n\\n\\nI consider the feature complete and committed.\\n\\nPatch 3 is optional and per discussion buried upthread we generally agreed\\nthat it'd be better to replace most user visible uses of the 'xid' data\\ntype with a new epoch extended 'xid64' or similar.\\n\\nPatch 4, txid_incinerate, has never been intended for commit. It's a\\ntesting tool.\\n\\nPatches 1 and 2 were the key parts and thanks to Robert's helpful review,\\nadvice and edits they're committed now.\\n\\nCommitted, done. Yay.\\n\\nOn 31 Mar. 2017 22:31, \\\"David Steele\\\" <david@pgmasters.net> wrote:On 3/25/17 12:12 AM, Peter Eisentraut wrote:\\n\\nI'm wondering if this is a perl version/platform issue around\\n\\n $tx->pump until $stdout =~ /[[:digit:]]+[\\\\r\\\\n]$/;\\n\\nwhere we're not recognising the required output from psql when we get it.\\n\\nWhat's in src/test/recovery/tmp_check/log/regress_log_011* ?\\n\\nI couldn't use PostgresNode::psql or PostgresNode::safe_psql here\\nbecause the session must stay open.\\n\\n\\nThe problem was that psql needs to be called with -X. Fix committed.\\n\\n\\nIt's not clear to me what remains to be done on this patch. I feel, at the least, that patches 3 and 4 need to be rebased and the status set back to \\\"Needs Review\\\".\\n\\nThis thread has been idle for six days. Please respond with a new patch by 2017-04-04 00:00 AoE (UTC-12) or this submission will be marked \\\"Returned with Feedback\\\".I consider the feature complete and committed.Patch 3 is optional and per discussion buried upthread we generally agreed that it'd be better to replace most user visible uses of the 'xid' data type with a new epoch extended 'xid64' or similar.Patch 4, txid_incinerate, has never been intended for commit. It's a testing tool.Patches 1 and 2 were the key parts and thanks to Robert's helpful review, advice and edits they're committed now.Committed, done. Yay.\",\n \"msg_date\": \"Fri, 31 Mar 2017 22:46:33 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On 3/31/17 10:46 AM, Craig Ringer wrote:\\n>\\n> Patches 1 and 2 were the key parts and thanks to Robert's helpful\\n> review, advice and edits they're committed now.\\n>\\n> Committed, done. Yay.\\n\\nExcellent. I have marked this a \\\"Committed\\\" by Robert.\\n\\nOne down...\\n\\n-- \\n-David\\ndavid@pgmasters.net\\n\\n\\n-- \\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, 31 Mar 2017 11:08:05 -0400\",\n \"msg_from\": \"David Steele <david@pgmasters.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 31, 2017 at 11:08 AM, David Steele <david@pgmasters.net> wrote:\\n> On 3/31/17 10:46 AM, Craig Ringer wrote:\\n>> Patches 1 and 2 were the key parts and thanks to Robert's helpful\\n>> review, advice and edits they're committed now.\\n>>\\n>> Committed, done. Yay.\\n>\\n> Excellent. I have marked this a \\\"Committed\\\" by Robert.\\n>\\n> One down...\\n\\n...71 to go?\\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, 31 Mar 2017 13:28:15 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Transaction traceability - txid_status(bigint)\"\n }\n]","isConversational":false}}},{"rowIdx":42996,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Make initdb's suggested \"pg_ctl start\" command line more reliable.\n\nThe original coding here was not nearly careful enough about quoting\nspecial characters, and it didn't get corner cases right for constructing\nthe pg_ctl path either. Use join_path_components() and appendShellString()\nto do it honestly, so that the string will more likely work if blindly\ncopied-and-pasted.\n\nWhile at it, teach appendShellString() not to quote strings that clearly\ndon't need it, so that the output from initdb doesn't become uglier than\nit was before in typical cases where quoting is not needed.\n\nRyan Murphy, reviewed by Michael Paquier and myself\n\nDiscussion: <CAHeEsBeAe1FeBypT3E8R1ZVZU0e8xv3A-7BHg6bEOi=jZny2Uw@mail.gmail.com>\n\nBranch\n------\nmaster\n\nDetails\n-------\nhttp://git.postgresql.org/pg/commitdiff/a00c58314745772f6c6a49b6d02a9572cd600bda\n\nModified Files\n--------------\nsrc/bin/initdb/Makefile | 3 +++\nsrc/bin/initdb/initdb.c | 43 ++++++++++++++++++++++++++++---------------\nsrc/fe_utils/string_utils.c | 17 +++++++++++++++--\n3 files changed, 46 insertions(+), 17 deletions(-)\n\n\n-- \nSent via pgsql-committers mailing list (pgsql-committers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-committers\n","msg_date":"Sat, 20 Aug 2016 19:05:30 +0000","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"pgsql: Make initdb's suggested \"pg_ctl start\" command line more\n reliabl"},{"msg_contents":"On 8/20/16 3:05 PM, Tom Lane wrote:\n> Make initdb's suggested \"pg_ctl start\" command line more reliable.\n> \n> The original coding here was not nearly careful enough about quoting\n> special characters, and it didn't get corner cases right for constructing\n> the pg_ctl path either. Use join_path_components() and appendShellString()\n> to do it honestly, so that the string will more likely work if blindly\n> copied-and-pasted.\n> \n> While at it, teach appendShellString() not to quote strings that clearly\n> don't need it, so that the output from initdb doesn't become uglier than\n> it was before in typical cases where quoting is not needed.\n\nA couple of problems with this:\n\nThe not-quoting-if-not-needed doesn't appear to do anything useful for me:\n\n 'pg-install/bin/pg_ctl' -D 'pg-install/var/data' -l logfile start\n\nThe indentation of that line was changed from 4 to 10. I don't think\nthat was a good change.\n\nAs just mentioned elsewhere, this accidentally introduces a failure if\nthe PostgreSQL installation path contains LF/CR, because of the use of\nappendShellString().\n\n-- \nPeter Eisentraut 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":"Tue, 6 Sep 2016 12:19:31 -0400","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: [COMMITTERS] pgsql: Make initdb's suggested \"pg_ctl start\"\n command line more reliabl"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n> On 8/20/16 3:05 PM, Tom Lane wrote:\n>> Make initdb's suggested \"pg_ctl start\" command line more reliable.\n\n> A couple of problems with this:\n\n> The not-quoting-if-not-needed doesn't appear to do anything useful for me:\n> 'pg-install/bin/pg_ctl' -D 'pg-install/var/data' -l logfile start\n\nDash is considered a character that needs quoting. It might be possible\nto avoid that if we could be certain that appendShellString's output would\nnever be placed in a spot where it could be taken for a switch, but that\nseems like a large assumption to me. Or maybe we could treat it as\nforcing quotes only if it starts the string; but I don't personally care\nenough to write the code for that. Feel free if you want to.\n\n> The indentation of that line was changed from 4 to 10. I don't think\n> that was a good change.\n\nHmm, not intentional ... [ looks closely... ] Looks like a tab got in\nthere somehow. Will fix.\n\n> As just mentioned elsewhere, this accidentally introduces a failure if\n> the PostgreSQL installation path contains LF/CR, because of the use of\n> appendShellString().\n\nI think that's intentional, not accidental. What actual use case is\nthere for allowing such paths?\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, 06 Sep 2016 13:08:51 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Make initdb's suggested \"pg_ctl start\"\n command line more reliabl"},{"msg_contents":"On Tue, Sep 6, 2016 at 2:08 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> The not-quoting-if-not-needed doesn't appear to do anything useful for me:\n>> 'pg-install/bin/pg_ctl' -D 'pg-install/var/data' -l logfile start\n>\n> Dash is considered a character that needs quoting. It might be possible\n> to avoid that if we could be certain that appendShellString's output would\n> never be placed in a spot where it could be taken for a switch, but that\n> seems like a large assumption to me. Or maybe we could treat it as\n> forcing quotes only if it starts the string; but I don't personally care\n> enough to write the code for that. Feel free if you want to.\n\n\nWouldn't it be taken for a switch even with quoting?\n\nQuoting \"-D\" seems to work fine, which would suggest the above is true.\n\n\n-- \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, 6 Sep 2016 14:20:08 -0300","msg_from":"Claudio Freire <klaussfreire@gmail.com>","msg_from_op":false,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Make initdb's suggested \"pg_ctl\n start\" command line more reliabl"},{"msg_contents":"On 2016-09-06 13:08:51 -0400, Tom Lane wrote:\n> Dash is considered a character that needs quoting. It might be possible\n> to avoid that if we could be certain that appendShellString's output would\n> never be placed in a spot where it could be taken for a switch, but that\n> seems like a large assumption to me. Or maybe we could treat it as\n> forcing quotes only if it starts the string; but I don't personally care\n> enough to write the code for that. Feel free if you want to.\n\nBut quotes are interpreted by the shell, not by the called program. So I\ndon't think putting --whatnot in quotes changes anything here?\n\n\n-- \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, 6 Sep 2016 10:23:11 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Make initdb's suggested\n \"pg_ctl start\" command line more reliabl"},{"msg_contents":"Claudio Freire <klaussfreire@gmail.com> writes:\n> On Tue, Sep 6, 2016 at 2:08 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> Dash is considered a character that needs quoting. It might be possible\n>> to avoid that if we could be certain that appendShellString's output would\n>> never be placed in a spot where it could be taken for a switch, but that\n>> seems like a large assumption to me.\n\n> Wouldn't it be taken for a switch even with quoting?\n> Quoting \"-D\" seems to work fine, which would suggest the above is true.\n\n[ thinks about that... ] Oh, you're right, brain fade on my part. The\nshell doesn't care whether words are switches or not. So actually the\nrisk-factor for us is whether we have designed any command-line syntaxes\nin a way that would allow a path starting with a dash to cause bad things\nto happen. I have a feeling the answer is \"yes\", even without considering\nthe prospect that GNU getopt will arbitrarily rearrange the command words\non us depending on what it thinks is a switch. (Maybe leading-dash is\nanother one of the things we'd better make a policy against.)\n\nBut meanwhile, yes, the argument for treating it as quotable in\nappendShellString seems completely bogus. I'll go change that.\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, 06 Sep 2016 13:41:26 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Make initdb's suggested \"pg_ctl start\"\n command line more reliabl"},{"msg_contents":"On 9/6/16 1:08 PM, Tom Lane wrote:\n>> As just mentioned elsewhere, this accidentally introduces a failure if\n>> > the PostgreSQL installation path contains LF/CR, because of the use of\n>> > appendShellString().\n> I think that's intentional, not accidental. What actual use case is\n> there for allowing such paths?\n\nThere probably isn't one. But we ought to be introducing this change in\na more intentional and consistent way.\n\nFor example, pg_basebackup has no such restriction. So using\npg_basebackup, then promote, then pg_upgrade will (probably) fail now\nfor some paths.\n\nMore generally, I'm concerned that appendShellString() looks pretty\nattractive for future use. It's not inconceivable that someone will\nwant to use it for say calling pg_dump from pg_dumpall or pg_upgrade at\nsome point, and then maybe we'll accidentally disallow LF/CR in\ntablespace names, say.\n\nAlso, if we're concerned about the potential for confusion that these\ncharacters can cause, maybe we should be disallowing more control\ncharacters in similar places.\n\n-- \nPeter Eisentraut 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, 8 Sep 2016 12:07:35 -0400","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Make initdb's suggested \"pg_ctl\n start\" command line more reliabl"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n> More generally, I'm concerned that appendShellString() looks pretty\n> attractive for future use. It's not inconceivable that someone will\n> want to use it for say calling pg_dump from pg_dumpall or pg_upgrade at\n> some point, and then maybe we'll accidentally disallow LF/CR in\n> tablespace names, say.\n\nThat's fair, but I'm not sure how you propose to avoid it, given that\nwe don't have a method for safely quoting those characters on Windows.\n\nI would certainly far rather find a way to do that and then remove the\nprohibition. But lacking such a solution, we're going to be forced into\nmessy compromises.\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, 08 Sep 2016 12:48:33 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Make initdb's suggested \"pg_ctl start\"\n command line more reliabl"}],"string":"[\n {\n \"msg_contents\": \"Make initdb's suggested \\\"pg_ctl start\\\" command line more reliable.\\n\\nThe original coding here was not nearly careful enough about quoting\\nspecial characters, and it didn't get corner cases right for constructing\\nthe pg_ctl path either. Use join_path_components() and appendShellString()\\nto do it honestly, so that the string will more likely work if blindly\\ncopied-and-pasted.\\n\\nWhile at it, teach appendShellString() not to quote strings that clearly\\ndon't need it, so that the output from initdb doesn't become uglier than\\nit was before in typical cases where quoting is not needed.\\n\\nRyan Murphy, reviewed by Michael Paquier and myself\\n\\nDiscussion: <CAHeEsBeAe1FeBypT3E8R1ZVZU0e8xv3A-7BHg6bEOi=jZny2Uw@mail.gmail.com>\\n\\nBranch\\n------\\nmaster\\n\\nDetails\\n-------\\nhttp://git.postgresql.org/pg/commitdiff/a00c58314745772f6c6a49b6d02a9572cd600bda\\n\\nModified Files\\n--------------\\nsrc/bin/initdb/Makefile | 3 +++\\nsrc/bin/initdb/initdb.c | 43 ++++++++++++++++++++++++++++---------------\\nsrc/fe_utils/string_utils.c | 17 +++++++++++++++--\\n3 files changed, 46 insertions(+), 17 deletions(-)\\n\\n\\n-- \\nSent via pgsql-committers mailing list (pgsql-committers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-committers\\n\",\n \"msg_date\": \"Sat, 20 Aug 2016 19:05:30 +0000\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"pgsql: Make initdb's suggested \\\"pg_ctl start\\\" command line more\\n reliabl\"\n },\n {\n \"msg_contents\": \"On 8/20/16 3:05 PM, Tom Lane wrote:\\n> Make initdb's suggested \\\"pg_ctl start\\\" command line more reliable.\\n> \\n> The original coding here was not nearly careful enough about quoting\\n> special characters, and it didn't get corner cases right for constructing\\n> the pg_ctl path either. Use join_path_components() and appendShellString()\\n> to do it honestly, so that the string will more likely work if blindly\\n> copied-and-pasted.\\n> \\n> While at it, teach appendShellString() not to quote strings that clearly\\n> don't need it, so that the output from initdb doesn't become uglier than\\n> it was before in typical cases where quoting is not needed.\\n\\nA couple of problems with this:\\n\\nThe not-quoting-if-not-needed doesn't appear to do anything useful for me:\\n\\n 'pg-install/bin/pg_ctl' -D 'pg-install/var/data' -l logfile start\\n\\nThe indentation of that line was changed from 4 to 10. I don't think\\nthat was a good change.\\n\\nAs just mentioned elsewhere, this accidentally introduces a failure if\\nthe PostgreSQL installation path contains LF/CR, because of the use of\\nappendShellString().\\n\\n-- \\nPeter Eisentraut 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\": \"Tue, 6 Sep 2016 12:19:31 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [COMMITTERS] pgsql: Make initdb's suggested \\\"pg_ctl start\\\"\\n command line more reliabl\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n> On 8/20/16 3:05 PM, Tom Lane wrote:\\n>> Make initdb's suggested \\\"pg_ctl start\\\" command line more reliable.\\n\\n> A couple of problems with this:\\n\\n> The not-quoting-if-not-needed doesn't appear to do anything useful for me:\\n> 'pg-install/bin/pg_ctl' -D 'pg-install/var/data' -l logfile start\\n\\nDash is considered a character that needs quoting. It might be possible\\nto avoid that if we could be certain that appendShellString's output would\\nnever be placed in a spot where it could be taken for a switch, but that\\nseems like a large assumption to me. Or maybe we could treat it as\\nforcing quotes only if it starts the string; but I don't personally care\\nenough to write the code for that. Feel free if you want to.\\n\\n> The indentation of that line was changed from 4 to 10. I don't think\\n> that was a good change.\\n\\nHmm, not intentional ... [ looks closely... ] Looks like a tab got in\\nthere somehow. Will fix.\\n\\n> As just mentioned elsewhere, this accidentally introduces a failure if\\n> the PostgreSQL installation path contains LF/CR, because of the use of\\n> appendShellString().\\n\\nI think that's intentional, not accidental. What actual use case is\\nthere for allowing such paths?\\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, 06 Sep 2016 13:08:51 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Make initdb's suggested \\\"pg_ctl start\\\"\\n command line more reliabl\"\n },\n {\n \"msg_contents\": \"On Tue, Sep 6, 2016 at 2:08 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> The not-quoting-if-not-needed doesn't appear to do anything useful for me:\\n>> 'pg-install/bin/pg_ctl' -D 'pg-install/var/data' -l logfile start\\n>\\n> Dash is considered a character that needs quoting. It might be possible\\n> to avoid that if we could be certain that appendShellString's output would\\n> never be placed in a spot where it could be taken for a switch, but that\\n> seems like a large assumption to me. Or maybe we could treat it as\\n> forcing quotes only if it starts the string; but I don't personally care\\n> enough to write the code for that. Feel free if you want to.\\n\\n\\nWouldn't it be taken for a switch even with quoting?\\n\\nQuoting \\\"-D\\\" seems to work fine, which would suggest the above is true.\\n\\n\\n-- \\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, 6 Sep 2016 14:20:08 -0300\",\n \"msg_from\": \"Claudio Freire <klaussfreire@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Make initdb's suggested \\\"pg_ctl\\n start\\\" command line more reliabl\"\n },\n {\n \"msg_contents\": \"On 2016-09-06 13:08:51 -0400, Tom Lane wrote:\\n> Dash is considered a character that needs quoting. It might be possible\\n> to avoid that if we could be certain that appendShellString's output would\\n> never be placed in a spot where it could be taken for a switch, but that\\n> seems like a large assumption to me. Or maybe we could treat it as\\n> forcing quotes only if it starts the string; but I don't personally care\\n> enough to write the code for that. Feel free if you want to.\\n\\nBut quotes are interpreted by the shell, not by the called program. So I\\ndon't think putting --whatnot in quotes changes anything here?\\n\\n\\n-- \\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, 6 Sep 2016 10:23:11 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Make initdb's suggested\\n \\\"pg_ctl start\\\" command line more reliabl\"\n },\n {\n \"msg_contents\": \"Claudio Freire <klaussfreire@gmail.com> writes:\\n> On Tue, Sep 6, 2016 at 2:08 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> Dash is considered a character that needs quoting. It might be possible\\n>> to avoid that if we could be certain that appendShellString's output would\\n>> never be placed in a spot where it could be taken for a switch, but that\\n>> seems like a large assumption to me.\\n\\n> Wouldn't it be taken for a switch even with quoting?\\n> Quoting \\\"-D\\\" seems to work fine, which would suggest the above is true.\\n\\n[ thinks about that... ] Oh, you're right, brain fade on my part. The\\nshell doesn't care whether words are switches or not. So actually the\\nrisk-factor for us is whether we have designed any command-line syntaxes\\nin a way that would allow a path starting with a dash to cause bad things\\nto happen. I have a feeling the answer is \\\"yes\\\", even without considering\\nthe prospect that GNU getopt will arbitrarily rearrange the command words\\non us depending on what it thinks is a switch. (Maybe leading-dash is\\nanother one of the things we'd better make a policy against.)\\n\\nBut meanwhile, yes, the argument for treating it as quotable in\\nappendShellString seems completely bogus. I'll go change that.\\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, 06 Sep 2016 13:41:26 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Make initdb's suggested \\\"pg_ctl start\\\"\\n command line more reliabl\"\n },\n {\n \"msg_contents\": \"On 9/6/16 1:08 PM, Tom Lane wrote:\\n>> As just mentioned elsewhere, this accidentally introduces a failure if\\n>> > the PostgreSQL installation path contains LF/CR, because of the use of\\n>> > appendShellString().\\n> I think that's intentional, not accidental. What actual use case is\\n> there for allowing such paths?\\n\\nThere probably isn't one. But we ought to be introducing this change in\\na more intentional and consistent way.\\n\\nFor example, pg_basebackup has no such restriction. So using\\npg_basebackup, then promote, then pg_upgrade will (probably) fail now\\nfor some paths.\\n\\nMore generally, I'm concerned that appendShellString() looks pretty\\nattractive for future use. It's not inconceivable that someone will\\nwant to use it for say calling pg_dump from pg_dumpall or pg_upgrade at\\nsome point, and then maybe we'll accidentally disallow LF/CR in\\ntablespace names, say.\\n\\nAlso, if we're concerned about the potential for confusion that these\\ncharacters can cause, maybe we should be disallowing more control\\ncharacters in similar places.\\n\\n-- \\nPeter Eisentraut 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, 8 Sep 2016 12:07:35 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Make initdb's suggested \\\"pg_ctl\\n start\\\" command line more reliabl\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n> More generally, I'm concerned that appendShellString() looks pretty\\n> attractive for future use. It's not inconceivable that someone will\\n> want to use it for say calling pg_dump from pg_dumpall or pg_upgrade at\\n> some point, and then maybe we'll accidentally disallow LF/CR in\\n> tablespace names, say.\\n\\nThat's fair, but I'm not sure how you propose to avoid it, given that\\nwe don't have a method for safely quoting those characters on Windows.\\n\\nI would certainly far rather find a way to do that and then remove the\\nprohibition. But lacking such a solution, we're going to be forced into\\nmessy compromises.\\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, 08 Sep 2016 12:48:33 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Make initdb's suggested \\\"pg_ctl start\\\"\\n command line more reliabl\"\n }\n]","isConversational":false}}},{"rowIdx":42997,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi all\n\nMy earlier efforts at logical decoding of sequence advances were too\nsimplistic[1], falling afoul of issues with sequences being both\ntransactional and not transactional depending on whether the sequence is\ncreated in the current xact or not.\n\nTL;DR of solution:\n\n* extend SeqTableData and xl_seq_rec with xid\n* set xid in SeqTableData during DefineSequence()\n* if SeqTableData set, write it in xl_seq_rec when writing xlog\n* assign sequence update to specified xid's reorder buffer in decoding if\nxid set, otherwise decode immediately\n\n\nThe problem:\n\nIf the sequence is created in the current xact (i.e. uncommitted) we have\nto add the sequence updates to that xact to be replayed only if it commits.\nThe sequence is visible only to the toplevel xact that created the sequence\nso advances of it can only come from that xact and its children. The actual\nCREATE SEQUENCE is presumed to be handled separately by an event trigger or\nsimilar.