\\r\\n> \\r\\n> \\r\\n> \\r\\n> Ok, changed.\\r\\n> \\r\\n> \\r\\n> \\tDavid, your patch avoids creating any mergefuncs for existing\\r\\n> \\taggregates. We would need to supply working examples for at least\\r\\n> a\\r\\n> \\tfew of the builtin aggregates, so we can test the infrastructure.\\r\\n> We\\r\\n> \\tcan add examples for all cases later.\\r\\n> \\r\\n> \\r\\n> \\r\\n> \\r\\n> In addition to MIN(), MAX(), BIT_AND(), BIT_OR, SUM() for floating point\\r\\n> types, cash and interval. I've now added combine functions for count(*)\\r\\n> and count(col). It seems that int8pl() is suitable for this.\\r\\n> \\r\\n> \\r\\n> Do you think it's worth adding any new functions at this stage, or is it\\r\\n> best to wait until there's a patch which can use these?\\r\\n> \\r\\n> Regards\\r\\n> \\r\\n> David Rowley\\r\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 18 Feb 2015 08:13:10 +0000\",\n \"msg_from\": \"Kouhei Kaigai <kaigai@ak.jp.nec.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 18.2.2015 09:13, Kouhei Kaigai wrote:\\n>>\\n>> In addition to MIN(), MAX(), BIT_AND(), BIT_OR, SUM() for floating\\n>> point types, cash and interval. I've now added combine functions\\n>> for count(*) and count(col). It seems that int8pl() is suitable for\\n>> this.\\n>>\\n>>\\n>> Do you think it's worth adding any new functions at this stage, or \\n>> is it best to wait until there's a patch which can use these?\\n\\nA few comments about this patch:\\n\\n1) another use case / grouping sets\\n-----------------------------------\\n\\nI agree this would be nice to have in-core.\\n\\nI remember discussing this functionality (combining partial aggregate\\nresults) as an alternative implementation to grouping sets. IIRC the\\ngrouping sets patch implements this by repeatedly sorting the tuples,\\nbut in some cases we could compute the aggregates at the most detailed\\nlevel, and then build the results by combining the partial results. Just\\nan idea, at this moment, though.\\n\\n\\n2) serialize/deserialize functions\\n----------------------------------\\n\\nThis thread mentions \\\"parallel queries\\\" as a use case, but that means\\npassing data between processes, and that requires being able to\\nserialize and deserialize the aggregate state somehow. For actual data\\ntypes that's not overly difficult I guess (we can use in/out functions),\\nbut what about aggretates using 'internal' state? I.e. aggregates\\npassing pointers that we can't serialize?\\n\\nAnd we do have plenty of those, including things like\\n\\n array_agg\\n avg\\n cume_dist\\n dense_rank\\n json_agg\\n jsonb_agg\\n jsonb_object_agg\\n json_object_agg\\n mode\\n percentile_cont\\n percentile_disc\\n percent_rank\\n rank\\n stddev\\n stddev_pop\\n stddev_samp\\n string_agg\\n sum\\n variance\\n var_pop\\n var_samp\\n\\nThose are pretty important aggregates IMHO, and ISTM we won't be able to\\nuse them with this patch. Or am I missing something?\\n\\nSo maybe this patch should include support for serialize/deserialize\\nfunctions too? Or maybe a follow-up patch ... I'm not entirely\\ncomfortable with a patch without an actual use case, except for a simple\\nexample. But maybe that's OK.\\n\\nFWIW the serialize/deserialize functions would be useful also for\\nimplementing a truly 'memory-bounded hash aggregate' (discussed\\nelsewhere, currently stuck because of difficulty with implementing\\nmemory accounting). So that's yet another use case for this (both the\\n'combine' function and the 'serialize/deserialize').\\n\\nregards\\n\\n-- \\nTomas Vondra 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, 20 Feb 2015 19:16:14 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"Is there a case where the combining function is different from the\\ntransition function, other than for count?\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 20 Feb 2015 15:01:52 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 20.2.2015 21:01, Peter Eisentraut wrote:\\n> Is there a case where the combining function is different from the\\n> transition function, other than for count?\\n\\nIt's different in all the cases when the aggregate state is not\\nidentical to a single value - for example the usual avg(), sum() and\\nstddev() aggregates keep state which is equal to\\n\\n {count(X), sum(X), sum(X*X)}\\n\\nThe 'combine' function gets two such 'state' values, while transition\\ngets 'state' + next value.\\n\\nI'm inclined to say that 'combinefn == transfn' is a minority case.\\n\\n-- \\nTomas Vondra 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, 20 Feb 2015 21:09:49 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 2/20/15 3:09 PM, Tomas Vondra wrote:\\n> On 20.2.2015 21:01, Peter Eisentraut wrote:\\n>> Is there a case where the combining function is different from the\\n>> transition function, other than for count?\\n> \\n> It's different in all the cases when the aggregate state is not\\n> identical to a single value - for example the usual avg(), sum() and\\n> stddev() aggregates keep state which is equal to\\n> \\n> {count(X), sum(X), sum(X*X)}\\n> \\n> The 'combine' function gets two such 'state' values, while transition\\n> gets 'state' + next value.\\n\\nThat's just because the count is hidden there in an opaque custom\\ntransition function. If, say, we had instead an array of transition\\nfunctions {inc, plus, plussq} and we knew that plus and plussq are\\nassociative operators, all we'd need to special case is the count case.\\n This would avoid a lot of repetitive code for stddev, avg, etc.\\n\\n(As a bonus, you could use this knowledge to compute count, sum, avg,\\nand stddev in parallel using only three counters.)\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 20 Feb 2015 15:23:24 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 20.2.2015 21:23, Peter Eisentraut wrote:\\n> On 2/20/15 3:09 PM, Tomas Vondra wrote:\\n>> On 20.2.2015 21:01, Peter Eisentraut wrote:\\n>>> Is there a case where the combining function is different from the\\n>>> transition function, other than for count?\\n>>\\n>> It's different in all the cases when the aggregate state is not\\n>> identical to a single value - for example the usual avg(), sum() and\\n>> stddev() aggregates keep state which is equal to\\n>>\\n>> {count(X), sum(X), sum(X*X)}\\n>>\\n>> The 'combine' function gets two such 'state' values, while transition\\n>> gets 'state' + next value.\\n> \\n> That's just because the count is hidden there in an opaque custom \\n> transition function. If, say, we had instead an array of transition \\n> functions {inc, plus, plussq} and we knew that plus and plussq are \\n> associative operators, all we'd need to special case is the count\\n> case. This would avoid a lot of repetitive code for stddev, avg, etc.\\n\\nUmmm, I'm not entirely sure I understand that, but the main point was\\nthat the current implementation does not work like that. We have no idea\\nwhat transition functions are transitive, and we do have opaque\\naggregate states.\\n\\nAlso, there are aggregate functions like array_agg() or string_agg()\\nthat make this impossible, just like for many custom aggregates (like\\nhyperloglog for example). Again, I might not understand the idea\\ncorrectly ...\\n\\n> (As a bonus, you could use this knowledge to compute count, sum, avg,\\n> and stddev in parallel using only three counters.)\\n\\nYeah, sharing the aggregate state by multiple aggregates would be nice.\\n\\nregards\\n\\n-- \\nTomas Vondra 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, 20 Feb 2015 21:32:50 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 2/20/15 3:32 PM, Tomas Vondra wrote:\\n>> That's just because the count is hidden there in an opaque custom \\n>> transition function. If, say, we had instead an array of transition \\n>> functions {inc, plus, plussq} and we knew that plus and plussq are \\n>> associative operators, all we'd need to special case is the count\\n>> case. This would avoid a lot of repetitive code for stddev, avg, etc.\\n> \\n> Ummm, I'm not entirely sure I understand that, but the main point was\\n> that the current implementation does not work like that. We have no idea\\n> what transition functions are transitive, and we do have opaque\\n> aggregate states.\\n\\nWell, my point is that you could make it work that way and make your\\ncurrent patch a lot smaller and simpler.\\n\\n> Also, there are aggregate functions like array_agg() or string_agg()\\n> that make this impossible, just like for many custom aggregates (like\\n> hyperloglog for example). Again, I might not understand the idea\\n> correctly ...\\n\\nHow would a combining function work for something like array_agg()? I\\ndon't think it would, at least if you want to preserve the ordering\\noption for the user.\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 24 Feb 2015 14:15:51 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 2/20/15 3:09 PM, Tomas Vondra wrote:\\n> The 'combine' function gets two such 'state' values, while transition\\n> gets 'state' + next value.\\n\\nI think the combine function is not actually a property of the\\naggregate, but a property of the transition function. If two aggregates\\nhave the same transition function, they will also have the same combine\\nfunction. The combine function really just says, how do you combine two\\nseries of these function calls. Say maybe this should be put into\\npg_proc instead. (Or you make the transition functions transition\\noperators and put the info into pg_operator instead, which is where\\nfunction call optimization information is usually kept.)\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 24 Feb 2015 14:20:32 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Tue, Feb 24, 2015 at 2:20 PM, Peter Eisentraut <peter_e@gmx.net> wrote:\\n> On 2/20/15 3:09 PM, Tomas Vondra wrote:\\n>> The 'combine' function gets two such 'state' values, while transition\\n>> gets 'state' + next value.\\n>\\n> I think the combine function is not actually a property of the\\n> aggregate, but a property of the transition function. If two aggregates\\n> have the same transition function, they will also have the same combine\\n> function. The combine function really just says, how do you combine two\\n> series of these function calls. Say maybe this should be put into\\n> pg_proc instead. (Or you make the transition functions transition\\n> operators and put the info into pg_operator instead, which is where\\n> function call optimization information is usually kept.)\\n\\nThis seems like a weird design to me. It's probably true that if the\\ntransition function is the same, the state-combiner function will also\\nbe the same. But the state-combiner function is only going to exist\\nfor aggregate transition functions, not functions or operators in\\ngeneral. So linking it from pg_proc or pg_operator feels wrong to me.\\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, 3 Mar 2015 19:00:16 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Tue, Feb 24, 2015 at 2:20 PM, Peter Eisentraut <peter_e@gmx.net> wrote:\\n>> I think the combine function is not actually a property of the\\n>> aggregate, but a property of the transition function. If two aggregates\\n>> have the same transition function, they will also have the same combine\\n>> function. The combine function really just says, how do you combine two\\n>> series of these function calls. Say maybe this should be put into\\n>> pg_proc instead. (Or you make the transition functions transition\\n>> operators and put the info into pg_operator instead, which is where\\n>> function call optimization information is usually kept.)\\n\\n> This seems like a weird design to me. It's probably true that if the\\n> transition function is the same, the state-combiner function will also\\n> be the same. But the state-combiner function is only going to exist\\n> for aggregate transition functions, not functions or operators in\\n> general. So linking it from pg_proc or pg_operator feels wrong to me.\\n\\nFWIW, I agree with Robert. It's better to keep this in pg_aggregate.\\nWe don't need yet another mostly-empty column in pg_proc; and as for\\npg_operator, there is no expectation that an aggregate transition function\\nis in pg_operator.\\n\\nAlso, I don't think it's a foregone conclusion that same transition\\nfunction implies same combiner function: that logic falls apart when\\nyou consider that one of them might be polymorphic and the other not.\\n(Admittedly, that probably means the aggregate creator is missing a bet;\\nbut we should not design the catalog schema in a way that says that only\\nmaximally efficient aggregate designs are allowed.)\\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, 03 Mar 2015 19:55:49 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 25 February 2015 at 08:15, Peter Eisentraut <peter_e@gmx.net> wrote:\\n\\n> On 2/20/15 3:32 PM, Tomas Vondra wrote:\\n>\\n> > Also, there are aggregate functions like array_agg() or string_agg()\\n> > that make this impossible, just like for many custom aggregates (like\\n> > hyperloglog for example). Again, I might not understand the idea\\n> > correctly ...\\n>\\n> How would a combining function work for something like array_agg()? I\\n> don't think it would, at least if you want to preserve the ordering\\n> option for the user.\\n>\\n>\\nThey just wouldn't work in that case. We'd simply not have a combine\\nfunction for that aggregate.\\n\\nThe yet to be written code, (say parallel hash aggregate), the planner\\nwould have to ensure that each aggregate function being used had a combine\\nfunction, if any aggregate in the current query level didn't have one then\\nit would not parallelise the query.\\n\\nRegards\\n\\nDavid Rowley\\n\\nOn 25 February 2015 at 08:15, Peter Eisentraut <peter_e@gmx.net> wrote:On 2/20/15 3:32 PM, Tomas Vondra wrote:\\n> Also, there are aggregate functions like array_agg() or string_agg()\\n> that make this impossible, just like for many custom aggregates (like\\n> hyperloglog for example). Again, I might not understand the idea\\n> correctly ...\\n\\nHow would a combining function work for something like array_agg()? I\\ndon't think it would, at least if you want to preserve the ordering\\noption for the user.\\nThey just wouldn't work in that case. We'd simply not have a combine function for that aggregate.The yet to be written code, (say parallel hash aggregate), the planner would have to ensure that each aggregate function being used had a combine function, if any aggregate in the current query level didn't have one then it would not parallelise the query.RegardsDavid Rowley\",\n \"msg_date\": \"Wed, 4 Mar 2015 22:23:29 +1300\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 18 February 2015 at 21:13, Kouhei Kaigai <kaigai@ak.jp.nec.com> wrote:\\n>\\n> This patch itself looks good as an infrastructure towards\\n> the big picture, however, we still don't reach the consensus\\n> how combined functions are used instead of usual translation\\n> functions.\\n\\n\\nThank you for taking the time to look at the patch.\\n\\n>\\n> Aggregate function usually consumes one or more values extracted\\n> from a tuple, then it accumulates its internal state according\\n> to the argument. Exiting transition function performs to update\\n> its internal state with assumption of a function call per records.\\n> On the other hand, new combined function allows to update its\\n> internal state with partial aggregated values which is processed\\n> by preprocessor node.\\n> An aggregate function is represented with Aggref node in plan tree,\\n> however, we have no certain way to determine which function shall\\n> be called to update internal state of aggregate.\\n>\\n>\\nThis is true, there's nothing done in the planner to set any sort of state\\nin the aggregation nodes to tell them weather to call the final function or\\nnot. It's quite hard to know how far to go with this patch. It's really\\nonly intended to provide the necessary infrastructure for things like\\nparallel query and various other possible usages of aggregate combine\\nfunctions. I don't think it's really appropriate for this patch to go\\nadding such a property to any nodes as there would still be nothing in the\\nplanner to actually set those properties... The only thing I can think of\\nto get around this is implement the most simple use for combine aggregate\\nfunctions, the problem with that is, that the most simple case is not at\\nall simple.\\n\\n\\n\\n> For example, avg(float) has an internal state with float[3] type\\n> for number of rows, sum of X and X^2. If combined function can\\n> update its internal state with partially aggregated values, its\\n> argument should be float[3]. It is obviously incompatible to\\n> float8_accum(float) that is transition function of avg(float).\\n> I think, we need a new flag on Aggref node to inform executor\\n> which function shall be called to update internal state of\\n> aggregate. Executor cannot decide it without this hint.\\n>\\n> Also, do you have idea to push down aggregate function across\\n> joins? Even though it is a bit old research, I could find\\n> a systematic approach to push down aggregate across join.\\n> https://cs.uwaterloo.ca/research/tr/1993/46/file.pdf\\n>\\n>\\nI've not read the paper yet, but I do have a very incomplete WIP patch to\\ndo this. I've just not had much time to work on it.\\n\\n\\n> I think, it is great if core functionality support this query\\n> rewriting feature based on cost estimation, without external\\n> modules.\\n>\\n\\nRegards\\n\\nDavid Rowley\\n\\nOn 18 February 2015 at 21:13, Kouhei Kaigai <kaigai@ak.jp.nec.com> wrote:\\nThis patch itself looks good as an infrastructure towards\\nthe big picture, however, we still don't reach the consensus\\nhow combined functions are used instead of usual translation\\nfunctions.Thank you for taking the time to look at the patch.\\n\\nAggregate function usually consumes one or more values extracted\\nfrom a tuple, then it accumulates its internal state according\\nto the argument. Exiting transition function performs to update\\nits internal state with assumption of a function call per records.\\nOn the other hand, new combined function allows to update its\\ninternal state with partial aggregated values which is processed\\nby preprocessor node.\\nAn aggregate function is represented with Aggref node in plan tree,\\nhowever, we have no certain way to determine which function shall\\nbe called to update internal state of aggregate.\\nThis is true, there's nothing done in the planner to set any sort of state in the aggregation nodes to tell them weather to call the final function or not. It's quite hard to know how far to go with this patch. It's really only intended to provide the necessary infrastructure for things like parallel query and various other possible usages of aggregate combine functions. I don't think it's really appropriate for this patch to go adding such a property to any nodes as there would still be nothing in the planner to actually set those properties... The only thing I can think of to get around this is implement the most simple use for combine aggregate functions, the problem with that is, that the most simple case is not at all simple. \\nFor example, avg(float) has an internal state with float[3] type\\nfor number of rows, sum of X and X^2. If combined function can\\nupdate its internal state with partially aggregated values, its\\nargument should be float[3]. It is obviously incompatible to\\nfloat8_accum(float) that is transition function of avg(float).\\nI think, we need a new flag on Aggref node to inform executor\\nwhich function shall be called to update internal state of\\naggregate. Executor cannot decide it without this hint.\\n\\nAlso, do you have idea to push down aggregate function across\\njoins? Even though it is a bit old research, I could find\\na systematic approach to push down aggregate across join.\\nhttps://cs.uwaterloo.ca/research/tr/1993/46/file.pdf\\nI've not read the paper yet, but I do have a very incomplete WIP patch to do this. I've just not had much time to work on it. \\nI think, it is great if core functionality support this query\\nrewriting feature based on cost estimation, without external\\nmodules. RegardsDavid Rowley\",\n \"msg_date\": \"Wed, 4 Mar 2015 22:37:19 +1300\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 21 February 2015 at 07:16, Tomas Vondra <tomas.vondra@2ndquadrant.com>\\nwrote:\\n\\n>\\n> 2) serialize/deserialize functions\\n> ----------------------------------\\n>\\n> This thread mentions \\\"parallel queries\\\" as a use case, but that means\\n> passing data between processes, and that requires being able to\\n> serialize and deserialize the aggregate state somehow. For actual data\\n> types that's not overly difficult I guess (we can use in/out functions),\\n> but what about aggretates using 'internal' state? I.e. aggregates\\n> passing pointers that we can't serialize?\\n>\\n>\\nThis is a good question. I really don't know the answer to it as I've not\\nlooked at the Robert's API for passing data between backends yet.\\n\\nMaybe Robert or Amit can answer this?\\n\\nRegards\\n\\nDavid Rowley\\n\\nOn 21 February 2015 at 07:16, Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote:\\n2) serialize/deserialize functions\\n----------------------------------\\n\\nThis thread mentions \\\"parallel queries\\\" as a use case, but that means\\npassing data between processes, and that requires being able to\\nserialize and deserialize the aggregate state somehow. For actual data\\ntypes that's not overly difficult I guess (we can use in/out functions),\\nbut what about aggretates using 'internal' state? I.e. aggregates\\npassing pointers that we can't serialize?\\nThis is a good question. I really don't know the answer to it as I've not looked at the Robert's API for passing data between backends yet.Maybe Robert or Amit can answer this?RegardsDavid Rowley\",\n \"msg_date\": \"Wed, 4 Mar 2015 22:41:12 +1300\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 4, 2015 at 4:41 AM, David Rowley <dgrowleyml@gmail.com> wrote:\\n>> This thread mentions \\\"parallel queries\\\" as a use case, but that means\\n>> passing data between processes, and that requires being able to\\n>> serialize and deserialize the aggregate state somehow. For actual data\\n>> types that's not overly difficult I guess (we can use in/out functions),\\n>> but what about aggretates using 'internal' state? I.e. aggregates\\n>> passing pointers that we can't serialize?\\n>\\n> This is a good question. I really don't know the answer to it as I've not\\n> looked at the Robert's API for passing data between backends yet.\\n>\\n> Maybe Robert or Amit can answer this?\\n\\nI think parallel aggregation will probably require both the\\ninfrastructure discussed here, namely the ability to combine two\\ntransition states into a single transition state, and also the ability\\nto serialize and de-serialize transition states, which has previously\\nbeen discussed in the context of letting hash aggregates spill to\\ndisk. My current thinking is that the parallel plan will look\\nsomething like this:\\n\\nHashAggregateFinish\\n-> Funnel\\n -> HashAggregatePartial\\n -> PartialHeapScan\\n\\nSo the workers will all pull from a partial heap scan and each worker\\nwill aggregate its own portion of the data. Then it'll need to pass\\nthe results of that step back to the master, which will aggregate the\\npartial results and produce the final results. I'm guessing that if\\nwe're grouping on, say, column a, the HashAggregatePartial nodes will\\nkick out 2-column tuples of the form (<value for a>, <serialized\\ntransition state for group with that value for a>).\\n\\nOf course this is all pie in the sky right now, but I think it's a\\npretty good bet that partial aggregation is a useful thing to be able\\nto do. Postgres-XC put a bunch of work into this, too, so there's\\nlots of people saying \\\"hey, we want that\\\".\\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, 4 Mar 2015 11:11:28 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"> On Wed, Mar 4, 2015 at 4:41 AM, David Rowley <dgrowleyml@gmail.com> wrote:\\r\\n> >> This thread mentions \\\"parallel queries\\\" as a use case, but that means\\r\\n> >> passing data between processes, and that requires being able to\\r\\n> >> serialize and deserialize the aggregate state somehow. For actual data\\r\\n> >> types that's not overly difficult I guess (we can use in/out functions),\\r\\n> >> but what about aggretates using 'internal' state? I.e. aggregates\\r\\n> >> passing pointers that we can't serialize?\\r\\n> >\\r\\n> > This is a good question. I really don't know the answer to it as I've not\\r\\n> > looked at the Robert's API for passing data between backends yet.\\r\\n> >\\r\\n> > Maybe Robert or Amit can answer this?\\r\\n> \\r\\n> I think parallel aggregation will probably require both the\\r\\n> infrastructure discussed here, namely the ability to combine two\\r\\n> transition states into a single transition state, and also the ability\\r\\n> to serialize and de-serialize transition states, which has previously\\r\\n> been discussed in the context of letting hash aggregates spill to\\r\\n> disk. My current thinking is that the parallel plan will look\\r\\n> something like this:\\r\\n> \\r\\n> HashAggregateFinish\\r\\n> -> Funnel\\r\\n> -> HashAggregatePartial\\r\\n> -> PartialHeapScan\\r\\n> \\r\\n> So the workers will all pull from a partial heap scan and each worker\\r\\n> will aggregate its own portion of the data. Then it'll need to pass\\r\\n> the results of that step back to the master, which will aggregate the\\r\\n> partial results and produce the final results. I'm guessing that if\\r\\n> we're grouping on, say, column a, the HashAggregatePartial nodes will\\r\\n> kick out 2-column tuples of the form (<value for a>, <serialized\\r\\n> transition state for group with that value for a>).\\r\\n> \\r\\nIt may not be an urgent topic to be discussed towards v9.5, however,\\r\\nI'd like to raise a topic about planner and aggregates.\\r\\n\\r\\nOnce we could get the two phase aggregation, planner needs to pay\\r\\nattention possibility of partial aggregate during plan construction,\\r\\neven though our current implementation attach Agg node after the\\r\\njoin/scan plan construction.\\r\\n\\r\\nProbably, a straightforward design is to add FunnelPath with some\\r\\nchild nodes on set_rel_pathlist() or add_paths_to_joinrel().\\r\\nIts child node may be PartialAggregate node (or some other parallel\\r\\nsafe node of course). In this case, it must inform the planner this\\r\\nnode (for more correctness, targetlist of the node) returns partial\\r\\naggregation, instead of the values row-by-row.\\r\\nThen, planner need to track and care about which type of data shall\\r\\nbe returned to the upper node. Path node may have a flag to show it,\\r\\nand we also may need to inject dummy PartialAggregate if we try to\\r\\njoin a relation that returns values row-by-row and another one that\\r\\nreturns partial aggregate.\\r\\nIt also leads another problem. The RelOptInfo->targetlist will\\r\\ndepend on the Path node chosen. Even if float datum is expected as\\r\\nan argument of AVG(), its state to be combined is float[3].\\r\\nSo, we will need to adjust the targetlist of RelOptInfo, once Path\\r\\ngot chosen.\\r\\n\\r\\nAnyway, I could find out, at least, these complicated issues around\\r\\ntwo-phase aggregate integration with planner. Someone can suggest\\r\\nminimum invasive way for these integration?\\r\\n\\r\\nThanks,\\r\\n--\\r\\nNEC OSS Promotion Center / PG-Strom Project\\r\\nKaiGai Kohei <kaigai@ak.jp.nec.com>\\r\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 5 Mar 2015 04:00:10 +0000\",\n \"msg_from\": \"Kouhei Kaigai <kaigai@ak.jp.nec.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 4, 2015 at 11:00 PM, Kouhei Kaigai <kaigai@ak.jp.nec.com> wrote:\\n> Anyway, I could find out, at least, these complicated issues around\\n> two-phase aggregate integration with planner. Someone can suggest\\n> minimum invasive way for these integration?\\n\\nI don't think I have the brain space to think about that until we get\\nthe basic parallelism stuff in.\\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, 5 Mar 2015 10:50:09 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 5, 2015 at 9:30 AM, Kouhei Kaigai <kaigai@ak.jp.nec.com> wrote:\\n\\n> > On Wed, Mar 4, 2015 at 4:41 AM, David Rowley <dgrowleyml@gmail.com>\\n> wrote:\\n> > >> This thread mentions \\\"parallel queries\\\" as a use case, but that means\\n> > >> passing data between processes, and that requires being able to\\n> > >> serialize and deserialize the aggregate state somehow. For actual data\\n> > >> types that's not overly difficult I guess (we can use in/out\\n> functions),\\n> > >> but what about aggretates using 'internal' state? I.e. aggregates\\n> > >> passing pointers that we can't serialize?\\n> > >\\n> > > This is a good question. I really don't know the answer to it as I've\\n> not\\n> > > looked at the Robert's API for passing data between backends yet.\\n> > >\\n> > > Maybe Robert or Amit can answer this?\\n> >\\n> > I think parallel aggregation will probably require both the\\n> > infrastructure discussed here, namely the ability to combine two\\n> > transition states into a single transition state, and also the ability\\n> > to serialize and de-serialize transition states, which has previously\\n> > been discussed in the context of letting hash aggregates spill to\\n> > disk. My current thinking is that the parallel plan will look\\n> > something like this:\\n> >\\n> > HashAggregateFinish\\n> > -> Funnel\\n> > -> HashAggregatePartial\\n> > -> PartialHeapScan\\n> >\\n> > So the workers will all pull from a partial heap scan and each worker\\n> > will aggregate its own portion of the data. Then it'll need to pass\\n> > the results of that step back to the master, which will aggregate the\\n> > partial results and produce the final results. I'm guessing that if\\n> > we're grouping on, say, column a, the HashAggregatePartial nodes will\\n> > kick out 2-column tuples of the form (<value for a>, <serialized\\n> > transition state for group with that value for a>).\\n> >\\n> It may not be an urgent topic to be discussed towards v9.5, however,\\n> I'd like to raise a topic about planner and aggregates.\\n>\\n> Once we could get the two phase aggregation, planner needs to pay\\n> attention possibility of partial aggregate during plan construction,\\n> even though our current implementation attach Agg node after the\\n> join/scan plan construction.\\n>\\n> Probably, a straightforward design is to add FunnelPath with some\\n> child nodes on set_rel_pathlist() or add_paths_to_joinrel().\\n> Its child node may be PartialAggregate node (or some other parallel\\n> safe node of course). In this case, it must inform the planner this\\n> node (for more correctness, targetlist of the node) returns partial\\n> aggregation, instead of the values row-by-row.\\n> Then, planner need to track and care about which type of data shall\\n> be returned to the upper node. Path node may have a flag to show it,\\n> and we also may need to inject dummy PartialAggregate if we try to\\n> join a relation that returns values row-by-row and another one that\\n> returns partial aggregate.\\n> It also leads another problem. The RelOptInfo->targetlist will\\n> depend on the Path node chosen. Even if float datum is expected as\\n> an argument of AVG(), its state to be combined is float[3].\\n> So, we will need to adjust the targetlist of RelOptInfo, once Path\\n> got chosen.\\n>\\n> Anyway, I could find out, at least, these complicated issues around\\n> two-phase aggregate integration with planner. Someone can suggest\\n> minimum invasive way for these integration?\\n>\\n>\\nPostgres-XC solved this question by creating a plan with two Agg/Group\\nnodes, one for combining transitioned result and one for creating the\\ndistributed transition results (one per distributed run per group). So,\\nAgg/Group for combining result had as many Agg/Group nodes as there are\\ndistributed/parallel runs. But XC chose this way to reduce the code\\nfootprint. In Postgres, we can have different nodes for combining and\\ntransitioning as you have specified above. Aggregation is not pathified in\\ncurrent planner, hence XC took the approach of pushing the Agg nodes down\\nthe plan tree when there was distributed/parallel execution possible. If we\\ncan get aggregation pathified, we can go by path-based approach which might\\ngive a better judgement of whether or not to distribute the aggregates\\nitself.\\n\\nLooking at Postgres-XC might be useful to get ideas. I can help you there.\\n\\n\\n> Thanks,\\n> --\\n> NEC OSS Promotion Center / PG-Strom Project\\n> KaiGai Kohei <kaigai@ak.jp.nec.com>\\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\\nOn Thu, Mar 5, 2015 at 9:30 AM, Kouhei Kaigai <kaigai@ak.jp.nec.com> wrote:> On Wed, Mar 4, 2015 at 4:41 AM, David Rowley <dgrowleyml@gmail.com> wrote:\\n> >> This thread mentions \\\"parallel queries\\\" as a use case, but that means\\n> >> passing data between processes, and that requires being able to\\n> >> serialize and deserialize the aggregate state somehow. For actual data\\n> >> types that's not overly difficult I guess (we can use in/out functions),\\n> >> but what about aggretates using 'internal' state? I.e. aggregates\\n> >> passing pointers that we can't serialize?\\n> >\\n> > This is a good question. I really don't know the answer to it as I've not\\n> > looked at the Robert's API for passing data between backends yet.\\n> >\\n> > Maybe Robert or Amit can answer this?\\n>\\n> I think parallel aggregation will probably require both the\\n> infrastructure discussed here, namely the ability to combine two\\n> transition states into a single transition state, and also the ability\\n> to serialize and de-serialize transition states, which has previously\\n> been discussed in the context of letting hash aggregates spill to\\n> disk. My current thinking is that the parallel plan will look\\n> something like this:\\n>\\n> HashAggregateFinish\\n> -> Funnel\\n> -> HashAggregatePartial\\n> -> PartialHeapScan\\n>\\n> So the workers will all pull from a partial heap scan and each worker\\n> will aggregate its own portion of the data. Then it'll need to pass\\n> the results of that step back to the master, which will aggregate the\\n> partial results and produce the final results. I'm guessing that if\\n> we're grouping on, say, column a, the HashAggregatePartial nodes will\\n> kick out 2-column tuples of the form (<value for a>, <serialized\\n> transition state for group with that value for a>).\\n>\\nIt may not be an urgent topic to be discussed towards v9.5, however,\\nI'd like to raise a topic about planner and aggregates.\\n\\nOnce we could get the two phase aggregation, planner needs to pay\\nattention possibility of partial aggregate during plan construction,\\neven though our current implementation attach Agg node after the\\njoin/scan plan construction.\\n\\nProbably, a straightforward design is to add FunnelPath with some\\nchild nodes on set_rel_pathlist() or add_paths_to_joinrel().\\nIts child node may be PartialAggregate node (or some other parallel\\nsafe node of course). In this case, it must inform the planner this\\nnode (for more correctness, targetlist of the node) returns partial\\naggregation, instead of the values row-by-row.\\nThen, planner need to track and care about which type of data shall\\nbe returned to the upper node. Path node may have a flag to show it,\\nand we also may need to inject dummy PartialAggregate if we try to\\njoin a relation that returns values row-by-row and another one that\\nreturns partial aggregate.\\nIt also leads another problem. The RelOptInfo->targetlist will\\ndepend on the Path node chosen. Even if float datum is expected as\\nan argument of AVG(), its state to be combined is float[3].\\nSo, we will need to adjust the targetlist of RelOptInfo, once Path\\ngot chosen.\\n\\nAnyway, I could find out, at least, these complicated issues around\\ntwo-phase aggregate integration with planner. Someone can suggest\\nminimum invasive way for these integration?\\nPostgres-XC solved this question by creating a plan with two Agg/Group nodes, one for combining transitioned result and one for creating the distributed transition results (one per distributed run per group). So, Agg/Group for combining result had as many Agg/Group nodes as there are distributed/parallel runs. But XC chose this way to reduce the code footprint. In Postgres, we can have different nodes for combining and transitioning as you have specified above. Aggregation is not pathified in current planner, hence XC took the approach of pushing the Agg nodes down the plan tree when there was distributed/parallel execution possible. If we can get aggregation pathified, we can go by path-based approach which might give a better judgement of whether or not to distribute the aggregates itself.Looking at Postgres-XC might be useful to get ideas. I can help you there. \\nThanks,\\n--\\nNEC OSS Promotion Center / PG-Strom Project\\nKaiGai Kohei <kaigai@ak.jp.nec.com>\\n\\n--\\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\": \"Fri, 6 Mar 2015 11:31:14 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 6 March 2015 at 19:01, Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\\nwrote:\\n\\n> Postgres-XC solved this question by creating a plan with two Agg/Group\\n> nodes, one for combining transitioned result and one for creating the\\n> distributed transition results (one per distributed run per group).\\n>\\n\\n\\n> So, Agg/Group for combining result had as many Agg/Group nodes as there\\n> are distributed/parallel runs.\\n>\\n\\nThis sounds quite like the planner must be forcing the executor to having\\nto execute the plan on a fixed number of worker processes.\\n\\nI really hoped that we could, one day, have a load monitor process that\\ndecided what might be the best number of threads to execute a parallel plan\\non. Otherwise how would we decide how many worker processes to allocate to\\na plan? Surely there must be times where only utilising half of the\\nprocessors for a query would be better than trying to use all processors\\nand having many more context switched to perform.\\n\\nProbably the harder part about dynamically deciding the number of workers\\nwould be around the costing. Where maybe the plan will execute the fastest\\nwith 32 workers, but if it was only given 2 workers then it might execute\\nbetter as a non-parallel plan.\\n\\n\\n> But XC chose this way to reduce the code footprint. In Postgres, we can\\n> have different nodes for combining and transitioning as you have specified\\n> above. Aggregation is not pathified in current planner, hence XC took the\\n> approach of pushing the Agg nodes down the plan tree when there was\\n> distributed/parallel execution possible. If we can get aggregation\\n> pathified, we can go by path-based approach which might give a better\\n> judgement of whether or not to distribute the aggregates itself.\\n>\\n> Looking at Postgres-XC might be useful to get ideas. I can help you there.\\n>\\n>\\n\\n Regards\\n\\nDavid Rowley\\n\\nOn 6 March 2015 at 19:01, Ashutosh Bapat <ashutosh.bapat@enterprisedb.com> wrote:Postgres-XC solved this question by creating a plan with two Agg/Group nodes, one for combining transitioned result and one for creating the distributed transition results (one per distributed run per group). So, Agg/Group for combining result had as many Agg/Group nodes as there are distributed/parallel runs. This sounds quite like the planner must be forcing the executor to having to execute the plan on a fixed number of worker processes.I really hoped that we could, one day, have a load monitor process that decided what might be the best number of threads to execute a parallel plan on. Otherwise how would we decide how many worker processes to allocate to a plan? Surely there must be times where only utilising half of the processors for a query would be better than trying to use all processors and having many more context switched to perform.Probably the harder part about dynamically deciding the number of workers would be around the costing. Where maybe the plan will execute the fastest with 32 workers, but if it was only given 2 workers then it might execute better as a non-parallel plan. But XC chose this way to reduce the code footprint. In Postgres, we can have different nodes for combining and transitioning as you have specified above. Aggregation is not pathified in current planner, hence XC took the approach of pushing the Agg nodes down the plan tree when there was distributed/parallel execution possible. If we can get aggregation pathified, we can go by path-based approach which might give a better judgement of whether or not to distribute the aggregates itself.Looking at Postgres-XC might be useful to get ideas. I can help you there. RegardsDavid Rowley\",\n \"msg_date\": \"Fri, 6 Mar 2015 20:11:09 +1300\",\n \"msg_from\": \"David Rowley <dgrowley@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 6, 2015 at 12:41 PM, David Rowley <dgrowley@gmail.com> wrote:\\n\\n> On 6 March 2015 at 19:01, Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\\n> wrote:\\n>\\n>> Postgres-XC solved this question by creating a plan with two Agg/Group\\n>> nodes, one for combining transitioned result and one for creating the\\n>> distributed transition results (one per distributed run per group).\\n>>\\n>\\n>\\n>> So, Agg/Group for combining result had as many Agg/Group nodes as there\\n>> are distributed/parallel runs.\\n>>\\n>\\n> This sounds quite like the planner must be forcing the executor to having\\n> to execute the plan on a fixed number of worker processes.\\n>\\n> I really hoped that we could, one day, have a load monitor process that\\n> decided what might be the best number of threads to execute a parallel plan\\n> on. Otherwise how would we decide how many worker processes to allocate to\\n> a plan? Surely there must be times where only utilising half of the\\n> processors for a query would be better than trying to use all processors\\n> and having many more context switched to perform.\\n>\\n> Probably the harder part about dynamically deciding the number of workers\\n> would be around the costing. Where maybe the plan will execute the fastest\\n> with 32 workers, but if it was only given 2 workers then it might execute\\n> better as a non-parallel plan.\\n>\\n\\nXC does that, because it knew on how many nodes it had to distribute the\\naggregation while creating the plan. To keep that dynamic, we can add a\\nplace-holder planner node for producing transitioned results for a given\\ndistributed run. At the time of execution, that node creates one executor\\nnode (corresponding to the place-holder node) per parallel run. I haven't\\nseen a precedence in PG code to create more than one executor node for a\\ngiven planner node, but is that a rule?\\n\\n\\n>\\n>\\n>> But XC chose this way to reduce the code footprint. In Postgres, we can\\n>> have different nodes for combining and transitioning as you have specified\\n>> above. Aggregation is not pathified in current planner, hence XC took the\\n>> approach of pushing the Agg nodes down the plan tree when there was\\n>> distributed/parallel execution possible. If we can get aggregation\\n>> pathified, we can go by path-based approach which might give a better\\n>> judgement of whether or not to distribute the aggregates itself.\\n>>\\n>> Looking at Postgres-XC might be useful to get ideas. I can help you there.\\n>>\\n>>\\n>\\n> Regards\\n>\\n> David Rowley\\n>\\n\\n\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database Company\\n\\nOn Fri, Mar 6, 2015 at 12:41 PM, David Rowley <dgrowley@gmail.com> wrote:On 6 March 2015 at 19:01, Ashutosh Bapat <ashutosh.bapat@enterprisedb.com> wrote:Postgres-XC solved this question by creating a plan with two Agg/Group nodes, one for combining transitioned result and one for creating the distributed transition results (one per distributed run per group). So, Agg/Group for combining result had as many Agg/Group nodes as there are distributed/parallel runs. This sounds quite like the planner must be forcing the executor to having to execute the plan on a fixed number of worker processes.I really hoped that we could, one day, have a load monitor process that decided what might be the best number of threads to execute a parallel plan on. Otherwise how would we decide how many worker processes to allocate to a plan? Surely there must be times where only utilising half of the processors for a query would be better than trying to use all processors and having many more context switched to perform.Probably the harder part about dynamically deciding the number of workers would be around the costing. Where maybe the plan will execute the fastest with 32 workers, but if it was only given 2 workers then it might execute better as a non-parallel plan.XC does that, because it knew on how many nodes it had to distribute the aggregation while creating the plan. To keep that dynamic, we can add a place-holder planner node for producing transitioned results for a given distributed run. At the time of execution, that node creates one executor node (corresponding to the place-holder node) per parallel run. I haven't seen a precedence in PG code to create more than one executor node for a given planner node, but is that a rule? But XC chose this way to reduce the code footprint. In Postgres, we can have different nodes for combining and transitioning as you have specified above. Aggregation is not pathified in current planner, hence XC took the approach of pushing the Agg nodes down the plan tree when there was distributed/parallel execution possible. If we can get aggregation pathified, we can go by path-based approach which might give a better judgement of whether or not to distribute the aggregates itself.Looking at Postgres-XC might be useful to get ideas. I can help you there. RegardsDavid Rowley\\n-- Best Wishes,Ashutosh BapatEnterpriseDB CorporationThe Postgres Database Company\",\n \"msg_date\": \"Fri, 6 Mar 2015 13:25:24 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 18 December 2014 at 02:48, Simon Riggs <simon@2ndquadrant.com> wrote:\\n\\n> David, if you can update your patch with some docs to explain the\\n> behaviour, it looks complete enough to think about committing it in\\n> early January, to allow other patches that depend upon it to stand a\\n> chance of getting into 9.5. (It is not yet ready, but I see it could\\n> be).\\n>\\n>\\nI've been thinking of bumping this patch to the June commitfest as the\\npatch only exists to provide the basic infrastructure for things like\\nparallel aggregation, aggregate before join, and perhaps auto updating\\nmaterialised views.\\n\\nIt seems unlikely that any of those things will happen for 9.5.\\nDoes anybody object to me moving this to June's commitfest?\\n\\nRegards\\n\\nDavid Rowley\\n\\nOn 18 December 2014 at 02:48, Simon Riggs <simon@2ndquadrant.com> wrote:David, if you can update your patch with some docs to explain the\\nbehaviour, it looks complete enough to think about committing it in\\nearly January, to allow other patches that depend upon it to stand a\\nchance of getting into 9.5. (It is not yet ready, but I see it could\\nbe).\\nI've been thinking of bumping this patch to the June commitfest as the patch only exists to provide the basic infrastructure for things like parallel aggregation, aggregate before join, and perhaps auto updating materialised views.It seems unlikely that any of those things will happen for 9.5.Does anybody object to me moving this to June's commitfest?RegardsDavid Rowley\",\n \"msg_date\": \"Mon, 30 Mar 2015 18:08:19 +1300\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 30, 2015 at 2:08 PM, David Rowley <dgrowleyml@gmail.com> wrote:\\n\\n>\\n> On 18 December 2014 at 02:48, Simon Riggs <simon@2ndquadrant.com> wrote:\\n>\\n>> David, if you can update your patch with some docs to explain the\\n>> behaviour, it looks complete enough to think about committing it in\\n>> early January, to allow other patches that depend upon it to stand a\\n>> chance of getting into 9.5. (It is not yet ready, but I see it could\\n>> be).\\n>>\\n>>\\n> I've been thinking of bumping this patch to the June commitfest as the\\n> patch only exists to provide the basic infrastructure for things like\\n> parallel aggregation, aggregate before join, and perhaps auto updating\\n> materialised views.\\n>\\n> It seems unlikely that any of those things will happen for 9.5.\\n>\\n\\nYeah, I guess so...\\n\\n\\n> Does anybody object to me moving this to June's commitfest?\\n>\\n\\nNot from my side FWIW. I think it actually makes sense.\\n-- \\nMichael\\n\\nOn Mon, Mar 30, 2015 at 2:08 PM, David Rowley <dgrowleyml@gmail.com> wrote:On 18 December 2014 at 02:48, Simon Riggs <simon@2ndquadrant.com> wrote:David, if you can update your patch with some docs to explain the\\nbehaviour, it looks complete enough to think about committing it in\\nearly January, to allow other patches that depend upon it to stand a\\nchance of getting into 9.5. (It is not yet ready, but I see it could\\nbe).\\nI've been thinking of bumping this patch to the June commitfest as the patch only exists to provide the basic infrastructure for things like parallel aggregation, aggregate before join, and perhaps auto updating materialised views.It seems unlikely that any of those things will happen for 9.5.Yeah, I guess so... Does anybody object to me moving this to June's commitfest?\\nNot from my side FWIW. I think it actually makes sense.-- Michael\",\n \"msg_date\": \"Mon, 30 Mar 2015 14:28:37 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 30 March 2015 at 01:08, David Rowley <dgrowleyml@gmail.com> wrote:\\n> On 18 December 2014 at 02:48, Simon Riggs <simon@2ndquadrant.com> wrote:\\n>>\\n>> David, if you can update your patch with some docs to explain the\\n>> behaviour, it looks complete enough to think about committing it in\\n>> early January, to allow other patches that depend upon it to stand a\\n>> chance of getting into 9.5. (It is not yet ready, but I see it could\\n>> be).\\n>>\\n>\\n> I've been thinking of bumping this patch to the June commitfest as the patch\\n> only exists to provide the basic infrastructure for things like parallel\\n> aggregation, aggregate before join, and perhaps auto updating materialised\\n> views.\\n>\\n> It seems unlikely that any of those things will happen for 9.5.\\n> Does anybody object to me moving this to June's commitfest?\\n\\nIf the patch is ready, it should stay in the queue.\\n\\nA global decision to move all uncommitted patches to June might occur\\nlater, but that's a different decision. I don't think we should be\\nprejudging that situation, all it would do is hide the extent of the\\nreal problem with reviews.\\n\\n-- \\n Simon Riggs http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, RemoteDBA, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 30 Mar 2015 08:25:14 -0400\",\n \"msg_from\": \"Simon Riggs <simon@2ndQuadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 30, 2015 at 1:28 AM, Michael Paquier\\n<michael.paquier@gmail.com> wrote:\\n>> I've been thinking of bumping this patch to the June commitfest as the\\n>> patch only exists to provide the basic infrastructure for things like\\n>> parallel aggregation, aggregate before join, and perhaps auto updating\\n>> materialised views.\\n>>\\n>> It seems unlikely that any of those things will happen for 9.5.\\n>\\n> Yeah, I guess so...\\n>\\n>> Does anybody object to me moving this to June's commitfest?\\n>\\n> Not from my side FWIW. I think it actually makes sense.\\n\\n+1. I'd like to devote some time to looking at this, but I don't have\\nthe time right now. The chances that we can do something useful with\\nit in 9.6 seem good.\\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, 1 Apr 2015 11:28:21 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 04/01/2015 06:28 PM, Robert Haas wrote:\\n> On Mon, Mar 30, 2015 at 1:28 AM, Michael Paquier\\n> <michael.paquier@gmail.com> wrote:\\n>>> I've been thinking of bumping this patch to the June commitfest as the\\n>>> patch only exists to provide the basic infrastructure for things like\\n>>> parallel aggregation, aggregate before join, and perhaps auto updating\\n>>> materialised views.\\n>>>\\n>>> It seems unlikely that any of those things will happen for 9.5.\\n>>\\n>> Yeah, I guess so...\\n>>\\n>>> Does anybody object to me moving this to June's commitfest?\\n>>\\n>> Not from my side FWIW. I think it actually makes sense.\\n>\\n> +1. I'd like to devote some time to looking at this, but I don't have\\n> the time right now. The chances that we can do something useful with\\n> it in 9.6 seem good.\\n\\nAnd the June commitfest is now in progress.\\n\\nThis patch seems sane to me, as far as it goes. However, there's no \\nplanner or executor code to use the aggregate combining for anything. \\nI'm not a big fan of dead code, I'd really like to see something to use \\nthis.\\n\\nThe main use case people have been talking about is parallel query, but \\nis there some other case this would be useful right now, without the \\nparallel query feature? You and Simon talked about this case:\\n\\n> 2. Queries such as:\\n>\\n> SELECT p.name, SUM(s.qty) FROM sales s INNER JOIN product p ON s.product_id\\n> = p.product_id GROUP BY p.name;\\n>\\n> Such a query could be transformed into:\\n>\\n> SELECT p.name,SUM(qty) FROM (SELECT product_id,SUM(qty) AS qty FROM sales\\n> GROUP BY product_id) s\\n> INNER JOIN product p ON p.product_id = s.product_id GROUP BY p_name;\\n>\\n> Of course the outer query's SUM and GROUP BY would not be required if there\\n> happened to be a UNIQUE index on product(name), but assuming there's not\\n> then the above should produce the results faster. This of course works ok\\n> for SUM(), but for something like AVG() or STDDEV() the combine/merge\\n> aggregate functions would be required to process those intermediate\\n> aggregate results that were produced by the sub-query.\\n\\nAny chance you could implement that in the planner?\\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\": \"Sun, 26 Jul 2015 19:58:41 +0300\",\n \"msg_from\": \"Heikki Linnakangas <hlinnaka@iki.fi>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"> On 04/01/2015 06:28 PM, Robert Haas wrote:\\r\\n> > On Mon, Mar 30, 2015 at 1:28 AM, Michael Paquier\\r\\n> > <michael.paquier@gmail.com> wrote:\\r\\n> >>> I've been thinking of bumping this patch to the June commitfest as the\\r\\n> >>> patch only exists to provide the basic infrastructure for things like\\r\\n> >>> parallel aggregation, aggregate before join, and perhaps auto updating\\r\\n> >>> materialised views.\\r\\n> >>>\\r\\n> >>> It seems unlikely that any of those things will happen for 9.5.\\r\\n> >>\\r\\n> >> Yeah, I guess so...\\r\\n> >>\\r\\n> >>> Does anybody object to me moving this to June's commitfest?\\r\\n> >>\\r\\n> >> Not from my side FWIW. I think it actually makes sense.\\r\\n> >\\r\\n> > +1. I'd like to devote some time to looking at this, but I don't have\\r\\n> > the time right now. The chances that we can do something useful with\\r\\n> > it in 9.6 seem good.\\r\\n> \\r\\n> And the June commitfest is now in progress.\\r\\n> \\r\\n> This patch seems sane to me, as far as it goes. However, there's no\\r\\n> planner or executor code to use the aggregate combining for anything.\\r\\n> I'm not a big fan of dead code, I'd really like to see something to use\\r\\n> this.\\r\\n>\\r\\n+1, this patch itself looks good for me, but...\\r\\n\\r\\n> The main use case people have been talking about is parallel query, but\\r\\n> is there some other case this would be useful right now, without the\\r\\n> parallel query feature? You and Simon talked about this case:\\r\\n>\\r\\n> > 2. Queries such as:\\r\\n> >\\r\\n> > SELECT p.name, SUM(s.qty) FROM sales s INNER JOIN product p ON s.product_id\\r\\n> > = p.product_id GROUP BY p.name;\\r\\n> >\\r\\n> > Such a query could be transformed into:\\r\\n> >\\r\\n> > SELECT p.name,SUM(qty) FROM (SELECT product_id,SUM(qty) AS qty FROM sales\\r\\n> > GROUP BY product_id) s\\r\\n> > INNER JOIN product p ON p.product_id = s.product_id GROUP BY p_name;\\r\\n> >\\r\\n> > Of course the outer query's SUM and GROUP BY would not be required if there\\r\\n> > happened to be a UNIQUE index on product(name), but assuming there's not\\r\\n> > then the above should produce the results faster. This of course works ok\\r\\n> > for SUM(), but for something like AVG() or STDDEV() the combine/merge\\r\\n> > aggregate functions would be required to process those intermediate\\r\\n> > aggregate results that were produced by the sub-query.\\r\\n> \\r\\n> Any chance you could implement that in the planner?\\r\\n>\\r\\nIt likely needs planner enhancement prior to other applications...\\r\\nhttp://www.postgresql.org/message-id/CA+TgmobgWKHfZc09B+s2LxJTwoRD5Ht-avoVDvaQ4+RpwrO4bg@mail.gmail.com\\r\\n\\r\\nOnce planner allowed to have both of normal path and partial aggregation\\r\\npaths to compare according to the cost, it is the straightforward way to\\r\\ndo.\\r\\n\\r\\nHere are various academic research, for example, below is the good starting\\r\\npoint to clarify aggregate queries that we can run with 2-phase approach.\\r\\nhttp://www.researchgate.net/publication/2715288_Performing_Group-By_before_Join\\r\\n\\r\\nThanks,\\r\\n--\\r\\nNEC Business Creation Division / PG-Strom Project\\r\\nKaiGai Kohei <kaigai@ak.jp.nec.com>\\r\\n\\r\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 27 Jul 2015 00:14:19 +0000\",\n \"msg_from\": \"Kouhei Kaigai <kaigai@ak.jp.nec.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 27 July 2015 at 04:58, Heikki Linnakangas <hlinnaka@iki.fi> wrote:\\n\\n> On 04/01/2015 06:28 PM, Robert Haas wrote:\\n>\\n>> On Mon, Mar 30, 2015 at 1:28 AM, Michael Paquier\\n>> <michael.paquier@gmail.com> wrote:\\n>>\\n>>> I've been thinking of bumping this patch to the June commitfest as the\\n>>>> patch only exists to provide the basic infrastructure for things like\\n>>>> parallel aggregation, aggregate before join, and perhaps auto updating\\n>>>> materialised views.\\n>>>>\\n>>>> It seems unlikely that any of those things will happen for 9.5.\\n>>>>\\n>>>\\n>>> Yeah, I guess so...\\n>>>\\n>>> Does anybody object to me moving this to June's commitfest?\\n>>>>\\n>>>\\n>>> Not from my side FWIW. I think it actually makes sense.\\n>>>\\n>>\\n>> +1. I'd like to devote some time to looking at this, but I don't have\\n>> the time right now. The chances that we can do something useful with\\n>> it in 9.6 seem good.\\n>>\\n>\\n> And the June commitfest is now in progress.\\n>\\n> This patch seems sane to me, as far as it goes. However, there's no\\n> planner or executor code to use the aggregate combining for anything. I'm\\n> not a big fan of dead code, I'd really like to see something to use this.\\n>\\n\\nThanks for looking. I partially agree with that, it is a little weird to\\nput in code that's yet to be used. I'd certainly agree 95% if this was the\\nfinal commitfest of 9.5, but we're in the first commitfest of 9.6 and I\\nthink there's a very high probability of this getting used in 9.6, and\\nlikely that probability would be even higher if the code is already in.\\nPerhaps it's a little bit in the same situation as to Robert's parallel\\nworker stuff?\\n\\n\\n\\n>\\n> The main use case people have been talking about is parallel query, but is\\n> there some other case this would be useful right now, without the parallel\\n> query feature? You and Simon talked about this case:\\n>\\n> 2. Queries such as:\\n>>\\n>> SELECT p.name, SUM(s.qty) FROM sales s INNER JOIN product p ON\\n>> s.product_id\\n>> = p.product_id GROUP BY p.name;\\n>>\\n>> Such a query could be transformed into:\\n>>\\n>> SELECT p.name,SUM(qty) FROM (SELECT product_id,SUM(qty) AS qty FROM sales\\n>> GROUP BY product_id) s\\n>> INNER JOIN product p ON p.product_id = s.product_id GROUP BY p_name;\\n>>\\n>> Of course the outer query's SUM and GROUP BY would not be required if\\n>> there\\n>> happened to be a UNIQUE index on product(name), but assuming there's not\\n>> then the above should produce the results faster. This of course works ok\\n>> for SUM(), but for something like AVG() or STDDEV() the combine/merge\\n>> aggregate functions would be required to process those intermediate\\n>> aggregate results that were produced by the sub-query.\\n>>\\n>\\n> Any chance you could implement that in the planner?\\n>\\n>\\nYes! I'm actually working on it now and so far have it partially working.\\nQuite likely I'll be able to submit for CF2. There's still some costing\\ntweaks to do. So far it just works for GROUP BY with no aggs. I plan to fix\\nthat later using this patch.\\n\\nI don't want to talk too much about it on this thread, but in a test query\\nwhich is the one in my example, minus the SUM(qty), with 1 million sales\\nrecords, and 100 products, performance goes from 350ms to 200ms on my\\nmachine, so looking good so far.\\n\\nRegards\\n\\nDavid Rowley\\n--\\n David Rowley http://www.2ndQuadrant.com/\\n<http://www.2ndquadrant.com/>\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 27 July 2015 at 04:58, Heikki Linnakangas <hlinnaka@iki.fi> wrote:On 04/01/2015 06:28 PM, Robert Haas wrote:\\n\\nOn Mon, Mar 30, 2015 at 1:28 AM, Michael Paquier\\n<michael.paquier@gmail.com> wrote:\\n\\nI've been thinking of bumping this patch to the June commitfest as the\\npatch only exists to provide the basic infrastructure for things like\\nparallel aggregation, aggregate before join, and perhaps auto updating\\nmaterialised views.\\n\\nIt seems unlikely that any of those things will happen for 9.5.\\n\\n\\nYeah, I guess so...\\n\\n\\nDoes anybody object to me moving this to June's commitfest?\\n\\n\\nNot from my side FWIW. I think it actually makes sense.\\n\\n\\n+1. I'd like to devote some time to looking at this, but I don't have\\nthe time right now. The chances that we can do something useful with\\nit in 9.6 seem good.\\n\\n\\nAnd the June commitfest is now in progress.\\n\\nThis patch seems sane to me, as far as it goes. However, there's no planner or executor code to use the aggregate combining for anything. I'm not a big fan of dead code, I'd really like to see something to use this.Thanks for looking. I partially agree with that, it is a little weird to put in code that's yet to be used. I'd certainly agree 95% if this was the final commitfest of 9.5, but we're in the first commitfest of 9.6 and I think there's a very high probability of this getting used in 9.6, and likely that probability would be even higher if the code is already in. Perhaps it's a little bit in the same situation as to Robert's parallel worker stuff? \\n\\nThe main use case people have been talking about is parallel query, but is there some other case this would be useful right now, without the parallel query feature? You and Simon talked about this case:\\n\\n\\n2. Queries such as:\\n\\nSELECT p.name, SUM(s.qty) FROM sales s INNER JOIN product p ON s.product_id\\n= p.product_id GROUP BY p.name;\\n\\nSuch a query could be transformed into:\\n\\nSELECT p.name,SUM(qty) FROM (SELECT product_id,SUM(qty) AS qty FROM sales\\nGROUP BY product_id) s\\nINNER JOIN product p ON p.product_id = s.product_id GROUP BY p_name;\\n\\nOf course the outer query's SUM and GROUP BY would not be required if there\\nhappened to be a UNIQUE index on product(name), but assuming there's not\\nthen the above should produce the results faster. This of course works ok\\nfor SUM(), but for something like AVG() or STDDEV() the combine/merge\\naggregate functions would be required to process those intermediate\\naggregate results that were produced by the sub-query.\\n\\n\\nAny chance you could implement that in the planner?Yes! I'm actually working on it now and so far have it partially working. Quite likely I'll be able to submit for CF2. There's still some costing tweaks to do. So far it just works for GROUP BY with no aggs. I plan to fix that later using this patch.I don't want to talk too much about it on this thread, but in a test query which is the one in my example, minus the SUM(qty), with 1 million sales records, and 100 products, performance goes from 350ms to 200ms on my machine, so looking good so far.RegardsDavid Rowley-- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Mon, 27 Jul 2015 14:59:32 +1200\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 27 July 2015 at 12:14, Kouhei Kaigai <kaigai@ak.jp.nec.com> wrote:\\n\\n> > The main use case people have been talking about is parallel query, but\\n> > is there some other case this would be useful right now, without the\\n> > parallel query feature? You and Simon talked about this case:\\n> >\\n> > > 2. Queries such as:\\n> > >\\n> > > SELECT p.name, SUM(s.qty) FROM sales s INNER JOIN product p ON\\n> s.product_id\\n> > > = p.product_id GROUP BY p.name;\\n> > >\\n> > > Such a query could be transformed into:\\n> > >\\n> > > SELECT p.name,SUM(qty) FROM (SELECT product_id,SUM(qty) AS qty FROM\\n> sales\\n> > > GROUP BY product_id) s\\n> > > INNER JOIN product p ON p.product_id = s.product_id GROUP BY p_name;\\n> > >\\n> > > Of course the outer query's SUM and GROUP BY would not be required if\\n> there\\n> > > happened to be a UNIQUE index on product(name), but assuming there's\\n> not\\n> > > then the above should produce the results faster. This of course works\\n> ok\\n> > > for SUM(), but for something like AVG() or STDDEV() the combine/merge\\n> > > aggregate functions would be required to process those intermediate\\n> > > aggregate results that were produced by the sub-query.\\n> >\\n> > Any chance you could implement that in the planner?\\n> >\\n>\\nIt likely needs planner enhancement prior to other applications...\\n>\\n> http://www.postgresql.org/message-id/CA+TgmobgWKHfZc09B+s2LxJTwoRD5Ht-avoVDvaQ4+RpwrO4bg@mail.gmail.com\\n>\\n>\\nI had thought this too, but I'm not sure that's 100% correct. As I just\\nsaid in the my previous email on this thread, I am now working on \\\"group by\\nbefore join\\\". I had started it with the intentions of path-ifying the\\ngrouping planner, but I soon realised that the grouping_planner() does\\nquite a bit more at that final stage that I propose to do to allow \\\"group\\nby before join\\\". This is mostly around handling of DISTINCT, HAVING and\\nLIMIT. I don't think those need to be handled in my patch, perhaps with the\\nexception of DISTINCT without GROUP BY, but not when both are present.\\n\\nInstead I've started by inventing GroupingPath and I'm now building these\\nnew path types once there's enough tables joined for all Vars of the GROUP\\nBY and agg parameters.\\n\\nI believe this optimisation needs to be costed as pushing the GROUP BY down\\nto happen as early as possible is *not* always a win. Perhaps the JOIN is\\nvery selective and eliminates many groups. Hence I've added costing and\\nallowed the planner to choose what it thinks is cheapest.\\n\\n\\n> Once planner allowed to have both of normal path and partial aggregation\\n> paths to compare according to the cost, it is the straightforward way to\\n> do.\\n>\\n\\nI've ended up with 2 boolean members to struct GroupingPath, combine_states\\nand finalize_aggs. I plan to modify nodeAgg.c to use these, and EXPLAIN to\\ngive a better description of what its doing.\\n\\n\\n>\\n> Here are various academic research, for example, below is the good starting\\n> point to clarify aggregate queries that we can run with 2-phase approach.\\n>\\n> http://www.researchgate.net/publication/2715288_Performing_Group-By_before_Join\\n>\\n>\\nThanks, I've been using that very paper.\\n\\nRegards\\n\\nDavid Rowley\\n\\n--\\n David Rowley http://www.2ndQuadrant.com/\\n<http://www.2ndquadrant.com/>\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 27 July 2015 at 12:14, Kouhei Kaigai <kaigai@ak.jp.nec.com> wrote:> The main use case people have been talking about is parallel query, but\\n> is there some other case this would be useful right now, without the\\n> parallel query feature? You and Simon talked about this case:\\n>\\n> > 2. Queries such as:\\n> >\\n> > SELECT p.name, SUM(s.qty) FROM sales s INNER JOIN product p ON s.product_id\\n> > = p.product_id GROUP BY p.name;\\n> >\\n> > Such a query could be transformed into:\\n> >\\n> > SELECT p.name,SUM(qty) FROM (SELECT product_id,SUM(qty) AS qty FROM sales\\n> > GROUP BY product_id) s\\n> > INNER JOIN product p ON p.product_id = s.product_id GROUP BY p_name;\\n> >\\n> > Of course the outer query's SUM and GROUP BY would not be required if there\\n> > happened to be a UNIQUE index on product(name), but assuming there's not\\n> > then the above should produce the results faster. This of course works ok\\n> > for SUM(), but for something like AVG() or STDDEV() the combine/merge\\n> > aggregate functions would be required to process those intermediate\\n> > aggregate results that were produced by the sub-query.\\n>\\n> Any chance you could implement that in the planner?\\n>\\nIt likely needs planner enhancement prior to other applications...\\nhttp://www.postgresql.org/message-id/CA+TgmobgWKHfZc09B+s2LxJTwoRD5Ht-avoVDvaQ4+RpwrO4bg@mail.gmail.com\\nI had thought this too, but I'm not sure that's 100% correct. As I just said in the my previous email on this thread, I am now working on \\\"group by before join\\\". I had started it with the intentions of path-ifying the grouping planner, but I soon realised that the grouping_planner() does quite a bit more at that final stage that I propose to do to allow \\\"group by before join\\\". This is mostly around handling of DISTINCT, HAVING and LIMIT. I don't think those need to be handled in my patch, perhaps with the exception of DISTINCT without GROUP BY, but not when both are present.Instead I've started by inventing GroupingPath and I'm now building these new path types once there's enough tables joined for all Vars of the GROUP BY and agg parameters.I believe this optimisation needs to be costed as pushing the GROUP BY down to happen as early as possible is *not* always a win. Perhaps the JOIN is very selective and eliminates many groups. Hence I've added costing and allowed the planner to choose what it thinks is cheapest. \\nOnce planner allowed to have both of normal path and partial aggregation\\npaths to compare according to the cost, it is the straightforward way to\\ndo.I've ended up with 2 boolean members to struct GroupingPath, combine_states and finalize_aggs. I plan to modify nodeAgg.c to use these, and EXPLAIN to give a better description of what its doing. \\n\\nHere are various academic research, for example, below is the good starting\\npoint to clarify aggregate queries that we can run with 2-phase approach.\\nhttp://www.researchgate.net/publication/2715288_Performing_Group-By_before_Join\\nThanks, I've been using that very paper.RegardsDavid Rowley-- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Mon, 27 Jul 2015 15:16:18 +1200\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 27 July 2015 at 04:58, Heikki Linnakangas <hlinnaka@iki.fi> wrote:\\n\\n>\\n> This patch seems sane to me, as far as it goes. However, there's no\\n> planner or executor code to use the aggregate combining for anything. I'm\\n> not a big fan of dead code, I'd really like to see something to use this.