\n\nIf the new sequence is committed we must replay sequence advances\nimmediately and non-transactionally to ensure they're not lost due to xact\nrollback or replayed in the wrong order due to xact commit order.\n\nIf the sequence is ALTERed in a way that changes pg_class that's event\ntriggers' job and sequence decoding doesn't care. If it's ALTERed in a way\nthat changes Form_pg_sequence we replay the change immediately, using the\nlast committed snapshot to get the sequence details, so the change will\ntake immediate effect and is retained whether or not any pg_class changes\nare committed. This reflects how it happens on the upstream.\n\n\nPlanned solution:\n\nExtend xl_seq_rec with a created_in_xid TransactionId field. If\ncreated_in_xid != InvalidTransactionId, logical decoding associates the\nsequence advance with the given toplevel xact and adds it to the reorder\nbuffer instead of immediately invoking the sequence decoding callback. The\ndecoding callback then gets invoked during ReorderBufferCommit processing\nat the appropriate time, like any other transactional change.\n\nTo determine whether to log an xid for the sequence advance we need some\nbackend local state to determine whether the sequence is new in this xact.\nHandily we already have one, the seqhashtab of SeqTableData in sequence.c,\njust where it's needed. So all that's needed is to add a TransactionId\nfield that we set if we created that sequence in this session. If it's set\nwe test it for TransactionIsInProgress() when xlog'ing a sequence advance;\nif it is, log that xid. If not in progress, clear the xid in SeqTableData\nentry so we don't check again.\n\n\n\nAnother approach would be, during decoding, to look up the relfilenode of\nthe sequence to get the sequence oid and do a pg_class lookup. Check to see\nwhether xmin is part of an in-progress xact. If so, add the sequence\nadvance to that xact's reorder buffer, otherwise decode it immediately. The\nproblem is that\n\n(a) I think we lack relfilenode-to-oid mapping information at\ndecoding-time. RelidByRelfilenode() needs a snapshot and is invoked\nduring ReorderBufferCommit(). We have make the transactional vs\nnontransactional decision in LogicalDecodingProcessRecord() when I'm pretty\nsure we don't have a snapshot.\n\n(b) It also has issues with ALTER TRANSACTION. We must replay decoded xact\nupdates immediately even if some in-flight xact has modified the pg_class\nentry for the sequence. So we can't just check whether the xmin is one of\nour xact's (sub)xids, we must also check whether some older tuple for the\nsame sequence oid has a corresponding xmax and keep walking backwards until\nwe determine whether we originally CREATEd the sequence in this xact or\nonly ALTERed it.\n\n\nSo yeah. I think extending SeqTableData and xl_seq_rec with xid is the way\nto go. Objections?\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\nHi allMy earlier efforts at logical decoding of sequence advances were too simplistic[1], falling afoul of issues with sequences being both transactional and not transactional depending on whether the sequence is created in the current xact or not.TL;DR of solution:* extend SeqTableData and xl_seq_rec with xid* set xid in SeqTableData during DefineSequence()* if SeqTableData set, write it in xl_seq_rec when writing xlog* assign sequence update to specified xid's reorder buffer in decoding if xid set, otherwise decode immediatelyThe problem:If the sequence is created in the current xact (i.e. uncommitted) we have to add the sequence updates to that xact to be replayed only if it commits. The sequence is visible only to the toplevel xact that created the sequence so advances of it can only come from that xact and its children. The actual CREATE SEQUENCE is presumed to be handled separately by an event trigger or similar.If the new sequence is committed we must replay sequence advances immediately and non-transactionally to ensure they're not lost due to xact rollback or replayed in the wrong order due to xact commit order.If the sequence is ALTERed in a way that changes pg_class that's event triggers' job and sequence decoding doesn't care. If it's ALTERed in a way that changes Form_pg_sequence we replay the change immediately, using the last committed snapshot to get the sequence details, so the change will take immediate effect and is retained whether or not any pg_class changes are committed. This reflects how it happens on the upstream.Planned solution:Extend xl_seq_rec with a created_in_xid TransactionId field. If created_in_xid != InvalidTransactionId, logical decoding associates the sequence advance with the given toplevel xact and adds it to the reorder buffer instead of immediately invoking the sequence decoding callback. The decoding callback then gets invoked during ReorderBufferCommit processing at the appropriate time, like any other transactional change.To determine whether to log an xid for the sequence advance we need some backend local state to determine whether the sequence is new in this xact. Handily we already have one, the seqhashtab of SeqTableData in sequence.c, just where it's needed. So all that's needed is to add a TransactionId field that we set if we created that sequence in this session. If it's set we test it for TransactionIsInProgress() when xlog'ing a sequence advance; if it is, log that xid. If not in progress, clear the xid in SeqTableData entry so we don't check again.Another approach would be, during decoding, to look up the relfilenode of the sequence to get the sequence oid and do a pg_class lookup. Check to see whether xmin is part of an in-progress xact. If so, add the sequence advance to that xact's reorder buffer, otherwise decode it immediately. The problem is that(a) I think we lack relfilenode-to-oid mapping information at decoding-time. RelidByRelfilenode() needs a snapshot and is invoked during ReorderBufferCommit(). We have make the transactional vs nontransactional decision in LogicalDecodingProcessRecord() when I'm pretty sure we don't have a snapshot.(b) It also has issues with ALTER TRANSACTION. We must replay decoded xact updates immediately even if some in-flight xact has modified the pg_class entry for the sequence. So we can't just check whether the xmin is one of our xact's (sub)xids, we must also check whether some older tuple for the same sequence oid has a corresponding xmax and keep walking backwards until we determine whether we originally CREATEd the sequence in this xact or only ALTERed it.So yeah. I think extending SeqTableData and xl_seq_rec with xid is the way to go. Objections?-- Craig Ringer http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services","msg_date":"Mon, 22 Aug 2016 11:13:31 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Logical decoding of sequence advances, part II"},{"msg_contents":"On 22 August 2016 at 11:13, Craig Ringer <craig@2ndquadrant.com> wrote:\n\n\n> So yeah. I think extending SeqTableData and xl_seq_rec with xid is the way\n> to go. Objections?\n>\n\n\nPrototyping this shows that re-using SeqTableData to store the creator xid\nwon't work out. We can't just store the toplevel xid because TRUNCATE ...\nRESTART IDENTITY is transactional, using a new relfilenode and new sequence\n\"timeline\". If we just store and xlog the toplevel xid as the sequence's\ncreator/restarter we'll fail to correctly handle it if a subxact that did\nTRUNCATE ... RESTART IDENTITY rolls back, e.g.\n\nCREATE TABLE x(id serial not null);\nSELECT nextval('x_id_seq'); => 1\nBEGIN;\nSELECT nextval('x_id_seq'); => 2\nSAVEPOINT sp1;\nSELECT nextval('x_id_seq'); => 3\nTRUNCATE TABLE x RESTART IDENTITY;\nSELECT nextval('x_id_seq'); => 1\nROLLBACK TO SAVEPOINT sp1;\nSELECT nextval('x_id_seq'); => 4\n\nsequence.c:init_sequence() detects this by noticing that the relfilenode\nhas changed and discarding cached values, resuming at last_value. Knowledge\nof whether we created the sequence relfilenode is not retained so we can't\ndo something similar.\n\nUnless anyone has any clever (or obvious but not to me) solutions to this,\nI'll probably need to maintain a separate map of sequence relfilenodes we\ncreated and which xid we created them in, so we can test whether that xid\nis still in progress when logging a change. It's still pretty much free\nwhen wal_level < logical or the current xact hasn't created any sequences.\n\nOtherwise I could store a List of xids in the SeqTableData for the sequence\nand check that for in-progress xids. It'd usually be NIL. If not, it'll\nalmost always be a 1-item List, the creating / resetting xid. If subxacts\nare involved it'll become a stack. We walk down the stack checking whether\nxacts are in progress and popping them if not until we find an in-progress\nentry or run out of stack and set it to NIL.\n\nEither will produce the same desired result: the correct subxact xid for\nthe innermost in-progress xact that created or reset this sequence, if any.\n\n(I initially planned to just punt on TRUNCATE and let event triggers handle\nit, but the need to roll back sequence advances if a TRUNCATE ... RESTART\nIDENTITY is rolled back means sequence decoding must pay attention to it).\n\n\nI'm also having trouble working out how to get a historical snapshot for\nthe most recent committed xact in a decoding session so the sequence's name\ncan be looked up by oid in the relcache during decoding. Advice would be\nwelcome if anyone can spare a moment.\n\nI'll keep working on this concurrent with some higher priority work.\nSuggestions, advice, or screams of horror welcomed. I think we really,\nreally need logical decoding of sequence advances...\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\nOn 22 August 2016 at 11:13, Craig Ringer <craig@2ndquadrant.com> wrote: So yeah. I think extending SeqTableData and xl_seq_rec with xid is the way to go. Objections?Prototyping this shows that re-using SeqTableData to store the creator xid won't work out. We can't just store the toplevel xid because TRUNCATE ... RESTART IDENTITY is transactional, using a new relfilenode and new sequence \"timeline\". If we just store and xlog the toplevel xid as the sequence's creator/restarter we'll fail to correctly handle it if a subxact that did TRUNCATE ... RESTART IDENTITY rolls back, e.g. CREATE TABLE x(id serial not null);SELECT nextval('x_id_seq'); => 1BEGIN;SELECT nextval('x_id_seq'); => 2SAVEPOINT sp1;SELECT nextval('x_id_seq'); => 3TRUNCATE TABLE x RESTART IDENTITY;SELECT nextval('x_id_seq'); => 1ROLLBACK TO SAVEPOINT sp1;SELECT nextval('x_id_seq'); => 4sequence.c:init_sequence() detects this by noticing that the relfilenode has changed and discarding cached values, resuming at last_value. Knowledge of whether we created the sequence relfilenode is not retained so we can't do something similar.Unless anyone has any clever (or obvious but not to me) solutions to this, I'll probably need to maintain a separate map of sequence relfilenodes we created and which xid we created them in, so we can test whether that xid is still in progress when logging a change. It's still pretty much free when wal_level < logical or the current xact hasn't created any sequences.Otherwise I could store a List of xids in the SeqTableData for the sequence and check that for in-progress xids. It'd usually be NIL. If not, it'll almost always be a 1-item List, the creating / resetting xid. If subxacts are involved it'll become a stack. We walk down the stack checking whether xacts are in progress and popping them if not until we find an in-progress entry or run out of stack and set it to NIL.Either will produce the same desired result: the correct subxact xid for the innermost in-progress xact that created or reset this sequence, if any.(I initially planned to just punt on TRUNCATE and let event triggers handle it, but the need to roll back sequence advances if a TRUNCATE ... RESTART IDENTITY is rolled back means sequence decoding must pay attention to it).I'm also having trouble working out how to get a historical snapshot for the most recent committed xact in a decoding session so the sequence's name can be looked up by oid in the relcache during decoding. Advice would be welcome if anyone can spare a moment.I'll keep working on this concurrent with some higher priority work. Suggestions, advice, or screams of horror welcomed. I think we really, really need logical decoding of sequence advances...-- Craig Ringer http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services","msg_date":"Mon, 22 Aug 2016 13:28:22 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Logical decoding of sequence advances, part II"},{"msg_contents":"On Sun, Aug 21, 2016 at 11:13 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> If the sequence is created in the current xact (i.e. uncommitted) we have to\n> add the sequence updates to that xact to be replayed only if it commits. The\n> sequence is visible only to the toplevel xact that created the sequence so\n> advances of it can only come from that xact and its children. The actual\n> CREATE SEQUENCE is presumed to be handled separately by an event trigger or\n> similar.\n>\n> If the new sequence is committed we must replay sequence advances\n> immediately and non-transactionally to ensure they're not lost due to xact\n> rollback or replayed in the wrong order due to xact commit order.\n\nSo, I wish I could give you some better advice on this topic, but\nsadly I am not an expert in this area. However, it seems to me that\nthis is just one facet of a much more general problem: given two\ntransactions T1 and T2, the order of replay must match the order of\ncommit unless you can prove that there are no dependencies between\nthem. I don't see why it matters whether the operations are sequence\noperations or data operations; it's just a question of whether they're\nmodifying the same \"stuff\".\n\nOf course, it's possible I'm missing something important 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":"Mon, 22 Aug 2016 16:29:12 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Logical decoding of sequence advances, part II"},{"msg_contents":"On 2016-08-22 16:29:12 -0400, Robert Haas wrote:\n> So, I wish I could give you some better advice on this topic, but\n> sadly I am not an expert in this area. However, it seems to me that\n> this is just one facet of a much more general problem: given two\n> transactions T1 and T2, the order of replay must match the order of\n> commit unless you can prove that there are no dependencies between\n> them. I don't see why it matters whether the operations are sequence\n> operations or data operations; it's just a question of whether they're\n> modifying the same \"stuff\".\n> \n> Of course, it's possible I'm missing something important here...\n\nMaybe that normally logical decoding outputs stuff in commit order?\n\nAndres\n\n\n-- \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 Aug 2016 13:32:29 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Logical decoding of sequence advances, part II"},{"msg_contents":"On Mon, Aug 22, 2016 at 3:29 PM, Robert Haas <robertmhaas@gmail.com> wrote:\n\n> it seems to me that\n> this is just one facet of a much more general problem: given two\n> transactions T1 and T2, the order of replay must match the order of\n> commit unless you can prove that there are no dependencies between\n> them. I don't see why it matters whether the operations are sequence\n> operations or data operations; it's just a question of whether they're\n> modifying the same \"stuff\".\n\nThe commit order is the simplest and safest *unless* there is a\nread-write anti-dependency a/k/a read-write dependency a/k/a\nrw-conflict: where a read from one transaction sees the \"before\"\nversion of data modified by the other transaction. In such a case\nit is necessary for correct serializable transaction behavior for\nthe transaction that read the \"before\" image to be replayed before\nthe write it didn't see, regardless of commit order. If you're not\ntrying to avoid serialization anomalies, it is less clear to me\nwhat is best.\n\n--\nKevin Grittner\nEDB: 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, 22 Aug 2016 16:43:40 -0500","msg_from":"Kevin Grittner <kgrittn@gmail.com>","msg_from_op":false,"msg_subject":"Re: Logical decoding of sequence advances, part II"},{"msg_contents":"On 23 Aug 2016 05:43, \"Kevin Grittner\" <kgrittn@gmail.com> wrote:\n>\n> On Mon, Aug 22, 2016 at 3:29 PM, Robert Haas <robertmhaas@gmail.com>\nwrote:\n>\n> > it seems to me that\n> > this is just one facet of a much more general problem: given two\n> > transactions T1 and T2, the order of replay must match the order of\n> > commit unless you can prove that there are no dependencies between\n> > them. I don't see why it matters whether the operations are sequence\n> > operations or data operations; it's just a question of whether they're\n> > modifying the same \"stuff\".\n\nIt matters because sequence operations aren't transactional in pg. Except\nwhen they are - operations on a newly CREATEd sequence or one where we did\na TRUNCATE ...RESTART IDENTITY.\n\nBut we don't store the xid of the xact associated with a transactional\nsequence update along with the sequence update anywhere. We just rely on nk\nother xact knowing to look at the sequence relfilenode we're changing.\nDoesn't work so well in logical rep.\n\nWe also don't store knowledge of whether or not the sequence advance is\ntransactional. Again important because for two xacts t1 and t2:\n\n* Sequence last value is 50\n\n* T1 calls nextval. Needs a new chunk because all cached values have been\nused. Writes sequence wal advancing seq last_value to 100, returns 51.\n\n* T2 calls nextval, gets cached value 52.\n\n* T2 commits\n\n* Master crashes and we fail over to replica.\n\nThis is fine for physical rep. We replay the sequence advance and all is\nwell.\n\nBut for logical rep the sequence can't be treated as part of t1. If t1\nrolls back or we fail over before replying it we might return value 52 from\nnextval even though we replayed and committed t2 that used value 52. Oops.\n\nHowever if some xact t3 creates a sequence we can't replay updates to it\nuntil the sequence relation is committed. And it's even more fun with\nTRUNCATE ... RESTART IDENTITY where we need rollback behaviour too.\n\nMake sense? It's hard because sequences are sometimes but not always exrmpt\nfrom transactional behaviour and pg doesn't record when, since it can rely\non physical wal redo order and can apply sequence advances before the\nsequence relation is committed yet.\n\n>\n> The commit order is the simplest and safest *unless* there is a\n> read-write anti-dependency a/k/a read-write dependency a/k/a\n> rw-conflict: where a read from one transaction sees the \"before\"\n> version of data modified by the other transaction. In such a case\n> it is necessary for correct serializable transaction behavior for\n> the transaction that read the \"before\" image to be replayed before\n> the write it didn't see, regardless of commit order. If you're not\n> trying to avoid serialization anomalies, it is less clear to me\n> what is best.\n\nCould you provide an example of a case where xacts replayed in commit order\nwill produce incorrect results?\n\nRemember that we aren't doing statement based replication in pg logical\ndecoding/replication. We don't care how a row got changed, only that we\nmake consistent transitions from before state to after state to for each\ntransaction, such that the data committed and visible on the master is\nvisible on the standby and no uncommitted or not yet visible data on the\nmaster is committed/visible on the replica. The replica should have visible\ncommitted data matching the master as it was when it originally executed\nthe xact we most recently replayed.\n\nNo locking is decoded or replayed. It is not expected that a normal non\nreplication client executing some other concurrent xact will have the same\neffect if run on standby as on master.\n\nIt's replication not tightly coupled clustering. If/when we have things\nlike parallel decoding and replay of concurrent xacts then issues like the\ndependencies you mention will start to become a concern. We are a long way\nfrom there.\n\nFor sequences the requirement IMO is that the sequence advances on the\nreplica to or past the position it was at on the master when the first xact\nthat saw those sequence values committed. We should never see the sequence\n'behind' such that calling nextval on the replica can produce a value\nalready seen and stored by some committed xact on the replica. Being a bit\nahead is ok, much like pg discards sequence values on crash.\n\nThat's not that hard. The problems arise when the sequence it's self isn't\ncommitted yet, per above.\n\n\nOn 23 Aug 2016 05:43, \"Kevin Grittner\" <kgrittn@gmail.com> wrote:\n>\n> On Mon, Aug 22, 2016 at 3:29 PM, Robert Haas <robertmhaas@gmail.com> wrote:\n>\n> > it seems to me that\n> > this is just one facet of a much more general problem: given two\n> > transactions T1 and T2, the order of replay must match the order of\n> > commit unless you can prove that there are no dependencies between\n> > them. I don't see why it matters whether the operations are sequence\n> > operations or data operations; it's just a question of whether they're\n> > modifying the same \"stuff\".\nIt matters because sequence operations aren't transactional in pg. Except when they are - operations on a newly CREATEd sequence or one where we did a TRUNCATE ...RESTART IDENTITY. \nBut we don't store the xid of the xact associated with a transactional sequence update along with the sequence update anywhere. We just rely on nk other xact knowing to look at the sequence relfilenode we're changing. Doesn't work so well in logical rep.\nWe also don't store knowledge of whether or not the sequence advance is transactional. Again important because for two xacts t1 and t2:\n* Sequence last value is 50\n* T1 calls nextval. Needs a new chunk because all cached values have been used. Writes sequence wal advancing seq last_value to 100, returns 51.\n* T2 calls nextval, gets cached value 52.\n* T2 commits\n* Master crashes and we fail over to replica.\nThis is fine for physical rep. We replay the sequence advance and all is well.\nBut for logical rep the sequence can't be treated as part of t1. If t1 rolls back or we fail over before replying it we might return value 52 from nextval even though we replayed and committed t2 that used value 52. Oops.\nHowever if some xact t3 creates a sequence we can't replay updates to it until the sequence relation is committed. And it's even more fun with TRUNCATE ... RESTART IDENTITY where we need rollback behaviour too.\nMake sense? It's hard because sequences are sometimes but not always exrmpt from transactional behaviour and pg doesn't record when, since it can rely on physical wal redo order and can apply sequence advances before the sequence relation is committed yet.\n>\n> The commit order is the simplest and safest *unless* there is a\n> read-write anti-dependency a/k/a read-write dependency a/k/a\n> rw-conflict: where a read from one transaction sees the \"before\"\n> version of data modified by the other transaction. In such a case\n> it is necessary for correct serializable transaction behavior for\n> the transaction that read the \"before\" image to be replayed before\n> the write it didn't see, regardless of commit order. If you're not\n> trying to avoid serialization anomalies, it is less clear to me\n> what is best.\nCould you provide an example of a case where xacts replayed in commit order will produce incorrect results? \nRemember that we aren't doing statement based replication in pg logical decoding/replication. We don't care how a row got changed, only that we make consistent transitions from before state to after state to for each transaction, such that the data committed and visible on the master is visible on the standby and no uncommitted or not yet visible data on the master is committed/visible on the replica. The replica should have visible committed data matching the master as it was when it originally executed the xact we most recently replayed. \nNo locking is decoded or replayed. It is not expected that a normal non replication client executing some other concurrent xact will have the same effect if run on standby as on master.