\\n>\\n\\nI've attached an updated version of the patch. The main change from last\\ntime is that I've added executor support and exposed this to the planner\\nvia two new parameters in make_agg().\\n\\nI've also added EXPLAIN support, this will display \\\"Partial\\n[Hash|Group]Aggregate\\\" for cases where the final function won't be called\\nand displays \\\"Finalize [Hash|Group]Aggregate\\\" when combining states and\\nfinalizing aggregates.\\n\\nThis patch is currently intended for foundation work for parallel\\naggregation.\\n\\n--\\n David Rowley http://www.2ndQuadrant.com/\\n<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, 3 Dec 2015 18:01:53 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 3, 2015 at 12:01 AM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> On 27 July 2015 at 04:58, Heikki Linnakangas <hlinnaka@iki.fi> wrote:\\n>> This patch seems sane to me, as far as it goes. However, there's no\\n>> planner or executor code to use the aggregate combining for anything. I'm\\n>> not a big fan of dead code, I'd really like to see something to use this.\\n>\\n> I've attached an updated version of the patch. The main change from last\\n> time is that I've added executor support and exposed this to the planner via\\n> two new parameters in make_agg().\\n>\\n> I've also added EXPLAIN support, this will display \\\"Partial\\n> [Hash|Group]Aggregate\\\" for cases where the final function won't be called\\n> and displays \\\"Finalize [Hash|Group]Aggregate\\\" when combining states and\\n> finalizing aggregates.\\n>\\n> This patch is currently intended for foundation work for parallel\\n> aggregation.\\n\\nGiven Tom's dislike of executor nodes that do more than one thing, I\\nfear he's not going to be very happy about combining Aggregate,\\nPartialAggregate, FinalizeAggregate, GroupAggregate,\\nPartialGroupAggregate, FinalizeGroupAggregate, HashAggregate,\\nPartialHashAggregate, and FinalizeHashAggregate under one roof.\\nHowever, it's not at all obvious to me what would be any better.\\nnodeAgg.c is already a very big file, and this patch adds a\\nsurprisingly small amount of code to it.\\n\\nI think the parameter in CREATE AGGREGATE ought to be called\\nCOMBINEFUNC rather than CFUNC. There are a lot of English words that\\nbegin with C, and it's not self-evident which one is meant.\\n\\n\\\"iif performing the final aggregate stage we'll check the qual\\\" should\\nprobably say \\\"If\\\" with a capital letter rather than \\\"iif\\\" without one.\\n\\nI think it would be useful to have a test patch associated with this\\npatch that generates a FinalizeAggregate + PartialAggregate combo\\ninstead of an aggregate, and run that through the regression tests.\\nThere will obviously be EXPLAIN differences, but in theory nothing\\nelse should blow up. Of course, such tests will become more\\nmeaningful as we add more combine-functions, but this would at least\\nbe a start.\\n\\nGenerally, I think that this patch will be excellent infrastructure\\nfor parallel aggregate and I think we should try to get it committed\\nsoon.\\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, 8 Dec 2015 12:18:05 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 9 December 2015 at 06:18, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On Thu, Dec 3, 2015 at 12:01 AM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n> > On 27 July 2015 at 04:58, Heikki Linnakangas <hlinnaka@iki.fi> wrote:\\n> >> This patch seems sane to me, as far as it goes. However, there's no\\n> >> planner or executor code to use the aggregate combining for anything.\\n> I'm\\n> >> not a big fan of dead code, I'd really like to see something to use\\n> this.\\n> >\\n> > I've attached an updated version of the patch. The main change from last\\n> > time is that I've added executor support and exposed this to the planner\\n> via\\n> > two new parameters in make_agg().\\n> >\\n> > I've also added EXPLAIN support, this will display \\\"Partial\\n> > [Hash|Group]Aggregate\\\" for cases where the final function won't be called\\n> > and displays \\\"Finalize [Hash|Group]Aggregate\\\" when combining states and\\n> > finalizing aggregates.\\n> >\\n> > This patch is currently intended for foundation work for parallel\\n> > aggregation.\\n>\\n> Given Tom's dislike of executor nodes that do more than one thing, I\\n> fear he's not going to be very happy about combining Aggregate,\\n> PartialAggregate, FinalizeAggregate, GroupAggregate,\\n> PartialGroupAggregate, FinalizeGroupAggregate, HashAggregate,\\n> PartialHashAggregate, and FinalizeHashAggregate under one roof.\\n> However, it's not at all obvious to me what would be any better.\\n> nodeAgg.c is already a very big file, and this patch adds a\\n> surprisingly small amount of code to it.\\n>\\n>\\nHmm yes. I read that too. It's a tricky one. I'm not sure where the line is\\nand when we go over it. At least nodeAgg.c does not need any additional\\nwork again for parallel enabling... This should be all that's required for\\nthat.\\n\\n\\n> I think the parameter in CREATE AGGREGATE ought to be called\\n> COMBINEFUNC rather than CFUNC. There are a lot of English words that\\n> begin with C, and it's not self-evident which one is meant.\\n>\\n>\\nGood idea, I've changed this in the attached.\\n\\n\\n> \\\"iif performing the final aggregate stage we'll check the qual\\\" should\\n> probably say \\\"If\\\" with a capital letter rather than \\\"iif\\\" without one.\\n>\\n>\\nWhile building the test code I ended up deciding it's best to not change\\nthis part at all and just always check the qual. In the case of partial\\naggregation I think it should just be up to the planner not to pass the\\nHAVING clause as the node's qual, and only pass this when Finalizing the\\naggregates. Seem fair?\\n\\n\\n> I think it would be useful to have a test patch associated with this\\n> patch that generates a FinalizeAggregate + PartialAggregate combo\\n> instead of an aggregate, and run that through the regression tests.\\n> There will obviously be EXPLAIN differences, but in theory nothing\\n> else should blow up. Of course, such tests will become more\\n> meaningful as we add more combine-functions, but this would at least\\n> be a start.\\n>\\n>\\nI've been working on this, and I've attached what I've got so far. The\\nplans will look something like:\\n\\n# explain select sum(a) from test;\\n QUERY PLAN\\n--------------------------------------------------------------------\\n Finalize Aggregate (cost=18.53..18.54 rows=1 width=4)\\n -> Partial Aggregate (cost=18.51..18.52 rows=1 width=4)\\n -> Seq Scan on test (cost=0.00..16.01 rows=1001 width=4)\\n\\nI seem to have got all of this working with the test patch, and the only\\nregression tests which failed were due to EXPLAIN output changing, rather\\nthan results changing. I *was* all quite happy with it until I thought that\\nI'd better write a aggregate combine function which has an INTERNAL state.\\nI had thought that I could get this working by insisting that the combine\\nfunction either update the existing state, or in the case there's no\\nexisting state, just write all the values from the combining state into a\\nnewly created state which is allocated in the correct memory context. I\\ntried this by implementing a combinefn for SUM(NUMERIC) and all seems to\\nwork fine for hash aggregates, but falls flat for sorted or plain\\naggregates.\\n\\n# select x,sum(x::numeric) from generate_series(1,3) x(x) group by x;\\n x | sum\\n---+-----\\n 1 | 1\\n 3 | 3\\n 2 | 2\\n(3 rows)\\n\\nThis one works ok using hash aggregate.\\n\\nBut sorted and plain aggregates fail:\\n\\n# select sum(x::numeric) from generate_series(1,3) x(x);\\nERROR: invalid memory alloc request size 18446744072025250716\\n\\nThe reason that this happens is down to the executor always thinking that\\nAggref returns the aggtype, but in cases where we're not finalizing the\\naggregate the executor needs to know that we're actually returning\\naggtranstype instead. Hash aggregates appear to work as the trans value is\\njust stuffed into a hash table, but with plain and sorted aggregates this\\ngets formed into a Tuple again, and forming a tuple naturally requires the\\ntypes to be set correctly! I'm not sure exactly what the best way to fix\\nthis will be. I've hacked something up in the attached test patch which\\ngets around the problem by adding a new aggtranstype to Aggref and also an\\n'aggskipfinalize' field which I manually set to true in a bit of a hacky\\nway inside the grouping planner. Then in exprType() for Aggref I\\nconditionally return the aggtype or aggtranstype based on the\\naggskipfinalize setting. This is most likely not the way to properly fix\\nthis, but I'm running out of steam today to think of how it should be done,\\nso I'm currently very open to ideas on this.\\n\\nI should also mention that the setrefs stuff in the test patch is borrowed\\nfrom Haribabu's Parallel Aggregate patch. I'm not quite clear on which\\npatch that part should go into at the moment.\\n\\nGenerally, I think that this patch will be excellent infrastructure\\n> for parallel aggregate and I think we should try to get it committed\\n> soon.\\n>\\n\\nThanks.\\n\\n-- \\n David Rowley 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, 18 Dec 2015 01:28:56 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 18 December 2015 at 01:28, David Rowley <david.rowley@2ndquadrant.com>\\nwrote:\\n\\n> # select sum(x::numeric) from generate_series(1,3) x(x);\\n> ERROR: invalid memory alloc request size 18446744072025250716\\n>\\n> The reason that this happens is down to the executor always thinking that\\n> Aggref returns the aggtype, but in cases where we're not finalizing the\\n> aggregate the executor needs to know that we're actually returning\\n> aggtranstype instead. Hash aggregates appear to work as the trans value is\\n> just stuffed into a hash table, but with plain and sorted aggregates this\\n> gets formed into a Tuple again, and forming a tuple naturally requires the\\n> types to be set correctly! I'm not sure exactly what the best way to fix\\n> this will be. I've hacked something up in the attached test patch which\\n> gets around the problem by adding a new aggtranstype to Aggref and also an\\n> 'aggskipfinalize' field which I manually set to true in a bit of a hacky\\n> way inside the grouping planner. Then in exprType() for Aggref I\\n> conditionally return the aggtype or aggtranstype based on the\\n> aggskipfinalize setting. This is most likely not the way to properly fix\\n> this, but I'm running out of steam today to think of how it should be done,\\n> so I'm currently very open to ideas on this.\\n>\\n\\n Ok, so it seems that my mindset was not in parallel process space when I\\nwas thinking about this. I think having the pointer in the Tuple is\\nprobably going to be fine for this multiple stage aggregation when that's\\noccurring in a single backend process, but obviously if the memory that the\\npointer points to belongs to a worker process in a parallel aggregate\\nsituation, then bad things will happen.\\n\\nNow, there has been talk of this previously, on various threads, but I\\ndon't believe any final decisions were made on how exactly it should be\\ndone. At the moment I plan to make changes as follows:\\n\\n\\n 1. Add 3 new columns to pg_aggregate, aggserialfn, aggdeserialfn and\\n aggserialtype These will only be required when aggtranstype is INTERNAL.\\n Perhaps we should disallow CREATE AGGREAGET from accepting them for any\\n other type... The return type of aggserialfn should be aggserialtype, and\\n it should take a single parameter of aggtranstype. aggdeserialfn will be\\n the reverse of that.\\n 2. Add a new bool field to nodeAgg's state named serialStates. If this\\n is field is set to true then when we're in finalizeAgg = false mode, we'll\\n call the serialfn on the agg state instead of the finalfn. This will allow\\n the serialized state to be stored in the tuple and sent off to the main\\n backend. The combine agg node should also be set to serialStates = true,\\n so that it knows to deserialize instead of just assuming that the agg state\\n is of type aggtranstype.\\n\\nI believe this should allow us to not cause any performance regressions by\\nmoving away from INTERNAL agg states. It should also be very efficient for\\ninternal states such as Int8TransTypeData, as this struct merely has 2\\nint64 fields which should be very simple to stuff into a bytea varlena\\ntype. We don't need to mess around converting the ->count and ->sum into a\\nstrings or anything.\\n\\nThen in order for the planner to allow parallel aggregation all aggregates\\nmust:\\n\\n 1. Not have a DISTINCT or ORDER BY clause\\n 2. Have a combinefn\\n 3. If aggtranstype = INTERNAL, must have a aggserialfn and aggdeserialfn.\\n\\nWe can relax the requirement on 3 if we're using 2-stage aggregation, but\\nnot parallel aggregation.\\n\\nAny objections, or better ideas?\\n\\nThat just leaves me to figure out how to set the correct Aggref->aggtype\\nduring planning, as now there's 3 possible types:\\n\\nif (finalizeAggs == false)\\n{\\n if (serialStates == true)\\n type = aggserialtype;\\n else\\n type = aggtranstype;\\n}\\nelse\\n type = prorettype; /* normal case */\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 18 December 2015 at 01:28, David Rowley <david.rowley@2ndquadrant.com> wrote:# select sum(x::numeric) from generate_series(1,3) x(x);ERROR: invalid memory alloc request size 18446744072025250716The reason that this happens is down to the executor always thinking that Aggref returns the aggtype, but in cases where we're not finalizing the aggregate the executor needs to know that we're actually returning aggtranstype instead. Hash aggregates appear to work as the trans value is just stuffed into a hash table, but with plain and sorted aggregates this gets formed into a Tuple again, and forming a tuple naturally requires the types to be set correctly! I'm not sure exactly what the best way to fix this will be. I've hacked something up in the attached test patch which gets around the problem by adding a new aggtranstype to Aggref and also an 'aggskipfinalize' field which I manually set to true in a bit of a hacky way inside the grouping planner. Then in exprType() for Aggref I conditionally return the aggtype or aggtranstype based on the aggskipfinalize setting. This is most likely not the way to properly fix this, but I'm running out of steam today to think of how it should be done, so I'm currently very open to ideas on this. Ok, so it seems that my mindset was not in parallel process space when I was thinking about this. I think having the pointer in the Tuple is probably going to be fine for this multiple stage aggregation when that's occurring in a single backend process, but obviously if the memory that the pointer points to belongs to a worker process in a parallel aggregate situation, then bad things will happen.Now, there has been talk of this previously, on various threads, but I don't believe any final decisions were made on how exactly it should be done. At the moment I plan to make changes as follows: Add 3 new columns to pg_aggregate, aggserialfn, aggdeserialfn and aggserialtype These will only be required when aggtranstype is INTERNAL. Perhaps we should disallow CREATE AGGREAGET from accepting them for any other type... The return type of aggserialfn should be aggserialtype, and it should take a single parameter of aggtranstype. aggdeserialfn will be the reverse of that. Add a new bool field to nodeAgg's state named serialStates. If this is field is set to true then when we're in finalizeAgg = false mode, we'll call the serialfn on the agg state instead of the finalfn. This will allow the serialized state to be stored in the tuple and sent off to the main backend. The combine agg node should also be set to serialStates = true, so that it knows to deserialize instead of just assuming that the agg state is of type aggtranstype.I believe this should allow us to not cause any performance regressions by moving away from INTERNAL agg states. It should also be very efficient for internal states such as Int8TransTypeData, as this struct merely has 2 int64 fields which should be very simple to stuff into a bytea varlena type. We don't need to mess around converting the ->count and ->sum into a strings or anything.Then in order for the planner to allow parallel aggregation all aggregates must:Not have a DISTINCT or ORDER BY clauseHave a combinefnIf aggtranstype = INTERNAL, must have a aggserialfn and aggdeserialfn.We can relax the requirement on 3 if we're using 2-stage aggregation, but not parallel aggregation.Any objections, or better ideas?That just leaves me to figure out how to set the correct Aggref->aggtype during planning, as now there's 3 possible types:if (finalizeAggs == false){ if (serialStates == true) type = aggserialtype; else type = aggtranstype;}else type = prorettype; /* normal case */-- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Mon, 21 Dec 2015 22:02:29 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 21, 2015 at 4:02 AM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> Ok, so it seems that my mindset was not in parallel process space when I\\n> was thinking about this. I think having the pointer in the Tuple is\\n> probably going to be fine for this multiple stage aggregation when that's\\n> occurring in a single backend process, but obviously if the memory that the\\n> pointer points to belongs to a worker process in a parallel aggregate\\n> situation, then bad things will happen.\\n>\\n> Now, there has been talk of this previously, on various threads, but I don't\\n> believe any final decisions were made on how exactly it should be done. At\\n> the moment I plan to make changes as follows:\\n>\\n> Add 3 new columns to pg_aggregate, aggserialfn, aggdeserialfn and\\n> aggserialtype These will only be required when aggtranstype is INTERNAL.\\n> Perhaps we should disallow CREATE AGGREAGET from accepting them for any\\n> other type... The return type of aggserialfn should be aggserialtype, and it\\n> should take a single parameter of aggtranstype. aggdeserialfn will be the\\n> reverse of that.\\n> Add a new bool field to nodeAgg's state named serialStates. If this is field\\n> is set to true then when we're in finalizeAgg = false mode, we'll call the\\n> serialfn on the agg state instead of the finalfn. This will allow the\\n> serialized state to be stored in the tuple and sent off to the main backend.\\n> The combine agg node should also be set to serialStates = true, so that it\\n> knows to deserialize instead of just assuming that the agg state is of type\\n> aggtranstype.\\n>\\n> I believe this should allow us to not cause any performance regressions by\\n> moving away from INTERNAL agg states. It should also be very efficient for\\n> internal states such as Int8TransTypeData, as this struct merely has 2 int64\\n> fields which should be very simple to stuff into a bytea varlena type. We\\n> don't need to mess around converting the ->count and ->sum into a strings or\\n> anything.\\n>\\n> Then in order for the planner to allow parallel aggregation all aggregates\\n> must:\\n>\\n> Not have a DISTINCT or ORDER BY clause\\n> Have a combinefn\\n> If aggtranstype = INTERNAL, must have a aggserialfn and aggdeserialfn.\\n>\\n> We can relax the requirement on 3 if we're using 2-stage aggregation, but\\n> not parallel aggregation.\\n>\\n> Any objections, or better ideas?\\n\\nCan we use Tom's expanded-object stuff instead of introducing\\naggserialfn and aggdeserialfn? In other words, if you have a\\naggtranstype = INTERNAL, then what we do is:\\n\\n1. Create a new data type that represents the transition state.\\n2. Use expanded-object notation for that data type when we're just\\nwithin a single process, and flatten it when we need to send it\\nbetween processes.\\n\\nOne thing to keep in mind is that we also want to be able to support a\\nplan that involves having one or more remote servers do partial\\naggregation, send us the partial values, combine them across servers\\nand possibly also with locally computed-values, and the finalize the\\naggregation. So it would be nice if there were a way to invoke the\\naggregate function from SQL and get a transition value back rather\\nthan a final value.\\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, 21 Dec 2015 07:30:48 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 22 December 2015 at 01:30, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> Can we use Tom's expanded-object stuff instead of introducing\\n> aggserialfn and aggdeserialfn? In other words, if you have a\\n> aggtranstype = INTERNAL, then what we do is:\\n>\\n> 1. Create a new data type that represents the transition state.\\n> 2. Use expanded-object notation for that data type when we're just\\n> within a single process, and flatten it when we need to send it\\n> between processes.\\n>\\n>\\nI'd not seen this before, but on looking at it I'm not sure if using it\\nwill be practical to use for this. I may have missed something, but it\\nseems that after each call of the transition function, I'd need to ensure\\nthat the INTERNAL state was in the varlana format. This might be ok for a\\nstate like Int8TransTypeData, since that struct has no pointers, but I\\ndon't see how that could be done efficiently for NumericAggState, which has\\ntwo NumericVar, which will have pointers to other memory. The trans\\nfunction also has no idea whether it'll be called again for this state, so\\nit does not seem possible to delay the conversion until the final call of\\nthe trans function.\\n\\n\\n> One thing to keep in mind is that we also want to be able to support a\\n> plan that involves having one or more remote servers do partial\\n> aggregation, send us the partial values, combine them across servers\\n> and possibly also with locally computed-values, and the finalize the\\n> aggregation. So it would be nice if there were a way to invoke the\\n> aggregate function from SQL and get a transition value back rather\\n> than a final value.\\n\\n\\nThis will be possible with what I proposed. The Agg Node will just need to\\nbe setup with finalizeAggs=false, serialState=true. That way the returned\\naggregate values will be the states converted into the serial type, to\\nwhich we can call the output function on and send where ever we like.\\n\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 22 December 2015 at 01:30, Robert Haas <robertmhaas@gmail.com> wrote:Can we use Tom's expanded-object stuff instead of introducing\\naggserialfn and aggdeserialfn? In other words, if you have a\\naggtranstype = INTERNAL, then what we do is:\\n\\n1. Create a new data type that represents the transition state.\\n2. Use expanded-object notation for that data type when we're just\\nwithin a single process, and flatten it when we need to send it\\nbetween processes.\\nI'd not seen this before, but on looking at it I'm not sure if using it will be practical to use for this. I may have missed something, but it seems that after each call of the transition function, I'd need to ensure that the INTERNAL state was in the varlana format. This might be ok for a state like Int8TransTypeData, since that struct has no pointers, but I don't see how that could be done efficiently for NumericAggState, which has two NumericVar, which will have pointers to other memory. The trans function also has no idea whether it'll be called again for this state, so it does not seem possible to delay the conversion until the final call of the trans function. \\nOne thing to keep in mind is that we also want to be able to support a\\nplan that involves having one or more remote servers do partial\\naggregation, send us the partial values, combine them across servers\\nand possibly also with locally computed-values, and the finalize the\\naggregation. So it would be nice if there were a way to invoke the\\naggregate function from SQL and get a transition value back rather\\nthan a final value.This will be possible with what I proposed. The Agg Node will just need to be setup with finalizeAggs=false, serialState=true. That way the returned aggregate values will be the states converted into the serial type, to which we can call the output function on and send where ever we like. -- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Tue, 22 Dec 2015 13:53:08 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 21 December 2015 at 22:02, David Rowley <david.rowley@2ndquadrant.com>\\nwrote:\\n\\n> At the moment I plan to make changes as follows:\\n>\\n>\\n> 1. Add 3 new columns to pg_aggregate, aggserialfn, aggdeserialfn and\\n> aggserialtype These will only be required when aggtranstype is INTERNAL.\\n> Perhaps we should disallow CREATE AGGREAGET from accepting them for any\\n> other type... The return type of aggserialfn should be aggserialtype, and\\n> it should take a single parameter of aggtranstype. aggdeserialfn will be\\n> the reverse of that.\\n> 2. Add a new bool field to nodeAgg's state named serialStates. If this\\n> is field is set to true then when we're in finalizeAgg = false mode, we'll\\n> call the serialfn on the agg state instead of the finalfn. This will allow\\n> the serialized state to be stored in the tuple and sent off to the main\\n> backend. The combine agg node should also be set to serialStates = true,\\n> so that it knows to deserialize instead of just assuming that the agg state\\n> is of type aggtranstype.\\n>\\n>\\nI've attached an updated patch which implements this.\\n\\ncombine_aggs_test_v3.patch can be applied on top of the other patch to have\\nthe planner generate partial aggregate plans. It is also possible to\\nfurther hack this test patch to remove the combine Agg node from planner.c\\nin order to have the intermediate states output. I've also invented a\\nserialize and deseriaize function for avg(NUMERIC), and SUM(NUMERIC). Quite\\npossible this is what we'd want a remote server to send us if we get remote\\naggregation on a server grid one day in the future. With the patch modified\\nto not generate the combine aggs node we get:\\n\\n# select sum(x::numeric) from generate_series(1,1000) x(x);\\n\\n----------------------\\n 1000 500500 0 1000 0\\n(1 row)\\n\\nThis is just my serial format that I've come up with for NumericAggState,\\nwhich is basically: \\\"N sumX maxScale maxScaleCount NaNcount\\\". Perhaps we\\ncan come up with something better, maybe just packing the ints and int64s\\ninto a bytea type and putting the text version of sumX on the end... I'm\\nsure we can think of something more efficient between us, but I think the\\nserial state should definitely be cross platform e.g if we do the bytea\\nthing, then the ints should be in network byte order so that a server\\ncluster can have a mix of little and big-endian processors.\\n\\nI've also left a failing regression test because I'm not quite sure how to\\ndeal with it:\\n\\n--- 283,292 ----\\n WHERE p1.prorettype = 'internal'::regtype AND NOT\\n 'internal'::regtype = ANY (p1.proargtypes);\\n oid | proname\\n! ------+-------------------------\\n! 3319 | numeric_avg_deserialize\\n 2304 | internal_in\\n\\nThis is basically making sure that there's only 1 function that takes zero\\ninternal types as parameters, but returns an INTERNAL type. I've made it so\\nthat numeric_avg_deserialize() can only be called from an aggregate\\ncontext, which hopefully makes that safe again, but I'm perhaps not\\nimagining hard enough for ways that this could be abused, so I've held off\\nfor now in updated the expected output on that to include the new function.\\n\\nOne other part that I'm not too sure on how to deal with is how to set the\\ndata type for the Aggrefs when we're not performing finalization on the\\naggregate node. The return type for the Aggref in this case will be either\\nthe transtype, or the serialtype, depending on if we're serializing the\\nstates or not. To do this, I've so far just come up\\nwith set_partialagg_aggref_types() which is called during setrefs. The only\\nother time that I can think to do this return type update would be when\\nbuilding the partial agg node's target list. I'm open to better ideas on\\nthis part.\\n\\nApart from the problems I listed above, I'm reasonably happy with the patch\\nnow, and I'm ready for someone else to take a serious look at it.\\n\\nMany thanks\\n\\n-- \\n David Rowley 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, 23 Dec 2015 21:50:03 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 23, 2015 at 7:50 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> One other part that I'm not too sure on how to deal with is how to set the\\n> data type for the Aggrefs when we're not performing finalization on the\\n> aggregate node. The return type for the Aggref in this case will be either\\n> the transtype, or the serialtype, depending on if we're serializing the\\n> states or not. To do this, I've so far just come up with\\n> set_partialagg_aggref_types() which is called during setrefs. The only other\\n> time that I can think to do this return type update would be when building\\n> the partial agg node's target list. I'm open to better ideas on this part.\\n\\n\\nThanks for the patch. I am not sure about the proper place of this change,\\nbut changing it with transtype will make all float4 and float8 aggregates to\\nwork in parallel. Most of these aggregates return type is typbyval and\\ntransition type is a variable length.\\n\\nwe may need to write better combine functions for these types to avoid wrong\\nresults because of parallel.\\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\": \"Thu, 24 Dec 2015 11:55:47 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 23, 2015 at 7:50 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> This is just my serial format that I've come up with for NumericAggState,\\n> which is basically: \\\"N sumX maxScale maxScaleCount NaNcount\\\". Perhaps we can\\n> come up with something better, maybe just packing the ints and int64s into a\\n> bytea type and putting the text version of sumX on the end... I'm sure we\\n> can think of something more efficient between us, but I think the serial\\n> state should definitely be cross platform e.g if we do the bytea thing, then\\n> the ints should be in network byte order so that a server cluster can have a\\n> mix of little and big-endian processors.\\n\\nInstead of adding serial and de-serial functions to all aggregates which have\\ntransition type as internal, how about adding these functions as send and\\nrecv functions for internal type? can be used only in aggregate context.\\nThe data can send and receive similar like other functions. Is it possible?\\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\": \"Thu, 24 Dec 2015 12:07:57 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 23 December 2015 at 21:50, David Rowley <david.rowley@2ndquadrant.com>\\nwrote:\\n\\n> Apart from the problems I listed above, I'm reasonably happy with the\\n> patch now, and I'm ready for someone else to take a serious look at it.\\n>\\n\\nI forgot to mention another situation where this patch might need a bit\\nmore thought. This does not affect any of the built in aggregate\\nfunctions, but if someone created a user defined aggregate such as:\\n\\nCREATE AGGREGATE sumplusten (int)\\n(\\nstype = int,\\nsfunc = int4pl,\\ncombinefunc = int4pl,\\ninitcond = '10'\\n);\\n\\nNote the initcond which initialises the state to 10. Then let's say we\\nlater add the ability to push aggregates down below a Append.\\n\\ncreate table t1 as select 10 as value from generate_series(1,10);\\ncreate table t2 as select 10 as value from generate_series(1,10);\\n\\nWith the following we get the correct result:\\n\\n# select sumplusten(value) from (select * from t1 union all select * from\\nt2) t;\\n sumplusten\\n------------\\n 210\\n(1 row)\\n\\nBut if we simulate how it would work in the aggregate push down situation:\\n\\n# select sum(value) from (select sumplusten(value) as value from t1 union\\nall select sumplusten(value) as value from t2) t;\\n sum\\n-----\\n 220\\n(1 row)\\n\\nHere I'm using sum() as a mock up of the combine function, but you get the\\nidea... Since we initialize the state twice, we get the wrong result.\\n\\nNow I'm not too sure if anyone would have an aggregate defined like this in\\nthe real world, but I don't think that'll give us a good enough excuse to\\ngo returning wrong results.\\n\\nIn the patch I could fix this by changing partial_aggregate_walker() to\\ndisable partial aggregation if the aggregate has an initvalue, but that\\nwould exclude a number of built-in aggregates unnecessarily. Alternatively\\nif there was some way to analyze the initvalue to see if it was \\\"zero\\\"...\\nI'm just not quite sure how we might do that at the moment.\\n\\nAny thoughts on a good way to fix this that does not exclude built-in\\naggregates with an initvalue are welcome.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 23 December 2015 at 21:50, David Rowley <david.rowley@2ndquadrant.com> wrote:Apart from the problems I listed above, I'm reasonably happy with the patch now, and I'm ready for someone else to take a serious look at it.I forgot to mention another situation where this patch might need a bit more thought. This does not affect any of the built in aggregate functions, but if someone created a user defined aggregate such as:CREATE AGGREGATE sumplusten (int)( stype = int, sfunc = int4pl, combinefunc = int4pl, initcond = '10');Note the initcond which initialises the state to 10. Then let's say we later add the ability to push aggregates down below a Append.create table t1 as select 10 as value from generate_series(1,10);create table t2 as select 10 as value from generate_series(1,10);With the following we get the correct result:# select sumplusten(value) from (select * from t1 union all select * from t2) t; sumplusten------------ 210(1 row)But if we simulate how it would work in the aggregate push down situation:# select sum(value) from (select sumplusten(value) as value from t1 union all select sumplusten(value) as value from t2) t; sum----- 220(1 row)Here I'm using sum() as a mock up of the combine function, but you get the idea... Since we initialize the state twice, we get the wrong result.Now I'm not too sure if anyone would have an aggregate defined like this in the real world, but I don't think that'll give us a good enough excuse to go returning wrong results.In the patch I could fix this by changing partial_aggregate_walker() to disable partial aggregation if the aggregate has an initvalue, but that would exclude a number of built-in aggregates unnecessarily. Alternatively if there was some way to analyze the initvalue to see if it was \\\"zero\\\"... I'm just not quite sure how we might do that at the moment.Any thoughts on a good way to fix this that does not exclude built-in aggregates with an initvalue are welcome. -- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Thu, 24 Dec 2015 14:28:59 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 24 December 2015 at 14:07, Haribabu Kommi <kommi.haribabu@gmail.com>\\nwrote:\\n\\n> On Wed, Dec 23, 2015 at 7:50 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n> > This is just my serial format that I've come up with for NumericAggState,\\n> > which is basically: \\\"N sumX maxScale maxScaleCount NaNcount\\\". Perhaps we\\n> can\\n> > come up with something better, maybe just packing the ints and int64s\\n> into a\\n> > bytea type and putting the text version of sumX on the end... I'm sure we\\n> > can think of something more efficient between us, but I think the serial\\n> > state should definitely be cross platform e.g if we do the bytea thing,\\n> then\\n> > the ints should be in network byte order so that a server cluster can\\n> have a\\n> > mix of little and big-endian processors.\\n>\\n> Instead of adding serial and de-serial functions to all aggregates which\\n> have\\n> transition type as internal, how about adding these functions as send and\\n> recv functions for internal type? can be used only in aggregate context.\\n> The data can send and receive similar like other functions. Is it possible?\\n\\n\\nThe requirement that needs to be met is that we have the ability to\\ncompletely represent the INTERNAL aggregate state inside a tuple. It's\\nthese tuples that are sent to the main process in the parallel aggregate\\nsituation, and also in Sorted and Plain Aggregate too, although in the case\\nof Sorted and Plain Aggregate we can simply put the pointer to the INTERNAL\\nstate in the tuple, and this can be dereferenced.\\n\\nPerhaps I'm lacking imagination, but right now I'm not sure how send and\\nrecv functions can be used to help us here. Although we perhaps could make\\nserial and deserial functions skip on type conversions by using similar\\nmethods as to what's used in send and recv.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 24 December 2015 at 14:07, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:On Wed, Dec 23, 2015 at 7:50 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> This is just my serial format that I've come up with for NumericAggState,\\n> which is basically: \\\"N sumX maxScale maxScaleCount NaNcount\\\". Perhaps we can\\n> come up with something better, maybe just packing the ints and int64s into a\\n> bytea type and putting the text version of sumX on the end... I'm sure we\\n> can think of something more efficient between us, but I think the serial\\n> state should definitely be cross platform e.g if we do the bytea thing, then\\n> the ints should be in network byte order so that a server cluster can have a\\n> mix of little and big-endian processors.\\n\\nInstead of adding serial and de-serial functions to all aggregates which have\\ntransition type as internal, how about adding these functions as send and\\nrecv functions for internal type? can be used only in aggregate context.\\nThe data can send and receive similar like other functions. Is it possible?The requirement that needs to be met is that we have the ability to completely represent the INTERNAL aggregate state inside a tuple. It's these tuples that are sent to the main process in the parallel aggregate situation, and also in Sorted and Plain Aggregate too, although in the case of Sorted and Plain Aggregate we can simply put the pointer to the INTERNAL state in the tuple, and this can be dereferenced.Perhaps I'm lacking imagination, but right now I'm not sure how send and recv functions can be used to help us here. Although we perhaps could make serial and deserial functions skip on type conversions by using similar methods as to what's used in send and recv.-- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Thu, 24 Dec 2015 14:56:12 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 24 December 2015 at 13:55, Haribabu Kommi <kommi.haribabu@gmail.com>\\nwrote:\\n\\n> On Wed, Dec 23, 2015 at 7:50 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n> > One other part that I'm not too sure on how to deal with is how to set\\n> the\\n> > data type for the Aggrefs when we're not performing finalization on the\\n> > aggregate node. The return type for the Aggref in this case will be\\n> either\\n> > the transtype, or the serialtype, depending on if we're serializing the\\n> > states or not. To do this, I've so far just come up with\\n> > set_partialagg_aggref_types() which is called during setrefs. The only\\n> other\\n> > time that I can think to do this return type update would be when\\n> building\\n> > the partial agg node's target list. I'm open to better ideas on this\\n> part.\\n>\\n>\\n> Thanks for the patch. I am not sure about the proper place of this change,\\n> but changing it with transtype will make all float4 and float8 aggregates\\n> to\\n> work in parallel. Most of these aggregates return type is typbyval and\\n> transition type is a variable length.\\n>\\n> we may need to write better combine functions for these types to avoid\\n> wrong\\n> results because of parallel.\\n>\\n\\nI might be misunderstanding you here, but yeah, well, if by \\\"write better\\\"\\nyou mean \\\"write some\\\", then yeah :) I only touched sum(), min() and max()\\nso far as I didn't need to do anything special with these. I'm not quite\\nsure what you mean with the \\\"wrong results\\\" part. Could you explain more?\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 24 December 2015 at 13:55, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:On Wed, Dec 23, 2015 at 7:50 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> One other part that I'm not too sure on how to deal with is how to set the\\n> data type for the Aggrefs when we're not performing finalization on the\\n> aggregate node. The return type for the Aggref in this case will be either\\n> the transtype, or the serialtype, depending on if we're serializing the\\n> states or not. To do this, I've so far just come up with\\n> set_partialagg_aggref_types() which is called during setrefs. The only other\\n> time that I can think to do this return type update would be when building\\n> the partial agg node's target list. I'm open to better ideas on this part.\\n\\n\\nThanks for the patch. I am not sure about the proper place of this change,\\nbut changing it with transtype will make all float4 and float8 aggregates to\\nwork in parallel. Most of these aggregates return type is typbyval and\\ntransition type is a variable length.\\nwe may need to write better combine functions for these types to avoid wrong\\nresults because of parallel.I might be misunderstanding you here, but yeah, well, if by \\\"write better\\\" you mean \\\"write some\\\", then yeah :) I only touched sum(), min() and max() so far as I didn't need to do anything special with these. I'm not quite sure what you mean with the \\\"wrong results\\\" part. Could you explain more?-- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Thu, 24 Dec 2015 15:12:52 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 24, 2015 at 1:12 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> On 24 December 2015 at 13:55, Haribabu Kommi <kommi.haribabu@gmail.com>\\n> wrote:\\n>>\\n>> On Wed, Dec 23, 2015 at 7:50 PM, David Rowley\\n>> <david.rowley@2ndquadrant.com> wrote:\\n>> > One other part that I'm not too sure on how to deal with is how to set\\n>> > the\\n>> > data type for the Aggrefs when we're not performing finalization on the\\n>> > aggregate node. The return type for the Aggref in this case will be\\n>> > either\\n>> > the transtype, or the serialtype, depending on if we're serializing the\\n>> > states or not. To do this, I've so far just come up with\\n>> > set_partialagg_aggref_types() which is called during setrefs. The only\\n>> > other\\n>> > time that I can think to do this return type update would be when\\n>> > building\\n>> > the partial agg node's target list. I'm open to better ideas on this\\n>> > part.\\n>>\\n>>\\n>> Thanks for the patch. I am not sure about the proper place of this change,\\n>> but changing it with transtype will make all float4 and float8 aggregates\\n>> to\\n>> work in parallel. Most of these aggregates return type is typbyval and\\n>> transition type is a variable length.\\n>>\\n>> we may need to write better combine functions for these types to avoid\\n>> wrong\\n>> results because of parallel.\\n>\\n>\\n> I might be misunderstanding you here, but yeah, well, if by \\\"write better\\\"\\n> you mean \\\"write some\\\", then yeah :) I only touched sum(), min() and max()\\n> so far as I didn't need to do anything special with these. I'm not quite\\n> sure what you mean with the \\\"wrong results\\\" part. Could you explain more?\\n\\nsorry for not providing clear information. To check the change of replacing\\naggtype with aggtranstype is working fine or not for float8 avg. I just tried\\nadding a float8_combine_accum function for float8 avg similar to\\nfloat8_acuum with a difference of adding the two transvalues instead\\nof newval to existing transval in float8_accum function as follows,\\n\\nN += transvalues1[0];\\nsumX += transvalues1[1];\\nsumX2 += transvalues1[2];\\n\\nBut the result came different compared to normal aggregate. The data in the\\ncolumn is 1.1\\n\\npostgres=# select avg(f3) from tbl where f1 < 1001 group by f3;\\n avg\\n-----------------------\\n 2.16921537594434e-316\\n(1 row)\\n\\npostgres=# set enable_parallelagg = false;\\nSET\\npostgres=# select avg(f3) from tbl where f1 < 1001 group by f3;\\n avg\\n------------------\\n 1.10000000000001\\n(1 row)\\n\\n\\nFirst i thought it is because of combine function problem, but it seems\\nsome where else the problem is present in parallel aggregate code. sorry\\nfor the noise.\\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\": \"Thu, 24 Dec 2015 13:30:01 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 21, 2015 at 4:53 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> On 22 December 2015 at 01:30, Robert Haas <robertmhaas@gmail.com> wrote:\\n>> Can we use Tom's expanded-object stuff instead of introducing\\n>> aggserialfn and aggdeserialfn? In other words, if you have a\\n>> aggtranstype = INTERNAL, then what we do is:\\n>>\\n>> 1. Create a new data type that represents the transition state.\\n>> 2. Use expanded-object notation for that data type when we're just\\n>> within a single process, and flatten it when we need to send it\\n>> between processes.\\n>>\\n>\\n> I'd not seen this before, but on looking at it I'm not sure if using it will\\n> be practical to use for this. I may have missed something, but it seems that\\n> after each call of the transition function, I'd need to ensure that the\\n> INTERNAL state was in the varlana format.\\n\\nNo, the idea I had in mind was to allow it to continue to exist in the\\nexpanded format until you really need it in the varlena format, and\\nthen serialize it at that point. You'd actually need to do the\\nopposite: if you get an input that is not in expanded format, expand\\nit.\\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, 24 Dec 2015 17:10:58 -0800\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 25 December 2015 at 14:10, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On Mon, Dec 21, 2015 at 4:53 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n> > On 22 December 2015 at 01:30, Robert Haas <robertmhaas@gmail.com> wrote:\\n> >> Can we use Tom's expanded-object stuff instead of introducing\\n> >> aggserialfn and aggdeserialfn? In other words, if you have a\\n> >> aggtranstype = INTERNAL, then what we do is:\\n> >>\\n> >> 1. Create a new data type that represents the transition state.\\n> >> 2. Use expanded-object notation for that data type when we're just\\n> >> within a single process, and flatten it when we need to send it\\n> >> between processes.\\n> >>\\n> >\\n> > I'd not seen this before, but on looking at it I'm not sure if using it\\n> will\\n> > be practical to use for this. I may have missed something, but it seems\\n> that\\n> > after each call of the transition function, I'd need to ensure that the\\n> > INTERNAL state was in the varlana format.\\n>\\n> No, the idea I had in mind was to allow it to continue to exist in the\\n> expanded format until you really need it in the varlena format, and\\n> then serialize it at that point. You'd actually need to do the\\n> opposite: if you get an input that is not in expanded format, expand\\n> it.\\n\\n\\nAdmittedly I'm struggling to see how this can be done. I've spent a good\\nbit of time analysing how the expanded object stuff works.\\n\\nHypothetically let's say we can make it work like:\\n\\n1. During partial aggregation (finalizeAggs = false), in\\nfinalize_aggregates(), where we'd normally call the final function, instead\\nflatten INTERNAL states and store the flattened Datum instead of the\\npointer to the INTERNAL state.\\n2. During combining aggregation (combineStates = true) have all the combine\\nfunctions written in such a ways that the INTERNAL states expand the\\nflattened states before combining the aggregate states.\\n\\nDoes that sound like what you had in mind?\\n\\nIf so I can't quite seem to wrap my head around 1. As I'm really not quite\\nsure how, from finalize_aggregates() we'd flatten the INTERNAL pointer. I\\nmean, how do we know which flatten function to call here? From reading the\\nexpanded-object code I see that its used in expand_array(), In this case we\\nknow we're working with arrays, so it just always uses the EA_methods\\nglobally scoped struct to get the function pointers it requires for\\nflattening the array. For the case of finalize_aggregates(), the best I can\\nthink of here is to have a bunch of global structs and then have a giant\\ncase statement to select the correct one. That's clearly horrid, and not\\ncommit worthy, and it does nothing to help user defined aggregates which\\nuse INTERNAL types. Am I missing something here?\\n\\nAs of the most recent patch I posted, having the serial and deserial\\nfunctions in the catalogs allows user defined aggregates with INTERNAL\\nstates to work just fine. Admittedly I'm not all that happy that I've had\\nto add 4 new columns to pg_aggregate to support this, but if I could think\\nof how to make it work without doing that, then I'd likely go and do that\\ninstead.\\n\\nIf your problem with the serialize and deserialize stuff is around the\\nserialized format, then can see no reason why we couldn't just invent some\\ncomposite types for the current INTERNAL aggregate states, and have the\\nserialfn convert the INTERNAL state into one of those, then have the\\ndeserialfn perform the opposite. Likely this would be neater than what I\\nhave at the moment with just converting the INTERNAL state into text.\\n\\nPlease let me know what I'm missing with the expanded-object code.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 25 December 2015 at 14:10, Robert Haas <robertmhaas@gmail.com> wrote:On Mon, Dec 21, 2015 at 4:53 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> On 22 December 2015 at 01:30, Robert Haas <robertmhaas@gmail.com> wrote:\\n>> Can we use Tom's expanded-object stuff instead of introducing\\n>> aggserialfn and aggdeserialfn? In other words, if you have a\\n>> aggtranstype = INTERNAL, then what we do is:\\n>>\\n>> 1. Create a new data type that represents the transition state.\\n>> 2. Use expanded-object notation for that data type when we're just\\n>> within a single process, and flatten it when we need to send it\\n>> between processes.\\n>>\\n>\\n> I'd not seen this before, but on looking at it I'm not sure if using it will\\n> be practical to use for this. I may have missed something, but it seems that\\n> after each call of the transition function, I'd need to ensure that the\\n> INTERNAL state was in the varlana format.\\n\\nNo, the idea I had in mind was to allow it to continue to exist in the\\nexpanded format until you really need it in the varlena format, and\\nthen serialize it at that point. You'd actually need to do the\\nopposite: if you get an input that is not in expanded format, expand\\nit.Admittedly I'm struggling to see how this can be done. I've spent a good bit of time analysing how the expanded object stuff works.Hypothetically let's say we can make it work like:1. During partial aggregation (finalizeAggs = false), in finalize_aggregates(), where we'd normally call the final function, instead flatten INTERNAL states and store the flattened Datum instead of the pointer to the INTERNAL state.2. During combining aggregation (combineStates = true) have all the combine functions written in such a ways that the INTERNAL states expand the flattened states before combining the aggregate states.Does that sound like what you had in mind?If so I can't quite seem to wrap my head around 1. As I'm really not quite sure how, from finalize_aggregates() we'd flatten the INTERNAL pointer. I mean, how do we know which flatten function to call here? >From reading the expanded-object code I see that its used in expand_array(), In this case we know we're working with arrays, so it just always uses the EA_methods globally scoped struct to get the function pointers it requires for flattening the array. For the case of finalize_aggregates(), the best I can think of here is to have a bunch of global structs and then have a giant case statement to select the correct one. That's clearly horrid, and not commit worthy, and it does nothing to help user defined aggregates which use INTERNAL types. Am I missing something here?As of the most recent patch I posted, having the serial and deserial functions in the catalogs allows user defined aggregates with INTERNAL states to work just fine. Admittedly I'm not all that happy that I've had to add 4 new columns to pg_aggregate to support this, but if I could think of how to make it work without doing that, then I'd likely go and do that instead.If your problem with the serialize and deserialize stuff is around the serialized format, then can see no reason why we couldn't just invent some composite types for the current INTERNAL aggregate states, and have the serialfn convert the INTERNAL state into one of those, then have the deserialfn perform the opposite. Likely this would be neater than what I have at the moment with just converting the INTERNAL state into text.Please let me know what I'm missing with the expanded-object code.-- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Wed, 30 Dec 2015 13:39:55 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"I've attached some re-based patched on current master. This is just to fix\\na duplicate OID problem.\\n\\n-- \\n David Rowley 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, 8 Jan 2016 22:43:28 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 8 January 2016 at 22:43, David Rowley <david.rowley@2ndquadrant.com>\\nwrote:\\n\\n> I've attached some re-based patched on current master. This is just to fix\\n> a duplicate OID problem.\\n>\\n>\\n\\nI've attached two updated patched to fix a conflict with a recent change to\\nplanner.c\\n\\n-- \\n David Rowley 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, 15 Jan 2016 17:34:07 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 15, 2016 at 3:34 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> On 8 January 2016 at 22:43, David Rowley <david.rowley@2ndquadrant.com>\\n> wrote:\\n>>\\n>> I've attached some re-based patched on current master. This is just to fix\\n>> a duplicate OID problem.\\n>>\\n>\\n> I've attached two updated patched to fix a conflict with a recent change to\\n> planner.c\\n\\nI am getting following compilation error and warning with the latest patch,\\nbecause of a function prototype mismatch.\\n\\naggregatecmds.c: In function ‘DefineAggregate’:\\naggregatecmds.c:93:8: warning: variable ‘serialTypeType’ set but not\\nused [-Wunused-but-set-variable]\\n char serialTypeType = 0;\\n ^\\nclauses.c:434:1: error: conflicting types for ‘partial_aggregate_walker’\\n partial_aggregate_walker(Node *node, partial_agg_context *context)\\n ^\\nclauses.c:100:13: note: previous declaration of\\n‘partial_aggregate_walker’ was here\\n static bool partial_aggregate_walker(Node *node, void *context);\\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\": \"Fri, 15 Jan 2016 15:46:22 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 15 January 2016 at 17:46, Haribabu Kommi <kommi.haribabu@gmail.com>\\nwrote:\\n\\n> On Fri, Jan 15, 2016 at 3:34 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n> > On 8 January 2016 at 22:43, David Rowley <david.rowley@2ndquadrant.com>\\n> > wrote:\\n> >>\\n> >> I've attached some re-based patched on current master. This is just to\\n> fix\\n> >> a duplicate OID problem.\\n> >>\\n> >\\n> > I've attached two updated patched to fix a conflict with a recent change\\n> to\\n> > planner.c\\n>\\n> I am getting following compilation error and warning with the latest patch,\\n> because of a function prototype mismatch.\\n>\\n> aggregatecmds.c: In function ‘DefineAggregate’:\\n> aggregatecmds.c:93:8: warning: variable ‘serialTypeType’ set but not\\n> used [-Wunused-but-set-variable]\\n> char serialTypeType = 0;\\n> ^\\n> clauses.c:434:1: error: conflicting types for ‘partial_aggregate_walker’\\n> partial_aggregate_walker(Node *node, partial_agg_context *context)\\n> ^\\n> clauses.c:100:13: note: previous declaration of\\n> ‘partial_aggregate_walker’ was here\\n> static bool partial_aggregate_walker(Node *node, void *context);\\n> ^\\n\\n\\nThanks for noticing that. The compiler I used didn't seem to mind that...\\nWhich is a bit of a worry.\\n\\nI've attached an updated patch, and the same test patch again.\\n\\n-- \\n David Rowley 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, 15 Jan 2016 18:40:26 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"Man, I really shouldn't go on vacation for Christmas or, like, ever.\\nMy email responses get way too slow. Sorry.\\n\\nOn Tue, Dec 29, 2015 at 7:39 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n>> No, the idea I had in mind was to allow it to continue to exist in the\\n>> expanded format until you really need it in the varlena format, and\\n>> then serialize it at that point. You'd actually need to do the\\n>> opposite: if you get an input that is not in expanded format, expand\\n>> it.\\n>\\n> Admittedly I'm struggling to see how this can be done. I've spent a good bit\\n> of time analysing how the expanded object stuff works.\\n>\\n> Hypothetically let's say we can make it work like:\\n>\\n> 1. During partial aggregation (finalizeAggs = false), in\\n> finalize_aggregates(), where we'd normally call the final function, instead\\n> flatten INTERNAL states and store the flattened Datum instead of the pointer\\n> to the INTERNAL state.\\n> 2. During combining aggregation (combineStates = true) have all the combine\\n> functions written in such a ways that the INTERNAL states expand the\\n> flattened states before combining the aggregate states.\\n>\\n> Does that sound like what you had in mind?\\n\\nMore or less. But what I was really imagining is that we'd get rid of\\nthe internal states and replace them with new datatypes built to\\npurpose. So, for example, for string_agg(text, text) you could make a\\nnew datatype that is basically a StringInfo. In expanded form, it\\nreally is a StringInfo. When you flatten it, you just get the string.\\nWhen somebody expands it again, they again have a StringInfo. So the\\nRW pointer to the expanded form supports append cheaply.\\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, 15 Jan 2016 09:03:14 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 16 January 2016 at 03:03, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On Tue, Dec 29, 2015 at 7:39 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n> >> No, the idea I had in mind was to allow it to continue to exist in the\\n> >> expanded format until you really need it in the varlena format, and\\n> >> then serialize it at that point. You'd actually need to do the\\n> >> opposite: if you get an input that is not in expanded format, expand\\n> >> it.\\n> >\\n> > Admittedly I'm struggling to see how this can be done. I've spent a good\\n> bit\\n> > of time analysing how the expanded object stuff works.\\n> >\\n> > Hypothetically let's say we can make it work like:\\n> >\\n> > 1. During partial aggregation (finalizeAggs = false), in\\n> > finalize_aggregates(), where we'd normally call the final function,\\n> instead\\n> > flatten INTERNAL states and store the flattened Datum instead of the\\n> pointer\\n> > to the INTERNAL state.\\n> > 2. During combining aggregation (combineStates = true) have all the\\n> combine\\n> > functions written in such a ways that the INTERNAL states expand the\\n> > flattened states before combining the aggregate states.\\n> >\\n> > Does that sound like what you had in mind?\\n>\\n> More or less. But what I was really imagining is that we'd get rid of\\n> the internal states and replace them with new datatypes built to\\n> purpose. So, for example, for string_agg(text, text) you could make a\\n> new datatype that is basically a StringInfo. In expanded form, it\\n> really is a StringInfo. When you flatten it, you just get the string.\\n> When somebody expands it again, they again have a StringInfo. So the\\n> RW pointer to the expanded form supports append cheaply.\\n\\n\\nThat sounds fine in theory, but where and how do you suppose we determine\\nwhich expand function to call? Nothing exists in the catalogs to decide\\nthis currently.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 16 January 2016 at 03:03, Robert Haas <robertmhaas@gmail.com> wrote:On Tue, Dec 29, 2015 at 7:39 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n>> No, the idea I had in mind was to allow it to continue to exist in the\\n>> expanded format until you really need it in the varlena format, and\\n>> then serialize it at that point. You'd actually need to do the\\n>> opposite: if you get an input that is not in expanded format, expand\\n>> it.\\n>\\n> Admittedly I'm struggling to see how this can be done. I've spent a good bit\\n> of time analysing how the expanded object stuff works.\\n>\\n> Hypothetically let's say we can make it work like:\\n>\\n> 1. During partial aggregation (finalizeAggs = false), in\\n> finalize_aggregates(), where we'd normally call the final function, instead\\n> flatten INTERNAL states and store the flattened Datum instead of the pointer\\n> to the INTERNAL state.\\n> 2. During combining aggregation (combineStates = true) have all the combine\\n> functions written in such a ways that the INTERNAL states expand the\\n> flattened states before combining the aggregate states.\\n>\\n> Does that sound like what you had in mind?\\n\\nMore or less. But what I was really imagining is that we'd get rid of\\nthe internal states and replace them with new datatypes built to\\npurpose. So, for example, for string_agg(text, text) you could make a\\nnew datatype that is basically a StringInfo. In expanded form, it\\nreally is a StringInfo. When you flatten it, you just get the string.\\nWhen somebody expands it again, they again have a StringInfo. So the\\nRW pointer to the expanded form supports append cheaply.That sounds fine in theory, but where and how do you suppose we determine which expand function to call? Nothing exists in the catalogs to decide this currently. -- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Sat, 16 Jan 2016 14:00:36 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Sat, Jan 16, 2016 at 12:00 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> On 16 January 2016 at 03:03, Robert Haas <robertmhaas@gmail.com> wrote:\\n>>\\n>> On Tue, Dec 29, 2015 at 7:39 PM, David Rowley\\n>> <david.rowley@2ndquadrant.com> wrote:\\n>> >> No, the idea I had in mind was to allow it to continue to exist in the\\n>> >> expanded format until you really need it in the varlena format, and\\n>> >> then serialize it at that point. You'd actually need to do the\\n>> >> opposite: if you get an input that is not in expanded format, expand\\n>> >> it.\\n>> >\\n>> > Admittedly I'm struggling to see how this can be done. I've spent a good\\n>> > bit\\n>> > of time analysing how the expanded object stuff works.\\n>> >\\n>> > Hypothetically let's say we can make it work like:\\n>> >\\n>> > 1. During partial aggregation (finalizeAggs = false), in\\n>> > finalize_aggregates(), where we'd normally call the final function,\\n>> > instead\\n>> > flatten INTERNAL states and store the flattened Datum instead of the\\n>> > pointer\\n>> > to the INTERNAL state.\\n>> > 2. During combining aggregation (combineStates = true) have all the\\n>> > combine\\n>> > functions written in such a ways that the INTERNAL states expand the\\n>> > flattened states before combining the aggregate states.\\n>> >\\n>> > Does that sound like what you had in mind?\\n>>\\n>> More or less. But what I was really imagining is that we'd get rid of\\n>> the internal states and replace them with new datatypes built to\\n>> purpose. So, for example, for string_agg(text, text) you could make a\\n>> new datatype that is basically a StringInfo. In expanded form, it\\n>> really is a StringInfo. When you flatten it, you just get the string.\\n>> When somebody expands it again, they again have a StringInfo. So the\\n>> RW pointer to the expanded form supports append cheaply.\\n>\\n>\\n> That sounds fine in theory, but where and how do you suppose we determine\\n> which expand function to call? Nothing exists in the catalogs to decide this\\n> currently.\\n\\nI am thinking of transition function returns and accepts the StringInfoData\\ninstead of PolyNumAggState internal data for int8_avg_accum for example.\\n\\nThe StringInfoData is formed with the members of the PolyNumAggState\\nstructure data. The input given StringInfoData is transformed into\\nPolyNumAggState data and finish the calculation and again form the\\nStringInfoData and return. Similar changes needs to be done for final\\nfunctions input type also. I am not sure whether this approach may have\\nsome impact on performance?\\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\": \"Mon, 18 Jan 2016 12:36:30 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 18 January 2016 at 14:36, Haribabu Kommi <kommi.haribabu@gmail.com>\\nwrote:\\n\\n> On Sat, Jan 16, 2016 at 12:00 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n> > On 16 January 2016 at 03:03, Robert Haas <robertmhaas@gmail.com> wrote:\\n> >>\\n> >> On Tue, Dec 29, 2015 at 7:39 PM, David Rowley\\n> >> <david.rowley@2ndquadrant.com> wrote:\\n> >> >> No, the idea I had in mind was to allow it to continue to exist in\\n> the\\n> >> >> expanded format until you really need it in the varlena format, and\\n> >> >> then serialize it at that point. You'd actually need to do the\\n> >> >> opposite: if you get an input that is not in expanded format, expand\\n> >> >> it.\\n> >> >\\n> >> > Admittedly I'm struggling to see how this can be done. I've spent a\\n> good\\n> >> > bit\\n> >> > of time analysing how the expanded object stuff works.\\n> >> >\\n> >> > Hypothetically let's say we can make it work like:\\n> >> >\\n> >> > 1. During partial aggregation (finalizeAggs = false), in\\n> >> > finalize_aggregates(), where we'd normally call the final function,\\n> >> > instead\\n> >> > flatten INTERNAL states and store the flattened Datum instead of the\\n> >> > pointer\\n> >> > to the INTERNAL state.\\n> >> > 2. During combining aggregation (combineStates = true) have all the\\n> >> > combine\\n> >> > functions written in such a ways that the INTERNAL states expand the\\n> >> > flattened states before combining the aggregate states.\\n> >> >\\n> >> > Does that sound like what you had in mind?\\n> >>\\n> >> More or less. But what I was really imagining is that we'd get rid of\\n> >> the internal states and replace them with new datatypes built to\\n> >> purpose. So, for example, for string_agg(text, text) you could make a\\n> >> new datatype that is basically a StringInfo. In expanded form, it\\n> >> really is a StringInfo. When you flatten it, you just get the string.\\n> >> When somebody expands it again, they again have a StringInfo. So the\\n> >> RW pointer to the expanded form supports append cheaply.\\n> >\\n> >\\n> > That sounds fine in theory, but where and how do you suppose we determine\\n> > which expand function to call? Nothing exists in the catalogs to decide\\n> this\\n> > currently.\\n>\\n> I am thinking of transition function returns and accepts the StringInfoData\\n> instead of PolyNumAggState internal data for int8_avg_accum for example.\\n>\\n\\nhmm, so wouldn't that mean that the transition function would need to (for\\neach input tuple):\\n\\n1. Parse that StringInfo into tokens.\\n2. Create a new aggregate state object.\\n3. Populate the new aggregate state based on the tokenised StringInfo, this\\nwould perhaps require that various *_in() functions are called on each\\ntoken.\\n4. Add the new tuple to the aggregate state.\\n5. Build a new StringInfo based on the aggregate state modified in 4.\\n\\n?\\n\\nCurrently the transition function only does 4, and performs 2 only if it's\\nthe first Tuple.\\n\\nIs that what you mean? as I'd say that would slow things down significantly!\\n\\nTo get a gauge on how much more CPU work that would be for some aggregates,\\nhave a look at how simple int8_avg_accum() is currently when we have\\nHAVE_INT128 defined. For the case of AVG(BIGINT) we just really have:\\n\\nstate->sumX += newval;\\nstate->N++;\\n\\nThe above code is step 4 only. So unless I've misunderstood you, that would\\nneed to turn into steps 1-5 above. Step 4 here is probably just a handful\\nof instructions right now, but adding code for steps 1,2,3 and 5 would turn\\nthat into hundreds.\\n\\nI've been trying to avoid any overhead by adding the serializeStates flag\\nto make_agg() so that we can maintain the same performance when we're just\\npassing internal states around in the same process. This keeps the\\nconversions between internal state and serialised state to a minimum.\\n\\nThe StringInfoData is formed with the members of the PolyNumAggState\\n> structure data. The input given StringInfoData is transformed into\\n> PolyNumAggState data and finish the calculation and again form the\\n> StringInfoData and return. Similar changes needs to be done for final\\n> functions input type also. I am not sure whether this approach may have\\n> some impact on performance?\\n\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 18 January 2016 at 14:36, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:On Sat, Jan 16, 2016 at 12:00 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> On 16 January 2016 at 03:03, Robert Haas <robertmhaas@gmail.com> wrote:\\n>>\\n>> On Tue, Dec 29, 2015 at 7:39 PM, David Rowley\\n>> <david.rowley@2ndquadrant.com> wrote:\\n>> >> No, the idea I had in mind was to allow it to continue to exist in the\\n>> >> expanded format until you really need it in the varlena format, and\\n>> >> then serialize it at that point. You'd actually need to do the\\n>> >> opposite: if you get an input that is not in expanded format, expand\\n>> >> it.\\n>> >\\n>> > Admittedly I'm struggling to see how this can be done. I've spent a good\\n>> > bit\\n>> > of time analysing how the expanded object stuff works.\\n>> >\\n>> > Hypothetically let's say we can make it work like:\\n>> >\\n>> > 1. During partial aggregation (finalizeAggs = false), in\\n>> > finalize_aggregates(), where we'd normally call the final function,\\n>> > instead\\n>> > flatten INTERNAL states and store the flattened Datum instead of the\\n>> > pointer\\n>> > to the INTERNAL state.\\n>> > 2. During combining aggregation (combineStates = true) have all the\\n>> > combine\\n>> > functions written in such a ways that the INTERNAL states expand the\\n>> > flattened states before combining the aggregate states.\\n>> >\\n>> > Does that sound like what you had in mind?\\n>>\\n>> More or less. But what I was really imagining is that we'd get rid of\\n>> the internal states and replace them with new datatypes built to\\n>> purpose. So, for example, for string_agg(text, text) you could make a\\n>> new datatype that is basically a StringInfo. In expanded form, it\\n>> really is a StringInfo. When you flatten it, you just get the string.\\n>> When somebody expands it again, they again have a StringInfo. So the\\n>> RW pointer to the expanded form supports append cheaply.\\n>\\n>\\n> That sounds fine in theory, but where and how do you suppose we determine\\n> which expand function to call? Nothing exists in the catalogs to decide this\\n> currently.