\nIt's replication not tightly coupled clustering. If/when we have things like parallel decoding and replay of concurrent xacts then issues like the dependencies you mention will start to become a concern. We are a long way from there.\nFor sequences the requirement IMO is that the sequence advances on the replica to or past the position it was at on the master when the first xact that saw those sequence values committed. We should never see the sequence 'behind' such that calling nextval on the replica can produce a value already seen and stored by some committed xact on the replica. Being a bit ahead is ok, much like pg discards sequence values on crash.\nThat's not that hard. The problems arise when the sequence it's self isn't committed yet, per above.","msg_date":"Tue, 23 Aug 2016 07:39:20 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Logical decoding of sequence advances, part II"},{"msg_contents":"On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> On 23 Aug 2016 05:43, \"Kevin Grittner\" <kgrittn@gmail.com> wrote:\n>> On Mon, Aug 22, 2016 at 3:29 PM, Robert Haas <robertmhaas@gmail.com> wrote:\n>>\n>>> it seems to me that\n>>> this is just one facet of a much more general problem: given two\n>>> transactions T1 and T2, the order of replay must match the order of\n>>> commit unless you can prove that there are no dependencies between\n>>> them. I don't see why it matters whether the operations are sequence\n>>> operations or data operations; it's just a question of whether they're\n>>> modifying the same \"stuff\".\n\n>> The commit order is the simplest and safest *unless* there is a\n>> read-write anti-dependency a/k/a read-write dependency a/k/a\n>> rw-conflict: where a read from one transaction sees the \"before\"\n>> version of data modified by the other transaction. In such a case\n>> it is necessary for correct serializable transaction behavior for\n>> the transaction that read the \"before\" image to be replayed before\n>> the write it didn't see, regardless of commit order. If you're not\n>> trying to avoid serialization anomalies, it is less clear to me\n>> what is best.\n>\n> Could you provide an example of a case where xacts replayed in\n> commit order will produce incorrect results?\n\nhttps://wiki.postgresql.org/wiki/SSI#Deposit_Report\n\n... where T3 is on the replication target.\n\n-- \nKevin Grittner\nEDB: 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 Aug 2016 07:10:35 -0500","msg_from":"Kevin Grittner <kgrittn@gmail.com>","msg_from_op":false,"msg_subject":"Re: Logical decoding of sequence advances, part II"},{"msg_contents":"On Tue, Aug 23, 2016 at 7:10 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\n> On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n\n>> Could you provide an example of a case where xacts replayed in\n>> commit order will produce incorrect results?\n>\n> https://wiki.postgresql.org/wiki/SSI#Deposit_Report\n>\n> ... where T3 is on the replication target.\n\nI should, perhaps, have mentioned that the cases where this is are\nproblem are \"eventually consistent\" -- it's a matter of being able\nto see a state that violates business rule invariants or where data\nwhich is \"locked down\" according to one part of the database is\nstill changing. Such problems are prevented on a single database,\nbut would not be prevented on a logical replica where transactions\nare applied in commit order. Given enough time, the replica would\neventually settle into a state without the anomalies, similar to\nsome other products with eventual consistency.\n\n--\nKevin Grittner\nEDB: 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 Aug 2016 07:26:31 -0500","msg_from":"Kevin Grittner <kgrittn@gmail.com>","msg_from_op":false,"msg_subject":"Re: Logical decoding of sequence advances, part II"},{"msg_contents":"On 23 Aug 2016 20:10, \"Kevin Grittner\" <kgrittn@gmail.com> wrote:\n>\n> On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com>\nwrote:\n> > On 23 Aug 2016 05:43, \"Kevin Grittner\" <kgrittn@gmail.com> wrote:\n> >> On Mon, Aug 22, 2016 at 3:29 PM, Robert Haas <robertmhaas@gmail.com>\nwrote:\n> >>\n> >>> it seems to me that\n> >>> this is just one facet of a much more general problem: given two\n> >>> transactions T1 and T2, the order of replay must match the order of\n> >>> commit unless you can prove that there are no dependencies between\n> >>> them. I don't see why it matters whether the operations are sequence\n> >>> operations or data operations; it's just a question of whether they're\n> >>> modifying the same \"stuff\".\n>\n> >> The commit order is the simplest and safest *unless* there is a\n> >> read-write anti-dependency a/k/a read-write dependency a/k/a\n> >> rw-conflict: where a read from one transaction sees the \"before\"\n> >> version of data modified by the other transaction. In such a case\n> >> it is necessary for correct serializable transaction behavior for\n> >> the transaction that read the \"before\" image to be replayed before\n> >> the write it didn't see, regardless of commit order. If you're not\n> >> trying to avoid serialization anomalies, it is less clear to me\n> >> what is best.\n> >\n> > Could you provide an example of a case where xacts replayed in\n> > commit order will produce incorrect results?\n>\n> https://wiki.postgresql.org/wiki/SSI#Deposit_Report\n>\n> ... where T3 is on the replication target.\n\nRight. But we don't attempt to replicate locking let alone SSI state. As I\nsaid this is expected. If T1, T2 and T3 run in the master in either READ\nCOMMITTED or SERIALIZABLE we will correctly replay whatever got committed\nand leave the replica in the same state as the master.\n\nIt is row level replication so there is no simple way to detect this\nanomaly. We would have to send a lot of co-ordination data *in both\ndirections*, right?\n\nSounds like a job for tightly coupled clustering with a GTM, GLM etc. We're\na very very long way from that.\n\nThe docs probably need to be fairly explicit about the guarantees made and\nnot made and how behavior of queries on replicas can differ from queries on\nmaster. The same is true for physical replication though, right?\n\nGood point that xacts run on replicas can differ in results from the same\nquery on master though. At least in read only SERIALIZABLE xacts which are\neffectively downgraded to REPEATABLE READ (snapshot) with respect to xacts\non master. But continue to get full SSI with respect to other (possibly\nr/w) xacts on the replica. We can't really just disallow SERIALIZABLE\nisolation xacts because they remain useful. But we might want to be able to\nat least warn if they touch replicated (subscriber) tables.\n\n\nOn 23 Aug 2016 20:10, \"Kevin Grittner\" <kgrittn@gmail.com> wrote:\n>\n> On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> > On 23 Aug 2016 05:43, \"Kevin Grittner\" <kgrittn@gmail.com> wrote:\n> >> On Mon, Aug 22, 2016 at 3:29 PM, Robert Haas <robertmhaas@gmail.com> wrote:\n> >>\n> >>> it seems to me that\n> >>> this is just one facet of a much more general problem: given two\n> >>> transactions T1 and T2, the order of replay must match the order of\n> >>> commit unless you can prove that there are no dependencies between\n> >>> them. I don't see why it matters whether the operations are sequence\n> >>> operations or data operations; it's just a question of whether they're\n> >>> modifying the same \"stuff\".\n>\n> >> The commit order is the simplest and safest *unless* there is a\n> >> read-write anti-dependency a/k/a read-write dependency a/k/a\n> >> rw-conflict: where a read from one transaction sees the \"before\"\n> >> version of data modified by the other transaction. In such a case\n> >> it is necessary for correct serializable transaction behavior for\n> >> the transaction that read the \"before\" image to be replayed before\n> >> the write it didn't see, regardless of commit order. If you're not\n> >> trying to avoid serialization anomalies, it is less clear to me\n> >> what is best.\n> >\n> > Could you provide an example of a case where xacts replayed in\n> > commit order will produce incorrect results?\n>\n> https://wiki.postgresql.org/wiki/SSI#Deposit_Report\n>\n> ... where T3 is on the replication target.\nRight. But we don't attempt to replicate locking let alone SSI state. As I said this is expected. If T1, T2 and T3 run in the master in either READ COMMITTED or SERIALIZABLE we will correctly replay whatever got committed and leave the replica in the same state as the master.\nIt is row level replication so there is no simple way to detect this anomaly. We would have to send a lot of co-ordination data *in both directions*, right?\nSounds like a job for tightly coupled clustering with a GTM, GLM etc. We're a very very long way from that.\nThe docs probably need to be fairly explicit about the guarantees made and not made and how behavior of queries on replicas can differ from queries on master. The same is true for physical replication though, right?\nGood point that xacts run on replicas can differ in results from the same query on master though. At least in read only SERIALIZABLE xacts which are effectively downgraded to REPEATABLE READ (snapshot) with respect to xacts on master. But continue to get full SSI with respect to other (possibly r/w) xacts on the replica. We can't really just disallow SERIALIZABLE isolation xacts because they remain useful. But we might want to be able to at least warn if they touch replicated (subscriber) tables.","msg_date":"Tue, 23 Aug 2016 20:40:17 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Logical decoding of sequence advances, part II"},{"msg_contents":"On Mon, Aug 22, 2016 at 4:32 PM, Andres Freund <andres@anarazel.de> wrote:\n> On 2016-08-22 16:29:12 -0400, Robert Haas wrote:\n>> So, I wish I could give you some better advice on this topic, but\n>> sadly I am not an expert in this area. However, it seems to me that\n>> this is just one facet of a much more general problem: given two\n>> transactions T1 and T2, the order of replay must match the order of\n>> commit unless you can prove that there are no dependencies between\n>> them. I don't see why it matters whether the operations are sequence\n>> operations or data operations; it's just a question of whether they're\n>> modifying the same \"stuff\".\n>>\n>> Of course, it's possible I'm missing something important here...\n>\n> Maybe that normally logical decoding outputs stuff in commit order?\n\nAs slow as I sometimes am, I did know that. :-)\n\nI think what I was missing is that nextval() operations are\nnon-transactional. Craig describes them as non-transactional except\nwhen they are transactional, but I think that's not really the right\nway of looking at it. Creating the sequence is transactional, and\nupdating the value is not. What seems to be causing trouble for Craig\nis that if the nextval() operation is inserted into the replication\nstream non-transactionally, it can happen before the sequence actually\ngets created. I'm wondering if we can't find a way to make it so that\nit's OK for those operations to happen out of order, rather than\ntrying to make the nextval() operation sometimes transactional and\nother times non-transactional.\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 Aug 2016 09:59:33 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Logical decoding of sequence advances, part II"},{"msg_contents":"On Tue, Aug 23, 2016 at 7:40 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> On 23 Aug 2016 20:10, \"Kevin Grittner\" <kgrittn@gmail.com> wrote:\n>>\n>> On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com>\n\n>>> Could you provide an example of a case where xacts replayed in\n>>> commit order will produce incorrect results?\n>>\n>> https://wiki.postgresql.org/wiki/SSI#Deposit_Report\n>>\n>> ... where T3 is on the replication target.\n>\n> Right. But we don't attempt to replicate locking let alone SSI state. As I\n> said this is expected. If T1, T2 and T3 run in the master in either READ\n> COMMITTED or SERIALIZABLE we will correctly replay whatever got committed\n> and leave the replica in the same state as the master.\n\nEventually. Between the commit of T3 and T2 a state can be seen on\nthe replica which would not have been allowed on the source.\n\n> It is row level replication so there is no simple way to detect this\n> anomaly.\n\nThat is probably true, but there is a way to *prevent* it.\n\n> We would have to send a lot of co-ordination data *in both\n> directions*, right?\n\nNo. The source has all the information about both commit order and\nread-write dependencies, and could produce a stream of transaction\nIDs to specify the safe order for applying transactions to prevent\nthe anomaly from appearing on the replica. In this case the commit\norder is T1->T3->T2, but the apparent order of execution (AOoE) is\nT1->T2->T3. If the source communicated that to the replica, and\nthe replica held up application of any changes from T3 until T2 was\ncommitted there would be no chance to read incorrect results. It\nwould not matter if T2 and T3 were committed on the replica\nsimultaneously or in AOoE, as long as the work of T3 does not\nappear before the work of T2.\n\nThe replica does not need to send anything back to the source for\nthis to work.\n\n--\nKevin Grittner\nEDB: 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 Aug 2016 09:07:50 -0500","msg_from":"Kevin Grittner <kgrittn@gmail.com>","msg_from_op":false,"msg_subject":"Re: Logical decoding of sequence advances, part II"},{"msg_contents":"On Tue, Aug 23, 2016 at 9:07 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\n> On Tue, Aug 23, 2016 at 7:40 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n>> On 23 Aug 2016 20:10, \"Kevin Grittner\" <kgrittn@gmail.com> wrote:\n>>>\n>>> On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com>\n>\n>>>> Could you provide an example of a case where xacts replayed in\n>>>> commit order will produce incorrect results?\n>>>\n>>> https://wiki.postgresql.org/wiki/SSI#Deposit_Report\n>>>\n>>> ... where T3 is on the replication target.\n>>\n>> Right. But we don't attempt to replicate locking let alone SSI state. As I\n>> said this is expected. If T1, T2 and T3 run in the master in either READ\n>> COMMITTED or SERIALIZABLE we will correctly replay whatever got committed\n>> and leave the replica in the same state as the master.\n>\n> Eventually. Between the commit of T3 and T2 a state can be seen on\n> the replica which would not have been allowed on the source.\n>\n>> It is row level replication so there is no simple way to detect this\n>> anomaly.\n>\n> That is probably true, but there is a way to *prevent* it.\n>\n>> We would have to send a lot of co-ordination data *in both\n>> directions*, right?\n>\n> No. The source has all the information about both commit order and\n> read-write dependencies, and could produce a stream of transaction\n> IDs to specify the safe order for applying transactions to prevent\n> the anomaly from appearing on the replica. In this case the commit\n> order is T1->T3->T2, but the apparent order of execution (AOoE) is\n> T1->T2->T3.\n\nSorry, trying to keep this conversation going while doing something\nelse and sent a response there which doesn't really make sense,\nsince the issue is whether to allow T3 to run *on the replica*.\nI'll take another look when I'm less distracted. You may be right.\n\n--\nKevin Grittner\nEDB: 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 Aug 2016 09:36:51 -0500","msg_from":"Kevin Grittner <kgrittn@gmail.com>","msg_from_op":false,"msg_subject":"Re: Logical decoding of sequence advances, part II"},{"msg_contents":"On 23/08/16 15:59, Robert Haas wrote:\n> On Mon, Aug 22, 2016 at 4:32 PM, Andres Freund <andres@anarazel.de> wrote:\n>> On 2016-08-22 16:29:12 -0400, Robert Haas wrote:\n>>> So, I wish I could give you some better advice on this topic, but\n>>> sadly I am not an expert in this area. However, it seems to me that\n>>> this is just one facet of a much more general problem: given two\n>>> transactions T1 and T2, the order of replay must match the order of\n>>> commit unless you can prove that there are no dependencies between\n>>> them. I don't see why it matters whether the operations are sequence\n>>> operations or data operations; it's just a question of whether they're\n>>> modifying the same \"stuff\".\n>>>\n>>> Of course, it's possible I'm missing something important here...\n>>\n>> Maybe that normally logical decoding outputs stuff in commit order?\n>\n> As slow as I sometimes am, I did know that. :-)\n>\n> I think what I was missing is that nextval() operations are\n> non-transactional. Craig describes them as non-transactional except\n> when they are transactional, but I think that's not really the right\n> way of looking at it. Creating the sequence is transactional, and\n> updating the value is not. What seems to be causing trouble for Craig\n> is that if the nextval() operation is inserted into the replication\n> stream non-transactionally, it can happen before the sequence actually\n> gets created. I'm wondering if we can't find a way to make it so that\n> it's OK for those operations to happen out of order, rather than\n> trying to make the nextval() operation sometimes transactional and\n> other times non-transactional.\n>\n\nWell, that's what Craig is trying to do by tracking if the transactional \nchange has happend on a sequence in current transaction, no?\n\n-- \n Petr Jelinek 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, 23 Aug 2016 17:03:03 +0200","msg_from":"Petr Jelinek <petr@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Logical decoding of sequence advances, part II"},{"msg_contents":"On Tue, Aug 23, 2016 at 9:36 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\n> On Tue, Aug 23, 2016 at 9:07 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\n>> On Tue, Aug 23, 2016 at 7:40 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n>>> On 23 Aug 2016 20:10, \"Kevin Grittner\" <kgrittn@gmail.com> wrote:\n>>>>\n>>>> On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com>\n>>\n>>>>> Could you provide an example of a case where xacts replayed in\n>>>>> commit order will produce incorrect results?\n>>>>\n>>>> https://wiki.postgresql.org/wiki/SSI#Deposit_Report\n>>>>\n>>>> ... where T3 is on the replication target.\n>>>\n>>> Right. But we don't attempt to replicate locking let alone SSI state. As I\n>>> said this is expected. If T1, T2 and T3 run in the master in either READ\n>>> COMMITTED or SERIALIZABLE we will correctly replay whatever got committed\n>>> and leave the replica in the same state as the master.\n>>\n>> Eventually. Between the commit of T3 and T2 a state can be seen on\n>> the replica which would not have been allowed on the source.\n>>\n>>> It is row level replication so there is no simple way to detect this\n>>> anomaly.\n>>\n>> That is probably true, but there is a way to *prevent* it.\n>>\n>>> We would have to send a lot of co-ordination data *in both\n>>> directions*, right?\n>>\n>> No. The source has all the information about both commit order and\n>> read-write dependencies, and could produce a stream of transaction\n>> IDs to specify the safe order for applying transactions to prevent\n>> the anomaly from appearing on the replica. In this case the commit\n>> order is T1->T3->T2, but the apparent order of execution (AOoE) is\n>> T1->T2->T3.\n>\n> Sorry, trying to keep this conversation going while doing something\n> else and sent a response there which doesn't really make sense,\n> since the issue is whether to allow T3 to run *on the replica*.\n> I'll take another look when I'm less distracted. You may be right.\n\nI had the right idea, but messed up the details. The source has\ncommit order T2->T1 and AOoE T1->T2. So as long as making any work\nfrom T2 visible is held up until at least the point where the work\nof T1 is visible, T3 on the replica cannot see an anomalous state.\n\nIt is still true that a one-way stream of information from the\nprimary to the replica regarding AOoE would be sufficient\ncommunication.\n\n--\nKevin Grittner\nEDB: 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 Aug 2016 11:26:35 -0500","msg_from":"Kevin Grittner <kgrittn@gmail.com>","msg_from_op":false,"msg_subject":"Re: Logical decoding of sequence advances, part II"},{"msg_contents":"On 24 August 2016 at 00:26, Kevin Grittner <kgrittn@gmail.com> wrote:\n\n> On Tue, Aug 23, 2016 at 9:36 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\n> > On Tue, Aug 23, 2016 at 9:07 AM, Kevin Grittner <kgrittn@gmail.com>\n> wrote:\n> >> On Tue, Aug 23, 2016 at 7:40 AM, Craig Ringer <craig@2ndquadrant.com>\n> wrote:\n> >>> On 23 Aug 2016 20:10, \"Kevin Grittner\" <kgrittn@gmail.com> wrote:\n> >>>>\n> >>>> On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com>\n> >>\n> >>>>> Could you provide an example of a case where xacts replayed in\n> >>>>> commit order will produce incorrect results?\n> >>>>\n> >>>> https://wiki.postgresql.org/wiki/SSI#Deposit_Report\n> >>>>\n> >>>> ... where T3 is on the replication target.\n> >>>\n> >>> Right. But we don't attempt to replicate locking let alone SSI state.\n> As I\n> >>> said this is expected. If T1, T2 and T3 run in the master in either\n> READ\n> >>> COMMITTED or SERIALIZABLE we will correctly replay whatever got\n> committed\n> >>> and leave the replica in the same state as the master.\n> >>\n> >> Eventually. Between the commit of T3 and T2 a state can be seen on\n> >> the replica which would not have been allowed on the source.\n> >>\n> >>> It is row level replication so there is no simple way to detect this\n> >>> anomaly.\n> >>\n> >> That is probably true, but there is a way to *prevent* it.\n> >>\n> >>> We would have to send a lot of co-ordination data *in both\n> >>> directions*, right?\n> >>\n> >> No. The source has all the information about both commit order and\n> >> read-write dependencies, and could produce a stream of transaction\n> >> IDs to specify the safe order for applying transactions to prevent\n> >> the anomaly from appearing on the replica. In this case the commit\n> >> order is T1->T3->T2, but the apparent order of execution (AOoE) is\n> >> T1->T2->T3.\n> >\n> > Sorry, trying to keep this conversation going while doing something\n> > else and sent a response there which doesn't really make sense,\n> > since the issue is whether to allow T3 to run *on the replica*.\n> > I'll take another look when I'm less distracted. You may be right.\n>\n> I had the right idea, but messed up the details. The source has\n> commit order T2->T1 and AOoE T1->T2. So as long as making any work\n> from T2 visible is held up until at least the point where the work\n> of T1 is visible, T3 on the replica cannot see an anomalous state.\n>\n> It is still true that a one-way stream of information from the\n> primary to the replica regarding AOoE would be sufficient\n> communication.\n>\n\nHm. That's really interesting. So we could achieve SSI-like properties\nfor read-only xacts on a replica if we could defer xact decoding, or buffer\nthe xact on the downstream and defer apply, based on SSI info from the\nupstream. That'd be pretty cool.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\nOn 24 August 2016 at 00:26, Kevin Grittner <kgrittn@gmail.com> wrote:On Tue, Aug 23, 2016 at 9:36 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\n> On Tue, Aug 23, 2016 at 9:07 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\n>> On Tue, Aug 23, 2016 at 7:40 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n>>> On 23 Aug 2016 20:10, \"Kevin Grittner\" <kgrittn@gmail.com> wrote:\n>>>>\n>>>> On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com>\n>>\n>>>>> Could you provide an example of a case where xacts replayed in\n>>>>> commit order will produce incorrect results?\n>>>>\n>>>> https://wiki.postgresql.org/wiki/SSI#Deposit_Report\n>>>>\n>>>> ... where T3 is on the replication target.\n>>>\n>>> Right. But we don't attempt to replicate locking let alone SSI state. As I\n>>> said this is expected. If T1, T2 and T3 run in the master in either READ\n>>> COMMITTED or SERIALIZABLE we will correctly replay whatever got committed\n>>> and leave the replica in the same state as the master.\n>>\n>> Eventually. Between the commit of T3 and T2 a state can be seen on\n>> the replica which would not have been allowed on the source.\n>>\n>>> It is row level replication so there is no simple way to detect this\n>>> anomaly.\n>>\n>> That is probably true, but there is a way to *prevent* it.\n>>\n>>> We would have to send a lot of co-ordination data *in both\n>>> directions*, right?\n>>\n>> No. The source has all the information about both commit order and\n>> read-write dependencies, and could produce a stream of transaction\n>> IDs to specify the safe order for applying transactions to prevent\n>> the anomaly from appearing on the replica. In this case the commit\n>> order is T1->T3->T2, but the apparent order of execution (AOoE) is\n>> T1->T2->T3.\n>\n> Sorry, trying to keep this conversation going while doing something\n> else and sent a response there which doesn't really make sense,\n> since the issue is whether to allow T3 to run *on the replica*.\n> I'll take another look when I'm less distracted. You may be right.\n\nI had the right idea, but messed up the details. The source has\ncommit order T2->T1 and AOoE T1->T2. So as long as making any work\nfrom T2 visible is held up until at least the point where the work\nof T1 is visible, T3 on the replica cannot see an anomalous state.\n\nIt is still true that a one-way stream of information from the\nprimary to the replica regarding AOoE would be sufficient\ncommunication. Hm. That's really interesting. So we could achieve SSI-like properties for read-only xacts on a replica if we could defer xact decoding, or buffer the xact on the downstream and defer apply, based on SSI info from the upstream. That'd be pretty cool.-- Craig Ringer http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services","msg_date":"Wed, 24 Aug 2016 00:38:20 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Logical decoding of sequence advances, part II"},{"msg_contents":"On 2016-08-23 07:26:31 -0500, Kevin Grittner wrote:\n> On Tue, Aug 23, 2016 at 7:10 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\n> > On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> \n> >> Could you provide an example of a case where xacts replayed in\n> >> commit order will produce incorrect results?\n> >\n> > https://wiki.postgresql.org/wiki/SSI#Deposit_Report\n> >\n> > ... where T3 is on the replication target.\n> \n> I should, perhaps, have mentioned that the cases where this is are\n> problem are \"eventually consistent\" -- it's a matter of being able\n> to see a state that violates business rule invariants or where data\n> which is \"locked down\" according to one part of the database is\n> still changing. Such problems are prevented on a single database,\n> but would not be prevented on a logical replica where transactions\n> are applied in commit order. Given enough time, the replica would\n> eventually settle into a state without the anomalies, similar to\n> some other products with eventual consistency.\n\nI've generally a bit of difficulty to see this as a significant problem\nfor logical rep, as long as hot-standby, and crash-recovery in general,\nalso has this problem...\n\n\n-- \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 Aug 2016 09:50:19 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Logical decoding of sequence advances, part II"},{"msg_contents":"On 24 August 2016 at 00:50, Andres Freund <andres@anarazel.de> wrote:\n\n> On 2016-08-23 07:26:31 -0500, Kevin Grittner wrote:\n> > On Tue, Aug 23, 2016 at 7:10 AM, Kevin Grittner <kgrittn@gmail.com>\n> wrote:\n> > [an explanation of SSI anomalies]\n>\n\nI've generally a bit of difficulty to see this as a significant problem\n> for logical rep, as long as hot-standby, and crash-recovery in general,\n> also has this problem...\n>\n\nSame here, as commented upthread. I think it'd be cool to be able to\ndeliver SSI-alike behaviour on a standby, but it's far from a priority, and\nit applies to phys ans well as logical rep.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\nOn 24 August 2016 at 00:50, Andres Freund <andres@anarazel.de> wrote:On 2016-08-23 07:26:31 -0500, Kevin Grittner wrote:\n> On Tue, Aug 23, 2016 at 7:10 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\n> [an explanation of SSI anomalies] I've generally a bit of difficulty to see this as a significant problem\nfor logical rep, as long as hot-standby, and crash-recovery in general,\nalso has this problem...\nSame here, as commented upthread. I think it'd be cool to be able to deliver SSI-alike behaviour on a standby, but it's far from a priority, and it applies to phys ans well as logical rep.-- Craig Ringer http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services","msg_date":"Wed, 24 Aug 2016 01:02:09 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Logical decoding of sequence advances, part II"},{"msg_contents":"On Tue, Aug 23, 2016 at 11:50 AM, Andres Freund <andres@anarazel.de> wrote:\n> On 2016-08-23 07:26:31 -0500, Kevin Grittner wrote:\n>> On Tue, Aug 23, 2016 at 7:10 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\n>> > On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\n>>\n>> >> Could you provide an example of a case where xacts replayed in\n>> >> commit order will produce incorrect results?\n>> >\n>> > https://wiki.postgresql.org/wiki/SSI#Deposit_Report\n>> >\n>> > ... where T3 is on the replication target.\n>>\n>> I should, perhaps, have mentioned that the cases where this is are\n>> problem are \"eventually consistent\" -- it's a matter of being able\n>> to see a state that violates business rule invariants or where data\n>> which is \"locked down\" according to one part of the database is\n>> still changing. Such problems are prevented on a single database,\n>> but would not be prevented on a logical replica where transactions\n>> are applied in commit order. Given enough time, the replica would\n>> eventually settle into a state without the anomalies, similar to\n>> some other products with eventual consistency.\n>\n> I've generally a bit of difficulty to see this as a significant problem\n> for logical rep, as long as hot-standby, and crash-recovery in general,\n> also has this problem...\n\nSerialization anomalies cannot be seen on a hot standby nor on\ncrash recovery. Granted, the mechanism which prevents it on the\nhot standby is that we don't allow the transaction isolation level\nto be set to SERIALIZABLE, to prevent the expectation that queries\nwill be free of anomalies, which is pretty crude. On crash\nrecovery you cannot see any anomalies as far as I'm aware -- what\nhas given you the idea that it is possible?\n\nI'm not sure that the same technique I described for logical\nreplication could be made to work for page-level physical\nreplication, but a \"safe snapshot\" technique has been previously\ndiscussed on the lists which would work for physical replication\n(although *that* technique seems unsuited to logical replication).\n\n--\nKevin Grittner\nEDB: 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 Aug 2016 13:03:48 -0500","msg_from":"Kevin Grittner <kgrittn@gmail.com>","msg_from_op":false,"msg_subject":"Re: Logical decoding of sequence advances, part II"}],"string":"[\n {\n \"msg_contents\": \"Hi all\\n\\nMy earlier efforts at logical decoding of sequence advances were too\\nsimplistic[1], falling afoul of issues with sequences being both\\ntransactional and not transactional depending on whether the sequence is\\ncreated in the current xact or not.\\n\\nTL;DR of solution:\\n\\n* extend SeqTableData and xl_seq_rec with xid\\n* set xid in SeqTableData during DefineSequence()\\n* if SeqTableData set, write it in xl_seq_rec when writing xlog\\n* assign sequence update to specified xid's reorder buffer in decoding if\\nxid set, otherwise decode immediately\\n\\n\\nThe problem:\\n\\nIf the sequence is created in the current xact (i.e. uncommitted) we have\\nto add the sequence updates to that xact to be replayed only if it commits.\\nThe sequence is visible only to the toplevel xact that created the sequence\\nso advances of it can only come from that xact and its children. The actual\\nCREATE SEQUENCE is presumed to be handled separately by an event trigger or\\nsimilar.\\n\\nIf the new sequence is committed we must replay sequence advances\\nimmediately and non-transactionally to ensure they're not lost due to xact\\nrollback or replayed in the wrong order due to xact commit order.\\n\\nIf the sequence is ALTERed in a way that changes pg_class that's event\\ntriggers' job and sequence decoding doesn't care. If it's ALTERed in a way\\nthat changes Form_pg_sequence we replay the change immediately, using the\\nlast committed snapshot to get the sequence details, so the change will\\ntake immediate effect and is retained whether or not any pg_class changes\\nare committed. This reflects how it happens on the upstream.\\n\\n\\nPlanned solution:\\n\\nExtend xl_seq_rec with a created_in_xid TransactionId field. If\\ncreated_in_xid != InvalidTransactionId, logical decoding associates the\\nsequence advance with the given toplevel xact and adds it to the reorder\\nbuffer instead of immediately invoking the sequence decoding callback. The\\ndecoding callback then gets invoked during ReorderBufferCommit processing\\nat the appropriate time, like any other transactional change.\\n\\nTo determine whether to log an xid for the sequence advance we need some\\nbackend local state to determine whether the sequence is new in this xact.\\nHandily we already have one, the seqhashtab of SeqTableData in sequence.c,\\njust where it's needed. So all that's needed is to add a TransactionId\\nfield that we set if we created that sequence in this session. If it's set\\nwe test it for TransactionIsInProgress() when xlog'ing a sequence advance;\\nif it is, log that xid. If not in progress, clear the xid in SeqTableData\\nentry so we don't check again.\\n\\n\\n\\nAnother approach would be, during decoding, to look up the relfilenode of\\nthe sequence to get the sequence oid and do a pg_class lookup. Check to see\\nwhether xmin is part of an in-progress xact. If so, add the sequence\\nadvance to that xact's reorder buffer, otherwise decode it immediately. The\\nproblem is that\\n\\n(a) I think we lack relfilenode-to-oid mapping information at\\ndecoding-time. RelidByRelfilenode() needs a snapshot and is invoked\\nduring ReorderBufferCommit(). We have make the transactional vs\\nnontransactional decision in LogicalDecodingProcessRecord() when I'm pretty\\nsure we don't have a snapshot.\\n\\n(b) It also has issues with ALTER TRANSACTION. We must replay decoded xact\\nupdates immediately even if some in-flight xact has modified the pg_class\\nentry for the sequence. So we can't just check whether the xmin is one of\\nour xact's (sub)xids, we must also check whether some older tuple for the\\nsame sequence oid has a corresponding xmax and keep walking backwards until\\nwe determine whether we originally CREATEd the sequence in this xact or\\nonly ALTERed it.\\n\\n\\nSo yeah. I think extending SeqTableData and xl_seq_rec with xid is the way\\nto go. Objections?\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nHi allMy earlier efforts at logical decoding of sequence advances were too simplistic[1], falling afoul of issues with sequences being both transactional and not transactional depending on whether the sequence is created in the current xact or not.TL;DR of solution:* extend SeqTableData and xl_seq_rec with xid* set xid in SeqTableData during DefineSequence()* if SeqTableData set, write it in xl_seq_rec when writing xlog* assign sequence update to specified xid's reorder buffer in decoding if xid set, otherwise decode immediatelyThe problem:If the sequence is created in the current xact (i.e. uncommitted) we have to add the sequence updates to that xact to be replayed only if it commits. The sequence is visible only to the toplevel xact that created the sequence so advances of it can only come from that xact and its children. The actual CREATE SEQUENCE is presumed to be handled separately by an event trigger or similar.If the new sequence is committed we must replay sequence advances immediately and non-transactionally to ensure they're not lost due to xact rollback or replayed in the wrong order due to xact commit order.If the sequence is ALTERed in a way that changes pg_class that's event triggers' job and sequence decoding doesn't care. If it's ALTERed in a way that changes Form_pg_sequence we replay the change immediately, using the last committed snapshot to get the sequence details, so the change will take immediate effect and is retained whether or not any pg_class changes are committed. This reflects how it happens on the upstream.Planned solution:Extend xl_seq_rec with a created_in_xid TransactionId field. If created_in_xid != InvalidTransactionId, logical decoding associates the sequence advance with the given toplevel xact and adds it to the reorder buffer instead of immediately invoking the sequence decoding callback. The decoding callback then gets invoked during ReorderBufferCommit processing at the appropriate time, like any other transactional change.To determine whether to log an xid for the sequence advance we need some backend local state to determine whether the sequence is new in this xact. Handily we already have one, the seqhashtab of SeqTableData in sequence.c, just where it's needed. So all that's needed is to add a TransactionId field that we set if we created that sequence in this session. If it's set we test it for TransactionIsInProgress() when xlog'ing a sequence advance; if it is, log that xid. If not in progress, clear the xid in SeqTableData entry so we don't check again.Another approach would be, during decoding, to look up the relfilenode of the sequence to get the sequence oid and do a pg_class lookup. Check to see whether xmin is part of an in-progress xact. If so, add the sequence advance to that xact's reorder buffer, otherwise decode it immediately. The problem is that(a) I think we lack relfilenode-to-oid mapping information at decoding-time. RelidByRelfilenode() needs a snapshot and is invoked during ReorderBufferCommit(). We have make the transactional vs nontransactional decision in LogicalDecodingProcessRecord() when I'm pretty sure we don't have a snapshot.(b) It also has issues with ALTER TRANSACTION. We must replay decoded xact updates immediately even if some in-flight xact has modified the pg_class entry for the sequence. So we can't just check whether the xmin is one of our xact's (sub)xids, we must also check whether some older tuple for the same sequence oid has a corresponding xmax and keep walking backwards until we determine whether we originally CREATEd the sequence in this xact or only ALTERed it.So yeah. I think extending SeqTableData and xl_seq_rec with xid is the way to go. Objections?-- Craig Ringer http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Mon, 22 Aug 2016 11:13:31 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Logical decoding of sequence advances, part II\"\n },\n {\n \"msg_contents\": \"On 22 August 2016 at 11:13, Craig Ringer <craig@2ndquadrant.com> wrote:\\n\\n\\n> So yeah. I think extending SeqTableData and xl_seq_rec with xid is the way\\n> to go. Objections?\\n>\\n\\n\\nPrototyping this shows that re-using SeqTableData to store the creator xid\\nwon't work out. We can't just store the toplevel xid because TRUNCATE ...\\nRESTART IDENTITY is transactional, using a new relfilenode and new sequence\\n\\\"timeline\\\". If we just store and xlog the toplevel xid as the sequence's\\ncreator/restarter we'll fail to correctly handle it if a subxact that did\\nTRUNCATE ... RESTART IDENTITY rolls back, e.g.\\n\\nCREATE TABLE x(id serial not null);\\nSELECT nextval('x_id_seq'); => 1\\nBEGIN;\\nSELECT nextval('x_id_seq'); => 2\\nSAVEPOINT sp1;\\nSELECT nextval('x_id_seq'); => 3\\nTRUNCATE TABLE x RESTART IDENTITY;\\nSELECT nextval('x_id_seq'); => 1\\nROLLBACK TO SAVEPOINT sp1;\\nSELECT nextval('x_id_seq'); => 4\\n\\nsequence.c:init_sequence() detects this by noticing that the relfilenode\\nhas changed and discarding cached values, resuming at last_value. Knowledge\\nof whether we created the sequence relfilenode is not retained so we can't\\ndo something similar.\\n\\nUnless anyone has any clever (or obvious but not to me) solutions to this,\\nI'll probably need to maintain a separate map of sequence relfilenodes we\\ncreated and which xid we created them in, so we can test whether that xid\\nis still in progress when logging a change. It's still pretty much free\\nwhen wal_level < logical or the current xact hasn't created any sequences.\\n\\nOtherwise I could store a List of xids in the SeqTableData for the sequence\\nand check that for in-progress xids. It'd usually be NIL. If not, it'll\\nalmost always be a 1-item List, the creating / resetting xid. If subxacts\\nare involved it'll become a stack. We walk down the stack checking whether\\nxacts are in progress and popping them if not until we find an in-progress\\nentry or run out of stack and set it to NIL.\\n\\nEither will produce the same desired result: the correct subxact xid for\\nthe innermost in-progress xact that created or reset this sequence, if any.\\n\\n(I initially planned to just punt on TRUNCATE and let event triggers handle\\nit, but the need to roll back sequence advances if a TRUNCATE ... RESTART\\nIDENTITY is rolled back means sequence decoding must pay attention to it).\\n\\n\\nI'm also having trouble working out how to get a historical snapshot for\\nthe most recent committed xact in a decoding session so the sequence's name\\ncan be looked up by oid in the relcache during decoding. Advice would be\\nwelcome if anyone can spare a moment.\\n\\nI'll keep working on this concurrent with some higher priority work.\\nSuggestions, advice, or screams of horror welcomed. I think we really,\\nreally need logical decoding of sequence advances...\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 22 August 2016 at 11:13, Craig Ringer <craig@2ndquadrant.com> wrote: So yeah. I think extending SeqTableData and xl_seq_rec with xid is the way to go. Objections?Prototyping this shows that re-using SeqTableData to store the creator xid won't work out. We can't just store the toplevel xid because TRUNCATE ... RESTART IDENTITY is transactional, using a new relfilenode and new sequence \\\"timeline\\\". If we just store and xlog the toplevel xid as the sequence's creator/restarter we'll fail to correctly handle it if a subxact that did TRUNCATE ... RESTART IDENTITY rolls back, e.g. CREATE TABLE x(id serial not null);SELECT nextval('x_id_seq'); => 1BEGIN;SELECT nextval('x_id_seq'); => 2SAVEPOINT sp1;SELECT nextval('x_id_seq'); => 3TRUNCATE TABLE x RESTART IDENTITY;SELECT nextval('x_id_seq'); => 1ROLLBACK TO SAVEPOINT sp1;SELECT nextval('x_id_seq'); => 4sequence.c:init_sequence() detects this by noticing that the relfilenode has changed and discarding cached values, resuming at last_value. Knowledge of whether we created the sequence relfilenode is not retained so we can't do something similar.Unless anyone has any clever (or obvious but not to me) solutions to this, I'll probably need to maintain a separate map of sequence relfilenodes we created and which xid we created them in, so we can test whether that xid is still in progress when logging a change. It's still pretty much free when wal_level < logical or the current xact hasn't created any sequences.Otherwise I could store a List of xids in the SeqTableData for the sequence and check that for in-progress xids. It'd usually be NIL. If not, it'll almost always be a 1-item List, the creating / resetting xid. If subxacts are involved it'll become a stack. We walk down the stack checking whether xacts are in progress and popping them if not until we find an in-progress entry or run out of stack and set it to NIL.Either will produce the same desired result: the correct subxact xid for the innermost in-progress xact that created or reset this sequence, if any.(I initially planned to just punt on TRUNCATE and let event triggers handle it, but the need to roll back sequence advances if a TRUNCATE ... RESTART IDENTITY is rolled back means sequence decoding must pay attention to it).I'm also having trouble working out how to get a historical snapshot for the most recent committed xact in a decoding session so the sequence's name can be looked up by oid in the relcache during decoding. Advice would be welcome if anyone can spare a moment.I'll keep working on this concurrent with some higher priority work. Suggestions, advice, or screams of horror welcomed. I think we really, really need logical decoding of sequence advances...-- Craig Ringer http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Mon, 22 Aug 2016 13:28:22 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Logical decoding of sequence advances, part II\"\n },\n {\n \"msg_contents\": \"On Sun, Aug 21, 2016 at 11:13 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> If the sequence is created in the current xact (i.e. uncommitted) we have to\\n> add the sequence updates to that xact to be replayed only if it commits. The\\n> sequence is visible only to the toplevel xact that created the sequence so\\n> advances of it can only come from that xact and its children. The actual\\n> CREATE SEQUENCE is presumed to be handled separately by an event trigger or\\n> similar.\\n>\\n> If the new sequence is committed we must replay sequence advances\\n> immediately and non-transactionally to ensure they're not lost due to xact\\n> rollback or replayed in the wrong order due to xact commit order.\\n\\nSo, I wish I could give you some better advice on this topic, but\\nsadly I am not an expert in this area. However, it seems to me that\\nthis is just one facet of a much more general problem: given two\\ntransactions T1 and T2, the order of replay must match the order of\\ncommit unless you can prove that there are no dependencies between\\nthem. I don't see why it matters whether the operations are sequence\\noperations or data operations; it's just a question of whether they're\\nmodifying the same \\\"stuff\\\".\\n\\nOf course, it's possible I'm missing something important 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\": \"Mon, 22 Aug 2016 16:29:12 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Logical decoding of sequence advances, part II\"\n },\n {\n \"msg_contents\": \"On 2016-08-22 16:29:12 -0400, Robert Haas wrote:\\n> So, I wish I could give you some better advice on this topic, but\\n> sadly I am not an expert in this area. However, it seems to me that\\n> this is just one facet of a much more general problem: given two\\n> transactions T1 and T2, the order of replay must match the order of\\n> commit unless you can prove that there are no dependencies between\\n> them. I don't see why it matters whether the operations are sequence\\n> operations or data operations; it's just a question of whether they're\\n> modifying the same \\\"stuff\\\".\\n> \\n> Of course, it's possible I'm missing something important here...\\n\\nMaybe that normally logical decoding outputs stuff in commit order?\\n\\nAndres\\n\\n\\n-- \\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 Aug 2016 13:32:29 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Logical decoding of sequence advances, part II\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 22, 2016 at 3:29 PM, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> it seems to me that\\n> this is just one facet of a much more general problem: given two\\n> transactions T1 and T2, the order of replay must match the order of\\n> commit unless you can prove that there are no dependencies between\\n> them. I don't see why it matters whether the operations are sequence\\n> operations or data operations; it's just a question of whether they're\\n> modifying the same \\\"stuff\\\".