\\n\\nI am thinking of transition function returns and accepts the StringInfoData\\ninstead of PolyNumAggState internal data for int8_avg_accum for example.hmm, so wouldn't that mean that the transition function would need to (for each input tuple):1. Parse that StringInfo into tokens.2. Create a new aggregate state object.3. Populate the new aggregate state based on the tokenised StringInfo, this would perhaps require that various *_in() functions are called on each token.4. Add the new tuple to the aggregate state.5. Build a new StringInfo based on the aggregate state modified in 4.?Currently the transition function only does 4, and performs 2 only if it's the first Tuple.Is that what you mean? as I'd say that would slow things down significantly!To get a gauge on how much more CPU work that would be for some aggregates, have a look at how simple int8_avg_accum() is currently when we have HAVE_INT128 defined. For the case of AVG(BIGINT) we just really have: state->sumX += newval; state->N++;The above code is step 4 only. So unless I've misunderstood you, that would need to turn into steps 1-5 above. Step 4 here is probably just a handful of instructions right now, but adding code for steps 1,2,3 and 5 would turn that into hundreds. I've been trying to avoid any overhead by adding the serializeStates flag to make_agg() so that we can maintain the same performance when we're just passing internal states around in the same process. This keeps the conversions between internal state and serialised state to a minimum.The StringInfoData is formed with the members of the PolyNumAggState\\nstructure data. The input given StringInfoData is transformed into\\nPolyNumAggState data and finish the calculation and again form the\\nStringInfoData and return. Similar changes needs to be done for final\\nfunctions input type also. I am not sure whether this approach may have\\nsome impact on performance?-- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Mon, 18 Jan 2016 15:26:30 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Sun, Jan 17, 2016 at 9:26 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> hmm, so wouldn't that mean that the transition function would need to (for\\n> each input tuple):\\n>\\n> 1. Parse that StringInfo into tokens.\\n> 2. Create a new aggregate state object.\\n> 3. Populate the new aggregate state based on the tokenised StringInfo, this\\n> would perhaps require that various *_in() functions are called on each\\n> token.\\n> 4. Add the new tuple to the aggregate state.\\n> 5. Build a new StringInfo based on the aggregate state modified in 4.\\n>\\n> ?\\n\\nI don't really know what you mean by parse the StringInfo into tokens.\\nThe whole point of the expanded-object interface is to be able to keep\\nthings in an expanded internal form so that you *don't* have to\\nrepeatedly construct and deconstruct internal data structures. I\\nworked up an example of this approach using string_agg(), which I\\nattach here. This changes the transition type of string_agg() from\\ninternal to text. The same code would work for bytea_string_agg(),\\nwhich would allow removal of some other code, but this patch doesn't\\ndo that, because the point of this is to elucidate the approach.\\n\\nIn my tests, this seems to be slightly slower than what we're doing\\ntoday; worst of all, it adds a handful of cycles to\\nadvance_transition_function() even when the aggregate is not an\\nexpanded object or, indeed, not even pass-by-reference. Some of this\\nmight be able to be fixed by a little massaging - in particular,\\nDatumIsReadWriteExpandedObject() seems like it could be partly or\\nentirely inlined, and maybe there's some other way to improve the\\ncoding here.\\n\\nGenerally, I think finding a way to pass expanded objects through\\nnodeAgg.c would be a good thing to pursue, if we can make it work.\\nThe immediate impetus for changing things this way would be that we\\nwouldn't need to add a mechanism for serializing and deserializing\\ninternal functions just to pass around partial aggregates. But\\nthere's another advantage, too: right now,\\nadvance_transition_function() does a lot of data copying to move data\\nfrom per-call context to the per-aggregate context. When an expanded\\nobject is in use, this can be skipped.\\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\": \"Sun, 17 Jan 2016 22:44:42 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 18 January 2016 at 16:44, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On Sun, Jan 17, 2016 at 9:26 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n> > hmm, so wouldn't that mean that the transition function would need to\\n> (for\\n> > each input tuple):\\n> >\\n> > 1. Parse that StringInfo into tokens.\\n> > 2. Create a new aggregate state object.\\n> > 3. Populate the new aggregate state based on the tokenised StringInfo,\\n> this\\n> > would perhaps require that various *_in() functions are called on each\\n> > token.\\n> > 4. Add the new tuple to the aggregate state.\\n> > 5. Build a new StringInfo based on the aggregate state modified in 4.\\n> >\\n> > ?\\n>\\n> I don't really know what you mean by parse the StringInfo into tokens.\\n> The whole point of the expanded-object interface is to be able to keep\\n> things in an expanded internal form so that you *don't* have to\\n> repeatedly construct and deconstruct internal data structures.\\n\\n\\nThat was a response to Haribabu proposal, although perhaps I misunderstood\\nthat. However I'm not sure how a PolyNumAggState could be converted to a\\nstring and back again without doing any string parsing.\\n\\nI worked up an example of this approach using string_agg(), which I\\n> attach here. This changes the transition type of string_agg() from\\n> internal to text. The same code would work for bytea_string_agg(),\\n> which would allow removal of some other code, but this patch doesn't\\n>\\ndo that, because the point of this is to elucidate the approach.\\n>\\n>\\nMany thanks for working up that patch. I was clearly missing what you meant\\npreviously. I understand it much better now. Thank you for taking the time\\non that.\\n\\n\\n> In my tests, this seems to be slightly slower than what we're doing\\n> today; worst of all, it adds a handful of cycles to\\n> advance_transition_function() even when the aggregate is not an\\n> expanded object or, indeed, not even pass-by-reference. Some of this\\n> might be able to be fixed by a little massaging - in particular,\\n> DatumIsReadWriteExpandedObject() seems like it could be partly or\\n> entirely inlined, and maybe there's some other way to improve the\\n> coding here.\\n>\\n\\nIt also seems that an expanded object requires a new memory context which\\nmust be malloc()d and free()d. This has added quite an overhead in my\\ntesting. I assume that we must be doing that so that we can ensure that all\\nmemory is properly free()d once we're done with the expanded-object.\\n\\ncreate table ab(a int, b text);\\ninsert into ab select x,'aaaaaaaaaaaaaaaaaaaaaaaaaaa' from\\ngenerate_Series(1,1000000) x(x);\\nset work_mem='1GB';\\nvacuum analyze ab;\\n\\nexplain analyze select a%1000000,length(string_agg(b,',')) from ab group by\\n1;\\n\\nPatched\\n1521.045 ms\\n1515.905 ms\\n1530.920 ms\\n\\nMaster\\n932.457 ms\\n959.285 ms\\n991.021 ms\\n\\nAlthough performance of the patched version is closer to master when we\\nhave less groups, I felt it necessary to show the extreme case. I feel this\\nputs a bit of a dampener on the future of this idea, as I've previously had\\npatches rejected for adding 2-5% on planning time alone, adding over 50%\\nexecution time seems like a dead-end.\\n\\nI've run profiles on this and malloc() and free() are both top of the\\nprofile with the patched version with about 30% CPU time each.\\n\\n\\n> Generally, I think finding a way to pass expanded objects through\\n> nodeAgg.c would be a good thing to pursue, if we can make it work.\\n> The immediate impetus for changing things this way would be that we\\n> wouldn't need to add a mechanism for serializing and deserializing\\n> internal functions just to pass around partial aggregates. But\\n> there's another advantage, too: right now,\\n> advance_transition_function() does a lot of data copying to move data\\n> from per-call context to the per-aggregate context. When an expanded\\n> object is in use, this can be skipped.\\n\\n\\nThe part I quite liked about the serialize/deserialize is that there's no\\nneed to add any overhead at all for serializing and deserializing the\\nstates when the aggregation is done in a single backend process. We'd\\nsimply just have the planner pass the make_agg()'s serializeStates as false\\nwhen we're working within a single backend. This does not appear possible\\nwith your proposed implementation, since it makes changes to each\\ntransition function. It is my understanding that we normally bend over\\nbackwards with new code to try and stop any performance regression. I'm not\\nquite sure the expanded-object idea works well in this regard, but I do\\nagree your approach seems neater. I just don't want to waste my time\\nwriting all the code required to replace all INTERNAL aggregate states when\\nI know the performance regression is going to be unacceptable.\\n\\nI also witnessed another regression with your patch when testing on another\\nmachine. It caused the plan to change to a HashAgg instead of GroupAgg\\ncausing a significant slowdown.\\n\\nUnpatched\\n\\n# explain analyze select a%1000000,length(string_agg(b,',')) from ab group\\nby 1;\\n QUERY PLAN\\n\\n---------------------------------------------------------------------------------------------------------------------------\\n GroupAggregate (cost=119510.84..144510.84 rows=1000000 width=32) (actual\\ntime=538.938..1015.278 rows=1000000 loops=1)\\n Group Key: ((a % 1000000))\\n -> Sort (cost=119510.84..122010.84 rows=1000000 width=32) (actual\\ntime=538.917..594.194 rows=1000000 loops=1)\\n Sort Key: ((a % 1000000))\\n Sort Method: quicksort Memory: 102702kB\\n -> Seq Scan on ab (cost=0.00..19853.00 rows=1000000 width=32)\\n(actual time=0.016..138.964 rows=1000000 loops=1)\\n Planning time: 0.146 ms\\n Execution time: 1047.511 ms\\n\\n\\nPatched\\n# explain analyze select a%1000000,length(string_agg(b,',')) from ab group\\nby 1;\\n QUERY PLAN\\n\\n------------------------------------------------------------------------------------------------------------------------\\n HashAggregate (cost=24853.00..39853.00 rows=1000000 width=32) (actual\\ntime=8072.346..144424.872 rows=1000000 loops=1)\\n Group Key: (a % 1000000)\\n -> Seq Scan on ab (cost=0.00..19853.00 rows=1000000 width=32) (actual\\ntime=0.025..481.332 rows=1000000 loops=1)\\n Planning time: 0.164 ms\\n Execution time: 263288.332 ms\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 18 January 2016 at 16:44, Robert Haas <robertmhaas@gmail.com> wrote:On Sun, Jan 17, 2016 at 9:26 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> hmm, so wouldn't that mean that the transition function would need to (for\\n> each input tuple):\\n>\\n> 1. Parse that StringInfo into tokens.\\n> 2. Create a new aggregate state object.\\n> 3. Populate the new aggregate state based on the tokenised StringInfo, this\\n> would perhaps require that various *_in() functions are called on each\\n> token.\\n> 4. Add the new tuple to the aggregate state.\\n> 5. Build a new StringInfo based on the aggregate state modified in 4.\\n>\\n> ?\\n\\nI don't really know what you mean by parse the StringInfo into tokens.\\nThe whole point of the expanded-object interface is to be able to keep\\nthings in an expanded internal form so that you *don't* have to\\nrepeatedly construct and deconstruct internal data structures. That was a response to Haribabu proposal, although perhaps I misunderstood that. However I'm not sure how a PolyNumAggState could be converted to a string and back again without doing any string parsing. I worked up an example of this approach using string_agg(), which I\\nattach here. This changes the transition type of string_agg() from\\ninternal to text. The same code would work for bytea_string_agg(),\\nwhich would allow removal of some other code, but this patch doesn't\\ndo that, because the point of this is to elucidate the approach.\\nMany thanks for working up that patch. I was clearly missing what you meant previously. I understand it much better now. Thank you for taking the time on that. \\nIn my tests, this seems to be slightly slower than what we're doing\\ntoday; worst of all, it adds a handful of cycles to\\nadvance_transition_function() even when the aggregate is not an\\nexpanded object or, indeed, not even pass-by-reference. Some of this\\nmight be able to be fixed by a little massaging - in particular,\\nDatumIsReadWriteExpandedObject() seems like it could be partly or\\nentirely inlined, and maybe there's some other way to improve the\\ncoding here.It also seems that an expanded object requires a new memory context which must be malloc()d and free()d. This has added quite an overhead in my testing. I assume that we must be doing that so that we can ensure that all memory is properly free()d once we're done with the expanded-object.create table ab(a int, b text);insert into ab select x,'aaaaaaaaaaaaaaaaaaaaaaaaaaa' from generate_Series(1,1000000) x(x);set work_mem='1GB';vacuum analyze ab;explain analyze select a%1000000,length(string_agg(b,',')) from ab group by 1;Patched1521.045 ms1515.905 ms1530.920 msMaster932.457 ms959.285 ms991.021 msAlthough performance of the patched version is closer to master when we have less groups, I felt it necessary to show the extreme case. I feel this puts a bit of a dampener on the future of this idea, as I've previously had patches rejected for adding 2-5% on planning time alone, adding over 50% execution time seems like a dead-end. I've run profiles on this and malloc() and free() are both top of the profile with the patched version with about 30% CPU time each.\\n\\nGenerally, I think finding a way to pass expanded objects through\\nnodeAgg.c would be a good thing to pursue, if we can make it work.\\nThe immediate impetus for changing things this way would be that we\\nwouldn't need to add a mechanism for serializing and deserializing\\ninternal functions just to pass around partial aggregates. But\\nthere's another advantage, too: right now,\\nadvance_transition_function() does a lot of data copying to move data\\nfrom per-call context to the per-aggregate context. When an expanded\\nobject is in use, this can be skipped.The part I quite liked about the serialize/deserialize is that there's no need to add any overhead at all for serializing and deserializing the states when the aggregation is done in a single backend process. We'd simply just have the planner pass the make_agg()'s serializeStates as false when we're working within a single backend. This does not appear possible with your proposed implementation, since it makes changes to each transition function. It is my understanding that we normally bend over backwards with new code to try and stop any performance regression. I'm not quite sure the expanded-object idea works well in this regard, but I do agree your approach seems neater. I just don't want to waste my time writing all the code required to replace all INTERNAL aggregate states when I know the performance regression is going to be unacceptable.I also witnessed another regression with your patch when testing on another machine. It caused the plan to change to a HashAgg instead of GroupAgg causing a significant slowdown.Unpatched# explain analyze select a%1000000,length(string_agg(b,',')) from ab group by 1; QUERY PLAN --------------------------------------------------------------------------------------------------------------------------- GroupAggregate (cost=119510.84..144510.84 rows=1000000 width=32) (actual time=538.938..1015.278 rows=1000000 loops=1) Group Key: ((a % 1000000)) -> Sort (cost=119510.84..122010.84 rows=1000000 width=32) (actual time=538.917..594.194 rows=1000000 loops=1) Sort Key: ((a % 1000000)) Sort Method: quicksort Memory: 102702kB -> Seq Scan on ab (cost=0.00..19853.00 rows=1000000 width=32) (actual time=0.016..138.964 rows=1000000 loops=1) Planning time: 0.146 ms Execution time: 1047.511 msPatched# explain analyze select a%1000000,length(string_agg(b,',')) from ab group by 1; QUERY PLAN ------------------------------------------------------------------------------------------------------------------------ HashAggregate (cost=24853.00..39853.00 rows=1000000 width=32) (actual time=8072.346..144424.872 rows=1000000 loops=1) Group Key: (a % 1000000) -> Seq Scan on ab (cost=0.00..19853.00 rows=1000000 width=32) (actual time=0.025..481.332 rows=1000000 loops=1) Planning time: 0.164 ms Execution time: 263288.332 ms-- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Tue, 19 Jan 2016 00:32:38 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 18, 2016 at 10:32 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> On 18 January 2016 at 16:44, Robert Haas <robertmhaas@gmail.com> wrote:\\n>>\\n>> On Sun, Jan 17, 2016 at 9:26 PM, David Rowley\\n>> <david.rowley@2ndquadrant.com> wrote:\\n>> > hmm, so wouldn't that mean that the transition function would need to\\n>> > (for\\n>> > each input tuple):\\n>> >\\n>> > 1. Parse that StringInfo into tokens.\\n>> > 2. Create a new aggregate state object.\\n>> > 3. Populate the new aggregate state based on the tokenised StringInfo,\\n>> > this\\n>> > would perhaps require that various *_in() functions are called on each\\n>> > token.\\n>> > 4. Add the new tuple to the aggregate state.\\n>> > 5. Build a new StringInfo based on the aggregate state modified in 4.\\n>> >\\n>> > ?\\n>>\\n>> I don't really know what you mean by parse the StringInfo into tokens.\\n>> The whole point of the expanded-object interface is to be able to keep\\n>> things in an expanded internal form so that you *don't* have to\\n>> repeatedly construct and deconstruct internal data structures.\\n>\\n>\\n> That was a response to Haribabu proposal, although perhaps I misunderstood\\n> that. However I'm not sure how a PolyNumAggState could be converted to a\\n> string and back again without doing any string parsing.\\n\\nI just thought like direct mapping of the structure with text pointer.\\nsomething like\\nthe below.\\n\\nresult = PG_ARGISNULL(0) ? NULL : (text *) PG_GETARG_POINTER(0);\\nstate = (PolyNumAggState *)VARDATA(result);\\n\\nTo handle the big-endian or little-endian, we may need some extra changes.\\n\\nInstead of adding 3 new columns to the pg_aggregate catalog table to handle\\nthe internal types, either something like the above to handle the internal types\\nor some other way is better IMO.\\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\": \"Tue, 19 Jan 2016 00:44:38 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 18, 2016 at 6:32 AM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n>> In my tests, this seems to be slightly slower than what we're doing\\n>> today; worst of all, it adds a handful of cycles to\\n>> advance_transition_function() even when the aggregate is not an\\n>> expanded object or, indeed, not even pass-by-reference. Some of this\\n>> might be able to be fixed by a little massaging - in particular,\\n>> DatumIsReadWriteExpandedObject() seems like it could be partly or\\n>> entirely inlined, and maybe there's some other way to improve the\\n>> coding here.\\n>\\n> It also seems that an expanded object requires a new memory context which\\n> must be malloc()d and free()d. This has added quite an overhead in my\\n> testing. I assume that we must be doing that so that we can ensure that all\\n> memory is properly free()d once we're done with the expanded-object.\\n\\nYeah. The API contract for an expanded object took me quite a while\\nto puzzle out, but it seems to be this: if somebody hands you an R/W\\npointer to an expanded object, you're entitled to assume that you can\\n\\\"take over\\\" that object and mutate it however you like. But the\\nobject might be in some other memory context, so you have to move it\\ninto your own memory context. That's implementing by reparenting the\\nobject's context under your context. This is nice because it's O(1) -\\nwhereas copying would be O(n) - but it's sort of aggravating, too. In\\nthis case, the transition function already knows that it wants\\neverything in the per-agg state, but it can't just create everything\\nthere and be done with it, because nodeAgg.c has to content with the\\npossibility that some other piece of code will return an expanded\\nobject that *isn't* parented to the aggcontext. And in fact one of\\nthe regression tests does exactly that, which caused me lots of\\nhead-banging yesterday.\\n\\nMaking it worse, the transition function isn't required to have the\\nsame behavior every time: the one in the regression tests sometimes\\nreturns an expanded-object pointer, and sometimes doesn't, and if\\nnodeAgg.c reparents that pointer to the aggcontext, the next call to\\nthe transition function reparents the pointer BACK to the per-tuple\\ncontext, so you can't even optimize away the repeated set-parent\\ncalls. Uggh.\\n\\n> create table ab(a int, b text);\\n> insert into ab select x,'aaaaaaaaaaaaaaaaaaaaaaaaaaa' from\\n> generate_Series(1,1000000) x(x);\\n> set work_mem='1GB';\\n> vacuum analyze ab;\\n>\\n> explain analyze select a%1000000,length(string_agg(b,',')) from ab group by\\n> 1;\\n>\\n> Patched\\n> 1521.045 ms\\n> 1515.905 ms\\n> 1530.920 ms\\n>\\n> Master\\n> 932.457 ms\\n> 959.285 ms\\n> 991.021 ms\\n>\\n> Although performance of the patched version is closer to master when we have\\n> less groups, I felt it necessary to show the extreme case. I feel this puts\\n> a bit of a dampener on the future of this idea, as I've previously had\\n> patches rejected for adding 2-5% on planning time alone, adding over 50%\\n> execution time seems like a dead-end.\\n\\nThanks for working up this test case. I certainly agree that adding\\n50% execution time is a dead-end, but I wonder if that problem can be\\nfixed somehow. It would be a shame to find out that expanded-objects\\ncan't be effectively used for anything other than the purposes for\\nwhich they were designed, namely speeding up PL/pgsql array\\noperations.\\n\\n> seems neater. I just don't want to waste my time writing all the code\\n> required to replace all INTERNAL aggregate states when I know the\\n> performance regression is going to be unacceptable.\\n\\nNo argument there. If the performance regression isn't fixable, this\\napproach is doomed.\\n\\n> I also witnessed another regression with your patch when testing on another\\n> machine. It caused the plan to change to a HashAgg instead of GroupAgg\\n> causing a significant slowdown.\\n>\\n> Unpatched\\n>\\n> # explain analyze select a%1000000,length(string_agg(b,',')) from ab group\\n> by 1;\\n> QUERY PLAN\\n> ---------------------------------------------------------------------------------------------------------------------------\\n> GroupAggregate (cost=119510.84..144510.84 rows=1000000 width=32) (actual\\n> time=538.938..1015.278 rows=1000000 loops=1)\\n> Group Key: ((a % 1000000))\\n> -> Sort (cost=119510.84..122010.84 rows=1000000 width=32) (actual\\n> time=538.917..594.194 rows=1000000 loops=1)\\n> Sort Key: ((a % 1000000))\\n> Sort Method: quicksort Memory: 102702kB\\n> -> Seq Scan on ab (cost=0.00..19853.00 rows=1000000 width=32)\\n> (actual time=0.016..138.964 rows=1000000 loops=1)\\n> Planning time: 0.146 ms\\n> Execution time: 1047.511 ms\\n>\\n>\\n> Patched\\n> # explain analyze select a%1000000,length(string_agg(b,',')) from ab group\\n> by 1;\\n> QUERY PLAN\\n> ------------------------------------------------------------------------------------------------------------------------\\n> HashAggregate (cost=24853.00..39853.00 rows=1000000 width=32) (actual\\n> time=8072.346..144424.872 rows=1000000 loops=1)\\n> Group Key: (a % 1000000)\\n> -> Seq Scan on ab (cost=0.00..19853.00 rows=1000000 width=32) (actual\\n> time=0.025..481.332 rows=1000000 loops=1)\\n> Planning time: 0.164 ms\\n> Execution time: 263288.332 ms\\n\\nWell, that's pretty odd. I guess the plan change must be a result of\\nswitching the transition type from internal to text, although I'm not\\nimmediately certain why that would make a difference.\\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, 18 Jan 2016 10:58:35 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \">\\n> > # explain analyze select a%1000000,length(string_agg(b,',')) from ab\\n> group\\n> > by 1;\\n> > QUERY PLAN\\n> >\\n> ---------------------------------------------------------------------------------------------------------------------------\\n> > GroupAggregate (cost=119510.84..144510.84 rows=1000000 width=32)\\n> (actual\\n> > time=538.938..1015.278 rows=1000000 loops=1)\\n> > Group Key: ((a % 1000000))\\n> > -> Sort (cost=119510.84..122010.84 rows=1000000 width=32) (actual\\n> > time=538.917..594.194 rows=1000000 loops=1)\\n> > Sort Key: ((a % 1000000))\\n> > Sort Method: quicksort Memory: 102702kB\\n> > -> Seq Scan on ab (cost=0.00..19853.00 rows=1000000 width=32)\\n> > (actual time=0.016..138.964 rows=1000000 loops=1)\\n> > Planning time: 0.146 ms\\n> > Execution time: 1047.511 ms\\n> >\\n> >\\n> > Patched\\n> > # explain analyze select a%1000000,length(string_agg(b,',')) from ab\\n> group\\n> > by 1;\\n> > QUERY PLAN\\n> >\\n> ------------------------------------------------------------------------------------------------------------------------\\n> > HashAggregate (cost=24853.00..39853.00 rows=1000000 width=32) (actual\\n> > time=8072.346..144424.872 rows=1000000 loops=1)\\n> > Group Key: (a % 1000000)\\n> > -> Seq Scan on ab (cost=0.00..19853.00 rows=1000000 width=32)\\n> (actual\\n> > time=0.025..481.332 rows=1000000 loops=1)\\n> > Planning time: 0.164 ms\\n> > Execution time: 263288.332 ms\\n>\\n> Well, that's pretty odd. I guess the plan change must be a result of\\n> switching the transition type from internal to text, although I'm not\\n> immediately certain why that would make a difference.\\n>\\n\\nIt is strange, why hashaggregate is too slow?\\n\\nPavel\\n\\n\\n\\n>\\n> --\\n> Robert Haas\\n> EnterpriseDB: http://www.enterprisedb.com\\n> The Enterprise PostgreSQL Company\\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> # explain analyze select a%1000000,length(string_agg(b,',')) from ab group\\n> by 1;\\n> QUERY PLAN\\n> ---------------------------------------------------------------------------------------------------------------------------\\n> GroupAggregate (cost=119510.84..144510.84 rows=1000000 width=32) (actual\\n> time=538.938..1015.278 rows=1000000 loops=1)\\n> Group Key: ((a % 1000000))\\n> -> Sort (cost=119510.84..122010.84 rows=1000000 width=32) (actual\\n> time=538.917..594.194 rows=1000000 loops=1)\\n> Sort Key: ((a % 1000000))\\n> Sort Method: quicksort Memory: 102702kB\\n> -> Seq Scan on ab (cost=0.00..19853.00 rows=1000000 width=32)\\n> (actual time=0.016..138.964 rows=1000000 loops=1)\\n> Planning time: 0.146 ms\\n> Execution time: 1047.511 ms\\n>\\n>\\n> Patched\\n> # explain analyze select a%1000000,length(string_agg(b,',')) from ab group\\n> by 1;\\n> QUERY PLAN\\n> ------------------------------------------------------------------------------------------------------------------------\\n> HashAggregate (cost=24853.00..39853.00 rows=1000000 width=32) (actual\\n> time=8072.346..144424.872 rows=1000000 loops=1)\\n> Group Key: (a % 1000000)\\n> -> Seq Scan on ab (cost=0.00..19853.00 rows=1000000 width=32) (actual\\n> time=0.025..481.332 rows=1000000 loops=1)\\n> Planning time: 0.164 ms\\n> Execution time: 263288.332 ms\\n\\nWell, that's pretty odd. I guess the plan change must be a result of\\nswitching the transition type from internal to text, although I'm not\\nimmediately certain why that would make a difference.It is strange, why hashaggregate is too slow?Pavel \\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, 18 Jan 2016 18:03:03 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> Yeah. The API contract for an expanded object took me quite a while\\n> to puzzle out, but it seems to be this: if somebody hands you an R/W\\n> pointer to an expanded object, you're entitled to assume that you can\\n> \\\"take over\\\" that object and mutate it however you like. But the\\n> object might be in some other memory context, so you have to move it\\n> into your own memory context.\\n\\nOnly if you intend to keep it --- for example, a function that is mutating\\nand returning an object isn't required to move it somewhere else, if the\\ninput is R/W, and I think it generally shouldn't.\\n\\nIn the context here, I'd think it is the responsibility of nodeAgg.c\\nnot individual datatype functions to make sure that expanded objects\\nlive where it wants them to.\\n\\n> Well, that's pretty odd. I guess the plan change must be a result of\\n> switching the transition type from internal to text, although I'm not\\n> immediately certain why that would make a difference.\\n\\nI'm pretty sure there's something in the planner that special-cases\\nits estimate of space consumed by hashtable entries when the data type\\nis \\\"internal\\\". You'd be wanting to fool with that estimate anyway\\nfor something like this ...\\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, 18 Jan 2016 12:09:07 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 18, 2016 at 12:09 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Robert Haas <robertmhaas@gmail.com> writes:\\n>> Yeah. The API contract for an expanded object took me quite a while\\n>> to puzzle out, but it seems to be this: if somebody hands you an R/W\\n>> pointer to an expanded object, you're entitled to assume that you can\\n>> \\\"take over\\\" that object and mutate it however you like. But the\\n>> object might be in some other memory context, so you have to move it\\n>> into your own memory context.\\n>\\n> Only if you intend to keep it --- for example, a function that is mutating\\n> and returning an object isn't required to move it somewhere else, if the\\n> input is R/W, and I think it generally shouldn't.\\n>\\n> In the context here, I'd think it is the responsibility of nodeAgg.c\\n> not individual datatype functions to make sure that expanded objects\\n> live where it wants them to.\\n\\nThat's how I did it in my prototype, but the problem with that is that\\nspinning up a memory context for every group sucks when there are many\\ngroups with only a small number of elements each - hence the 50%\\nregression that David Rowley observed. If we're going to use expanded\\nobjects here, which seems like a good idea in principle, that's going\\nto have to be fixed somehow. We're flogging the heck out of malloc by\\nrepeatedly creating a context, doing one or two allocations in it, and\\nthen destroying the context.\\n\\nI think that, in general, the memory context machinery wasn't really\\ndesigned to manage lots of small contexts. The overhead of spinning\\nup a new context for just a few allocations is substantial. That\\nmatters in some other situations too, I think - I've commonly seen\\nAllocSetContextCreate taking several percent of runtime in profiles.\\nBut here it's much exacerbated. I'm not sure whether it's better to\\nattack that problem at the root and try to make AllocSetContextCreate\\ncheaper, or whether we should try to figure out some change to the\\nexpanded-object machinery to address the issue.\\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, 18 Jan 2016 12:34:13 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 19 January 2016 at 02:44, Haribabu Kommi <kommi.haribabu@gmail.com>\\nwrote:\\n\\n> On Mon, Jan 18, 2016 at 10:32 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n>\\n> I just thought like direct mapping of the structure with text pointer.\\n> something like\\n> the below.\\n>\\n> result = PG_ARGISNULL(0) ? NULL : (text *) PG_GETARG_POINTER(0);\\n> state = (PolyNumAggState *)VARDATA(result);\\n>\\n> To handle the big-endian or little-endian, we may need some extra changes.\\n>\\n> Instead of adding 3 new columns to the pg_aggregate catalog table to handle\\n> the internal types, either something like the above to handle the internal\\n> types\\n> or some other way is better IMO.\\n\\n\\nThe problem with that is that most of these internal structs for the\\naggregate states have pointers to other memory, so even if we laid those\\nbytes down into a bytea or something, then doing so is not going to\\ndereference the pointers to the other memory, and when we dereference those\\npointers in the other process, we'll have problems as these addresses\\nbelong to the other process.\\n\\nFor example PolyNumAggState is defined as:\\n\\ntypedef NumericAggState PolyNumAggState;\\n\\nand NumericAggState has:\\n\\nNumericVar sumX; /* sum of processed numbers */\\nNumericVar sumX2; /* sum of squares of processed numbers */\\n\\nAnd NumericVar has:\\n\\nNumericDigit *buf; /* start of palloc'd space for digits[] */\\nNumericDigit *digits; /* base-NBASE digits */\\n\\nBoth of these point to other memory which won't be in the varlena type.\\n\\nSerialization is the process of collecting all of these pointers up in to\\nsome consecutive bytes.\\n\\nOf course, that's not to say that there's never Aggregate State structs\\nwhich don't have any pointers, I've not checked, but in these cases we\\ncould (perhaps) just make the serialize and deserialize function a simple\\nmemcpy() into a bytea array, although in reality, as you mentioned, we'd\\nlikely want to agree on some format that's cross platform for different\\nbyte orders, as we'll probably, one day, want to forward these values over\\nto some other server to finish off the aggregation.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 19 January 2016 at 02:44, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:On Mon, Jan 18, 2016 at 10:32 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\nI just thought like direct mapping of the structure with text pointer.\\nsomething like\\nthe below.\\n\\nresult = PG_ARGISNULL(0) ? NULL : (text *) PG_GETARG_POINTER(0);\\nstate = (PolyNumAggState *)VARDATA(result);\\n\\nTo handle the big-endian or little-endian, we may need some extra changes.\\n\\nInstead of adding 3 new columns to the pg_aggregate catalog table to handle\\nthe internal types, either something like the above to handle the internal types\\nor some other way is better IMO.The problem with that is that most of these internal structs for the aggregate states have pointers to other memory, so even if we laid those bytes down into a bytea or something, then doing so is not going to dereference the pointers to the other memory, and when we dereference those pointers in the other process, we'll have problems as these addresses belong to the other process. For example PolyNumAggState is defined as:typedef NumericAggState PolyNumAggState;and NumericAggState has: NumericVar sumX; /* sum of processed numbers */ NumericVar sumX2; /* sum of squares of processed numbers */And NumericVar has: NumericDigit *buf; /* start of palloc'd space for digits[] */ NumericDigit *digits; /* base-NBASE digits */Both of these point to other memory which won't be in the varlena type.Serialization is the process of collecting all of these pointers up in to some consecutive bytes.Of course, that's not to say that there's never Aggregate State structs which don't have any pointers, I've not checked, but in these cases we could (perhaps) just make the serialize and deserialize function a simple memcpy() into a bytea array, although in reality, as you mentioned, we'd likely want to agree on some format that's cross platform for different byte orders, as we'll probably, one day, want to forward these values over to some other server to finish off the aggregation.-- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Tue, 19 Jan 2016 15:48:59 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 19 January 2016 at 06:03, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n\\n>\\n>\\n> >\\n>> > # explain analyze select a%1000000,length(string_agg(b,',')) from ab\\n>> group\\n>> > by 1;\\n>> > QUERY PLAN\\n>> >\\n>> ---------------------------------------------------------------------------------------------------------------------------\\n>> > GroupAggregate (cost=119510.84..144510.84 rows=1000000 width=32)\\n>> (actual\\n>> > time=538.938..1015.278 rows=1000000 loops=1)\\n>> > Group Key: ((a % 1000000))\\n>> > -> Sort (cost=119510.84..122010.84 rows=1000000 width=32) (actual\\n>> > time=538.917..594.194 rows=1000000 loops=1)\\n>> > Sort Key: ((a % 1000000))\\n>> > Sort Method: quicksort Memory: 102702kB\\n>> > -> Seq Scan on ab (cost=0.00..19853.00 rows=1000000 width=32)\\n>> > (actual time=0.016..138.964 rows=1000000 loops=1)\\n>> > Planning time: 0.146 ms\\n>> > Execution time: 1047.511 ms\\n>> >\\n>> >\\n>> > Patched\\n>> > # explain analyze select a%1000000,length(string_agg(b,',')) from ab\\n>> group\\n>> > by 1;\\n>> > QUERY PLAN\\n>> >\\n>> ------------------------------------------------------------------------------------------------------------------------\\n>> > HashAggregate (cost=24853.00..39853.00 rows=1000000 width=32) (actual\\n>> > time=8072.346..144424.872 rows=1000000 loops=1)\\n>> > Group Key: (a % 1000000)\\n>> > -> Seq Scan on ab (cost=0.00..19853.00 rows=1000000 width=32)\\n>> (actual\\n>> > time=0.025..481.332 rows=1000000 loops=1)\\n>> > Planning time: 0.164 ms\\n>> > Execution time: 263288.332 ms\\n>>\\n>> Well, that's pretty odd. I guess the plan change must be a result of\\n>> switching the transition type from internal to text, although I'm not\\n>> immediately certain why that would make a difference.\\n>>\\n>\\n> It is strange, why hashaggregate is too slow?\\n>\\n\\nGood question. I looked at this and found my VM was swapping like crazy.\\nUpon investigation it appears that's because, since the patch creates a\\nmemory context per aggregated group, and in this case I've got 1 million of\\nthem, it means we create 1 million context, which are\\nALLOCSET_SMALL_INITSIZE (1KB) in size, which means about 1GB of memory,\\nwhich is more than my VM likes.\\n\\nset work_mem = '130MB' does coax the planner into a GroupAggregate plan,\\nwhich is faster, but due to the the hash agg executor code not giving any\\nregard to work_mem. If I set work_mem to 140MB (which is more realistic for\\nthis VM), it does cause the same swapping problems to occur. Probably\\nsetting aggtransspace for this aggregate to 1024 would help the costing\\nproblem, but it would also cause hashagg to be a less chosen option during\\nplanning.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 19 January 2016 at 06:03, Pavel Stehule <pavel.stehule@gmail.com> wrote:\\n>\\n> # explain analyze select a%1000000,length(string_agg(b,',')) from ab group\\n> by 1;\\n> QUERY PLAN\\n> ---------------------------------------------------------------------------------------------------------------------------\\n> GroupAggregate (cost=119510.84..144510.84 rows=1000000 width=32) (actual\\n> time=538.938..1015.278 rows=1000000 loops=1)\\n> Group Key: ((a % 1000000))\\n> -> Sort (cost=119510.84..122010.84 rows=1000000 width=32) (actual\\n> time=538.917..594.194 rows=1000000 loops=1)\\n> Sort Key: ((a % 1000000))\\n> Sort Method: quicksort Memory: 102702kB\\n> -> Seq Scan on ab (cost=0.00..19853.00 rows=1000000 width=32)\\n> (actual time=0.016..138.964 rows=1000000 loops=1)\\n> Planning time: 0.146 ms\\n> Execution time: 1047.511 ms\\n>\\n>\\n> Patched\\n> # explain analyze select a%1000000,length(string_agg(b,',')) from ab group\\n> by 1;\\n> QUERY PLAN\\n> ------------------------------------------------------------------------------------------------------------------------\\n> HashAggregate (cost=24853.00..39853.00 rows=1000000 width=32) (actual\\n> time=8072.346..144424.872 rows=1000000 loops=1)\\n> Group Key: (a % 1000000)\\n> -> Seq Scan on ab (cost=0.00..19853.00 rows=1000000 width=32) (actual\\n> time=0.025..481.332 rows=1000000 loops=1)\\n> Planning time: 0.164 ms\\n> Execution time: 263288.332 ms\\n\\nWell, that's pretty odd. I guess the plan change must be a result of\\nswitching the transition type from internal to text, although I'm not\\nimmediately certain why that would make a difference.It is strange, why hashaggregate is too slow?Good question. I looked at this and found my VM was swapping like crazy. Upon investigation it appears that's because, since the patch creates a memory context per aggregated group, and in this case I've got 1 million of them, it means we create 1 million context, which are ALLOCSET_SMALL_INITSIZE (1KB) in size, which means about 1GB of memory, which is more than my VM likes. set work_mem = '130MB' does coax the planner into a GroupAggregate plan, which is faster, but due to the the hash agg executor code not giving any regard to work_mem. If I set work_mem to 140MB (which is more realistic for this VM), it does cause the same swapping problems to occur. Probably setting aggtransspace for this aggregate to 1024 would help the costing problem, but it would also cause hashagg to be a less chosen option during planning. -- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Tue, 19 Jan 2016 17:00:21 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 18, 2016 at 12:34 PM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Mon, Jan 18, 2016 at 12:09 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> Robert Haas <robertmhaas@gmail.com> writes:\\n>>> Yeah. The API contract for an expanded object took me quite a while\\n>>> to puzzle out, but it seems to be this: if somebody hands you an R/W\\n>>> pointer to an expanded object, you're entitled to assume that you can\\n>>> \\\"take over\\\" that object and mutate it however you like. But the\\n>>> object might be in some other memory context, so you have to move it\\n>>> into your own memory context.\\n>>\\n>> Only if you intend to keep it --- for example, a function that is mutating\\n>> and returning an object isn't required to move it somewhere else, if the\\n>> input is R/W, and I think it generally shouldn't.\\n>>\\n>> In the context here, I'd think it is the responsibility of nodeAgg.c\\n>> not individual datatype functions to make sure that expanded objects\\n>> live where it wants them to.\\n>\\n> That's how I did it in my prototype, but the problem with that is that\\n> spinning up a memory context for every group sucks when there are many\\n> groups with only a small number of elements each - hence the 50%\\n> regression that David Rowley observed. If we're going to use expanded\\n> objects here, which seems like a good idea in principle, that's going\\n> to have to be fixed somehow. We're flogging the heck out of malloc by\\n> repeatedly creating a context, doing one or two allocations in it, and\\n> then destroying the context.\\n>\\n> I think that, in general, the memory context machinery wasn't really\\n> designed to manage lots of small contexts. The overhead of spinning\\n> up a new context for just a few allocations is substantial. That\\n> matters in some other situations too, I think - I've commonly seen\\n> AllocSetContextCreate taking several percent of runtime in profiles.\\n> But here it's much exacerbated. I'm not sure whether it's better to\\n> attack that problem at the root and try to make AllocSetContextCreate\\n> cheaper, or whether we should try to figure out some change to the\\n> expanded-object machinery to address the issue.\\n\\nHere is a patch that helps a good deal. I changed things so that when\\nwe create a context, we always allocate at least 1kB. If that's more\\nthan we need for the node itself and the name, then we use the rest of\\nthe space as a sort of keeper block. I think there's more that can be\\ndone to improve this, but I'm wimping out for now because it's late\\nhere.\\n\\nI suspect something like this is a good idea even if we don't end up\\nusing it for aggregate transition functions.\\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\": \"Mon, 18 Jan 2016 23:14:32 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> Here is a patch that helps a good deal. I changed things so that when\\n> we create a context, we always allocate at least 1kB.\\n\\nThat's going to kill performance in some other cases; subtransactions\\nin particular rely on the subtransaction's TransactionContext not causing\\nany actual malloc unless something gets put into the TransactionContext.\\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, 18 Jan 2016 23:24:51 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 01/19/2016 05:00 AM, David Rowley wrote:\\n> On 19 January 2016 at 06:03, Pavel Stehule <pavel.stehule@gmail.com\\n> <mailto:pavel.stehule@gmail.com>> wrote:\\n>\\n...\\n>\\n> It is strange, why hashaggregate is too slow?\\n>\\n>\\n> Good question. I looked at this and found my VM was swapping like crazy.\\n> Upon investigation it appears that's because, since the patch creates a\\n> memory context per aggregated group, and in this case I've got 1 million\\n> of them, it means we create 1 million context, which are\\n> ALLOCSET_SMALL_INITSIZE (1KB) in size, which means about 1GB of memory,\\n> which is more than my VM likes.\\n\\nReally? Where do we create the memory context? IIRC string_agg uses the \\naggcontext directly, and indeed that's what I see in string_agg_transfn \\nand makeStringAggState.\\n\\nPerhaps you mean that initStringInfo() allocates 1kB buffers by default?\\n\\n>\\n> set work_mem = '130MB' does coax the planner into a GroupAggregate plan,\\n> which is faster, but due to the the hash agg executor code not giving\\n> any regard to work_mem. If I set work_mem to 140MB (which is more\\n> realistic for this VM), it does cause the same swapping problems to\\n> occur. Probably setting aggtransspace for this aggregate to 1024 would\\n> help the costing problem, but it would also cause hashagg to be a less\\n> chosen option during planning.\\n\\nI'm not quite sure I understand - the current code ends up using 8192 \\nfor the transition space (per count_agg_clauses_walker). Are you \\nsuggesting lowering the value, despite the danger of OOM issues?\\n\\nregards\\n\\n--\\nTomas Vondra 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, 19 Jan 2016 06:04:36 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 01/19/2016 06:04 AM, Tomas Vondra wrote:\\n> Hi,\\n>\\n> On 01/19/2016 05:00 AM, David Rowley wrote:\\n>> Good question. I looked at this and found my VM was swapping like crazy.\\n>> Upon investigation it appears that's because, since the patch creates a\\n>> memory context per aggregated group, and in this case I've got 1 million\\n>> of them, it means we create 1 million context, which are\\n>> ALLOCSET_SMALL_INITSIZE (1KB) in size, which means about 1GB of memory,\\n>> which is more than my VM likes.\\n>\\n> Really? Where do we create the memory context? IIRC string_agg uses the\\n> aggcontext directly, and indeed that's what I see in string_agg_transfn\\n> and makeStringAggState.\\n>\\n> Perhaps you mean that initStringInfo() allocates 1kB buffers by default?\\n\\n...\\n\\n> I'm not quite sure I understand - the current code ends up using 8192\\n> for the transition space (per count_agg_clauses_walker). Are you\\n> suggesting lowering the value, despite the danger of OOM issues?\\n\\nMeh, right after sending the message I realized that perhaps this \\nrelates to Robert's patch and that maybe it changes this. And indeed, it \\nchanges the type from internal to text, and thus the special case in \\ncount_agg_clauses_walker no longer applies.\\n\\nregards\\n\\n--\\nTomas Vondra 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, 19 Jan 2016 06:11:40 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 19 January 2016 at 18:04, Tomas Vondra <tomas.vondra@2ndquadrant.com>\\nwrote:\\n\\n> Hi,\\n>\\n> On 01/19/2016 05:00 AM, David Rowley wrote:\\n>\\n>> On 19 January 2016 at 06:03, Pavel Stehule <pavel.stehule@gmail.com\\n>> <mailto:pavel.stehule@gmail.com>> wrote:\\n>>\\n>> ...\\n>\\n>>\\n>> It is strange, why hashaggregate is too slow?\\n>>\\n>>\\n>> Good question. I looked at this and found my VM was swapping like crazy.\\n>> Upon investigation it appears that's because, since the patch creates a\\n>> memory context per aggregated group, and in this case I've got 1 million\\n>> of them, it means we create 1 million context, which are\\n>> ALLOCSET_SMALL_INITSIZE (1KB) in size, which means about 1GB of memory,\\n>> which is more than my VM likes.\\n>>\\n>\\n> Really? Where do we create the memory context? IIRC string_agg uses the\\n> aggcontext directly, and indeed that's what I see in string_agg_transfn and\\n> makeStringAggState.\\n>\\n>\\nYeah, all this is talk is relating to Robert's expandedstring-v1.patch\\nwhich changes string_agg to use text and expanded-objects. This also means\\nthat a memory context is created per group, which is rather a big overhead.\\n\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 19 January 2016 at 18:04, Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote:Hi,\\n\\nOn 01/19/2016 05:00 AM, David Rowley wrote:\\n\\nOn 19 January 2016 at 06:03, Pavel Stehule <pavel.stehule@gmail.com\\n<mailto:pavel.stehule@gmail.com>> wrote:\\n\\n\\n...\\n\\n\\n It is strange, why hashaggregate is too slow?\\n\\n\\nGood question. I looked at this and found my VM was swapping like crazy.\\nUpon investigation it appears that's because, since the patch creates a\\nmemory context per aggregated group, and in this case I've got 1 million\\nof them, it means we create 1 million context, which are\\nALLOCSET_SMALL_INITSIZE (1KB) in size, which means about 1GB of memory,\\nwhich is more than my VM likes.\\n\\n\\nReally? Where do we create the memory context? IIRC string_agg uses the aggcontext directly, and indeed that's what I see in string_agg_transfn and makeStringAggState.\\nYeah, all this is talk is relating to Robert's expandedstring-v1.patch which changes string_agg to use text and expanded-objects. This also means that a memory context is created per group, which is rather a big overhead. -- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Tue, 19 Jan 2016 18:16:26 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"\\nOn 01/19/2016 05:14 AM, Robert Haas wrote:\\n> On Mon, Jan 18, 2016 at 12:34 PM, Robert Haas <robertmhaas@gmail.com> wrote:\\n>> On Mon, Jan 18, 2016 at 12:09 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>>> Robert Haas <robertmhaas@gmail.com> writes:\\n>>>> Yeah. The API contract for an expanded object took me quite a while\\n>>>> to puzzle out, but it seems to be this: if somebody hands you an R/W\\n>>>> pointer to an expanded object, you're entitled to assume that you can\\n>>>> \\\"take over\\\" that object and mutate it however you like. But the\\n>>>> object might be in some other memory context, so you have to move it\\n>>>> into your own memory context.\\n>>>\\n>>> Only if you intend to keep it --- for example, a function that is\\n>>> mutating and returning an object isn't required to move it\\n>>> somewhere else, if the input is R/W, and I think it generally\\n>>> shouldn't.\\n>>>\\n>>> In the context here, I'd think it is the responsibility of\\n>>> nodeAgg.c not individual datatype functions to make sure that\\n>>> expanded objects live where it wants them to.\\n>>\\n>> That's how I did it in my prototype, but the problem with that is that\\n>> spinning up a memory context for every group sucks when there are many\\n>> groups with only a small number of elements each - hence the 50%\\n>> regression that David Rowley observed. If we're going to use expanded\\n>> objects here, which seems like a good idea in principle, that's going\\n>> to have to be fixed somehow. We're flogging the heck out of malloc by\\n>> repeatedly creating a context, doing one or two allocations in it, and\\n>> then destroying the context.\\n>>\\n>> I think that, in general, the memory context machinery wasn't really\\n>> designed to manage lots of small contexts. The overhead of spinning\\n>> up a new context for just a few allocations is substantial. That\\n>> matters in some other situations too, I think - I've commonly seen\\n>> AllocSetContextCreate taking several percent of runtime in profiles.\\n>> But here it's much exacerbated. I'm not sure whether it's better to\\n>> attack that problem at the root and try to make AllocSetContextCreate\\n>> cheaper, or whether we should try to figure out some change to the\\n>> expanded-object machinery to address the issue.\\n>\\n> Here is a patch that helps a good deal. I changed things so that when\\n> we create a context, we always allocate at least 1kB. If that's more\\n> than we need for the node itself and the name, then we use the rest of\\n> the space as a sort of keeper block. I think there's more that can be\\n> done to improve this, but I'm wimping out for now because it's late\\n> here.\\n>\\n> I suspect something like this is a good idea even if we don't end up\\n> using it for aggregate transition functions.\\n\\nI dare to claim that if expanded objects require a separate memory \\ncontext per aggregate state (I don't see why would be the case), it's a \\ndead end. Not so long ago we've fixed exactly this issue in array_agg(), \\nwhich addressed a bunch of reported OOM issues and improved array_agg() \\nperformance quite a bit. It'd be rather crazy introducing the same \\nproblem to all aggregate functions.\\n\\nregards\\n\\n--\\nTomas Vondra 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, 19 Jan 2016 06:22:08 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 19 January 2016 at 17:14, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On Mon, Jan 18, 2016 at 12:34 PM, Robert Haas <robertmhaas@gmail.com>\\n> wrote:\\n> > On Mon, Jan 18, 2016 at 12:09 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> >> Robert Haas <robertmhaas@gmail.com> writes:\\n> >>> Yeah. The API contract for an expanded object took me quite a while\\n> >>> to puzzle out, but it seems to be this: if somebody hands you an R/W\\n> >>> pointer to an expanded object, you're entitled to assume that you can\\n> >>> \\\"take over\\\" that object and mutate it however you like. But the\\n> >>> object might be in some other memory context, so you have to move it\\n> >>> into your own memory context.\\n> >>\\n> >> Only if you intend to keep it --- for example, a function that is\\n> mutating\\n> >> and returning an object isn't required to move it somewhere else, if the\\n> >> input is R/W, and I think it generally shouldn't.\\n> >>\\n> >> In the context here, I'd think it is the responsibility of nodeAgg.c\\n> >> not individual datatype functions to make sure that expanded objects\\n> >> live where it wants them to.\\n> >\\n> > That's how I did it in my prototype, but the problem with that is that\\n> > spinning up a memory context for every group sucks when there are many\\n> > groups with only a small number of elements each - hence the 50%\\n> > regression that David Rowley observed. If we're going to use expanded\\n> > objects here, which seems like a good idea in principle, that's going\\n> > to have to be fixed somehow. We're flogging the heck out of malloc by\\n> > repeatedly creating a context, doing one or two allocations in it, and\\n> > then destroying the context.\\n> >\\n> > I think that, in general, the memory context machinery wasn't really\\n> > designed to manage lots of small contexts. The overhead of spinning\\n> > up a new context for just a few allocations is substantial. That\\n> > matters in some other situations too, I think - I've commonly seen\\n> > AllocSetContextCreate taking several percent of runtime in profiles.\\n> > But here it's much exacerbated. I'm not sure whether it's better to\\n> > attack that problem at the root and try to make AllocSetContextCreate\\n> > cheaper, or whether we should try to figure out some change to the\\n> > expanded-object machinery to address the issue.\\n>\\n> Here is a patch that helps a good deal. I changed things so that when\\n> we create a context, we always allocate at least 1kB. If that's more\\n> than we need for the node itself and the name, then we use the rest of\\n> the space as a sort of keeper block. I think there's more that can be\\n> done to improve this, but I'm wimping out for now because it's late\\n> here.\\n>\\n> I suspect something like this is a good idea even if we don't end up\\n> using it for aggregate transition functions.\\n\\n\\nThanks for writing this up, but I think the key issue is not addressed,\\nwhich is the memory context per aggregate group just being a performance\\nkiller. I ran the test query again with the attached patch and the hashagg\\nquery still took 300 seconds on my VM with 4GB ram.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 19 January 2016 at 17:14, Robert Haas <robertmhaas@gmail.com> wrote:On Mon, Jan 18, 2016 at 12:34 PM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Mon, Jan 18, 2016 at 12:09 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> Robert Haas <robertmhaas@gmail.com> writes:\\n>>> Yeah. The API contract for an expanded object took me quite a while\\n>>> to puzzle out, but it seems to be this: if somebody hands you an R/W\\n>>> pointer to an expanded object, you're entitled to assume that you can\\n>>> \\\"take over\\\" that object and mutate it however you like. But the\\n>>> object might be in some other memory context, so you have to move it\\n>>> into your own memory context.\\n>>\\n>> Only if you intend to keep it --- for example, a function that is mutating\\n>> and returning an object isn't required to move it somewhere else, if the\\n>> input is R/W, and I think it generally shouldn't.\\n>>\\n>> In the context here, I'd think it is the responsibility of nodeAgg.c\\n>> not individual datatype functions to make sure that expanded objects\\n>> live where it wants them to.\\n>\\n> That's how I did it in my prototype, but the problem with that is that\\n> spinning up a memory context for every group sucks when there are many\\n> groups with only a small number of elements each - hence the 50%\\n> regression that David Rowley observed. If we're going to use expanded\\n> objects here, which seems like a good idea in principle, that's going\\n> to have to be fixed somehow. We're flogging the heck out of malloc by\\n> repeatedly creating a context, doing one or two allocations in it, and\\n> then destroying the context.\\n>\\n> I think that, in general, the memory context machinery wasn't really\\n> designed to manage lots of small contexts. The overhead of spinning\\n> up a new context for just a few allocations is substantial. That\\n> matters in some other situations too, I think - I've commonly seen\\n> AllocSetContextCreate taking several percent of runtime in profiles.\\n> But here it's much exacerbated. I'm not sure whether it's better to\\n> attack that problem at the root and try to make AllocSetContextCreate\\n> cheaper, or whether we should try to figure out some change to the\\n> expanded-object machinery to address the issue.\\n\\nHere is a patch that helps a good deal. I changed things so that when\\nwe create a context, we always allocate at least 1kB. If that's more\\nthan we need for the node itself and the name, then we use the rest of\\nthe space as a sort of keeper block. I think there's more that can be\\ndone to improve this, but I'm wimping out for now because it's late\\nhere.\\n\\nI suspect something like this is a good idea even if we don't end up\\nusing it for aggregate transition functions.Thanks for writing this up, but I think the key issue is not addressed, which is the memory context per aggregate group just being a performance killer. I ran the test query again with the attached patch and the hashagg query still took 300 seconds on my VM with 4GB ram. -- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Tue, 19 Jan 2016 18:32:36 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 18, 2016 at 11:24 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Robert Haas <robertmhaas@gmail.com> writes:\\n>> Here is a patch that helps a good deal. I changed things so that when\\n>> we create a context, we always allocate at least 1kB.\\n>\\n> That's going to kill performance in some other cases; subtransactions\\n> in particular rely on the subtransaction's TransactionContext not causing\\n> any actual malloc unless something gets put into the TransactionContext.\\n\\nSorry, I guess I wasn't clear: it increases the allocation for the\\ncontext node itself to 1kB and uses the extra space to store a few\\nallocations.\\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, 19 Jan 2016 07:51:55 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 19, 2016 at 12:22 AM, Tomas Vondra\\n<tomas.vondra@2ndquadrant.com> wrote:\\n> I dare to claim that if expanded objects require a separate memory context\\n> per aggregate state (I don't see why would be the case), it's a dead end.\\n> Not so long ago we've fixed exactly this issue in array_agg(), which\\n> addressed a bunch of reported OOM issues and improved array_agg()\\n> performance quite a bit. It'd be rather crazy introducing the same problem\\n> to all aggregate functions.\\n\\nExpanded objects require a separate memory context per Datum. I think\\nI know how to rejigger things so that the overhead of that is\\nminimized as much as possible, but it's still 200 bytes for the\\nAllocSetContext + ~16 bytes for the name + 32 bytes for an AllocBlock\\n+ 48 bytes for an ExpandedObjectHeader. That's about 300 bytes of\\noverhead per expanded datum, which means on this test case the\\ntheoretical minimum memory burn for this approach is about 300 MB (and\\nin practice somewhat more). Ouch.\\n\\nThe upshot seems to be that expanded objects aren't going to actually\\nbe useful in any situation where you might have very many of them,\\nbecause the memory overhead is just awful. That's rather\\ndisappointing. Even if you could get rid of the requirement for a\\nmemory context per Datum - and it seems like you could if you added a\\nmethod defined to reparent the object to a designated context, which\\nlooks like a totally reasonable innovation - somebody's probably going\\nto say, not without some justification, that 48 bytes of\\nExpandedObjectHeader per object is an unaffordable luxury here. The\\nstructure has 8 bytes of unnecessary padding in it that we could elide\\neasily enough, but after that there's not really much way to squeeze\\nit down without hurting performance in some case cases where this\\nmechanism is fast today.\\n\\nSo I guess I'm going to agree that this approach is doomed. Rats.\\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, 19 Jan 2016 11:42:30 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"[ rewinding to here from the detour I led us on ]\\n\\nOn Mon, Dec 21, 2015 at 4:02 AM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> Now, there has been talk of this previously, on various threads, but I don't\\n> believe any final decisions were made on how exactly it should be done. At\\n> the moment I plan to make changes as follows:\\n>\\n> Add 3 new columns to pg_aggregate, aggserialfn, aggdeserialfn and\\n> aggserialtype These will only be required when aggtranstype is INTERNAL.\\n\\nCheck.\\n\\n> Perhaps we should disallow CREATE AGGREAGET from accepting them for any\\n> other type...\\n\\nWell, we should definitely not accept them and have them be\\nmeaningless. We should either reject them or accept them and then use\\nthem. I can't immediately think of a reason to serialize one\\nnon-internal type as another, but maybe there is one.\\n\\n> The return type of aggserialfn should be aggserialtype, and it\\n> should take a single parameter of aggtranstype. aggdeserialfn will be the\\n> reverse of that.\\n\\nCheck.\\n\\n> Add a new bool field to nodeAgg's state named serialStates. If this is field\\n> is set to true then when we're in finalizeAgg = false mode, we'll call the\\n> serialfn on the agg state instead of the finalfn. This will allow the\\n> serialized state to be stored in the tuple and sent off to the main backend.\\n> The combine agg node should also be set to serialStates = true, so that it\\n> knows to deserialize instead of just assuming that the agg state is of type\\n> aggtranstype.\\n\\nI'm not quite sure, but it sounds like you might be overloading\\nserialStates with two different meanings here.\\n\\n> I believe this should allow us to not cause any performance regressions by\\n> moving away from INTERNAL agg states. It should also be very efficient for\\n> internal states such as Int8TransTypeData, as this struct merely has 2 int64\\n> fields which should be very simple to stuff into a bytea varlena type. We\\n> don't need to mess around converting the ->count and ->sum into a strings or\\n> anything.\\n\\nI think it would be more user-friendly to emit these as 2-element\\ninteger arrays rather than bytea. Sure, bytea is fine when PostgreSQL\\nis talking to itself, but if you are communicating with an external\\nsystem, it's a different situation. If the remote system is\\nPostgreSQL then you are again OK, except for the data going over the\\nwire being incomprehensible and maybe byte-order-dependent, but what\\nif you want some other database system to do partial aggregation and\\nthen further aggregate the result? You don't want the intermediate\\nstate to be some kooky thing that only another PostgreSQL database has\\na chance of generating correctly.\\n\\n> Then in order for the planner to allow parallel aggregation all aggregates\\n> must:\\n>\\n> Not have a DISTINCT or ORDER BY clause\\n> Have a combinefn\\n> If aggtranstype = INTERNAL, must have a aggserialfn and aggdeserialfn.\\n>\\n> We can relax the requirement on 3 if we're using 2-stage aggregation, but\\n> not parallel aggregation.\\n\\nWhen would we do that?\\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, 19 Jan 2016 11:56:50 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 19, 2016 at 11:56 AM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> [ rewinding to here from the detour I led us on ]\\n>\\n> On Mon, Dec 21, 2015 at 4:02 AM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n>> Now, there has been talk of this previously, on various threads, but I don't\\n>> believe any final decisions were made on how exactly it should be done. At\\n>> the moment I plan to make changes as follows:\\n\\nOh, one more point: is there any reason why all of this needs to be a\\nsingle (giant) patch? I feel like the handling of internal states\\ncould be a separate patch from the basic patch to allow partial\\naggregation and aggregate-combining, and that the latter could be\\ncommitted first if we had a reasonably final version of it. That\\nseems like it would be easier from a review standpoint, and might\\nallow more development to proceed in parallel, 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\": \"Tue, 19 Jan 2016 11:58:09 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 20 January 2016 at 05:56, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n>\\n> On Mon, Dec 21, 2015 at 4:02 AM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n> > Now, there has been talk of this previously, on various threads, but I\\n> don't\\n> > believe any final decisions were made on how exactly it should be done.\\n> At\\n> > the moment I plan to make changes as follows:\\n> >\\n> > Add 3 new columns to pg_aggregate, aggserialfn, aggdeserialfn and\\n> > aggserialtype These will only be required when aggtranstype is INTERNAL.\\n>\\n> Check.\\n>\\n> > Perhaps we should disallow CREATE AGGREAGET from accepting them for any\\n> > other type...\\n>\\n> Well, we should definitely not accept them and have them be\\n> meaningless. We should either reject them or accept them and then use\\n> them. I can't immediately think of a reason to serialize one\\n> non-internal type as another, but maybe there is one.\\n>\\n>\\nIn the latest patch I'm disallowing serial and deserial functions for non\\nINTERNAL state aggregates during CREATE AGGREGATE.\\nI think it's best to keep this disabled for now, and do so until we\\ndiscover some reason that we might want want to enable it. If we enabled\\nit, and later decided that was a dumb idea, then it'll be much harder to\\nadd the restriction later, since it may cause errors for users who have\\ncreated their own aggregates.\\n\\n\\n> > Add a new bool field to nodeAgg's state named serialStates. If this is\\n> field\\n> > is set to true then when we're in finalizeAgg = false mode, we'll call\\n> the\\n> > serialfn on the agg state instead of the finalfn. This will allow the\\n> > serialized state to be stored in the tuple and sent off to the main\\n> backend.\\n> > The combine agg node should also be set to serialStates = true, so that\\n> it\\n> > knows to deserialize instead of just assuming that the agg state is of\\n> type\\n> > aggtranstype.\\n>\\n> I'm not quite sure, but it sounds like you might be overloading\\n> serialStates with two different meanings here.\\n>\\n\\nHmm, only in the sense that serialStates means \\\"serialise\\\" and\\n\\\"deserialise\\\". The only time that both could occur in the same node is if\\ncombineStates=true and finalizeAggs=false (in other words 3 or more\\naggregation stages).\\n\\nLet's say we are performing aggregation in 3 stages, stage 1 is operating\\non normal values and uses the transfn on these values, but does not\\nfinalise the states. If serialStates = true here then, if one exists, we\\ncall the serialfn, passing in the aggregated state, and pass the return\\nvalue of that function up to the next node. Now, this next (middle) node is\\nimportant, serialStates must also be true here so that the executor knows\\nto deserialise the previously serialised states. Now, this node uses the\\ncombinefn to merge states which require, and then since serialStates is\\ntrue, it also (re)serialises those new states again. Now if there was some\\nreason that this middle node should deserialise, but not re-serialise those\\nstates, then we may need an extra parameter to instruct it to do so. I\\nguess this may be required if we were to perform some partial aggregation\\nand combining again within a single process (in which case we'd not need to\\nserialise INTERNAL states, we can just pass the pointer to them in the\\nTuple), but then we might require the states to be serialised in order to\\nhand them over to the main process, from a worker process.\\n\\nI can imagine cases where we might want to do this in the future, so\\nperhaps it is worth it:\\n\\nImagine:\\n\\nSELECT COUNT(*) FROM (SELECT * FROM (SELECT * FROM a UNION ALL SELECT *\\nFROM b) AS ab UNION ALL (SELECT * FROM c UNION ALL SELECT * FROM d)) abcd;\\n\\nWhere the planner may decide to, on 1 worker perform count(*) on a then b,\\nand combine that into ab, while doing the same for c and d on some other\\nworker process.\\n\\n\\n> I believe this should allow us to not cause any performance regressions by\\n> > moving away from INTERNAL agg states. It should also be very efficient\\n> for\\n> > internal states such as Int8TransTypeData, as this struct merely has 2\\n> int64\\n> > fields which should be very simple to stuff into a bytea varlena type. We\\n> > don't need to mess around converting the ->count and ->sum into a\\n> strings or\\n> > anything.