\\n\\nThe commit order is the simplest and safest *unless* there is a\\nread-write anti-dependency a/k/a read-write dependency a/k/a\\nrw-conflict: where a read from one transaction sees the \\\"before\\\"\\nversion of data modified by the other transaction. In such a case\\nit is necessary for correct serializable transaction behavior for\\nthe transaction that read the \\\"before\\\" image to be replayed before\\nthe write it didn't see, regardless of commit order. If you're not\\ntrying to avoid serialization anomalies, it is less clear to me\\nwhat is best.\\n\\n--\\nKevin Grittner\\nEDB: 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, 22 Aug 2016 16:43:40 -0500\",\n \"msg_from\": \"Kevin Grittner <kgrittn@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Logical decoding of sequence advances, part II\"\n },\n {\n \"msg_contents\": \"On 23 Aug 2016 05:43, \\\"Kevin Grittner\\\" <kgrittn@gmail.com> wrote:\\n>\\n> On Mon, Aug 22, 2016 at 3:29 PM, Robert Haas <robertmhaas@gmail.com>\\nwrote:\\n>\\n> > it seems to me that\\n> > this is just one facet of a much more general problem: given two\\n> > transactions T1 and T2, the order of replay must match the order of\\n> > commit unless you can prove that there are no dependencies between\\n> > them. I don't see why it matters whether the operations are sequence\\n> > operations or data operations; it's just a question of whether they're\\n> > modifying the same \\\"stuff\\\".\\n\\nIt matters because sequence operations aren't transactional in pg. Except\\nwhen they are - operations on a newly CREATEd sequence or one where we did\\na TRUNCATE ...RESTART IDENTITY.\\n\\nBut we don't store the xid of the xact associated with a transactional\\nsequence update along with the sequence update anywhere. We just rely on nk\\nother xact knowing to look at the sequence relfilenode we're changing.\\nDoesn't work so well in logical rep.\\n\\nWe also don't store knowledge of whether or not the sequence advance is\\ntransactional. Again important because for two xacts t1 and t2:\\n\\n* Sequence last value is 50\\n\\n* T1 calls nextval. Needs a new chunk because all cached values have been\\nused. Writes sequence wal advancing seq last_value to 100, returns 51.\\n\\n* T2 calls nextval, gets cached value 52.\\n\\n* T2 commits\\n\\n* Master crashes and we fail over to replica.\\n\\nThis is fine for physical rep. We replay the sequence advance and all is\\nwell.\\n\\nBut for logical rep the sequence can't be treated as part of t1. If t1\\nrolls back or we fail over before replying it we might return value 52 from\\nnextval even though we replayed and committed t2 that used value 52. Oops.\\n\\nHowever if some xact t3 creates a sequence we can't replay updates to it\\nuntil the sequence relation is committed. And it's even more fun with\\nTRUNCATE ... RESTART IDENTITY where we need rollback behaviour too.\\n\\nMake sense? It's hard because sequences are sometimes but not always exrmpt\\nfrom transactional behaviour and pg doesn't record when, since it can rely\\non physical wal redo order and can apply sequence advances before the\\nsequence relation is committed yet.\\n\\n>\\n> The commit order is the simplest and safest *unless* there is a\\n> read-write anti-dependency a/k/a read-write dependency a/k/a\\n> rw-conflict: where a read from one transaction sees the \\\"before\\\"\\n> version of data modified by the other transaction. In such a case\\n> it is necessary for correct serializable transaction behavior for\\n> the transaction that read the \\\"before\\\" image to be replayed before\\n> the write it didn't see, regardless of commit order. If you're not\\n> trying to avoid serialization anomalies, it is less clear to me\\n> what is best.\\n\\nCould you provide an example of a case where xacts replayed in commit order\\nwill produce incorrect results?\\n\\nRemember that we aren't doing statement based replication in pg logical\\ndecoding/replication. We don't care how a row got changed, only that we\\nmake consistent transitions from before state to after state to for each\\ntransaction, such that the data committed and visible on the master is\\nvisible on the standby and no uncommitted or not yet visible data on the\\nmaster is committed/visible on the replica. The replica should have visible\\ncommitted data matching the master as it was when it originally executed\\nthe xact we most recently replayed.\\n\\nNo locking is decoded or replayed. It is not expected that a normal non\\nreplication client executing some other concurrent xact will have the same\\neffect if run on standby as on master.\\n\\nIt's replication not tightly coupled clustering. If/when we have things\\nlike parallel decoding and replay of concurrent xacts then issues like the\\ndependencies you mention will start to become a concern. We are a long way\\nfrom there.\\n\\nFor sequences the requirement IMO is that the sequence advances on the\\nreplica to or past the position it was at on the master when the first xact\\nthat saw those sequence values committed. We should never see the sequence\\n'behind' such that calling nextval on the replica can produce a value\\nalready seen and stored by some committed xact on the replica. Being a bit\\nahead is ok, much like pg discards sequence values on crash.\\n\\nThat's not that hard. The problems arise when the sequence it's self isn't\\ncommitted yet, per above.\\n\\n\\nOn 23 Aug 2016 05:43, \\\"Kevin Grittner\\\" <kgrittn@gmail.com> wrote:\\n>\\n> On Mon, Aug 22, 2016 at 3:29 PM, Robert Haas <robertmhaas@gmail.com> wrote:\\n>\\n> > it seems to me that\\n> > this is just one facet of a much more general problem: given two\\n> > transactions T1 and T2, the order of replay must match the order of\\n> > commit unless you can prove that there are no dependencies between\\n> > them. I don't see why it matters whether the operations are sequence\\n> > operations or data operations; it's just a question of whether they're\\n> > modifying the same \\\"stuff\\\".\\nIt matters because sequence operations aren't transactional in pg. Except when they are - operations on a newly CREATEd sequence or one where we did a TRUNCATE ...RESTART IDENTITY. \\nBut we don't store the xid of the xact associated with a transactional sequence update along with the sequence update anywhere. We just rely on nk other xact knowing to look at the sequence relfilenode we're changing. Doesn't work so well in logical rep.\\nWe also don't store knowledge of whether or not the sequence advance is transactional. Again important because for two xacts t1 and t2:\\n* Sequence last value is 50\\n* T1 calls nextval. Needs a new chunk because all cached values have been used. Writes sequence wal advancing seq last_value to 100, returns 51.\\n* T2 calls nextval, gets cached value 52.\\n* T2 commits\\n* Master crashes and we fail over to replica.\\nThis is fine for physical rep. We replay the sequence advance and all is well.\\nBut for logical rep the sequence can't be treated as part of t1. If t1 rolls back or we fail over before replying it we might return value 52 from nextval even though we replayed and committed t2 that used value 52. Oops.\\nHowever if some xact t3 creates a sequence we can't replay updates to it until the sequence relation is committed. And it's even more fun with TRUNCATE ... RESTART IDENTITY where we need rollback behaviour too.\\nMake sense? It's hard because sequences are sometimes but not always exrmpt from transactional behaviour and pg doesn't record when, since it can rely on physical wal redo order and can apply sequence advances before the sequence relation is committed yet.\\n>\\n> The commit order is the simplest and safest *unless* there is a\\n> read-write anti-dependency a/k/a read-write dependency a/k/a\\n> rw-conflict: where a read from one transaction sees the \\\"before\\\"\\n> version of data modified by the other transaction. In such a case\\n> it is necessary for correct serializable transaction behavior for\\n> the transaction that read the \\\"before\\\" image to be replayed before\\n> the write it didn't see, regardless of commit order. If you're not\\n> trying to avoid serialization anomalies, it is less clear to me\\n> what is best.\\nCould you provide an example of a case where xacts replayed in commit order will produce incorrect results? \\nRemember that we aren't doing statement based replication in pg logical decoding/replication. We don't care how a row got changed, only that we make consistent transitions from before state to after state to for each transaction, such that the data committed and visible on the master is visible on the standby and no uncommitted or not yet visible data on the master is committed/visible on the replica. The replica should have visible committed data matching the master as it was when it originally executed the xact we most recently replayed. \\nNo locking is decoded or replayed. It is not expected that a normal non replication client executing some other concurrent xact will have the same effect if run on standby as on master.\\nIt's replication not tightly coupled clustering. If/when we have things like parallel decoding and replay of concurrent xacts then issues like the dependencies you mention will start to become a concern. We are a long way from there.\\nFor sequences the requirement IMO is that the sequence advances on the replica to or past the position it was at on the master when the first xact that saw those sequence values committed. We should never see the sequence 'behind' such that calling nextval on the replica can produce a value already seen and stored by some committed xact on the replica. Being a bit ahead is ok, much like pg discards sequence values on crash.\\nThat's not that hard. The problems arise when the sequence it's self isn't committed yet, per above.\",\n \"msg_date\": \"Tue, 23 Aug 2016 07:39:20 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Logical decoding of sequence advances, part II\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> On 23 Aug 2016 05:43, \\\"Kevin Grittner\\\" <kgrittn@gmail.com> wrote:\\n>> On Mon, Aug 22, 2016 at 3:29 PM, Robert Haas <robertmhaas@gmail.com> wrote:\\n>>\\n>>> it seems to me that\\n>>> this is just one facet of a much more general problem: given two\\n>>> transactions T1 and T2, the order of replay must match the order of\\n>>> commit unless you can prove that there are no dependencies between\\n>>> them. I don't see why it matters whether the operations are sequence\\n>>> operations or data operations; it's just a question of whether they're\\n>>> modifying the same \\\"stuff\\\".\\n\\n>> The commit order is the simplest and safest *unless* there is a\\n>> read-write anti-dependency a/k/a read-write dependency a/k/a\\n>> rw-conflict: where a read from one transaction sees the \\\"before\\\"\\n>> version of data modified by the other transaction. In such a case\\n>> it is necessary for correct serializable transaction behavior for\\n>> the transaction that read the \\\"before\\\" image to be replayed before\\n>> the write it didn't see, regardless of commit order. If you're not\\n>> trying to avoid serialization anomalies, it is less clear to me\\n>> what is best.\\n>\\n> Could you provide an example of a case where xacts replayed in\\n> commit order will produce incorrect results?\\n\\nhttps://wiki.postgresql.org/wiki/SSI#Deposit_Report\\n\\n... where T3 is on the replication target.\\n\\n-- \\nKevin Grittner\\nEDB: 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 Aug 2016 07:10:35 -0500\",\n \"msg_from\": \"Kevin Grittner <kgrittn@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Logical decoding of sequence advances, part II\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 23, 2016 at 7:10 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\\n> On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n\\n>> Could you provide an example of a case where xacts replayed in\\n>> commit order will produce incorrect results?\\n>\\n> https://wiki.postgresql.org/wiki/SSI#Deposit_Report\\n>\\n> ... where T3 is on the replication target.\\n\\nI should, perhaps, have mentioned that the cases where this is are\\nproblem are \\\"eventually consistent\\\" -- it's a matter of being able\\nto see a state that violates business rule invariants or where data\\nwhich is \\\"locked down\\\" according to one part of the database is\\nstill changing. Such problems are prevented on a single database,\\nbut would not be prevented on a logical replica where transactions\\nare applied in commit order. Given enough time, the replica would\\neventually settle into a state without the anomalies, similar to\\nsome other products with eventual consistency.\\n\\n--\\nKevin Grittner\\nEDB: 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 Aug 2016 07:26:31 -0500\",\n \"msg_from\": \"Kevin Grittner <kgrittn@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Logical decoding of sequence advances, part II\"\n },\n {\n \"msg_contents\": \"On 23 Aug 2016 20:10, \\\"Kevin Grittner\\\" <kgrittn@gmail.com> wrote:\\n>\\n> On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com>\\nwrote:\\n> > On 23 Aug 2016 05:43, \\\"Kevin Grittner\\\" <kgrittn@gmail.com> wrote:\\n> >> On Mon, Aug 22, 2016 at 3:29 PM, Robert Haas <robertmhaas@gmail.com>\\nwrote:\\n> >>\\n> >>> it seems to me that\\n> >>> this is just one facet of a much more general problem: given two\\n> >>> transactions T1 and T2, the order of replay must match the order of\\n> >>> commit unless you can prove that there are no dependencies between\\n> >>> them. I don't see why it matters whether the operations are sequence\\n> >>> operations or data operations; it's just a question of whether they're\\n> >>> modifying the same \\\"stuff\\\".\\n>\\n> >> The commit order is the simplest and safest *unless* there is a\\n> >> read-write anti-dependency a/k/a read-write dependency a/k/a\\n> >> rw-conflict: where a read from one transaction sees the \\\"before\\\"\\n> >> version of data modified by the other transaction. In such a case\\n> >> it is necessary for correct serializable transaction behavior for\\n> >> the transaction that read the \\\"before\\\" image to be replayed before\\n> >> the write it didn't see, regardless of commit order. If you're not\\n> >> trying to avoid serialization anomalies, it is less clear to me\\n> >> what is best.\\n> >\\n> > Could you provide an example of a case where xacts replayed in\\n> > commit order will produce incorrect results?\\n>\\n> https://wiki.postgresql.org/wiki/SSI#Deposit_Report\\n>\\n> ... where T3 is on the replication target.\\n\\nRight. But we don't attempt to replicate locking let alone SSI state. As I\\nsaid this is expected. If T1, T2 and T3 run in the master in either READ\\nCOMMITTED or SERIALIZABLE we will correctly replay whatever got committed\\nand leave the replica in the same state as the master.\\n\\nIt is row level replication so there is no simple way to detect this\\nanomaly. We would have to send a lot of co-ordination data *in both\\ndirections*, right?\\n\\nSounds like a job for tightly coupled clustering with a GTM, GLM etc. We're\\na very very long way from that.\\n\\nThe docs probably need to be fairly explicit about the guarantees made and\\nnot made and how behavior of queries on replicas can differ from queries on\\nmaster. The same is true for physical replication though, right?\\n\\nGood point that xacts run on replicas can differ in results from the same\\nquery on master though. At least in read only SERIALIZABLE xacts which are\\neffectively downgraded to REPEATABLE READ (snapshot) with respect to xacts\\non master. But continue to get full SSI with respect to other (possibly\\nr/w) xacts on the replica. We can't really just disallow SERIALIZABLE\\nisolation xacts because they remain useful. But we might want to be able to\\nat least warn if they touch replicated (subscriber) tables.\\n\\n\\nOn 23 Aug 2016 20:10, \\\"Kevin Grittner\\\" <kgrittn@gmail.com> wrote:\\n>\\n> On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> > On 23 Aug 2016 05:43, \\\"Kevin Grittner\\\" <kgrittn@gmail.com> wrote:\\n> >> On Mon, Aug 22, 2016 at 3:29 PM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> >>\\n> >>> it seems to me that\\n> >>> this is just one facet of a much more general problem: given two\\n> >>> transactions T1 and T2, the order of replay must match the order of\\n> >>> commit unless you can prove that there are no dependencies between\\n> >>> them. I don't see why it matters whether the operations are sequence\\n> >>> operations or data operations; it's just a question of whether they're\\n> >>> modifying the same \\\"stuff\\\".\\n>\\n> >> The commit order is the simplest and safest *unless* there is a\\n> >> read-write anti-dependency a/k/a read-write dependency a/k/a\\n> >> rw-conflict: where a read from one transaction sees the \\\"before\\\"\\n> >> version of data modified by the other transaction. In such a case\\n> >> it is necessary for correct serializable transaction behavior for\\n> >> the transaction that read the \\\"before\\\" image to be replayed before\\n> >> the write it didn't see, regardless of commit order. If you're not\\n> >> trying to avoid serialization anomalies, it is less clear to me\\n> >> what is best.\\n> >\\n> > Could you provide an example of a case where xacts replayed in\\n> > commit order will produce incorrect results?\\n>\\n> https://wiki.postgresql.org/wiki/SSI#Deposit_Report\\n>\\n> ... where T3 is on the replication target.\\nRight. But we don't attempt to replicate locking let alone SSI state. As I said this is expected. If T1, T2 and T3 run in the master in either READ COMMITTED or SERIALIZABLE we will correctly replay whatever got committed and leave the replica in the same state as the master.\\nIt is row level replication so there is no simple way to detect this anomaly. We would have to send a lot of co-ordination data *in both directions*, right?\\nSounds like a job for tightly coupled clustering with a GTM, GLM etc. We're a very very long way from that.\\nThe docs probably need to be fairly explicit about the guarantees made and not made and how behavior of queries on replicas can differ from queries on master. The same is true for physical replication though, right?\\nGood point that xacts run on replicas can differ in results from the same query on master though. At least in read only SERIALIZABLE xacts which are effectively downgraded to REPEATABLE READ (snapshot) with respect to xacts on master. But continue to get full SSI with respect to other (possibly r/w) xacts on the replica. We can't really just disallow SERIALIZABLE isolation xacts because they remain useful. But we might want to be able to at least warn if they touch replicated (subscriber) tables.\",\n \"msg_date\": \"Tue, 23 Aug 2016 20:40:17 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Logical decoding of sequence advances, part II\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 22, 2016 at 4:32 PM, Andres Freund <andres@anarazel.de> wrote:\\n> On 2016-08-22 16:29:12 -0400, Robert Haas wrote:\\n>> So, I wish I could give you some better advice on this topic, but\\n>> sadly I am not an expert in this area. However, it seems to me that\\n>> this is just one facet of a much more general problem: given two\\n>> transactions T1 and T2, the order of replay must match the order of\\n>> commit unless you can prove that there are no dependencies between\\n>> them. I don't see why it matters whether the operations are sequence\\n>> operations or data operations; it's just a question of whether they're\\n>> modifying the same \\\"stuff\\\".\\n>>\\n>> Of course, it's possible I'm missing something important here...\\n>\\n> Maybe that normally logical decoding outputs stuff in commit order?\\n\\nAs slow as I sometimes am, I did know that. :-)\\n\\nI think what I was missing is that nextval() operations are\\nnon-transactional. Craig describes them as non-transactional except\\nwhen they are transactional, but I think that's not really the right\\nway of looking at it. Creating the sequence is transactional, and\\nupdating the value is not. What seems to be causing trouble for Craig\\nis that if the nextval() operation is inserted into the replication\\nstream non-transactionally, it can happen before the sequence actually\\ngets created. I'm wondering if we can't find a way to make it so that\\nit's OK for those operations to happen out of order, rather than\\ntrying to make the nextval() operation sometimes transactional and\\nother times non-transactional.\\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 Aug 2016 09:59:33 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Logical decoding of sequence advances, part II\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 23, 2016 at 7:40 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> On 23 Aug 2016 20:10, \\\"Kevin Grittner\\\" <kgrittn@gmail.com> wrote:\\n>>\\n>> On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com>\\n\\n>>> Could you provide an example of a case where xacts replayed in\\n>>> commit order will produce incorrect results?\\n>>\\n>> https://wiki.postgresql.org/wiki/SSI#Deposit_Report\\n>>\\n>> ... where T3 is on the replication target.\\n>\\n> Right. But we don't attempt to replicate locking let alone SSI state. As I\\n> said this is expected. If T1, T2 and T3 run in the master in either READ\\n> COMMITTED or SERIALIZABLE we will correctly replay whatever got committed\\n> and leave the replica in the same state as the master.\\n\\nEventually. Between the commit of T3 and T2 a state can be seen on\\nthe replica which would not have been allowed on the source.\\n\\n> It is row level replication so there is no simple way to detect this\\n> anomaly.\\n\\nThat is probably true, but there is a way to *prevent* it.\\n\\n> We would have to send a lot of co-ordination data *in both\\n> directions*, right?\\n\\nNo. The source has all the information about both commit order and\\nread-write dependencies, and could produce a stream of transaction\\nIDs to specify the safe order for applying transactions to prevent\\nthe anomaly from appearing on the replica. In this case the commit\\norder is T1->T3->T2, but the apparent order of execution (AOoE) is\\nT1->T2->T3. If the source communicated that to the replica, and\\nthe replica held up application of any changes from T3 until T2 was\\ncommitted there would be no chance to read incorrect results. It\\nwould not matter if T2 and T3 were committed on the replica\\nsimultaneously or in AOoE, as long as the work of T3 does not\\nappear before the work of T2.\\n\\nThe replica does not need to send anything back to the source for\\nthis to work.\\n\\n--\\nKevin Grittner\\nEDB: 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 Aug 2016 09:07:50 -0500\",\n \"msg_from\": \"Kevin Grittner <kgrittn@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Logical decoding of sequence advances, part II\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 23, 2016 at 9:07 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\\n> On Tue, Aug 23, 2016 at 7:40 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>> On 23 Aug 2016 20:10, \\\"Kevin Grittner\\\" <kgrittn@gmail.com> wrote:\\n>>>\\n>>> On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com>\\n>\\n>>>> Could you provide an example of a case where xacts replayed in\\n>>>> commit order will produce incorrect results?\\n>>>\\n>>> https://wiki.postgresql.org/wiki/SSI#Deposit_Report\\n>>>\\n>>> ... where T3 is on the replication target.\\n>>\\n>> Right. But we don't attempt to replicate locking let alone SSI state. As I\\n>> said this is expected. If T1, T2 and T3 run in the master in either READ\\n>> COMMITTED or SERIALIZABLE we will correctly replay whatever got committed\\n>> and leave the replica in the same state as the master.\\n>\\n> Eventually. Between the commit of T3 and T2 a state can be seen on\\n> the replica which would not have been allowed on the source.\\n>\\n>> It is row level replication so there is no simple way to detect this\\n>> anomaly.