\\n>\\n> I think it would be more user-friendly to emit these as 2-element\\n> integer arrays rather than bytea. Sure, bytea is fine when PostgreSQL\\n> is talking to itself, but if you are communicating with an external\\n> system, it's a different situation. If the remote system is\\n> PostgreSQL then you are again OK, except for the data going over the\\n> wire being incomprehensible and maybe byte-order-dependent, but what\\n> if you want some other database system to do partial aggregation and\\n> then further aggregate the result? You don't want the intermediate\\n> state to be some kooky thing that only another PostgreSQL database has\\n> a chance of generating correctly.\\n>\\n\\nIf we do that then we also need to invent composite database types for the\\nmore complicated internal states such as NumericAggState.\\nWe should probably also consider performance here too. I had been\\nconsidering just using the *send() functions to build a bytea.\\n\\n> Then in order for the planner to allow parallel aggregation all aggregates\\n> > must:\\n> >\\n> > Not have a DISTINCT or ORDER BY clause\\n> > Have a combinefn\\n> > If aggtranstype = INTERNAL, must have a aggserialfn and aggdeserialfn.\\n> >\\n> > We can relax the requirement on 3 if we're using 2-stage aggregation, but\\n> > not parallel aggregation.\\n>\\n> When would we do that?\\n\\n\\nThis is probably best explained in one of the code comments:\\n\\n+ * states as input rather than input tuples. This mode facilitates\\nmultiple\\n+ * aggregate stages which allows us to support pushing aggregation\\ndown\\n+ * deeper into the plan rather than leaving it for the final stage.\\nFor\\n+ * example with a query such as:\\n+ *\\n+ * SELECT count(*) FROM (SELECT * FROM a UNION ALL SELECT * FROM b);\\n+ *\\n+ * with this functionality the planner has the flexibility to\\ngenerate a\\n+ * plan which performs count(*) on table a and table b separately\\nand then\\n+ * add a combine phase to combine both results. In this case the\\ncombine\\n+ * function would simply add both counts together.\\n\\nThere's also the possibility to use this functionality later to allow GROUP\\nBY to occur before the final plan stage, e.g before joining to some\\nrelation.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 20 January 2016 at 05:56, Robert Haas <robertmhaas@gmail.com> wrote:\\nOn Mon, Dec 21, 2015 at 4:02 AM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> Now, there has been talk of this previously, on various threads, but I don't\\n> believe any final decisions were made on how exactly it should be done. At\\n> the moment I plan to make changes as follows:\\n>\\n> Add 3 new columns to pg_aggregate, aggserialfn, aggdeserialfn and\\n> aggserialtype These will only be required when aggtranstype is INTERNAL.\\n\\nCheck.\\n\\n> Perhaps we should disallow CREATE AGGREAGET from accepting them for any\\n> other type...\\n\\nWell, we should definitely not accept them and have them be\\nmeaningless. We should either reject them or accept them and then use\\nthem. I can't immediately think of a reason to serialize one\\nnon-internal type as another, but maybe there is one.\\nIn the latest patch I'm disallowing serial and deserial functions for non INTERNAL state aggregates during CREATE AGGREGATE.I think it's best to keep this disabled for now, and do so until we discover some reason that we might want want to enable it. If we enabled it, and later decided that was a dumb idea, then it'll be much harder to add the restriction later, since it may cause errors for users who have created their own aggregates. \\n> Add a new bool field to nodeAgg's state named serialStates. If this is field\\n> is set to true then when we're in finalizeAgg = false mode, we'll call the\\n> serialfn on the agg state instead of the finalfn. This will allow the\\n> serialized state to be stored in the tuple and sent off to the main backend.\\n> The combine agg node should also be set to serialStates = true, so that it\\n> knows to deserialize instead of just assuming that the agg state is of type\\n> aggtranstype.\\n\\nI'm not quite sure, but it sounds like you might be overloading\\nserialStates with two different meanings here.Hmm, only in the sense that serialStates means \\\"serialise\\\" and \\\"deserialise\\\". The only time that both could occur in the same node is if combineStates=true and finalizeAggs=false (in other words 3 or more aggregation stages).Let's say we are performing aggregation in 3 stages, stage 1 is operating on normal values and uses the transfn on these values, but does not finalise the states. If serialStates = true here then, if one exists, we call the serialfn, passing in the aggregated state, and pass the return value of that function up to the next node. Now, this next (middle) node is important, serialStates must also be true here so that the executor knows to deserialise the previously serialised states. Now, this node uses the combinefn to merge states which require, and then since serialStates is true, it also (re)serialises those new states again. Now if there was some reason that this middle node should deserialise, but not re-serialise those states, then we may need an extra parameter to instruct it to do so. I guess this may be required if we were to perform some partial aggregation and combining again within a single process (in which case we'd not need to serialise INTERNAL states, we can just pass the pointer to them in the Tuple), but then we might require the states to be serialised in order to hand them over to the main process, from a worker process.I can imagine cases where we might want to do this in the future, so perhaps it is worth it:Imagine:SELECT COUNT(*) FROM (SELECT * FROM (SELECT * FROM a UNION ALL SELECT * FROM b) AS ab UNION ALL (SELECT * FROM c UNION ALL SELECT * FROM d)) abcd; Where the planner may decide to, on 1 worker perform count(*) on a then b, and combine that into ab, while doing the same for c and d on some other worker process.\\n> I believe this should allow us to not cause any performance regressions by\\n> moving away from INTERNAL agg states. It should also be very efficient for\\n> internal states such as Int8TransTypeData, as this struct merely has 2 int64\\n> fields which should be very simple to stuff into a bytea varlena type. We\\n> don't need to mess around converting the ->count and ->sum into a strings or\\n> anything.\\n\\nI think it would be more user-friendly to emit these as 2-element\\ninteger arrays rather than bytea. Sure, bytea is fine when PostgreSQL\\nis talking to itself, but if you are communicating with an external\\nsystem, it's a different situation. If the remote system is\\nPostgreSQL then you are again OK, except for the data going over the\\nwire being incomprehensible and maybe byte-order-dependent, but what\\nif you want some other database system to do partial aggregation and\\nthen further aggregate the result? You don't want the intermediate\\nstate to be some kooky thing that only another PostgreSQL database has\\na chance of generating correctly.If we do that then we also need to invent composite database types for the more complicated internal states such as NumericAggState.We should probably also consider performance here too. I had been considering just using the *send() functions to build a bytea.\\n> Then in order for the planner to allow parallel aggregation all aggregates\\n> must:\\n>\\n> Not have a DISTINCT or ORDER BY clause\\n> Have a combinefn\\n> If aggtranstype = INTERNAL, must have a aggserialfn and aggdeserialfn.\\n>\\n> We can relax the requirement on 3 if we're using 2-stage aggregation, but\\n> not parallel aggregation.\\n\\nWhen would we do that?This is probably best explained in one of the code comments:+ * states as input rather than input tuples. This mode facilitates multiple+ * aggregate stages which allows us to support pushing aggregation down+ * deeper into the plan rather than leaving it for the final stage. For+ * example with a query such as:+ *+ * SELECT count(*) FROM (SELECT * FROM a UNION ALL SELECT * FROM b);+ *+ * with this functionality the planner has the flexibility to generate a+ * plan which performs count(*) on table a and table b separately and then+ * add a combine phase to combine both results. In this case the combine+ * function would simply add both counts together.There's also the possibility to use this functionality later to allow GROUP BY to occur before the final plan stage, e.g before joining to some relation.-- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Wed, 20 Jan 2016 10:44:16 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 20 January 2016 at 05:58, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> > On Mon, Dec 21, 2015 at 4:02 AM, David Rowley\\n> > <david.rowley@2ndquadrant.com> wrote:\\n> >> Now, there has been talk of this previously, on various threads, but I\\n> don't\\n> >> believe any final decisions were made on how exactly it should be done.\\n> At\\n> >> the moment I plan to make changes as follows:\\n>\\n> Oh, one more point: is there any reason why all of this needs to be a\\n> single (giant) patch? I feel like the handling of internal states\\n> could be a separate patch from the basic patch to allow partial\\n> aggregation and aggregate-combining, and that the latter could be\\n> committed first if we had a reasonably final version of it. That\\n> seems like it would be easier from a review standpoint, and might\\n> allow more development to proceed in parallel, too.\\n\\n\\nI didn't ever really imagine that I'd include any actual new combinefns or\\nserialfn/deserialfn in this patch. One set has of course now ended up in\\nthere, I can move these off to the test patch for now.\\n\\nDid you imagine that the first patch in the series would only add the\\naggcombinefn column to pg_aggregate and leave the aggserialfn stuff until\\nlater? I thought it seemed better to get the infrastructure committed in\\none hit, then add a bunch of new combinefn, serialfn, deserialfns for\\nexisting aggregates in follow-on commits.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 20 January 2016 at 05:58, Robert Haas <robertmhaas@gmail.com> wrote:> On Mon, Dec 21, 2015 at 4:02 AM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n>> Now, there has been talk of this previously, on various threads, but I don't\\n>> believe any final decisions were made on how exactly it should be done. At\\n>> the moment I plan to make changes as follows:\\n\\nOh, one more point: is there any reason why all of this needs to be a\\nsingle (giant) patch? I feel like the handling of internal states\\ncould be a separate patch from the basic patch to allow partial\\naggregation and aggregate-combining, and that the latter could be\\ncommitted first if we had a reasonably final version of it. That\\nseems like it would be easier from a review standpoint, and might\\nallow more development to proceed in parallel, too.I didn't ever really imagine that I'd include any actual new combinefns or serialfn/deserialfn in this patch. One set has of course now ended up in there, I can move these off to the test patch for now. Did you imagine that the first patch in the series would only add the aggcombinefn column to pg_aggregate and leave the aggserialfn stuff until later? I thought it seemed better to get the infrastructure committed in one hit, then add a bunch of new combinefn, serialfn, deserialfns for existing aggregates in follow-on commits. -- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Wed, 20 Jan 2016 10:50:20 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 19, 2016 at 4:50 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n>> Oh, one more point: is there any reason why all of this needs to be a\\n>> single (giant) patch? I feel like the handling of internal states\\n>> could be a separate patch from the basic patch to allow partial\\n>> aggregation and aggregate-combining, and that the latter could be\\n>> committed first if we had a reasonably final version of it. That\\n>> seems like it would be easier from a review standpoint, and might\\n>> allow more development to proceed in parallel, too.\\n>\\n> I didn't ever really imagine that I'd include any actual new combinefns or\\n> serialfn/deserialfn in this patch. One set has of course now ended up in\\n> there, I can move these off to the test patch for now.\\n>\\n> Did you imagine that the first patch in the series would only add the\\n> aggcombinefn column to pg_aggregate and leave the aggserialfn stuff until\\n> later?\\n\\nI think that would be a sensible way to proceed.\\n\\n> I thought it seemed better to get the infrastructure committed in one\\n> hit, then add a bunch of new combinefn, serialfn, deserialfns for existing\\n> aggregates in follow-on commits.\\n\\nTo my mind, priority #1 ought to be putting this fine new\\nfunctionality to some use. Expanding it to every aggregate we've got\\nseems like a distinctly second priority. That's not to say that it's\\nabsolutely gotta go down that way, but those would be my priorities.\\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, 19 Jan 2016 16:54:26 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 20 January 2016 at 10:54, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On Tue, Jan 19, 2016 at 4:50 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n> >> Oh, one more point: is there any reason why all of this needs to be a\\n> >> single (giant) patch? I feel like the handling of internal states\\n> >> could be a separate patch from the basic patch to allow partial\\n> >> aggregation and aggregate-combining, and that the latter could be\\n> >> committed first if we had a reasonably final version of it. That\\n> >> seems like it would be easier from a review standpoint, and might\\n> >> allow more development to proceed in parallel, too.\\n> >\\n> > I didn't ever really imagine that I'd include any actual new combinefns\\n> or\\n> > serialfn/deserialfn in this patch. One set has of course now ended up in\\n> > there, I can move these off to the test patch for now.\\n> >\\n> > Did you imagine that the first patch in the series would only add the\\n> > aggcombinefn column to pg_aggregate and leave the aggserialfn stuff until\\n> > later?\\n>\\n> I think that would be a sensible way to proceed.\\n>\\n> > I thought it seemed better to get the infrastructure committed in one\\n> > hit, then add a bunch of new combinefn, serialfn, deserialfns for\\n> existing\\n> > aggregates in follow-on commits.\\n>\\n> To my mind, priority #1 ought to be putting this fine new\\n> functionality to some use. Expanding it to every aggregate we've got\\n> seems like a distinctly second priority. That's not to say that it's\\n> absolutely gotta go down that way, but those would be my priorities.\\n\\n\\nAgreed. So I've attached a version of the patch which does not have any of\\nthe serialise/deserialise stuff in it.\\n\\nI've also attached a test patch which modifies the grouping planner to add\\na Partial Aggregate node, and a final aggregate node when it's possible.\\nRunning the regression tests with this patch only shows up variances in the\\nEXPLAIN outputs, which is of course expected.\\n\\n-- \\n David Rowley 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, 21 Jan 2016 01:38:39 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 20, 2016 at 7:38 AM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n>> To my mind, priority #1 ought to be putting this fine new\\n>> functionality to some use. Expanding it to every aggregate we've got\\n>> seems like a distinctly second priority. That's not to say that it's\\n>> absolutely gotta go down that way, but those would be my priorities.\\n>\\n> Agreed. So I've attached a version of the patch which does not have any of\\n> the serialise/deserialise stuff in it.\\n>\\n> I've also attached a test patch which modifies the grouping planner to add a\\n> Partial Aggregate node, and a final aggregate node when it's possible.\\n> Running the regression tests with this patch only shows up variances in the\\n> EXPLAIN outputs, which is of course expected.\\n\\nThat seems great as a test, but what's the first patch that can put\\nthis to real and permanent use?\\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, 20 Jan 2016 07:44:42 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 21 January 2016 at 01:44, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On Wed, Jan 20, 2016 at 7:38 AM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n> >> To my mind, priority #1 ought to be putting this fine new\\n> >> functionality to some use. Expanding it to every aggregate we've got\\n> >> seems like a distinctly second priority. That's not to say that it's\\n> >> absolutely gotta go down that way, but those would be my priorities.\\n> >\\n> > Agreed. So I've attached a version of the patch which does not have any\\n> of\\n> > the serialise/deserialise stuff in it.\\n> >\\n> > I've also attached a test patch which modifies the grouping planner to\\n> add a\\n> > Partial Aggregate node, and a final aggregate node when it's possible.\\n> > Running the regression tests with this patch only shows up variances in\\n> the\\n> > EXPLAIN outputs, which is of course expected.\\n>\\n> That seems great as a test, but what's the first patch that can put\\n> this to real and permanent use?\\n\\n\\nThere's no reason why parallel aggregates can't use\\nthe combine_aggregate_state_d6d480b_2016-01-21.patch patch.\\n\\nIn this patch I've changed aggregates_allow_partial() so that it properly\\ndetermines what is possible based on the aggregates which are in the query.\\nThis now, of course restricts the aggregates to \\\"internal only\\\" when the\\nagg state type is INTERNAL, providing there's a combine function, of course.\\n\\nParallel aggregate should work with all the MAX() and MIN() functions and a\\nhandful of other ones, I've managed to borrow various existing function as\\nthe combine function for many aggregates:\\n\\n# select count(*) from pg_aggregate where aggcombinefn <> 0;\\n count\\n-------\\n 58\\n(1 row)\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 21 January 2016 at 01:44, Robert Haas <robertmhaas@gmail.com> wrote:On Wed, Jan 20, 2016 at 7:38 AM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n>> To my mind, priority #1 ought to be putting this fine new\\n>> functionality to some use. Expanding it to every aggregate we've got\\n>> seems like a distinctly second priority. That's not to say that it's\\n>> absolutely gotta go down that way, but those would be my priorities.\\n>\\n> Agreed. So I've attached a version of the patch which does not have any of\\n> the serialise/deserialise stuff in it.\\n>\\n> I've also attached a test patch which modifies the grouping planner to add a\\n> Partial Aggregate node, and a final aggregate node when it's possible.\\n> Running the regression tests with this patch only shows up variances in the\\n> EXPLAIN outputs, which is of course expected.\\n\\nThat seems great as a test, but what's the first patch that can put\\nthis to real and permanent use?There's no reason why parallel aggregates can't use the combine_aggregate_state_d6d480b_2016-01-21.patch patch. In this patch I've changed aggregates_allow_partial() so that it properly determines what is possible based on the aggregates which are in the query. This now, of course restricts the aggregates to \\\"internal only\\\" when the agg state type is INTERNAL, providing there's a combine function, of course.Parallel aggregate should work with all the MAX() and MIN() functions and a handful of other ones, I've managed to borrow various existing function as the combine function for many aggregates:# select count(*) from pg_aggregate where aggcombinefn <> 0; count------- 58(1 row)-- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Thu, 21 Jan 2016 01:53:39 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 20, 2016 at 7:53 AM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> On 21 January 2016 at 01:44, Robert Haas <robertmhaas@gmail.com> wrote:\\n>>\\n>> On Wed, Jan 20, 2016 at 7:38 AM, David Rowley\\n>> <david.rowley@2ndquadrant.com> wrote:\\n>> >> To my mind, priority #1 ought to be putting this fine new\\n>> >> functionality to some use. Expanding it to every aggregate we've got\\n>> >> seems like a distinctly second priority. That's not to say that it's\\n>> >> absolutely gotta go down that way, but those would be my priorities.\\n>> >\\n>> > Agreed. So I've attached a version of the patch which does not have any\\n>> > of\\n>> > the serialise/deserialise stuff in it.\\n>> >\\n>> > I've also attached a test patch which modifies the grouping planner to\\n>> > add a\\n>> > Partial Aggregate node, and a final aggregate node when it's possible.\\n>> > Running the regression tests with this patch only shows up variances in\\n>> > the\\n>> > EXPLAIN outputs, which is of course expected.\\n>>\\n>> That seems great as a test, but what's the first patch that can put\\n>> this to real and permanent use?\\n>\\n> There's no reason why parallel aggregates can't use the\\n> combine_aggregate_state_d6d480b_2016-01-21.patch patch.\\n\\nI agree. Are you going to work on that? Are you expecting me to work\\non that? Do you think we can use Haribabu's patch? What other\\napplications are in play in the near term, if any?\\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, 20 Jan 2016 10:59:56 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 20, 2016 at 7:38 AM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> Agreed. So I've attached a version of the patch which does not have any of\\n> the serialise/deserialise stuff in it.\\n\\nI re-reviewed this and have committed most of it with only minor\\nkibitizing. A few notes:\\n\\n- I changed the EXPLAIN code, since it failed to display the aggregate\\nnode's mode of operation in non-text format.\\n\\n- I removed some of the planner support, since I'm not sure that it's\\ncompletely correct and I'm very sure that it contains more code\\nduplication than I'm comfortable with (set_combineagg_references in\\nparticular). Please feel free to resubmit this part, perhaps in\\ncombination with code that actually uses it for something. But please\\nalso think about whether there's a way to reduce the duplication, and\\nmaybe consider adding some comments, too.\\n\\n- I'm not sure that you made the right call regarding reusing\\ntransfn_oid and transfn for combine functions, vs. adding separate\\nfields. It is sort of logical and has the advantage of keeping the\\ndata structure smaller, but it might also be thought confusing or\\ninappropriate on other grounds. But I think we can change it later if\\nthat comes to seem like the right thing to do, so I've left it the way\\nyou had it for 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\": \"Wed, 20 Jan 2016 14:06:41 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 21 January 2016 at 04:59, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On Wed, Jan 20, 2016 at 7:53 AM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n> > On 21 January 2016 at 01:44, Robert Haas <robertmhaas@gmail.com> wrote:\\n> >>\\n> >> On Wed, Jan 20, 2016 at 7:38 AM, David Rowley\\n> >> <david.rowley@2ndquadrant.com> wrote:\\n> >> >> To my mind, priority #1 ought to be putting this fine new\\n> >> >> functionality to some use. Expanding it to every aggregate we've got\\n> >> >> seems like a distinctly second priority. That's not to say that it's\\n> >> >> absolutely gotta go down that way, but those would be my priorities.\\n> >> >\\n> >> > Agreed. So I've attached a version of the patch which does not have\\n> any\\n> >> > of\\n> >> > the serialise/deserialise stuff in it.\\n> >> >\\n> >> > I've also attached a test patch which modifies the grouping planner to\\n> >> > add a\\n> >> > Partial Aggregate node, and a final aggregate node when it's possible.\\n> >> > Running the regression tests with this patch only shows up variances\\n> in\\n> >> > the\\n> >> > EXPLAIN outputs, which is of course expected.\\n> >>\\n> >> That seems great as a test, but what's the first patch that can put\\n> >> this to real and permanent use?\\n> >\\n> > There's no reason why parallel aggregates can't use the\\n> > combine_aggregate_state_d6d480b_2016-01-21.patch patch.\\n>\\n> I agree. Are you going to work on that? Are you expecting me to work\\n> on that? Do you think we can use Haribabu's patch? What other\\n> applications are in play in the near term, if any?\\n\\n\\nAt the moment I think everything which will use this is queued up behind\\nthe pathification of the grouping planner which Tom is working on. I think\\nnaturally Parallel Aggregate makes sense to work on first, given all the\\nother parallel stuff in this release. I plan on working on that that by\\neither assisting Haribabu, or... whatever else it takes.\\n\\nThe other two usages which I have thought of are;\\n\\n1) Aggregating before UNION ALL, which might be fairly simple after the\\ngrouping planner changes, as it may just be a matter of considering another\\n\\\"grouping path\\\" which partially aggregates before the UNION ALL, and\\nperforms the final grouping stage after UNION ALL. At this stage it's hard\\nto say how that will work as I'm not sure how far changes to the grouping\\nplanner will go. Perhaps Tom can comment?\\n\\n2) Group before join. e.g select p.description,sum(s.qty) from sale s inner\\njoin s.product_id = p.product_id group by s.product_id group by\\np.description; I have a partial patch which implements this, although I\\nwas a bit stuck on if I should invent the concept of \\\"GroupingPaths\\\", or\\njust inject alternative subquery relations which are already grouped by the\\ncorrect clause. The problem with \\\"GroupingPaths\\\" was down to the row\\nestimates currently come from the RelOptInfo and is set\\nin set_baserel_size_estimates() which always assumes the ungrouped number\\nof rows, which is not what's needed if the grouping is already performed. I\\nwas holding off to see how Tom does this in the grouping planner changes.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 21 January 2016 at 04:59, Robert Haas <robertmhaas@gmail.com> wrote:On Wed, Jan 20, 2016 at 7:53 AM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> On 21 January 2016 at 01:44, Robert Haas <robertmhaas@gmail.com> wrote:\\n>>\\n>> On Wed, Jan 20, 2016 at 7:38 AM, David Rowley\\n>> <david.rowley@2ndquadrant.com> wrote:\\n>> >> To my mind, priority #1 ought to be putting this fine new\\n>> >> functionality to some use. Expanding it to every aggregate we've got\\n>> >> seems like a distinctly second priority. That's not to say that it's\\n>> >> absolutely gotta go down that way, but those would be my priorities.\\n>> >\\n>> > Agreed. So I've attached a version of the patch which does not have any\\n>> > of\\n>> > the serialise/deserialise stuff in it.\\n>> >\\n>> > I've also attached a test patch which modifies the grouping planner to\\n>> > add a\\n>> > Partial Aggregate node, and a final aggregate node when it's possible.\\n>> > Running the regression tests with this patch only shows up variances in\\n>> > the\\n>> > EXPLAIN outputs, which is of course expected.\\n>>\\n>> That seems great as a test, but what's the first patch that can put\\n>> this to real and permanent use?\\n>\\n> There's no reason why parallel aggregates can't use the\\n> combine_aggregate_state_d6d480b_2016-01-21.patch patch.\\n\\nI agree. Are you going to work on that? Are you expecting me to work\\non that? Do you think we can use Haribabu's patch? What other\\napplications are in play in the near term, if any?At the moment I think everything which will use this is queued up behind the pathification of the grouping planner which Tom is working on. I think naturally Parallel Aggregate makes sense to work on first, given all the other parallel stuff in this release. I plan on working on that that by either assisting Haribabu, or... whatever else it takes.The other two usages which I have thought of are;1) Aggregating before UNION ALL, which might be fairly simple after the grouping planner changes, as it may just be a matter of considering another \\\"grouping path\\\" which partially aggregates before the UNION ALL, and performs the final grouping stage after UNION ALL. At this stage it's hard to say how that will work as I'm not sure how far changes to the grouping planner will go. Perhaps Tom can comment?2) Group before join. e.g select p.description,sum(s.qty) from sale s inner join s.product_id = p.product_id group by s.product_id group by p.description; I have a partial patch which implements this, although I was a bit stuck on if I should invent the concept of \\\"GroupingPaths\\\", or just inject alternative subquery relations which are already grouped by the correct clause. The problem with \\\"GroupingPaths\\\" was down to the row estimates currently come from the RelOptInfo and is set in set_baserel_size_estimates() which always assumes the ungrouped number of rows, which is not what's needed if the grouping is already performed. I was holding off to see how Tom does this in the grouping planner changes.-- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Thu, 21 Jan 2016 14:32:52 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 21 January 2016 at 08:06, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On Wed, Jan 20, 2016 at 7:38 AM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n> > Agreed. So I've attached a version of the patch which does not have any\\n> of\\n> > the serialise/deserialise stuff in it.\\n>\\n> I re-reviewed this and have committed most of it with only minor\\n> kibitizing. A few notes:\\n>\\n\\nGreat! Thank you for committing.\\n\\n- I changed the EXPLAIN code, since it failed to display the aggregate\\n> node's mode of operation in non-text format.\\n>\\n\\nOops, thanks for fixing.\\n\\n\\n> - I removed some of the planner support, since I'm not sure that it's\\n> completely correct and I'm very sure that it contains more code\\n> duplication than I'm comfortable with (set_combineagg_references in\\n> particular). Please feel free to resubmit this part, perhaps in\\n> combination with code that actually uses it for something. But please\\n> also think about whether there's a way to reduce the duplication, and\\n> maybe consider adding some comments, too.\\n>\\n\\nThat part is a gray area. Wasn't that sure it belonged to this patch\\neither. The problem is, that without it an Aggref's return type is wrong\\nwhen the node is in combine mode, which might not be a problem now as we\\ndon't have a planner yet that generates these nodes.\\n\\n\\n> - I'm not sure that you made the right call regarding reusing\\n> transfn_oid and transfn for combine functions, vs. adding separate\\n> fields. It is sort of logical and has the advantage of keeping the\\n> data structure smaller, but it might also be thought confusing or\\n> inappropriate on other grounds. But I think we can change it later if\\n> that comes to seem like the right thing to do, so I've left it the way\\n> you had it for now.\\n\\n\\nSharing the variable, as horrid as that might seem does simplify some code.\\nFor example if you look at find_compatible_pertrans(), then you see:\\n\\nif (aggtransfn != pertrans->transfn_oid ||\\naggtranstype != pertrans->aggtranstype)\\ncontinue;\\n\\nThis would need to be changed to something like:\\n\\nif (!aggstate->combineStates &&\\n(aggtransfn != pertrans->transfn_oid ||\\naggtranstype != pertrans->aggtranstype))\\ncontinue;\\nelse if (aggstate->combineStates &&\\n(aggcombinefn != pertrans->combinefn_oid ||\\naggtranstype != pertrans->aggtranstype))\\ncontinue;\\n\\nAnd we'd then have to pass aggcombinefn to that function too.\\n\\nWhat might be better would be to rename the variable to something name that\\nscreams something a bit more generic.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 21 January 2016 at 08:06, Robert Haas <robertmhaas@gmail.com> wrote:On Wed, Jan 20, 2016 at 7:38 AM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> Agreed. So I've attached a version of the patch which does not have any of\\n> the serialise/deserialise stuff in it.\\n\\nI re-reviewed this and have committed most of it with only minor\\nkibitizing. A few notes:Great! Thank you for committing. \\n- I changed the EXPLAIN code, since it failed to display the aggregate\\nnode's mode of operation in non-text format.Oops, thanks for fixing. \\n- I removed some of the planner support, since I'm not sure that it's\\ncompletely correct and I'm very sure that it contains more code\\nduplication than I'm comfortable with (set_combineagg_references in\\nparticular). Please feel free to resubmit this part, perhaps in\\ncombination with code that actually uses it for something. But please\\nalso think about whether there's a way to reduce the duplication, and\\nmaybe consider adding some comments, too.That part is a gray area. Wasn't that sure it belonged to this patch either. The problem is, that without it an Aggref's return type is wrong when the node is in combine mode, which might not be a problem now as we don't have a planner yet that generates these nodes. \\n- I'm not sure that you made the right call regarding reusing\\ntransfn_oid and transfn for combine functions, vs. adding separate\\nfields. It is sort of logical and has the advantage of keeping the\\ndata structure smaller, but it might also be thought confusing or\\ninappropriate on other grounds. But I think we can change it later if\\nthat comes to seem like the right thing to do, so I've left it the way\\nyou had it for now.Sharing the variable, as horrid as that might seem does simplify some code. For example if you look at find_compatible_pertrans(), then you see: if (aggtransfn != pertrans->transfn_oid || aggtranstype != pertrans->aggtranstype) continue;This would need to be changed to something like: if (!aggstate->combineStates && (aggtransfn != pertrans->transfn_oid || aggtranstype != pertrans->aggtranstype)) continue; else if (aggstate->combineStates && (aggcombinefn != pertrans->combinefn_oid || aggtranstype != pertrans->aggtranstype)) continue;And we'd then have to pass aggcombinefn to that function too.What might be better would be to rename the variable to something name that screams something a bit more generic.-- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Thu, 21 Jan 2016 14:41:46 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 21, 2016 at 12:32 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> On 21 January 2016 at 04:59, Robert Haas <robertmhaas@gmail.com> wrote:\\n>>\\n>> On Wed, Jan 20, 2016 at 7:53 AM, David Rowley\\n>> <david.rowley@2ndquadrant.com> wrote:\\n>> > On 21 January 2016 at 01:44, Robert Haas <robertmhaas@gmail.com> wrote:\\n>> >>\\n>> >> On Wed, Jan 20, 2016 at 7:38 AM, David Rowley\\n>> >> <david.rowley@2ndquadrant.com> wrote:\\n>> >> >> To my mind, priority #1 ought to be putting this fine new\\n>> >> >> functionality to some use. Expanding it to every aggregate we've\\n>> >> >> got\\n>> >> >> seems like a distinctly second priority. That's not to say that\\n>> >> >> it's\\n>> >> >> absolutely gotta go down that way, but those would be my priorities.\\n>> >> >\\n>> >> > Agreed. So I've attached a version of the patch which does not have\\n>> >> > any\\n>> >> > of\\n>> >> > the serialise/deserialise stuff in it.\\n>> >> >\\n>> >> > I've also attached a test patch which modifies the grouping planner\\n>> >> > to\\n>> >> > add a\\n>> >> > Partial Aggregate node, and a final aggregate node when it's\\n>> >> > possible.\\n>> >> > Running the regression tests with this patch only shows up variances\\n>> >> > in\\n>> >> > the\\n>> >> > EXPLAIN outputs, which is of course expected.\\n>> >>\\n>> >> That seems great as a test, but what's the first patch that can put\\n>> >> this to real and permanent use?\\n>> >\\n>> > There's no reason why parallel aggregates can't use the\\n>> > combine_aggregate_state_d6d480b_2016-01-21.patch patch.