\\n>\\n> That is probably true, but there is a way to *prevent* it.\\n>\\n>> We would have to send a lot of co-ordination data *in both\\n>> directions*, right?\\n>\\n> No. The source has all the information about both commit order and\\n> read-write dependencies, and could produce a stream of transaction\\n> IDs to specify the safe order for applying transactions to prevent\\n> the anomaly from appearing on the replica. In this case the commit\\n> order is T1->T3->T2, but the apparent order of execution (AOoE) is\\n> T1->T2->T3.\\n\\nSorry, trying to keep this conversation going while doing something\\nelse and sent a response there which doesn't really make sense,\\nsince the issue is whether to allow T3 to run *on the replica*.\\nI'll take another look when I'm less distracted. You may be right.\\n\\n--\\nKevin Grittner\\nEDB: 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 Aug 2016 09:36:51 -0500\",\n \"msg_from\": \"Kevin Grittner <kgrittn@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Logical decoding of sequence advances, part II\"\n },\n {\n \"msg_contents\": \"On 23/08/16 15:59, Robert Haas wrote:\\n> On Mon, Aug 22, 2016 at 4:32 PM, Andres Freund <andres@anarazel.de> wrote:\\n>> On 2016-08-22 16:29:12 -0400, Robert Haas wrote:\\n>>> So, I wish I could give you some better advice on this topic, but\\n>>> sadly I am not an expert in this area. However, it seems to me that\\n>>> this is just one facet of a much more general problem: given two\\n>>> transactions T1 and T2, the order of replay must match the order of\\n>>> commit unless you can prove that there are no dependencies between\\n>>> them. I don't see why it matters whether the operations are sequence\\n>>> operations or data operations; it's just a question of whether they're\\n>>> modifying the same \\\"stuff\\\".\\n>>>\\n>>> Of course, it's possible I'm missing something important here...\\n>>\\n>> Maybe that normally logical decoding outputs stuff in commit order?\\n>\\n> As slow as I sometimes am, I did know that. :-)\\n>\\n> I think what I was missing is that nextval() operations are\\n> non-transactional. Craig describes them as non-transactional except\\n> when they are transactional, but I think that's not really the right\\n> way of looking at it. Creating the sequence is transactional, and\\n> updating the value is not. What seems to be causing trouble for Craig\\n> is that if the nextval() operation is inserted into the replication\\n> stream non-transactionally, it can happen before the sequence actually\\n> gets created. I'm wondering if we can't find a way to make it so that\\n> it's OK for those operations to happen out of order, rather than\\n> trying to make the nextval() operation sometimes transactional and\\n> other times non-transactional.\\n>\\n\\nWell, that's what Craig is trying to do by tracking if the transactional \\nchange has happend on a sequence in current transaction, no?\\n\\n-- \\n Petr Jelinek 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, 23 Aug 2016 17:03:03 +0200\",\n \"msg_from\": \"Petr Jelinek <petr@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Logical decoding of sequence advances, part II\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 23, 2016 at 9:36 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\\n> On Tue, Aug 23, 2016 at 9:07 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\\n>> On Tue, Aug 23, 2016 at 7:40 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>>> On 23 Aug 2016 20:10, \\\"Kevin Grittner\\\" <kgrittn@gmail.com> wrote:\\n>>>>\\n>>>> On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com>\\n>>\\n>>>>> Could you provide an example of a case where xacts replayed in\\n>>>>> commit order will produce incorrect results?\\n>>>>\\n>>>> https://wiki.postgresql.org/wiki/SSI#Deposit_Report\\n>>>>\\n>>>> ... where T3 is on the replication target.\\n>>>\\n>>> Right. But we don't attempt to replicate locking let alone SSI state. As I\\n>>> said this is expected. If T1, T2 and T3 run in the master in either READ\\n>>> COMMITTED or SERIALIZABLE we will correctly replay whatever got committed\\n>>> and leave the replica in the same state as the master.\\n>>\\n>> Eventually. Between the commit of T3 and T2 a state can be seen on\\n>> the replica which would not have been allowed on the source.\\n>>\\n>>> It is row level replication so there is no simple way to detect this\\n>>> anomaly.\\n>>\\n>> That is probably true, but there is a way to *prevent* it.\\n>>\\n>>> We would have to send a lot of co-ordination data *in both\\n>>> directions*, right?\\n>>\\n>> No. The source has all the information about both commit order and\\n>> read-write dependencies, and could produce a stream of transaction\\n>> IDs to specify the safe order for applying transactions to prevent\\n>> the anomaly from appearing on the replica. In this case the commit\\n>> order is T1->T3->T2, but the apparent order of execution (AOoE) is\\n>> T1->T2->T3.\\n>\\n> Sorry, trying to keep this conversation going while doing something\\n> else and sent a response there which doesn't really make sense,\\n> since the issue is whether to allow T3 to run *on the replica*.\\n> I'll take another look when I'm less distracted. You may be right.\\n\\nI had the right idea, but messed up the details. The source has\\ncommit order T2->T1 and AOoE T1->T2. So as long as making any work\\nfrom T2 visible is held up until at least the point where the work\\nof T1 is visible, T3 on the replica cannot see an anomalous state.\\n\\nIt is still true that a one-way stream of information from the\\nprimary to the replica regarding AOoE would be sufficient\\ncommunication.\\n\\n--\\nKevin Grittner\\nEDB: 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 Aug 2016 11:26:35 -0500\",\n \"msg_from\": \"Kevin Grittner <kgrittn@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Logical decoding of sequence advances, part II\"\n },\n {\n \"msg_contents\": \"On 24 August 2016 at 00:26, Kevin Grittner <kgrittn@gmail.com> wrote:\\n\\n> On Tue, Aug 23, 2016 at 9:36 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\\n> > On Tue, Aug 23, 2016 at 9:07 AM, Kevin Grittner <kgrittn@gmail.com>\\n> wrote:\\n> >> On Tue, Aug 23, 2016 at 7:40 AM, Craig Ringer <craig@2ndquadrant.com>\\n> wrote:\\n> >>> On 23 Aug 2016 20:10, \\\"Kevin Grittner\\\" <kgrittn@gmail.com> wrote:\\n> >>>>\\n> >>>> On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com>\\n> >>\\n> >>>>> Could you provide an example of a case where xacts replayed in\\n> >>>>> commit order will produce incorrect results?\\n> >>>>\\n> >>>> https://wiki.postgresql.org/wiki/SSI#Deposit_Report\\n> >>>>\\n> >>>> ... where T3 is on the replication target.\\n> >>>\\n> >>> Right. But we don't attempt to replicate locking let alone SSI state.\\n> As I\\n> >>> said this is expected. If T1, T2 and T3 run in the master in either\\n> READ\\n> >>> COMMITTED or SERIALIZABLE we will correctly replay whatever got\\n> committed\\n> >>> and leave the replica in the same state as the master.\\n> >>\\n> >> Eventually. Between the commit of T3 and T2 a state can be seen on\\n> >> the replica which would not have been allowed on the source.\\n> >>\\n> >>> It is row level replication so there is no simple way to detect this\\n> >>> anomaly.\\n> >>\\n> >> That is probably true, but there is a way to *prevent* it.\\n> >>\\n> >>> We would have to send a lot of co-ordination data *in both\\n> >>> directions*, right?\\n> >>\\n> >> No. The source has all the information about both commit order and\\n> >> read-write dependencies, and could produce a stream of transaction\\n> >> IDs to specify the safe order for applying transactions to prevent\\n> >> the anomaly from appearing on the replica. In this case the commit\\n> >> order is T1->T3->T2, but the apparent order of execution (AOoE) is\\n> >> T1->T2->T3.\\n> >\\n> > Sorry, trying to keep this conversation going while doing something\\n> > else and sent a response there which doesn't really make sense,\\n> > since the issue is whether to allow T3 to run *on the replica*.\\n> > I'll take another look when I'm less distracted. You may be right.\\n>\\n> I had the right idea, but messed up the details. The source has\\n> commit order T2->T1 and AOoE T1->T2. So as long as making any work\\n> from T2 visible is held up until at least the point where the work\\n> of T1 is visible, T3 on the replica cannot see an anomalous state.\\n>\\n> It is still true that a one-way stream of information from the\\n> primary to the replica regarding AOoE would be sufficient\\n> communication.\\n>\\n\\nHm. That's really interesting. So we could achieve SSI-like properties\\nfor read-only xacts on a replica if we could defer xact decoding, or buffer\\nthe xact on the downstream and defer apply, based on SSI info from the\\nupstream. That'd be pretty cool.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 24 August 2016 at 00:26, Kevin Grittner <kgrittn@gmail.com> wrote:On Tue, Aug 23, 2016 at 9:36 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\\n> On Tue, Aug 23, 2016 at 9:07 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\\n>> On Tue, Aug 23, 2016 at 7:40 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>>> On 23 Aug 2016 20:10, \\\"Kevin Grittner\\\" <kgrittn@gmail.com> wrote:\\n>>>>\\n>>>> On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com>\\n>>\\n>>>>> Could you provide an example of a case where xacts replayed in\\n>>>>> commit order will produce incorrect results?\\n>>>>\\n>>>> https://wiki.postgresql.org/wiki/SSI#Deposit_Report\\n>>>>\\n>>>> ... where T3 is on the replication target.\\n>>>\\n>>> Right. But we don't attempt to replicate locking let alone SSI state. As I\\n>>> said this is expected. If T1, T2 and T3 run in the master in either READ\\n>>> COMMITTED or SERIALIZABLE we will correctly replay whatever got committed\\n>>> and leave the replica in the same state as the master.\\n>>\\n>> Eventually. Between the commit of T3 and T2 a state can be seen on\\n>> the replica which would not have been allowed on the source.\\n>>\\n>>> It is row level replication so there is no simple way to detect this\\n>>> anomaly.\\n>>\\n>> That is probably true, but there is a way to *prevent* it.\\n>>\\n>>> We would have to send a lot of co-ordination data *in both\\n>>> directions*, right?\\n>>\\n>> No. The source has all the information about both commit order and\\n>> read-write dependencies, and could produce a stream of transaction\\n>> IDs to specify the safe order for applying transactions to prevent\\n>> the anomaly from appearing on the replica. In this case the commit\\n>> order is T1->T3->T2, but the apparent order of execution (AOoE) is\\n>> T1->T2->T3.\\n>\\n> Sorry, trying to keep this conversation going while doing something\\n> else and sent a response there which doesn't really make sense,\\n> since the issue is whether to allow T3 to run *on the replica*.\\n> I'll take another look when I'm less distracted. You may be right.\\n\\nI had the right idea, but messed up the details. The source has\\ncommit order T2->T1 and AOoE T1->T2. So as long as making any work\\nfrom T2 visible is held up until at least the point where the work\\nof T1 is visible, T3 on the replica cannot see an anomalous state.\\n\\nIt is still true that a one-way stream of information from the\\nprimary to the replica regarding AOoE would be sufficient\\ncommunication. Hm. That's really interesting. So we could achieve SSI-like properties for read-only xacts on a replica if we could defer xact decoding, or buffer the xact on the downstream and defer apply, based on SSI info from the upstream. That'd be pretty cool.-- Craig Ringer http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Wed, 24 Aug 2016 00:38:20 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Logical decoding of sequence advances, part II\"\n },\n {\n \"msg_contents\": \"On 2016-08-23 07:26:31 -0500, Kevin Grittner wrote:\\n> On Tue, Aug 23, 2016 at 7:10 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\\n> > On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> \\n> >> Could you provide an example of a case where xacts replayed in\\n> >> commit order will produce incorrect results?\\n> >\\n> > https://wiki.postgresql.org/wiki/SSI#Deposit_Report\\n> >\\n> > ... where T3 is on the replication target.\\n> \\n> I should, perhaps, have mentioned that the cases where this is are\\n> problem are \\\"eventually consistent\\\" -- it's a matter of being able\\n> to see a state that violates business rule invariants or where data\\n> which is \\\"locked down\\\" according to one part of the database is\\n> still changing. Such problems are prevented on a single database,\\n> but would not be prevented on a logical replica where transactions\\n> are applied in commit order. Given enough time, the replica would\\n> eventually settle into a state without the anomalies, similar to\\n> some other products with eventual consistency.\\n\\nI've generally a bit of difficulty to see this as a significant problem\\nfor logical rep, as long as hot-standby, and crash-recovery in general,\\nalso has this problem...\\n\\n\\n-- \\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 Aug 2016 09:50:19 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Logical decoding of sequence advances, part II\"\n },\n {\n \"msg_contents\": \"On 24 August 2016 at 00:50, Andres Freund <andres@anarazel.de> wrote:\\n\\n> On 2016-08-23 07:26:31 -0500, Kevin Grittner wrote:\\n> > On Tue, Aug 23, 2016 at 7:10 AM, Kevin Grittner <kgrittn@gmail.com>\\n> wrote:\\n> > [an explanation of SSI anomalies]\\n>\\n\\nI've generally a bit of difficulty to see this as a significant problem\\n> for logical rep, as long as hot-standby, and crash-recovery in general,\\n> also has this problem...\\n>\\n\\nSame here, as commented upthread. I think it'd be cool to be able to\\ndeliver SSI-alike behaviour on a standby, but it's far from a priority, and\\nit applies to phys ans well as logical rep.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 24 August 2016 at 00:50, Andres Freund <andres@anarazel.de> wrote:On 2016-08-23 07:26:31 -0500, Kevin Grittner wrote:\\n> On Tue, Aug 23, 2016 at 7:10 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\\n> [an explanation of SSI anomalies] I've generally a bit of difficulty to see this as a significant problem\\nfor logical rep, as long as hot-standby, and crash-recovery in general,\\nalso has this problem...\\nSame here, as commented upthread. I think it'd be cool to be able to deliver SSI-alike behaviour on a standby, but it's far from a priority, and it applies to phys ans well as logical rep.-- Craig Ringer http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Wed, 24 Aug 2016 01:02:09 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Logical decoding of sequence advances, part II\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 23, 2016 at 11:50 AM, Andres Freund <andres@anarazel.de> wrote:\\n> On 2016-08-23 07:26:31 -0500, Kevin Grittner wrote:\\n>> On Tue, Aug 23, 2016 at 7:10 AM, Kevin Grittner <kgrittn@gmail.com> wrote:\\n>> > On Mon, Aug 22, 2016 at 6:39 PM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>>\\n>> >> Could you provide an example of a case where xacts replayed in\\n>> >> commit order will produce incorrect results?\\n>> >\\n>> > https://wiki.postgresql.org/wiki/SSI#Deposit_Report\\n>> >\\n>> > ... where T3 is on the replication target.\\n>>\\n>> I should, perhaps, have mentioned that the cases where this is are\\n>> problem are \\\"eventually consistent\\\" -- it's a matter of being able\\n>> to see a state that violates business rule invariants or where data\\n>> which is \\\"locked down\\\" according to one part of the database is\\n>> still changing. Such problems are prevented on a single database,\\n>> but would not be prevented on a logical replica where transactions\\n>> are applied in commit order. Given enough time, the replica would\\n>> eventually settle into a state without the anomalies, similar to\\n>> some other products with eventual consistency.\\n>\\n> I've generally a bit of difficulty to see this as a significant problem\\n> for logical rep, as long as hot-standby, and crash-recovery in general,\\n> also has this problem...\\n\\nSerialization anomalies cannot be seen on a hot standby nor on\\ncrash recovery. Granted, the mechanism which prevents it on the\\nhot standby is that we don't allow the transaction isolation level\\nto be set to SERIALIZABLE, to prevent the expectation that queries\\nwill be free of anomalies, which is pretty crude. On crash\\nrecovery you cannot see any anomalies as far as I'm aware -- what\\nhas given you the idea that it is possible?\\n\\nI'm not sure that the same technique I described for logical\\nreplication could be made to work for page-level physical\\nreplication, but a \\\"safe snapshot\\\" technique has been previously\\ndiscussed on the lists which would work for physical replication\\n(although *that* technique seems unsuited to logical replication).\\n\\n--\\nKevin Grittner\\nEDB: 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 Aug 2016 13:03:48 -0500\",\n \"msg_from\": \"Kevin Grittner <kgrittn@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Logical decoding of sequence advances, part II\"\n }\n]","isConversational":false}}},{"rowIdx":42998,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Just out of curiopusity, I wonder why we can't make the encoding of\nSGML docs to be UTF-8, rather than current ISO-8859-1.\n\nAs long as everything is written in ASCII, the size of docs will be\nalmost same even if UTF-8 is used. Plus, if the encoding is changed to\nUTF-8, it is very easy to translate the doc to local languages. As far\nas I know, al of local language docs under\nhttps://www.postgresql.org/docs/ are using UTF-8.\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, 22 Aug 2016 14:16:45 +0900 (JST)","msg_from":"Tatsuo Ishii <ishii@sraoss.co.jp>","msg_from_op":true,"msg_subject":"UTF-8 docs?"},{"msg_contents":"On Mon, 22 Aug 2016 14:16:45 +0900 (JST)\nTatsuo Ishii <ishii@sraoss.co.jp> wrote:\n\n> Just out of curiopusity, I wonder why we can't make the encoding of\n> SGML docs to be UTF-8, rather than current ISO-8859-1.\n\n\nWhat a reason of \"make the encoding of sgml docs\" to be something?\nWhat actual change should be made and what problems it would solve?\n\nThere are various translations of postgreSQL docs, and they use various\nencodings. Translated versions of docs on http://postgresql.org/docs\nare just links to external sites where translations are maintained. \n\nEnglish documentation uses ISO-8859-1 (actually ASCII),\nRussian uses UTF-8 (you can download our source tarball from\nhttp://repo.postgrespro.ru/pgpro-9.5/src and see postgres source\ndistribution with UTF-8 sgmls inside). \n\nJapanese documentation in HTML form is served from\nhttp://www.postgresql.jp/document/9.5/html/\nin utf-8 too.\n\nI.e. everybody who need utf-8 to represent translation of\ndocumentation, already uses it.\n\nWhat exatly you proposes do be done?\n\nReally, what change we need, it is conversion from SGML to XML format. \nIt would solve some real problems, such as ability to include diagrams\nin the docs, and also let everyone to explicitely specify encoding in\nXML declaration (and probably cause switch to UTF-8 as side effect,\nbecause most XML-based tools use UTF-8 as default).\n\n\n-- \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 Aug 2016 10:13:48 +0300","msg_from":"Victor Wagner <vitus@wagner.pp.ru>","msg_from_op":false,"msg_subject":"Re: UTF-8 docs?"},{"msg_contents":"> On Mon, 22 Aug 2016 14:16:45 +0900 (JST)\n> Tatsuo Ishii <ishii@sraoss.co.jp> wrote:\n> \n>> Just out of curiopusity, I wonder why we can't make the encoding of\n>> SGML docs to be UTF-8, rather than current ISO-8859-1.\n> \n> \n> What a reason of \"make the encoding of sgml docs\" to be something?\n> What actual change should be made and what problems it would solve?\n\nProblem is, the PostgreSQL docs is fixed to ISO-8859-1, and if I want\nto use other encoding, I need to change the build system, which is\nannoying. Ideally, if someone wants to use other than ISO-8859-1, then\nhe/she just change the contents of sgml files. Just changing the\nISO-8859-1 to UTF-8 will solve most problems.\n\n(Probably allowing to specify arbitrary encoding is better but it\nneeds some work).\n\n> There are various translations of postgreSQL docs, and they use various\n> encodings. Translated versions of docs on http://postgresql.org/docs\n> are just links to external sites where translations are maintained. \n> English documentation uses ISO-8859-1 (actually ASCII),\n> Russian uses UTF-8 (you can download our source tarball from\n> http://repo.postgrespro.ru/pgpro-9.5/src and see postgres source\n> distribution with UTF-8 sgmls inside). \n> \n> Japanese documentation in HTML form is served from\n> http://www.postgresql.jp/document/9.5/html/\n> in utf-8 too.\n> \n> I.e. everybody who need utf-8 to represent translation of\n> documentation, already uses it.\n> \n> What exatly you proposes do be done?\n\nSee above.\n\n> Really, what change we need, it is conversion from SGML to XML format. \n> It would solve some real problems, such as ability to include diagrams\n> in the docs, and also let everyone to explicitely specify encoding in\n> XML declaration (and probably cause switch to UTF-8 as side effect,\n> because most XML-based tools use UTF-8 as default).\n\nThat's another story.\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, 22 Aug 2016 17:28:31 +0900 (JST)","msg_from":"Tatsuo Ishii <ishii@sraoss.co.jp>","msg_from_op":true,"msg_subject":"Re: UTF-8 docs?"},{"msg_contents":"On 8/22/16 1:16 AM, Tatsuo Ishii wrote:\n> Just out of curiopusity, I wonder why we can't make the encoding of\n> SGML docs to be UTF-8, rather than current ISO-8859-1.\n\nEncoding handling in DocBook SGML is weird, and making it work robustly\nwill either fail or might be more work than just completing the\nconversion to XML.\n\n-- \nPeter Eisentraut 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":"Mon, 22 Aug 2016 08:46:32 -0400","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: UTF-8 docs?"},{"msg_contents":"> On 8/22/16 1:16 AM, Tatsuo Ishii wrote:\n>> Just out of curiopusity, I wonder why we can't make the encoding of\n>> SGML docs to be UTF-8, rather than current ISO-8859-1.\n> \n> Encoding handling in DocBook SGML is weird, and making it work robustly\n> will either fail or might be more work than just completing the\n> conversion to XML.\n\nI don't know what kind of problem you are seeing with encoding\nhandling, but at least UTF-8 is working for Japanese, French and\nRussian.\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, 22 Aug 2016 22:32:15 +0900 (JST)","msg_from":"Tatsuo Ishii <ishii@sraoss.co.jp>","msg_from_op":true,"msg_subject":"Re: UTF-8 docs?"},{"msg_contents":"On 8/22/16 9:32 AM, Tatsuo Ishii wrote:\n> I don't know what kind of problem you are seeing with encoding\n> handling, but at least UTF-8 is working for Japanese, French and\n> Russian.\n\nThose translations are using DocBook XML.\n\n-- \nPeter Eisentraut 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":"Mon, 22 Aug 2016 10:53:25 -0400","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: UTF-8 docs?"},{"msg_contents":"> On 8/22/16 9:32 AM, Tatsuo Ishii wrote:\n>> I don't know what kind of problem you are seeing with encoding\n>> handling, but at least UTF-8 is working for Japanese, French and\n>> Russian.\n> \n> Those translations are using DocBook XML.\n\nBut in the mean time I can create UTF-8 HTML files like this:\n\nmake html\n[snip]\n/bin/mkdir -p html\nSP_CHARSET_FIXED=1 SP_ENCODING=UTF-8 openjade -wall -wno-unused-param -wno-empty -wfully-tagged -D . -D . -c /usr/share/sgml/docbook/stylesheet/dsssl/modular/catalog -d stylesheet.dsl -t sgml -i output-html -i include-index postgres.sgml\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, 23 Aug 2016 07:47:37 +0900 (JST)","msg_from":"Tatsuo Ishii <ishii@sraoss.co.jp>","msg_from_op":true,"msg_subject":"Re: UTF-8 docs?"},{"msg_contents":"On Mon, 22 Aug 2016 10:53:25 -0400\nPeter Eisentraut <peter.eisentraut@2ndquadrant.com> wrote:\n\n> On 8/22/16 9:32 AM, Tatsuo Ishii wrote:\n> > I don't know what kind of problem you are seeing with encoding\n> > handling, but at least UTF-8 is working for Japanese, French and\n> > Russian. \n> \n> Those translations are using DocBook XML.\n> \n\nRussian translation by Postgres Professional does use DocBook SGML,\nalthough it uses xml as intermediate representation when applying\ngettext to the documentation. I've already posted URL where sources of\npostgresql with russian documentation in SGML format included can be\ndownloaded.\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":"Tue, 23 Aug 2016 17:47:54 +0300","msg_from":"Victor Wagner <vitus@wagner.pp.ru>","msg_from_op":false,"msg_subject":"Re: UTF-8 docs?"}],"string":"[\n {\n \"msg_contents\": \"Just out of curiopusity, I wonder why we can't make the encoding of\\nSGML docs to be UTF-8, rather than current ISO-8859-1.\\n\\nAs long as everything is written in ASCII, the size of docs will be\\nalmost same even if UTF-8 is used. Plus, if the encoding is changed to\\nUTF-8, it is very easy to translate the doc to local languages. As far\\nas I know, al of local language docs under\\nhttps://www.postgresql.org/docs/ are using UTF-8.\\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, 22 Aug 2016 14:16:45 +0900 (JST)\",\n \"msg_from\": \"Tatsuo Ishii <ishii@sraoss.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"UTF-8 docs?\"\n },\n {\n \"msg_contents\": \"On Mon, 22 Aug 2016 14:16:45 +0900 (JST)\\nTatsuo Ishii <ishii@sraoss.co.jp> wrote:\\n\\n> Just out of curiopusity, I wonder why we can't make the encoding of\\n> SGML docs to be UTF-8, rather than current ISO-8859-1.\\n\\n\\nWhat a reason of \\\"make the encoding of sgml docs\\\" to be something?\\nWhat actual change should be made and what problems it would solve?\\n\\nThere are various translations of postgreSQL docs, and they use various\\nencodings. Translated versions of docs on http://postgresql.org/docs\\nare just links to external sites where translations are maintained. \\n\\nEnglish documentation uses ISO-8859-1 (actually ASCII),\\nRussian uses UTF-8 (you can download our source tarball from\\nhttp://repo.postgrespro.ru/pgpro-9.5/src and see postgres source\\ndistribution with UTF-8 sgmls inside). \\n\\nJapanese documentation in HTML form is served from\\nhttp://www.postgresql.jp/document/9.5/html/\\nin utf-8 too.\\n\\nI.e. everybody who need utf-8 to represent translation of\\ndocumentation, already uses it.\\n\\nWhat exatly you proposes do be done?\\n\\nReally, what change we need, it is conversion from SGML to XML format. \\nIt would solve some real problems, such as ability to include diagrams\\nin the docs, and also let everyone to explicitely specify encoding in\\nXML declaration (and probably cause switch to UTF-8 as side effect,\\nbecause most XML-based tools use UTF-8 as default).\\n\\n\\n-- \\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 Aug 2016 10:13:48 +0300\",\n \"msg_from\": \"Victor Wagner <vitus@wagner.pp.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: UTF-8 docs?\"\n },\n {\n \"msg_contents\": \"> On Mon, 22 Aug 2016 14:16:45 +0900 (JST)\\n> Tatsuo Ishii <ishii@sraoss.co.jp> wrote:\\n> \\n>> Just out of curiopusity, I wonder why we can't make the encoding of\\n>> SGML docs to be UTF-8, rather than current ISO-8859-1.\\n> \\n> \\n> What a reason of \\\"make the encoding of sgml docs\\\" to be something?\\n> What actual change should be made and what problems it would solve?\\n\\nProblem is, the PostgreSQL docs is fixed to ISO-8859-1, and if I want\\nto use other encoding, I need to change the build system, which is\\nannoying. Ideally, if someone wants to use other than ISO-8859-1, then\\nhe/she just change the contents of sgml files. Just changing the\\nISO-8859-1 to UTF-8 will solve most problems.\\n\\n(Probably allowing to specify arbitrary encoding is better but it\\nneeds some work).\\n\\n> There are various translations of postgreSQL docs, and they use various\\n> encodings. Translated versions of docs on http://postgresql.org/docs\\n> are just links to external sites where translations are maintained. \\n> English documentation uses ISO-8859-1 (actually ASCII),\\n> Russian uses UTF-8 (you can download our source tarball from\\n> http://repo.postgrespro.ru/pgpro-9.5/src and see postgres source\\n> distribution with UTF-8 sgmls inside). \\n> \\n> Japanese documentation in HTML form is served from\\n> http://www.postgresql.jp/document/9.5/html/\\n> in utf-8 too.\\n> \\n> I.e. everybody who need utf-8 to represent translation of\\n> documentation, already uses it.\\n> \\n> What exatly you proposes do be done?\\n\\nSee above.\\n\\n> Really, what change we need, it is conversion from SGML to XML format. \\n> It would solve some real problems, such as ability to include diagrams\\n> in the docs, and also let everyone to explicitely specify encoding in\\n> XML declaration (and probably cause switch to UTF-8 as side effect,\\n> because most XML-based tools use UTF-8 as default).\\n\\nThat's another story.\\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, 22 Aug 2016 17:28:31 +0900 (JST)\",\n \"msg_from\": \"Tatsuo Ishii <ishii@sraoss.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: UTF-8 docs?\"\n },\n {\n \"msg_contents\": \"On 8/22/16 1:16 AM, Tatsuo Ishii wrote:\\n> Just out of curiopusity, I wonder why we can't make the encoding of\\n> SGML docs to be UTF-8, rather than current ISO-8859-1.\\n\\nEncoding handling in DocBook SGML is weird, and making it work robustly\\nwill either fail or might be more work than just completing the\\nconversion to XML.\\n\\n-- \\nPeter Eisentraut 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\": \"Mon, 22 Aug 2016 08:46:32 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: UTF-8 docs?\"\n },\n {\n \"msg_contents\": \"> On 8/22/16 1:16 AM, Tatsuo Ishii wrote:\\n>> Just out of curiopusity, I wonder why we can't make the encoding of\\n>> SGML docs to be UTF-8, rather than current ISO-8859-1.\\n> \\n> Encoding handling in DocBook SGML is weird, and making it work robustly\\n> will either fail or might be more work than just completing the\\n> conversion to XML.\\n\\nI don't know what kind of problem you are seeing with encoding\\nhandling, but at least UTF-8 is working for Japanese, French and\\nRussian.\\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, 22 Aug 2016 22:32:15 +0900 (JST)\",\n \"msg_from\": \"Tatsuo Ishii <ishii@sraoss.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: UTF-8 docs?\"\n },\n {\n \"msg_contents\": \"On 8/22/16 9:32 AM, Tatsuo Ishii wrote:\\n> I don't know what kind of problem you are seeing with encoding\\n> handling, but at least UTF-8 is working for Japanese, French and\\n> Russian.\\n\\nThose translations are using DocBook XML.\\n\\n-- \\nPeter Eisentraut 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\": \"Mon, 22 Aug 2016 10:53:25 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: UTF-8 docs?\"\n },\n {\n \"msg_contents\": \"> On 8/22/16 9:32 AM, Tatsuo Ishii wrote:\\n>> I don't know what kind of problem you are seeing with encoding\\n>> handling, but at least UTF-8 is working for Japanese, French and\\n>> Russian.\\n> \\n> Those translations are using DocBook XML.\\n\\nBut in the mean time I can create UTF-8 HTML files like this:\\n\\nmake html\\n[snip]\\n/bin/mkdir -p html\\nSP_CHARSET_FIXED=1 SP_ENCODING=UTF-8 openjade -wall -wno-unused-param -wno-empty -wfully-tagged -D . -D . -c /usr/share/sgml/docbook/stylesheet/dsssl/modular/catalog -d stylesheet.dsl -t sgml -i output-html -i include-index postgres.sgml\\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, 23 Aug 2016 07:47:37 +0900 (JST)\",\n \"msg_from\": \"Tatsuo Ishii <ishii@sraoss.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: UTF-8 docs?\"\n },\n {\n \"msg_contents\": \"On Mon, 22 Aug 2016 10:53:25 -0400\\nPeter Eisentraut <peter.eisentraut@2ndquadrant.com> wrote:\\n\\n> On 8/22/16 9:32 AM, Tatsuo Ishii wrote:\\n> > I don't know what kind of problem you are seeing with encoding\\n> > handling, but at least UTF-8 is working for Japanese, French and\\n> > Russian. \\n> \\n> Those translations are using DocBook XML.\\n> \\n\\nRussian translation by Postgres Professional does use DocBook SGML,\\nalthough it uses xml as intermediate representation when applying\\ngettext to the documentation. I've already posted URL where sources of\\npostgresql with russian documentation in SGML format included can be\\ndownloaded.\\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\": \"Tue, 23 Aug 2016 17:47:54 +0300\",\n \"msg_from\": \"Victor Wagner <vitus@wagner.pp.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: UTF-8 docs?\"\n }\n]","isConversational":false}}},{"rowIdx":42999,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"While hacking on the CSN patch, I spotted a race condition between \nGetOldestActiveTransactionId() and GetNewTransactionId(). \nGetOldestActiveTransactionId() calculates the oldest XID that's still \nrunning, by doing:\n\n1. Read nextXid, without a lock. This is the upper bound, if there are \nno active XIDs in the proc array.\n2. Loop through the proc array, making note of the oldest XID.\n\nWhile GetNewTransactionId() does:\n\n1. Read and Increment nextXid\n2. Set MyPgXact->xid.\n\nIt seems possible that if you call GetNewTransactionId() concurrently \nwith GetOldestActiveTransactionId(), GetOldestActiveTransactionId() sees \nthe new nextXid value that the concurrent GetNewTransactionId() set, but \ndoesn't see the old XID in the proc array. It will return a value that \ndoesn't cover the old XID, i.e. it won't consider the just-assigned XID \nas in-progress.\n\nAm I missing something? Commit c6d76d7c added a comment to \nGetOldestActiveTransactionId() explaining why it's not necessary to \nacquire XidGenLock there, but I think it missed the above race condition.\n\nGetOldestActiveTransactionId() is not performance-critical, it's only \ncalled when performing a checkpoint, so I think we should just bite the \nbullet and grab the lock. Per attached patch.\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":"Mon, 22 Aug 2016 13:46:08 +0300","msg_from":"Heikki Linnakangas <hlinnaka@iki.fi>","msg_from_op":true,"msg_subject":"Race condition in GetOldestActiveTransactionId()"},{"msg_contents":"On 08/22/2016 01:46 PM, Heikki Linnakangas wrote:\n> While hacking on the CSN patch, I spotted a race condition between\n> GetOldestActiveTransactionId() and GetNewTransactionId().\n> GetOldestActiveTransactionId() calculates the oldest XID that's still\n> running, by doing:\n>\n> 1. Read nextXid, without a lock. This is the upper bound, if there are\n> no active XIDs in the proc array.\n> 2. Loop through the proc array, making note of the oldest XID.\n>\n> While GetNewTransactionId() does:\n>\n> 1. Read and Increment nextXid\n> 2. Set MyPgXact->xid.\n>\n> It seems possible that if you call GetNewTransactionId() concurrently\n> with GetOldestActiveTransactionId(), GetOldestActiveTransactionId() sees\n> the new nextXid value that the concurrent GetNewTransactionId() set, but\n> doesn't see the old XID in the proc array. It will return a value that\n> doesn't cover the old XID, i.e. it won't consider the just-assigned XID\n> as in-progress.\n>\n> Am I missing something? Commit c6d76d7c added a comment to\n> GetOldestActiveTransactionId() explaining why it's not necessary to\n> acquire XidGenLock there, but I think it missed the above race condition.\n>\n> GetOldestActiveTransactionId() is not performance-critical, it's only\n> called when performing a checkpoint, so I think we should just bite the\n> bullet and grab the lock. Per attached patch.\n\nI did some testing, and was able to indeed construct a case where \noldestActiveXid was off-by-one in an online checkpoint record. However, \nlooking at how it's used, I think it happened to not have any \nuser-visible effect. The oldestActiveXid value of an online checkpoint \nis only used to determine the point where pg_subtrans is initialized, \nand the XID missed in the computation could not be a subtransaction.\n\nNevertheless, it's clearly a bug and the fix is simple, so I committed a \nfix.\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, 13 Jul 2017 15:52:44 +0300","msg_from":"Heikki Linnakangas <hlinnaka@iki.fi>","msg_from_op":true,"msg_subject":"Re: Race condition in GetOldestActiveTransactionId()"}],"string":"[\n {\n \"msg_contents\": \"While hacking on the CSN patch, I spotted a race condition between \\nGetOldestActiveTransactionId() and GetNewTransactionId(). \\nGetOldestActiveTransactionId() calculates the oldest XID that's still \\nrunning, by doing:\\n\\n1. Read nextXid, without a lock. This is the upper bound, if there are \\nno active XIDs in the proc array.\\n2. Loop through the proc array, making note of the oldest XID.\\n\\nWhile GetNewTransactionId() does:\\n\\n1. Read and Increment nextXid\\n2. Set MyPgXact->xid.\\n\\nIt seems possible that if you call GetNewTransactionId() concurrently \\nwith GetOldestActiveTransactionId(), GetOldestActiveTransactionId() sees \\nthe new nextXid value that the concurrent GetNewTransactionId() set, but \\ndoesn't see the old XID in the proc array. It will return a value that \\ndoesn't cover the old XID, i.e. it won't consider the just-assigned XID \\nas in-progress.\\n\\nAm I missing something? Commit c6d76d7c added a comment to \\nGetOldestActiveTransactionId() explaining why it's not necessary to \\nacquire XidGenLock there, but I think it missed the above race condition.\\n\\nGetOldestActiveTransactionId() is not performance-critical, it's only \\ncalled when performing a checkpoint, so I think we should just bite the \\nbullet and grab the lock. Per attached patch.\\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\": \"Mon, 22 Aug 2016 13:46:08 +0300\",\n \"msg_from\": \"Heikki Linnakangas <hlinnaka@iki.fi>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Race condition in GetOldestActiveTransactionId()\"\n },\n {\n \"msg_contents\": \"On 08/22/2016 01:46 PM, Heikki Linnakangas wrote:\\n> While hacking on the CSN patch, I spotted a race condition between\\n> GetOldestActiveTransactionId() and GetNewTransactionId().\\n> GetOldestActiveTransactionId() calculates the oldest XID that's still\\n> running, by doing:\\n>\\n> 1. Read nextXid, without a lock. This is the upper bound, if there are\\n> no active XIDs in the proc array.\\n> 2. Loop through the proc array, making note of the oldest XID.\\n>\\n> While GetNewTransactionId() does:\\n>\\n> 1. Read and Increment nextXid\\n> 2. Set MyPgXact->xid.\\n>\\n> It seems possible that if you call GetNewTransactionId() concurrently\\n> with GetOldestActiveTransactionId(), GetOldestActiveTransactionId() sees\\n> the new nextXid value that the concurrent GetNewTransactionId() set, but\\n> doesn't see the old XID in the proc array. It will return a value that\\n> doesn't cover the old XID, i.e. it won't consider the just-assigned XID\\n> as in-progress.\\n>\\n> Am I missing something? Commit c6d76d7c added a comment to\\n> GetOldestActiveTransactionId() explaining why it's not necessary to\\n> acquire XidGenLock there, but I think it missed the above race condition.\\n>\\n> GetOldestActiveTransactionId() is not performance-critical, it's only\\n> called when performing a checkpoint, so I think we should just bite the\\n> bullet and grab the lock. Per attached patch.\\n\\nI did some testing, and was able to indeed construct a case where \\noldestActiveXid was off-by-one in an online checkpoint record. However, \\nlooking at how it's used, I think it happened to not have any \\nuser-visible effect. The oldestActiveXid value of an online checkpoint \\nis only used to determine the point where pg_subtrans is initialized, \\nand the XID missed in the computation could not be a subtransaction.\\n\\nNevertheless, it's clearly a bug and the fix is simple, so I committed a \\nfix.\\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, 13 Jul 2017 15:52:44 +0300\",\n \"msg_from\": \"Heikki Linnakangas <hlinnaka@iki.fi>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Race condition in GetOldestActiveTransactionId()\"\n }\n]","isConversational":false}}}],"truncated":false,"partial":false},"paginationData":{"pageIndex":429,"numItemsPerPage":100,"numTotalItems":58371,"offset":42900,"length":100}},"jwt":"eyJhbGciOiJFZERTQSIsImtpZCI6IjVHZDBvd0g5MTM2eDZjc1FvbE1zcktNWUZoRVFoUm5PVVVybEpjOUhaUEEifQ.eyJyZWFkIjp0cnVlLCJwZXJtaXNzaW9ucyI6eyJyZXBvLmNvbnRlbnQucmVhZCI6dHJ1ZX0sImlhdCI6MTc4NTQ2NDY1NiwianRpIjoiYTQ4MGU2NzAtZjNkZC00NDQ4LWIyODEtYzllNmVlNmZmZmRkIiwic3ViIjoiL2RhdGFzZXRzL3dhbnNoZW5sL3Bnc3FsLWhhY2tlcnMiLCJleHAiOjE3ODU0NjgyNTYsImlzcyI6Imh0dHBzOi8vaHVnZ2luZ2ZhY2UuY28ifQ.iqde-GZuEBfil3Rz4lH6_6QMDVJrEWPXTKv6pt9m0fSDfnXMJ5nW7ncvuWQhm58HxtMpq0Mj3nlcYP6TtE3zCg","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": "I've got a customer that discovered something odd...\n\nSELECT f1 FROM v1 WHERE f2 not in (SELECT bad FROM v2 WHERE f3 = 1);\n\ndoes not error, even though bad doesn't exist, but\n\nSELECT bad FROM v2 WHERE f3 = 1;\ngives\n\nERROR: column \"bad\" does not exist\n\nIs that expected?\n\nThis i... |
[
{
"msg_contents": "I looked into the problem described at\nhttps://www.postgresql.org/message-id/flat/VisenaEmail.26.df42f82acae38a58.156463942b8%40tc7-visena\nand I believe I've reproduced it: the requirement is that the inner join\ncolumn for the antijoin must contain a lot of NULL values, and what isn't\nNUL... |
[
{
"msg_contents": "Working on page compression and some other issues related to\naccess methods, I found out that the code related to heap\nlooks too complicated. Much more complicated, than it should be.\nSince I anyway got into this area, I want to suggest a set of improvements.\n\nThere is a number of proble... |
[
{
"msg_contents": "Hello\n\nI recently hit a road blocker when I tried to create a truncate trigger on\na foreign table. trigger.c::CreateTrigger() function has explicit check to\nblock truncate trigger on foreign tables.\n\nHowever, trigger.c::ExecuteTruncate() does not seem to have any problems\nissuing befor... |
[
{
"msg_contents": "Hello Developers,\n\nI have been following the recent discussions on increasing the\nsize of the hash function used in Postgres and the work to\nprovide WAL and performance improvements for hash indexes. \nI know it was mentioned when we moved to the new hashing\nfunctions, but the existing f... |
[
{
"msg_contents": "Hi,\n\nas promised here is WIP version of logical replication patch.\n\nThis is by no means anywhere close to be committable, but it should be\nenough for discussion on the approaches chosen. I do plan to give this\nsome more time before September CF as well as during the CF itself.\n\nYou've... |
[
{
"msg_contents": "The input function of the money type has no overflow checks:\n\n=> select '12345678901234567890'::money;\n money\n-----------------------------\n -$13,639,628,150,831,692.72\n(1 row)\n\nThe tests in the regression test file money.sql are bogus because they\nonly test the overflow c... |
[
{
"msg_contents": "One time too many, I ran some minor change using psql on a production\nserver and was wondering why it was taking so much longer than it did\non the test server. Only to discover, after messing around with\nopening new windows and running queries against pg_stat_activity and\npg_locks and so... |
[
{
"msg_contents": "Basically,\n\ndiff --git a/src/backend/parser/parse_param.c\nb/src/backend/parser/parse_param.c\nindex b402843..97064fc 100644\n--- a/src/backend/parser/parse_param.c\n+++ b/src/backend/parser/parse_param.c\n@@ -108,6 +108,9 @@ fixed_paramref_hook(ParseState *pstate, ParamRef *pref)\n erepor... |
[
{
"msg_contents": "Hi,\n\nsince updating master from c93d873..fc509cd, I see crashes in\nGetOldestSnapshot() on update/delete returning statements.\n\nI reduced the triggering statements down to this:\n\n update clstr_tst set d = d returning d;\n\nBacktrace below.\n\nregards,\nAndreas\n\nProgram received sig... |
[
{
"msg_contents": "I've put up draft notes for 9.5.4 at\n\nhttps://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=3676631c687009c2fadcb35e7d312e9eb8a98182\n\nThe website should have them after guaibasaurus's next run, due a couple\nhours from now.\n\nAs usual, the older branches will have subsets of... |
[
{
"msg_contents": "Hi,\n\nLooking at commits f10eab73d and c50d192c, I wondered why we don't\nhave a reusable in-place unique function. It may be trivial, but we\nseem to have a lot of copies and variations in the tree.\n\nHere's a sketch patch that creates a function array_unique which takes\nthe same argumen... |
[
{
"msg_contents": "Hi\r\n\r\nI am sending a initial implementation of xmltable function:\r\n\r\nThe code is not clean now, but it does almost of expected work. The usage\r\nis simple. It is fast - 16K entries in 400ms.\r\n\r\nI invite any help with documentation and testing.\r\n\r\nThe full ANSI/SQL, or Oracle ... |
[
{
"msg_contents": "Hi,\n\nI've summarized Wait events monitoring discussion at Developer unconference\nin Ottawa this year on wiki:\n\nhttps://wiki.postgresql.org/wiki/PgCon_2016_Developer_Unconference/Wait_events_monitoring\n\n\n(Thanks to Alexander Korotkov for patiently pushing me to make this thing\nfinally... |
[
{
"msg_contents": "Hi\n\nWhen I checked result of xml test I found some strange lines in\nexpected/xml.out file\n\nIt is little bit strange - regress test reports ok, but diff is not empty\n\n[pavel@nemesis regress]$ diff results/xml.out expected/xml.out\n11a12,13\n> <wrong\n> ^\n63a66,67\n> <syntax\n> ... |
[
{
"msg_contents": "Hi,\n\nsqlsmith just triggered a crash in pg_get_viewdef_name_ext(). Looks\nlike commit 976b24fb4 failed to update this caller of\npg_get_viewdef_worker(). Backtrace below. Patch attached.\n\nregards,\nAndreas\n\nProgram terminated with signal SIGSEGV, Segmentation fault.\n(gdb) bt\n#0 st... |
[
{
"msg_contents": "Hi, I have found a basic use case which is supported by the xml2 module,\nbut is unsupported by the new XML API.\n\nIt is not possible to correctly extract text (either from text nodes or\nattribute values) which contains the characters '<', '&', or '>'. \nxpath() (correctly) returns XML text... |
[
{
"msg_contents": "Hi hackers,\n\nI'm asking out of curiosity, do anyone know why we don't have\nper-statement-level INSTEAD OF triggers? I looked into the \nstandard SQL (20xx draft), but I can't find the restriction\nsuch that INSTEAD OF triggers must be row-level. Is there\nany technical difficulties, or oth... |
[
{
"msg_contents": "Over on the \"Parallel tuplesort (for parallel B-Tree index creation)\"\nthread [1], there has been some discussion of merging vs.\npartitioning. There is a concern about the fact the merge of the\ntuplesort used to build a B-Tree is not itself parallelized. There is\na weak consensus that we... |
[
{
"msg_contents": "Hi hackers,\n\nDSM segments have a concept of 'pinning'. Normally, segments are\ndestroyed when they are no longer mapped by any backend, using a\nreference counting scheme. If you call dsm_pin_segment(segment), that\nis disabled so that the segment won't be destroyed until the cluster\nis ... |
[
{
"msg_contents": "Hello,\n\nSince 14e8803f1, it's not necessary to acquire the SyncRepLock to see up\nto date data. But it looks like this commit didn't update all the\ncomment around MyProc->syncRepState, which still mention retrieving the\nvalue without and without lock. Also, there's I think a now unneeded... |
[
{
"msg_contents": "Here is a small patch that sets log_line_prefix and application name in\npg_regress and the TAP tests, to make analyzing the server log output\neasier.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\... |
[
{
"msg_contents": "Does anyone know why the phrase distance \"<3>\" was changed from \"at most\nthree tokens away\" to \"exactly three tokens away\"? I looked at the\nthread at:\n\n\thttps://www.postgresql.org/message-id/flat/33828354.WrrSMviC7Y%40abook\n\nand didn't see the answer. I assume if you are lookin... |
[
{
"msg_contents": "Hackers,\n\nThe attached attempts to make comprehension of the code in\n\"TS_phrase_execute\" a bit easier. I posted similar on the \"typo patch\"\nthread of July 2nd/5th but my comments there reflected my mis-understanding\nof the distance operator being exact as opposed to the expected\nle... |
[
{
"msg_contents": "Hi,\n\nAttached is the latest set of patches to implement declarative\npartitioning. There is already a commitfest entry for the same:\nhttps://commitfest.postgresql.org/10/611/\n\nThe old discussion is here:\nhttp://www.postgresql.org/message-id/55D3093C.5010800@lab.ntt.co.jp/\n\nAttached p... |
[
{
"msg_contents": "Hi,\n\nRunning one specific test from our application against REL9_5_STABLE \n(current as of today) gives me this:\n\n#2 0x00007effb59595be in ExceptionalCondition (\n conditionName=conditionName@entry=0x7effb5b27a88 \n\"!(CritSectionCount > 0 || TransactionIdIsCurrentTransactionId(( \n(... |
[
{
"msg_contents": "Hi,\n\nWould anyone mind too much if I refreshed typedefs.list and\nre-indented the whole tree before we branch? There's not too much\nchurn at the moment, and I feel like it would be nice to start the\ncycle in as clean a state as possible.\n\nCurrent results of this attached.\n\n-- \nRober... |
[
{
"msg_contents": "I wanted to create a new relopt named something like\nautovacuum_vacuum_pagevisible_factor which would cause autovacuum to\nvacuum a table once less than a certain fraction of the relation's\npages are marked allvisible.\n\nI wanted some feedback on some things.\n\n1) One issue is that pg_cla... |
[
{
"msg_contents": "Hi,\n\nI've run YCSB(Yahoo! Cloud Service Benchmark) on PostgreSQL and MongoDB\nwith WiredTiger.\nAnd I found some interesting results and some issues(maybe) on Btree index\nof PostgreSQL.\n\nHere is my experiments and results.\nYCSB is for document store benchmark and I build following schem... |
[
{
"msg_contents": "Hello.\n\nSuggested patch introduces an %r substitution in psql's prompt. This\nsubstitution allows to display whether user is connected to master or\nreplica right in a prompt.\n\nUsage example:\n\n```\n$ cat ~/.psqlrc \n\\set PROMPT1 '%p (%r) =# '\n\n$ psql -p 5432\npsql (9.6beta4)\nType \"... |
[
{
"msg_contents": "> I have found a basic use case which is supported by the xml2 module,\n> but is unsupported by the new XML API. \n> It is not possible to correctly extract text\n\nIndeed. I came accross this shortcomming some months ago myself but still manage an item on my ToDo list to report it here as th... |
[
{
"msg_contents": "In\nhttps://www.postgresql.org/message-id/15293.1466536829@sss.pgh.pa.us\nI speculated that it might not take too much to replace all the variants\nof GIN array_ops with a single polymorphic opclass over anyarray.\nAttached is a proposed patch that does that.\n\nThere are two bits of added fu... |
[
{
"msg_contents": "CREATE FUNCTION pg_temp.is() RETURNS text LANGUAGE sql AS $$SELECT \n'x'::text$$;\nSELECT 'x'||is();\nERROR: syntax error at or near \"(\"\nLINE 1: SELECT 'x'||is();\n\nI was finally able to figure out this was because \"is\" needs to be \nquoted; is there a way this could be hinted?\n\nFWIW... |
[
{
"msg_contents": "Hi,\n\nSome of my EnterpriseDB colleagues and I have been working on various\nparallel query projects, all of which have been previously disclosed\nhere:\n\nhttps://wiki.postgresql.org/wiki/EnterpriseDB_database_server_roadmap\n\nOne issue we've encountered is that it's not very easy for one ... |
[
{
"msg_contents": "Hi all,\n\nAs CVE-2016-5424 has put recently in light, using LF and CR in\ndatabase and role names can lead to unexpected problems in the way\nthey are handled in logical backups or generated command lines. There\nis as well a comment in the code mentioning a potential restriction\nfor that, ... |
[
{
"msg_contents": "Hi all\n\nI see this sort of question quite a bit:\n\nhttp://stackoverflow.com/q/38903811/398670\n\nwhere the user wonders why\n\nCOPY gemeenten\n FROM 'D:\\CBS_woningcijfers_2014.csv'\n DELIMITER ';' CSV\n\nfails with\n\nERROR: could not open file \"D:\\CBS_woningcijfers_2014.csv\" for re... |
[
{
"msg_contents": "Hello, hackers!\n\n\nAt this moment EXPLAIN ANALYZE with turned off timing option after \nexecution query that expires trigger prints time of total execution of \ntrigger function:\n\npostgres=# EXPLAIN (ANALYZE, timing false) insert INTO foo \nvalues(1000001, '');\n ... |
[
{
"msg_contents": "Hackers,\n\nI'm now exploring code working with heap tuples. The following code\nin heap_update() catch my eyes.\n\nif (DoesMultiXactIdConflict((MultiXactId) xwait, infomask,\n> *lockmode))\n> {\n> LockBuffer(buffer, BUFFER_LOCK_UNLOCK);\n> /* acquire tuple lock, if necessary */\n> heap_acqu... |
[
{
"msg_contents": "Hi All,\nWhen I am calling dsm_create on Linux using the POSIX DSM implementation can succeed, but result in SIGBUS when later try to access the memory. This happens because of my system does not have enough shm space & current allocation in dsm_impl_posix does not allocate disk blocks[1]. ... |
[
{
"msg_contents": "point <-> point, circle <-> point and polygon <-> point all exist as\norderable-by-operator operators (in fact they are the only ones by\ndefault). But there's no box <-> point operator at all, and no index\nsupport for box <-> box.\n\nWas this intentional, or just a strange oversight?\n\n-- ... |
[
{
"msg_contents": "Hi,\n\nAbout a month or two ago I reported a pg_dump bug regarding tables (and\nother objects) created inside a schema from an extension.\n\nObjects created by the extensions are not dumped, as they will be\ncreated once again with the CREATE EXTENSION call, but and other objects\nwhich might... |
[
{
"msg_contents": "Many have expressed their interest in this topic, but I haven't seen any\ndesign of how it should work. Here's my attempt; I've been playing with\nthis for some time now and I think what I propose here is a good initial\nplan. This will allow us to write permanent table storage that works\n... |
[
{
"msg_contents": "Doesn't tuplesort_heap_siftup() actually shift-down?\n\nThe Wikipedia article on heaps [1] lists \"shift-down\" (never \"sift\ndown\", FWIW) as a common operation on a heap:\n\n\"shift-down: move a node down in the tree, similar to shift-up; used\nto restore heap condition after deletion or r... |
[
{
"msg_contents": "Hi All,\n\nRecently while running tpc-h queries on postgresql master branch, I am\nnoticed\nrandom server crash. Most of the time server crash coming while turn tpch\nquery\nnumber 3 - (but its very random).\n\nCall Stack of server crash:\n\n(gdb) bt\n#0 0x00000000102aa9ac in ExplainNode (pl... |
[
{
"msg_contents": "I am getting corrupted Bloom indexes in which a tuple in the table\nheap is not in the index.\n\nI see it as early as commit a9284849b48b, with commit e13ac5586c49c\ncherry picked onto it. I don't see it before a9284849b48b because the\ntest-case seg faults before anything interesting can ha... |
[
{
"msg_contents": "Just to make sure that everyone's on the same page: the schedule the\nrelease team has agreed to is that we'll branch off REL9_6_STABLE on\nAug 15 (Monday!) and issue 9.6rc1 the week of Aug 29. Barring scary\nbugs emerging, this should keep us on track for 9.6.0 release in\nlate September.\n... |
[
{
"msg_contents": "Hi hackers,\n\nI would like to propose \"barriers\" for Postgres processes. A barrier\nis a very simple mechanism for coordinating parallel computation, as\nfound in many threading libraries.\n\nFirst, you initialise a Barrier object somewhere in shared memory,\nmost likely in the DSM segmen... |
[
{
"msg_contents": "Hi,\n\nwhen I do\n\n select 1.0e+1001::numeric;\n\nI get\n\ninvalid input syntax for type numeric: \"1.0e+1001\"\n\nBut\n\nselect 1.0e+1000::numeric;\n\nseems to be valid.\n\nPostgresql documentation tells me that numeric data type has an upper\nlimit of 131072 digits before decimal point.\n\... |
[
{
"msg_contents": "Hello Postgres Team!\n\nThis is to fix an issue that came up for me when running initdb.\n\nAt the end of a successful initdb it says:\n\n Success. You can now start the database server using:\n pg_ctl -D /some/path/to/data -l logfile start\n\nbut this command did not work for me be... |
[
{
"msg_contents": "Hello Postgres Hackers!\n\nI sent this already a few hours ago but it got blocked because I had not\nyet joined the mailing list. Trying again, sorry for any redundancy or\nconfusion!\n\nThis is to fix an issue that came up for me when running initdb.\n\nAt the end of a successful initdb it ... |
[
{
"msg_contents": "I did a trial run following the current pgindent README procedure, and\nnoticed that the perltidy step left me with a pile of '.bak' files\nlittering the entire tree. This seems like a pretty bad idea because\na naive \"git add .\" would have committed them. It's evidently because\nsrc/tool... |
[
{
"msg_contents": "Hi folks --\n\nWe have ported Postgres over to the C++ language (C++11 standard).\n\nhttps://github.com/jarulraj/postgresql-cpp\n\nOur goal is to use certain features of the C++ language and its standard\nlibrary to simplify coding, improve code reuse, and avoid bugs. Peter's\narticle titled ... |
[
{
"msg_contents": "I recently had a client contact me thinking that a CREATE MATERIALIZED \nVIEW had somehow managed to keep running in the background after being \n^C'd in psql. Turns out this confusion was because their monitoring \neventually complained about the still open (but now aborted) \ntransaction. T... |
[
{
"msg_contents": "Hi,\n\nDuring the recovery process, It would be nice if PostgreSQL generates an\nerror by aborting the recovery process (instead of starting-up the cluster)\nif the intended recovery target point is not reached and give an option to\nDBA to resume the recovery process from where it exactly st... |
[
{
"msg_contents": "Hi All,\n\nWhile working on pg_hba_lookup function that can be used to lookup for an client\nauthentication that can be matched for given input parameters, Tom raised some\nconcrete use case issues in the following mail [1]. In this same\nthread, he raised\nsome advantages of having a view si... |
[
{
"msg_contents": "After doing the tedious and easily forgotten (I almost did forget)\nminor version bumps for our shared libraries,\nhttps://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=0b9358d4406b9b45a06855d53f491cc7ce9550a9\n\nit suddenly struck me that in the brave new two-part-version-number... |
[
{
"msg_contents": "src/timezone/README saith\n\n When there has been a new release of Windows (probably including Service\n Packs), the list of matching timezones need to be updated. Run the\n script in src/tools/win32tzlist.pl on a Windows machine running this new\n release and apply any new timezones that... |
[
{
"msg_contents": "Recently a hacker proposed a patch to add pg_dynshmem to the list of\ndirectories whose contents are excluded in pg_basebackup. I wasn't able\nto find the original email despite several attempts.\n\nThat patch got me thinking about what else could be excluded and after\nsome investigation I ... |
[
{
"msg_contents": "Hi, hackers!\n\nThere is a known issue that index only scan (IOS) can only work with \nsimple index keys based on single attributes and doesn't work with index \nexpressions. In this patch I propose a solution that adds support of IOS \nfor index expressions. Here's an example:\n\ncreate tabl... |
[
{
"msg_contents": "Any reason why we can create a function that accepts anyelement and \nreturns anyarray, but can't do the same with anyrange? Could we attempt \nto match each range subtype looking for a match?\n\ncreate function range__create(anyelement,anyelement,text = '[]') RETURNS \nanyrange LANGUAGE plpg... |
[
{
"msg_contents": "Hi all\n\nWhile implementing support for traceable transactions (finding out after\nthe fact whether an xact committed or aborted), I've found that Pg is very\ninconsistent with what it considers a transaction ID from a user facing\npoint of view, to the point where I think it's hard for user... |
[
{
"msg_contents": "There is something rotten in the state of Denmark. Here are four recent\nruns that failed with unexpected \"cache lookup failed for type nnnn\"\nerrors:\n\nhttp://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=grouse&dt=2016-08-16%2008%3A39%3A03\nhttp://buildfarm.postgresql.org/cgi-bin/show... |
[
{
"msg_contents": "I propose the attached patch to clean up the comment formatting in\nplpgsql.h.\n\n-- \nPeter Eisentraut 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... |
[
{
"msg_contents": "\nHi all, been awhile! Some may be saying \"not long enough\" but eh.\n\nI recently moved a mybb forum away from mysql to postgres. Along the way \nI encountered a couple of things that either didn't seem to be supported \nor I'm just not doing it right.\n\nFirst, the server this is on is r... |
[
{
"msg_contents": "Here is a patch for implementing the NEXT VALUE FOR expression. This is\nthe SQL-standard conforming version of our nextval() function, and it's\nalso used by Oracle, MS SQL, DB2. Example:\n\nSELECT NEXT VALUE FOR foo_seq;\n\nThe second patch changes the serial column to use this new expres... |
[
{
"msg_contents": "From: https://www.postgresql.org/docs/9.4/static/sql-createtable.html\n\n\"The access method must support amgettuple (see Chapter 55); at\npresent this means GIN cannot be used. Although it's allowed, there is\nlittle point in using B-tree or hash indexes with an exclusion\nconstraint, becaus... |
[
{
"msg_contents": "Hi,\n\nThe Uber blog post, among other things, pointed out that PG uses lseek + \nread instead of pread. I didn't see any discussion around that and my \nGoogle searches didn't find any posts about pread / pwrite for the past \n10 years.\n\nWith that plus the \"C++ port\" thread in mind, I w... |
[
{
"msg_contents": "Hi,\n\nI am currently exploring pglogical replication for my db servers. I would\nlike to know how can I automatically failover from Provider Node to\nSubscriber Node, if the Provider node goes down for some reasons. How can I\nredirect all the traffic to SubscriberNode automatically ? In the... |
[
{
"msg_contents": "/* only needed in OS X 10.1 and possibly early 10.2 releases */\n\nMaybe it's time to let it go?\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hacker... |
[
{
"msg_contents": "https://wiki.postgresql.org/wiki/Binary_Replication_Tutorial does not \nspecify -c for any of the rsync commands. That's maybe safe for WAL, but \nI don't think it's safe for any of the other uses, right? I'd like \nsomeone to confirm before I just change the page... my intention is to \njust... |
[
{
"msg_contents": "During preliminary analysis of what it would take to produce a\nparallel CLUSTER patch that is analogous of what I came up with for\nCREATE INDEX, which in general seems quite possible, I identified\nreform_and_rewrite_tuple() as a major bottleneck for the current\nCLUSTER implementation.\n\n... |
[
{
"msg_contents": "Fix deletion of speculatively inserted TOAST on conflict\n\nINSERT .. ON CONFLICT runs a pre-check of the possible conflicting\nconstraints before performing the actual speculative insertion. In case\nthe inserted tuple included TOASTed columns the ON CONFLICT condition\nwould be handled co... |
[
{
"msg_contents": "Fix deletion of speculatively inserted TOAST on conflict\n\nINSERT .. ON CONFLICT runs a pre-check of the possible conflicting\nconstraints before performing the actual speculative insertion. In case\nthe inserted tuple included TOASTed columns the ON CONFLICT condition\nwould be handled co... |
[
{
"msg_contents": "Hi,\n\nI was trying to implement a middleware that lies between client and\npostgres.\n\nSo, this middleware is supposed to run query with libpq, do its job on\nthem, and then serialize the result of query, and send it to the client !\n(client deserializes to PGresult)\n\nI could simply itera... |
[
{
"msg_contents": "The attached patch tries to maintain the initial status of B-Tree\nindexes, which are created with equal-key runs in physical order,\nduring the whole life of the B-Tree, and make key-tid pairs\nefficiently searchable in the process.\n\nThe patch doesn't add an interface to perform the key-ti... |
[
{
"msg_contents": "Hi hackers,\n\nTo do something about the confusion I keep seeing about what exactly\n\"on\" means, I've often wished we had \"remote_flush\". But it's not\nobvious how the backwards compatibility could work, ie how to keep the\npeople happy who use \"local\" vs \"on\" to control syncrep, and... |
[
{
"msg_contents": "Attached is a patch I'd fooled around with back in July but not submitted.\nThe idea is that, if our initial scan of the query tree found only\nparallel-safe functions, there is no need to rescan subsets of the tree\nlooking for parallel-restricted functions. We can mechanize that by\nsaving... |
[
{
"msg_contents": "The regex library used to have a notion of a \"collating element\"\nthat was distinct from a \"character\", but Henry Spencer never\nactually implemented support for multi-character collating elements,\nand the Tcl crew ripped out most of the stubs for it years ago.\nThe only thing left that ... |
[
{
"msg_contents": "I was looking at this TODO item from 2009:\n\n\thttps://www.postgresql.org/message-id/4AA7B197.70002%40usit.uio.no\n\nI have implemented this in the attached patch. It does two things:\n\no compares words in columns that can only support keywords as\ncase-insensitive, double-quoted or not\n... |
[
{
"msg_contents": "I have a query which contains a where clause like:\n\n aid =ANY(VALUES (1),(45),(87), <6948 more>, (4999947))\n\nfor example:\n\nperl -le 'print \"explain (analyze) select sum(abalance) from\npgbench_accounts where aid=ANY(VALUES \"; print join \",\", map \"($_)\", sort\n{$a<=>$b} map int(ran... |
[
{
"msg_contents": "While researching a customer issue with BDR I noticed that one ereport()\ncall happens after clobbering errno, leading to the wrong strerror being\nreported. This patch fixes it by saving before calling\nCloseTransientFile and restoring afterwards.\n\nI also threw in a missing errcode I noti... |
[
{
"msg_contents": "Hackers,\n\noriginally this idea was proposed by Andres Freund while experimenting with\nlockfree Pin/UnpinBuffer [1].\nThe patch is attached as well as results of pgbench -S on 72-cores\nmachine. As before it shows huge benefit in this case.\nFor sure, we should validate that it doesn't cau... |
[
{
"msg_contents": "I would like to proppse to export these functions in libpq.\n\npqPutMsgStart\npqPutMsgEnd\npqPutc\npqPuts\npqPutInt\npqPutnchar\npqFlush\npqHandleSendFailure\n\nI think this would be useful to create a tool/library which needs to\nhandle frontend/backend protocol messages in detail.\n\nBest r... |
[
{
"msg_contents": "Hi hackers,\n\nI would like to propose a new subsystem called Dynamic Shared [Memory]\nAreas, or \"DSA\". It provides an object called a \"dsa_area\" which can\nbe used by multiple backends to share data. Under the covers, a\ndsa_area is made up of some number of DSM segments, but it appear... |
[
{
"msg_contents": "Hi,\n\nWe are using logical decoding in multimaster and we are faced with the problem that inconsistent transactions are sent to replica.\nBriefly, multimaster is using logical decoding in this way:\n1. Each multimaster node is connected with each other using logical decoding channel and so e... |
[
{
"msg_contents": "Hi\n\nI am sending implementation of xmltable function. The code should to have\nnear to final quality and it is available for testing.\n\nI invite any help with documentation and testing.\n\nRegards\n\nPavel\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo mak... |
[
{
"msg_contents": "Hi,\n\nThe postgres_fdw join pushdown in 9.6 is great, but it can't handle full \njoins on relations with restrictions. The reason for that is, \npostgres_fdw can't support deparsing subqueries when creating a remote \njoin query. So, by adding the deparsing logic to it, I removed that ... |
[
{
"msg_contents": "plpgsql.h defines a number of enums, but most of the code passes them\naround as ints. The attached patch updates structs and function\nprototypes to take enum types instead. This clarifies the struct\ndefinitions in plpgsql.h in particular.\n\nI didn't deal with the PLPGSQL_RC_* symbols, s... |
[
{
"msg_contents": "Use LEFT JOINs in some system views in case referenced row doesn't exist.\n\nIn particular, left join to pg_authid so that rows in pg_stat_activity\ndon't disappear if the session's owning user has been dropped.\nAlso convert a few joins to pg_database to left joins, in the same spirit,\nthou... |
[
{
"msg_contents": "I found another bug as a result of using amcheck on Heroku customer\ndatabases. This time, the bug is in core Postgres. It's one of mine.\n\nThere was a thinko in tuplesort's abbreviation abort logic, causing\ncertain SortTuples to be spuriously marked NULL (and so, subsequently\nsorted as a ... |
[
{
"msg_contents": "Hello,\ni'm testing the new row security level functionality in postgresql 9.5.\nTo do that i run this script:\n\n-----------cut here ----------------------\n\nCREATE TABLE public.policy_tab\n(\n id bigint NOT NULL,\n description character varying(160) NOT NULL,\n usr name NOT NULL,\n CO... |
[
{
"msg_contents": "I have done my yearly TODO list cleanup, and I did more extensive item\nremoval this time. There were a number of entries I could not figure out\nso if people want to review what is left and remove items that are\nundesired or done, please do that. Thanks.\n\n\thttps://wiki.postgresql.org/w... |
[
{
"msg_contents": "Someone reported that a replication slot that existed at the time a base\nbackup was done on the master was copied to the standby. Because they\ndidn't realize it, their WAL was not being recycled on the standby.\n\nIs that possible? Is it a known behavior? I don't see it documented.\n\n--... |
[
{
"msg_contents": "Hi Yury,\n\n\nOn Fri, Aug 19, 2016 at 9:46 AM, Yury Zhuravlev\n<u.zhuravlev@postgrespro.ru> wrote:\n> Christian Convey wrote:\n>>\n>> I'm interested in helping with your CMake effort. I don't have any\n>> experience contributing to PG, but I do have some free time at the\n>> moment. Please ... |
[
{
"msg_contents": "This is a new statistics view that is used to provide the number of\nSQL operations that are\nhappened on a particular interval of time. This view is useful for the\nsystem to find out the\npattern of the operations that are happening in the instance during\nparticular interval of\ntime.\n\nF... |
[
{
"msg_contents": "Hi all\n\nFollowing on from\n\nbigint txids vs 'xid' type, new txid_recent(bigint) => xid\n\nhttps://www.postgresql.org/message-id/CAMsr+YFDZMN_iZ7KrRoe+j0KVLQvFVgvZxbcVxR-MLjgtoZugA@mail.gmail.com\n\n\nhere's a patch that implements a txid_status(bigint) function to report the\ncommit status... |
[
{
"msg_contents": "Make initdb's suggested \"pg_ctl start\" command line more reliable.\n\nThe original coding here was not nearly careful enough about quoting\nspecial characters, and it didn't get corner cases right for constructing\nthe pg_ctl path either. Use join_path_components() and appendShellString()\... |
[
{
"msg_contents": "Hi all\n\nMy earlier efforts at logical decoding of sequence advances were too\nsimplistic[1], falling afoul of issues with sequences being both\ntransactional and not transactional depending on whether the sequence is\ncreated in the current xact or not.\n\nTL;DR of solution:\n\n* extend Seq... |
[
{
"msg_contents": "Just out of curiopusity, I wonder why we can't make the encoding of\nSGML docs to be UTF-8, rather than current ISO-8859-1.\n\nAs long as everything is written in ASCII, the size of docs will be\nalmost same even if UTF-8 is used. Plus, if the encoding is changed to\nUTF-8, it is very easy to... |
[
{
"msg_contents": "While hacking on the CSN patch, I spotted a race condition between \nGetOldestActiveTransactionId() and GetNewTransactionId(). \nGetOldestActiveTransactionId() calculates the oldest XID that's still \nrunning, by doing:\n\n1. Read nextXid, without a lock. This is the upper bound, if there are... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.