\\n>>\\n>> I agree. Are you going to work on that? Are you expecting me to work\\n>> on that? Do you think we can use Haribabu's patch? What other\\n>> applications are in play in the near term, if any?\\n>\\n>\\n> At the moment I think everything which will use this is queued up behind the\\n> pathification of the grouping planner which Tom is working on. I think\\n> naturally Parallel Aggregate makes sense to work on first, given all the\\n> other parallel stuff in this release. I plan on working on that that by\\n> either assisting Haribabu, or... whatever else it takes.\\n>\\n\\nHere I attached updated patch of parallel aggregate based on the latest\\nchanges in master. Still it lack of cost comparison of normal aggregate to\\nparallel aggregate because of difficulty. This cost comparison is required\\nin parallel aggregate as this is having some regression when the number\\nof groups are less in the query plan.\\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\": \"Thu, 21 Jan 2016 13:33:50 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 21, 2016 at 1:33 PM, Haribabu Kommi\\n<kommi.haribabu@gmail.com> wrote:\\n>\\n> Here I attached updated patch of parallel aggregate based on the latest\\n> changes in master. Still it lack of cost comparison of normal aggregate to\\n> parallel aggregate because of difficulty. This cost comparison is required\\n> in parallel aggregate as this is having some regression when the number\\n> of groups are less in the query plan.\\n>\\n\\nUpdated patch is attached after removing a warning in building group\\naggregate path.\\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\": \"Thu, 21 Jan 2016 13:53:01 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 21 January 2016 at 08:06, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On Wed, Jan 20, 2016 at 7:38 AM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n> > Agreed. So I've attached a version of the patch which does not have any\\n> of\\n> > the serialise/deserialise stuff in it.\\n>\\n> I re-reviewed this and have committed most of it with only minor\\n> kibitizing. A few notes:\\n\\n\\nI've attached the re-based remainder, which includes the serial/deserial\\nagain.\\n\\nI'll submit this part to March 'fest, where hopefully we'll also have\\nsomething to utilise it.\\n\\n-- \\n David Rowley 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, 21 Jan 2016 17:42:58 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 21 January 2016 at 15:53, Haribabu Kommi <kommi.haribabu@gmail.com>\\nwrote:\\n\\n> On Thu, Jan 21, 2016 at 1:33 PM, Haribabu Kommi\\n> <kommi.haribabu@gmail.com> wrote:\\n> >\\n> > Here I attached updated patch of parallel aggregate based on the latest\\n> > changes in master. Still it lack of cost comparison of normal aggregate\\n> to\\n> > parallel aggregate because of difficulty. This cost comparison is\\n> required\\n> > in parallel aggregate as this is having some regression when the number\\n> > of groups are less in the query plan.\\n> >\\n>\\n> Updated patch is attached after removing a warning in building group\\n> aggregate path.\\n\\n\\nHi,\\n\\nThanks for updating the patch. I'd like to look at this with priority, but\\ncan you post it on the Parallel Agg thread? that way anyone following there\\ncan chime in over there rather than here. I've still got a bit of work to\\ndo (in the not too distant future) on the serial/deserial part, so would be\\nbetter to keep this thread for discussion on that.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\nOn 21 January 2016 at 15:53, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:On Thu, Jan 21, 2016 at 1:33 PM, Haribabu Kommi\\n<kommi.haribabu@gmail.com> wrote:\\n>\\n> Here I attached updated patch of parallel aggregate based on the latest\\n> changes in master. Still it lack of cost comparison of normal aggregate to\\n> parallel aggregate because of difficulty. This cost comparison is required\\n> in parallel aggregate as this is having some regression when the number\\n> of groups are less in the query plan.\\n>\\n\\nUpdated patch is attached after removing a warning in building group\\naggregate path.Hi,Thanks for updating the patch. I'd like to look at this with priority, but can you post it on the Parallel Agg thread? that way anyone following there can chime in over there rather than here. I've still got a bit of work to do (in the not too distant future) on the serial/deserial part, so would be better to keep this thread for discussion on that.-- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services\",\n \"msg_date\": \"Thu, 21 Jan 2016 17:52:47 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 20, 2016 at 8:32 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> At the moment I think everything which will use this is queued up behind the\\n> pathification of the grouping planner which Tom is working on. I think\\n> naturally Parallel Aggregate makes sense to work on first, given all the\\n> other parallel stuff in this release. I plan on working on that that by\\n> either assisting Haribabu, or... whatever else it takes.\\n\\nGreat!\\n\\n> The other two usages which I have thought of are;\\n>\\n> 1) Aggregating before UNION ALL, which might be fairly simple after the\\n> grouping planner changes, as it may just be a matter of considering another\\n> \\\"grouping path\\\" which partially aggregates before the UNION ALL, and\\n> performs the final grouping stage after UNION ALL. At this stage it's hard\\n> to say how that will work as I'm not sure how far changes to the grouping\\n> planner will go. Perhaps Tom can comment?\\n\\nI hope he will, but in the meantime let me ask how this does us any\\ngood. UNION ALL turns into an Append node. Pushing aggregation\\nthrough an Append doesn't make anything faster AFAICS *unless* you can\\nfurther optimize beginning at that point. For example, if one or both\\nsides of the Append node have a Gather under them, and you can push\\nthe partial aggregation down into the Gather; or if you hit a Foreign\\nScan and can have it do partial aggregation on the remote side, you\\nwin. But if you just have:\\n\\nFinalize Aggregate\\n-> Append\\n -> Partial Aggregate\\n -> Thing One\\n -> Partial Aggregate\\n -> Thing Two\\n\\nInstead of:\\n\\nAggregate\\n-> Append\\n -> Thing One\\n -> Thing Two\\n\\n...then I don't see the point, at least not for a single-group\\nAggregate or HashAggregate. For a GroupAggregate, Thing One and Thing\\nTwo might need to be sorted, and sorting two or more smaller data sets\\nmight be faster than sorting one larger data set, but I can't see us\\nwinning anything very big here.\\n\\nTo be clear, I'm not saying we shouldn't do this. I just don't think\\nit does much *by itself*. If you combine it with other optimizations\\nthat let the aggregation be pushed further down the plan tree, it\\ncould win big.\\n\\n> 2) Group before join. e.g select p.description,sum(s.qty) from sale s inner\\n> join s.product_id = p.product_id group by s.product_id group by\\n> p.description; I have a partial patch which implements this, although I was\\n> a bit stuck on if I should invent the concept of \\\"GroupingPaths\\\", or just\\n> inject alternative subquery relations which are already grouped by the\\n> correct clause. The problem with \\\"GroupingPaths\\\" was down to the row\\n> estimates currently come from the RelOptInfo and is set in\\n> set_baserel_size_estimates() which always assumes the ungrouped number of\\n> rows, which is not what's needed if the grouping is already performed. I was\\n> holding off to see how Tom does this in the grouping planner changes.\\n\\nYour sample query has two GROUP BY clauses; I think you meant to\\ninclude only the second one. This seems like it can win big, because\\nit's possible that joining first means joining against a much larger\\nrelation, whereas aggregating first might reduce \\\"sale\\\" to a volume of\\ndata that fits in work_mem, after which you can hash join. On the\\nother hand, if you add WHERE p.description LIKE '%snuffleupagus%' then\\nthe aggregate-first strategy figures to lose, assuming things are\\nindexed properly.\\n\\nOn the substantive question you raise, I hope Tom will weigh in here.\\nHowever, I'll give you my take. It seems to me that you are likely to\\nrun into problems not only with the row counts but also with target\\nlists and width estimates. All of those things change once you\\naggregate. It seems clear that you are going to need a GroupingPath\\nhere, but it also seems really clear that you can't take a path where\\nsome aggregation has been done and use add_path to toss it on the pile\\nfor the same RelOptInfo as unaggregated paths to which it is not\\ncomparable. Right now, there's a RelOptInfo corresponding to each\\nsubset of the joinrels for which we build paths; but in this new\\nworld, I think you'll end up with multiple joinrels for each\\nRelOptInfo, one per\\n\\nIn your example above, you'd end up with RelOptInfos for (1) {s} with\\nno aggregation, (2) {p} with no aggregation, (3) {s p} with no\\naggregation, (4) {s} aggregated by s.product_id and (5) {s p}\\naggregated by p.description. You can generate paths for (5) either by\\njoining (2) to (4) or by aggregating (3), and the cheapest path for\\n(5) is the overall winner. It seems a bit tricky to determine which\\naggregations are worth considering, and how to represent that in the\\nRelOptInfo, but overall that feels like the right idea.\\n\\nI'm not sure how much this is going to overlap with the work Tom is\\nalready doing. I think his principal goal is to let the planning of a\\nsubquery return multiple candidate paths. That's almost certain to\\ninvolve creating something like GroupingPath, though I can't predict\\nthe details, and I suspect it's also likely that he'll create some new\\nRelOptInfo that represents the output of the grouping stage. However,\\nI'm not sure that'll have the aggregation that has been performed\\nrepresented in any principled way that would lend itself to having\\nmore than one RelOptInfo of that kind. That's probably another\\nproblem altogether.\\n\\nHowever, my Tom-predictive capabilities are far from perfect.\\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, 21 Jan 2016 12:56:23 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 20, 2016 at 9:33 PM, Haribabu Kommi\\n<kommi.haribabu@gmail.com> wrote:\\n> Here I attached updated patch of parallel aggregate based on the latest\\n> changes in master. Still it lack of cost comparison of normal aggregate to\\n> parallel aggregate because of difficulty. This cost comparison is required\\n> in parallel aggregate as this is having some regression when the number\\n> of groups are less in the query plan.\\n\\nPlease keep this on its own thread rather than colonizing this 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\": \"Thu, 21 Jan 2016 12:56:57 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 22 January 2016 at 06:56, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Wed, Jan 20, 2016 at 8:32 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n>> The other two usages which I have thought of are;\\n>>\\n>> 1) Aggregating before UNION ALL, which might be fairly simple after the\\n>> grouping planner changes, as it may just be a matter of considering another\\n>> \\\"grouping path\\\" which partially aggregates before the UNION ALL, and\\n>> performs the final grouping stage after UNION ALL. At this stage it's hard\\n>> to say how that will work as I'm not sure how far changes to the grouping\\n>> planner will go. Perhaps Tom can comment?\\n>\\n> I hope he will, but in the meantime let me ask how this does us any\\n> good. UNION ALL turns into an Append node. Pushing aggregation\\n> through an Append doesn't make anything faster AFAICS *unless* you can\\n> further optimize beginning at that point. For example, if one or both\\n> sides of the Append node have a Gather under them, and you can push\\n> the partial aggregation down into the Gather; or if you hit a Foreign\\n> Scan and can have it do partial aggregation on the remote side, you\\n> win. But if you just have:\\n>\\n> Finalize Aggregate\\n> -> Append\\n> -> Partial Aggregate\\n> -> Thing One\\n> -> Partial Aggregate\\n> -> Thing Two\\n>\\n> Instead of:\\n>\\n> Aggregate\\n> -> Append\\n> -> Thing One\\n> -> Thing Two\\n>\\n> ...then I don't see the point, at least not for a single-group\\n> Aggregate or HashAggregate. For a GroupAggregate, Thing One and Thing\\n> Two might need to be sorted, and sorting two or more smaller data sets\\n> might be faster than sorting one larger data set, but I can't see us\\n> winning anything very big here.\\n>\\n> To be clear, I'm not saying we shouldn't do this. I just don't think\\n> it does much *by itself*. If you combine it with other optimizations\\n> that let the aggregation be pushed further down the plan tree, it\\n> could win big.\\n\\nYes, I agree, it's not a big win, at least not in the case of a serial\\nplan. If each branch of the UNION ALL could be processed in parallel,\\nthen that's different.\\n\\nIt's quite simple to test how much of a win it'll be in the serial\\ncase today, and yes, it's not much, but it's a bit.\\n\\ncreate table t1 as select x from generate_series(1,1000000) x(x);\\nvacuum analyze t1;\\nselect count(*) from (select * from t1 union all select * from t1) t;\\n count\\n---------\\n 2000000\\n(1 row)\\n\\nTime: 185.793 ms\\n\\n-- Mock up pushed down aggregation by using sum() as a combine\\nfunction for count(*)\\nselect sum(c) from (select count(*) c from t1 union all select\\ncount(*) from t1) t;\\n sum\\n---------\\n 2000000\\n(1 row)\\n\\nTime: 162.076 ms\\n\\nNot particularly incredible, but we don't normally turn our noses up\\nat a 14% improvement, so let's just see how complex it will be to\\nimplement, once the upper planner changes are done.\\n\\nBut as you mention about lack of ability to make use of pre-sorted\\nPath for each branch of the UNION ALL; I was really hoping Tom's patch\\nwill improve that part by allowing the planner to choose a pre-sorted\\nPath and perform a MergeAppend instead of an Append, which would allow\\npre-sorted input into a GroupAggregate node.\\n\\n-- \\n David Rowley 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, 22 Jan 2016 10:08:26 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 21, 2016 at 3:52 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> On 21 January 2016 at 15:53, Haribabu Kommi <kommi.haribabu@gmail.com>\\n> wrote:\\n>>\\n>> On Thu, Jan 21, 2016 at 1:33 PM, Haribabu Kommi\\n>> <kommi.haribabu@gmail.com> wrote:\\n>> >\\n>> > Here I attached updated patch of parallel aggregate based on the latest\\n>> > changes in master. Still it lack of cost comparison of normal aggregate\\n>> > to\\n>> > parallel aggregate because of difficulty. This cost comparison is\\n>> > required\\n>> > in parallel aggregate as this is having some regression when the number\\n>> > of groups are less in the query plan.\\n>> >\\n>>\\n>> Updated patch is attached after removing a warning in building group\\n>> aggregate path.\\n>\\n>\\n> Hi,\\n>\\n> Thanks for updating the patch. I'd like to look at this with priority, but\\n> can you post it on the Parallel Agg thread? that way anyone following there\\n> can chime in over there rather than here. I've still got a bit of work to\\n> do (in the not too distant future) on the serial/deserial part, so would be\\n> better to keep this thread for discussion on that.\\n\\nThanks for the details. Sorry for sending parallel aggregate patch in\\nthis thread.\\nI will take care of it from next time onward.\\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\": \"Fri, 22 Jan 2016 10:24:06 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 21, 2016 at 4:08 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> It's quite simple to test how much of a win it'll be in the serial\\n> case today, and yes, it's not much, but it's a bit.\\n>\\n> create table t1 as select x from generate_series(1,1000000) x(x);\\n> vacuum analyze t1;\\n> select count(*) from (select * from t1 union all select * from t1) t;\\n> count\\n> ---------\\n> 2000000\\n> (1 row)\\n>\\n> Time: 185.793 ms\\n>\\n> -- Mock up pushed down aggregation by using sum() as a combine\\n> function for count(*)\\n> select sum(c) from (select count(*) c from t1 union all select\\n> count(*) from t1) t;\\n> sum\\n> ---------\\n> 2000000\\n> (1 row)\\n>\\n> Time: 162.076 ms\\n>\\n> Not particularly incredible, but we don't normally turn our noses up\\n> at a 14% improvement, so let's just see how complex it will be to\\n> implement, once the upper planner changes are done.\\n\\nMumble mumble. Why is that even any faster? Just because we avoid\\nthe projection overhead of the Append node, which is a no-op here\\nanyway? If so, a planner change is one thought, but perhaps we also\\nought to look at whether we can't reduce the execution-time overhead.\\n\\n> But as you mention about lack of ability to make use of pre-sorted\\n> Path for each branch of the UNION ALL; I was really hoping Tom's patch\\n> will improve that part by allowing the planner to choose a pre-sorted\\n> Path and perform a MergeAppend instead of an Append, which would allow\\n> pre-sorted input into a GroupAggregate node.\\n\\nI won't hazard a guess on that point...\\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, 22 Jan 2016 09:55:12 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 20, 2016 at 11:06 AM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Wed, Jan 20, 2016 at 7:38 AM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n>> Agreed. So I've attached a version of the patch which does not have any of\\n>> the serialise/deserialise stuff in it.\\n>\\n> I re-reviewed this and have committed most of it with only minor\\n> kibitizing. A few notes:\\n\\n\\nThis commit has broken pg_dump. At least, I think this is the thread\\nreferencing this commit:\\n\\n\\ncommit a7de3dc5c346e07e0439275982569996e645b3c2\\nAuthor: Robert Haas <rhaas@postgresql.org>\\nDate: Wed Jan 20 13:46:50 2016 -0500\\n Support multi-stage aggregation.\\n\\n\\nIf I add an aggregate to an otherwise empty 9.4.5 database:\\n\\n\\nCREATE OR REPLACE FUNCTION first_agg ( anyelement, anyelement )\\nRETURNS anyelement LANGUAGE sql IMMUTABLE STRICT AS $$\\n SELECT $1;\\n$$;\\n\\n-- And then wrap an aggregate around it\\n-- aggregates do not support create or replace, alas.\\nCREATE AGGREGATE first (\\n sfunc = first_agg,\\n basetype = anyelement,\\n stype = anyelement\\n);\\n\\n\\nAnd then run pg_dump from 9.6.this on it, I get:\\n\\n\\npg_dump: column number -1 is out of range 0..17\\nSegmentation fault (core dumped)\\n\\nProgram terminated with signal 11, Segmentation fault.\\n#0 0x0000000000416b0b in dumpAgg (fout=0x1e551e0, agginfo=0x1e65ec0)\\nat pg_dump.c:12670\\n12670 if (strcmp(aggcombinefn, \\\"-\\\") != 0)\\n(gdb) bt\\n#0 0x0000000000416b0b in dumpAgg (fout=0x1e551e0, agginfo=0x1e65ec0)\\nat pg_dump.c:12670\\n#1 0x000000000041df7a in main (argc=<value optimized out>,\\nargv=<value optimized out>) at pg_dump.c:810\\n\\n\\nCheers,\\n\\nJeff\\n\\n\\n-- \\nSent via pgsql-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, 22 Jan 2016 12:17:05 -0800\",\n \"msg_from\": \"Jeff Janes <jeff.janes@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 23 January 2016 at 09:17, Jeff Janes <jeff.janes@gmail.com> wrote:\\n> On Wed, Jan 20, 2016 at 11:06 AM, Robert Haas <robertmhaas@gmail.com> wrote:\\n>> On Wed, Jan 20, 2016 at 7:38 AM, David Rowley\\n>> <david.rowley@2ndquadrant.com> wrote:\\n>>> Agreed. So I've attached a version of the patch which does not have any of\\n>>> the serialise/deserialise stuff in it.\\n>>\\n>> I re-reviewed this and have committed most of it with only minor\\n>> kibitizing. A few notes:\\n>\\n>\\n> This commit has broken pg_dump. At least, I think this is the thread\\n> referencing this commit:\\n>\\n...\\n> pg_dump: column number -1 is out of range 0..17\\n> Segmentation fault (core dumped)\\n>\\n> Program terminated with signal 11, Segmentation fault.\\n> #0 0x0000000000416b0b in dumpAgg (fout=0x1e551e0, agginfo=0x1e65ec0)\\n> at pg_dump.c:12670\\n> 12670 if (strcmp(aggcombinefn, \\\"-\\\") != 0)\\n> (gdb) bt\\n> #0 0x0000000000416b0b in dumpAgg (fout=0x1e551e0, agginfo=0x1e65ec0)\\n> at pg_dump.c:12670\\n> #1 0x000000000041df7a in main (argc=<value optimized out>,\\n> argv=<value optimized out>) at pg_dump.c:810\\n\\nYikes :-( I will look at this with priority in the next few hours.\\n\\nThanks for the report.\\n\\n-- \\n David Rowley 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, 23 Jan 2016 09:44:47 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 23 January 2016 at 09:44, David Rowley <david.rowley@2ndquadrant.com> wrote:\\n> On 23 January 2016 at 09:17, Jeff Janes <jeff.janes@gmail.com> wrote:\\n>> On Wed, Jan 20, 2016 at 11:06 AM, Robert Haas <robertmhaas@gmail.com> wrote:\\n>>> On Wed, Jan 20, 2016 at 7:38 AM, David Rowley\\n>>> <david.rowley@2ndquadrant.com> wrote:\\n>>>> Agreed. So I've attached a version of the patch which does not have any of\\n>>>> the serialise/deserialise stuff in it.\\n>>>\\n>>> I re-reviewed this and have committed most of it with only minor\\n>>> kibitizing. A few notes:\\n>>\\n>>\\n>> This commit has broken pg_dump. At least, I think this is the thread\\n>> referencing this commit:\\n>>\\n> ...\\n>> pg_dump: column number -1 is out of range 0..17\\n>> Segmentation fault (core dumped)\\n>>\\n>> Program terminated with signal 11, Segmentation fault.\\n>> #0 0x0000000000416b0b in dumpAgg (fout=0x1e551e0, agginfo=0x1e65ec0)\\n>> at pg_dump.c:12670\\n>> 12670 if (strcmp(aggcombinefn, \\\"-\\\") != 0)\\n>> (gdb) bt\\n>> #0 0x0000000000416b0b in dumpAgg (fout=0x1e551e0, agginfo=0x1e65ec0)\\n>> at pg_dump.c:12670\\n>> #1 0x000000000041df7a in main (argc=<value optimized out>,\\n>> argv=<value optimized out>) at pg_dump.c:810\\n>\\n> Yikes :-( I will look at this with priority in the next few hours.\\n>\\n> Thanks for the report.\\n\\nIt seems that I must have mistakenly believed that non-existing\\ncolumns for previous versions were handled using the power of magic.\\nTurns out that I was wrong, and they need to be included as dummy\\ncolumns in the queries for previous versions.\\n\\nThe attached fixes the problem.\\n\\n-- \\n David Rowley 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\": \"Sat, 23 Jan 2016 13:53:05 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 22, 2016 at 4:53 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n>\\n> It seems that I must have mistakenly believed that non-existing\\n> columns for previous versions were handled using the power of magic.\\n> Turns out that I was wrong, and they need to be included as dummy\\n> columns in the queries for previous versions.\\n>\\n> The attached fixes the problem.\\n\\nYep, that fixes it.\\n\\nThanks,\\n\\nJeff\\n\\n\\n-- \\nSent via pgsql-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, 23 Jan 2016 15:26:12 -0800\",\n \"msg_from\": \"Jeff Janes <jeff.janes@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Sat, Jan 23, 2016 at 6:26 PM, Jeff Janes <jeff.janes@gmail.com> wrote:\\n> On Fri, Jan 22, 2016 at 4:53 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n>>\\n>> It seems that I must have mistakenly believed that non-existing\\n>> columns for previous versions were handled using the power of magic.\\n>> Turns out that I was wrong, and they need to be included as dummy\\n>> columns in the queries for previous versions.\\n>>\\n>> The attached fixes the problem.\\n>\\n> Yep, that fixes it.\\n\\nCommitted. Apologies for the delay.\\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, 27 Jan 2016 21:48:56 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 21, 2016 at 3:42 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> On 21 January 2016 at 08:06, Robert Haas <robertmhaas@gmail.com> wrote:\\n>>\\n>> On Wed, Jan 20, 2016 at 7:38 AM, David Rowley\\n>> <david.rowley@2ndquadrant.com> wrote:\\n>> > Agreed. So I've attached a version of the patch which does not have any\\n>> > of\\n>> > the serialise/deserialise stuff in it.\\n>>\\n>> I re-reviewed this and have committed most of it with only minor\\n>> kibitizing. A few notes:\\n>\\n>\\n> I've attached the re-based remainder, which includes the serial/deserial\\n> again.\\n>\\n> I'll submit this part to March 'fest, where hopefully we'll also have\\n> something to utilise it.\\n>\\n\\nWhile testing parallel aggregate with float4 and float8 types based on\\nthe latest patch,\\nI found the following problems,\\n\\n+ /*\\n+ * For partial aggregation we must adjust the return types of\\n+ * the Aggrefs\\n+ */\\n+ if (!aggplan->finalizeAggs)\\n+ set_partialagg_aggref_types(root, plan,\\n+ aggplan->serialStates);\\n\\n[...]\\n\\n+ aggref->aggtype = aggform->aggserialtype;\\n+ else\\n+ aggref->aggtype = aggform->aggtranstype;\\n\\nChanging the aggref->aggtype with aggtranstype or aggserialtype will\\nonly gets it changed in\\npartial aggregate plan, as set_upper_references starts from the top\\nplan and goes\\nfurther. Because of this reason, the targetlist contains for the node\\nbelow finalize\\naggregate are still points to original type only.\\n\\nTo fix this problem, I tried updating the targetlist aggref->aggtype\\nwith transtype during\\naggregate plan itself, that leads to a problem in setting upper plan\\nreferences. This is\\nbecause, while fixing the aggregate reference of upper plans after\\npartial aggregate,\\nthe aggref at upper plan nodes doesn't match with aggref that is\\ncoming from partial\\naggregate node because of aggtype difference in _equalAggref function.\\n\\nCOMPARE_SCALAR_FIELD(aggtype);\\n\\nTemporarily i corrected it compare it against aggtranstype and\\naggserialtype also then\\nit works fine. I don't see that change as correct approach. Do you\\nhave any better idea\\nto solve this problem?\\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, 8 Feb 2016 12:37:02 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 1/20/16 11:42 PM, David Rowley wrote:\\n> On 21 January 2016 at 08:06, Robert Haas <robertmhaas@gmail.com\\n> <mailto:robertmhaas@gmail.com>> wrote:\\n> \\n> On Wed, Jan 20, 2016 at 7:38 AM, David Rowley\\n> <david.rowley@2ndquadrant.com <mailto:david.rowley@2ndquadrant.com>>\\n> wrote:\\n> > Agreed. So I've attached a version of the patch which does not have any of\\n> > the serialise/deserialise stuff in it.\\n> \\n> I re-reviewed this and have committed most of it with only minor\\n> kibitizing. A few notes:\\n> \\n> \\n> I've attached the re-based remainder, which includes the serial/deserial\\n> again.\\n> \\n> I'll submit this part to March 'fest, where hopefully we'll also have\\n> something to utilise it.\\n\\nAs far as I can tell this is the most recent version of the patch but it\\ndoesn't apply on master. I'm guessing that's mostly because of Tom's\\nUPP patch.\\n\\nThis is a long and confusing thread and I should know since I just read\\nthrough the entire thing. I'm setting this back to \\\"waiting for author\\\"\\nand I'd like see a rebased patch and a summary of what's in the patch\\nbefore setting it back to \\\"needs review\\\".\\n\\n-- \\n-David\\ndavid@pgmasters.net\",\n \"msg_date\": \"Fri, 11 Mar 2016 12:28:51 -0500\",\n \"msg_from\": \"David Steele <david@pgmasters.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 12 March 2016 at 06:28, David Steele <david@pgmasters.net> wrote:\\n> As far as I can tell this is the most recent version of the patch but it\\n> doesn't apply on master. I'm guessing that's mostly because of Tom's\\n> UPP patch.\\n>\\n> This is a long and confusing thread and I should know since I just read\\n> through the entire thing. I'm setting this back to \\\"waiting for author\\\"\\n> and I'd like see a rebased patch and a summary of what's in the patch\\n> before setting it back to \\\"needs review\\\".\\n\\nMany thank for taking the time to read through this thread.\\n\\nLet me try to summarise the current status.\\n\\nOverview:\\nCombine Aggregate states is required to allow aggregate functions to\\nbe calculated in multiple stages. The primary intended use for this,\\nas of now is for allowing parallel aggregate. A simple combine\\nfunction for count(*) would perform something very similar to sum() of\\nall of the partially calculated aggregate states. In this case\\n\\\"partially\\\" means aggregated states without the final function called.\\n\\nStatus:\\nAs of [1] we have basic infrastructure in place to allow this, but\\nthat commit only added the infrastructure for aggregate functions\\nwhich don't have an INTERNAL state type. Many aggregates do have an\\ninternal state type, and for the case of parallel aggregate, since\\nthese internal states are just passed around as pointers to memory\\nwhich belongs to the process wherever the member was allocated, we're\\nunable to pass these back to the main process for the final aggregate\\nstage. This patch allows this to happen by way of serializing these\\ninternal states into some type that can be copied into the main\\nbackend process.\\nRobert was quite right to leave this part out of the commit as at the\\ntime the method I was using was under debate, and there was no reason\\nto hold things up, as the more simple case could be committed to allow\\nwork to continue on features which would use this new infrastructure.\\n\\nThese threads are a bit of a mess as there's often confusion (from me\\ntoo) about which code belongs in which patch. Quite possibly the\\nsetrefs.c stuff of the parallel aggregate patch actually do belong\\nhere, but it's probably not worth shuffling things around as I think\\nthe parallel agg stuff should go in before the additional serial\\nstates stuff.\\n\\nCurrent patch:\\nI've now updated the patch to base it on top of the parallel aggregate\\npatch in [2]. To apply the attached, you must apply [2] first!\\nThe changes to numeric.c are not really intended for commit for now. I\\nplan on spending a bit of time first on looking at using send\\nfunctions and bytea serial types, which if that works should be far\\nmore efficient than converting the serialized output into a string.\\nThe example combine function in numeric.c is just intended to show\\nthat this does actually work.\\n\\nI've also left one opr_sanity test failing for now. This test is\\ncomplaining that the de-serialisation function returns an INTERNAL\\ntype and does not accept one (which would disallow it from being\\ncalled from SQL interface). I think this is safe to allow as I've\\ncoded the function to fail if it's not called in aggregate context.\\nPerhaps I'm wrong though, so I left the test failing to highlight that\\nI'd like someone else to look and maybe think about this a bit. It\\nseems very important not to get this part wrong.\\n\\nComments and reviews etc all welcome. Please remember to apply [2]\\nbefore the attached.\\n\\n[1] http://git.postgresql.org/gitweb/?p=postgresql.git;a=commit;h=a7de3dc5c346e07e0439275982569996e645b3c2\\n[2] http://www.postgresql.org/message-id/CAKJS1f9Tr-+9aKWZ1XuHUnEJZ7GKKfo45b+4fCNj8DkrDZYK4g@mail.gmail.com\\n\\n-- \\n David Rowley 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, 14 Mar 2016 15:20:08 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 14 March 2016 at 15:20, David Rowley <david.rowley@2ndquadrant.com> wrote:\\n> Current patch:\\n> I've now updated the patch to base it on top of the parallel aggregate\\n> patch in [2]. To apply the attached, you must apply [2] first!\\n\\n...\\n> [2] http://www.postgresql.org/message-id/CAKJS1f9Tr-+9aKWZ1XuHUnEJZ7GKKfo45b+4fCNj8DkrDZYK4g@mail.gmail.com\\n\\nThe attached fixes a small thinko in apply_partialaggref_nodes().\\n\\n\\n-- \\n David Rowley 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, 14 Mar 2016 17:45:55 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 03/14/2016 05:45 AM, David Rowley wrote:\\n> On 14 March 2016 at 15:20, David Rowley <david.rowley@2ndquadrant.com> wrote:\\n>> Current patch:\\n>> I've now updated the patch to base it on top of the parallel aggregate\\n>> patch in [2]. To apply the attached, you must apply [2] first!\\n>\\n> ...\\n>> [2] http://www.postgresql.org/message-id/CAKJS1f9Tr-+9aKWZ1XuHUnEJZ7GKKfo45b+4fCNj8DkrDZYK4g@mail.gmail.com\\n>\\n> The attached fixes a small thinko in apply_partialaggref_nodes().\\n\\nAfter looking at the parallel aggregate patch, I also looked at this \\none, as it's naturally related. Sadly I haven't found any issue that I \\ncould nag about ;-) The patch seems well baked, as it was in the oven \\nfor quite a long time already.\\n\\nThe one concern I do have is that it only adds (de)serialize functions \\nfor SUM(numeric) and AVG(numeric). I think it's reasonable not to \\ninclude that into the patch, but it will look a bit silly if that's all \\nthat gets into 9.6.\\n\\nIt's true plenty of aggregates is supported out of the box, as they use \\nfixed-length data types for the aggregate state. But imagine users happy \\nthat their aggregate queries get parallelized, only to find out that \\nsum(int8) disables that :-/\\n\\nSo I think we should try to support as many of the aggregates using \\ninternal as possible - it seems quite straight-forward to do that at \\nleast for the aggregates using the same internal representation (avg, \\nsum, var_samp, var_pop, variance, ...). That's 26 aggregates out of the \\n45 with aggtranstype=INTERNAL.\\n\\nFor the remaining aggregates (jsonb_*. json_*, string_agg, array_agg, \\npercentile_) it's probably more complicated to serialize the state, and \\nmaybe even impossible to do that correctly (e.g. string_agg accumulates \\nthe strings in the order as received, but partial paths would build \\nprivate lists - not sure if this can actually happen in practice).\\n\\nI think it would be good to actually document which aggregates support \\nparallel execution, and which don't. Currently the docs don't mention \\nthis at all, and it's tricky for regular users to deduce that as it's \\nrelated to the type of the internal state and not to the input types. An \\nexample of that is the SUM(bigint) example mentioned above.\\n\\nThat's actually the one thing I think the patch is missing.\\n\\nregards\\n\\n-- \\nTomas Vondra 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, 15 Mar 2016 18:39:54 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 16 March 2016 at 06:39, Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote:\\n> After looking at the parallel aggregate patch, I also looked at this one, as\\n> it's naturally related. Sadly I haven't found any issue that I could nag\\n> about ;-) The patch seems well baked, as it was in the oven for quite a long\\n> time already.\\n\\nThanks for looking at this.\\n\\n> The one concern I do have is that it only adds (de)serialize functions for\\n> SUM(numeric) and AVG(numeric). I think it's reasonable not to include that\\n> into the patch, but it will look a bit silly if that's all that gets into\\n> 9.6.\\n\\nYes me too, so I spent several hours yesterday writing all of the\\ncombine functions and serialisation/deserialisation that are required\\nfor all of SUM(), AVG() STDDEV*(). I also noticed that I had missed\\nusing some existing functions for bool_and() and bool_or() so I added\\nthose to pg_aggregate.h. I'm just chasing down a crash bug on\\nHAVE_INT128 enabled builds, so should be posting a patch quite soon. I\\ndidn't touch the FLOAT4 and FLOAT8 aggregates as I believe Haribabu\\nhas a patch for that over on the parallel aggregate thread. I've not\\nlooked at it in detail yet.\\n\\nIf Haribabu's patch does all that's required for the numerical\\naggregates for floating point types then the status of covered\\naggregates is (in order of pg_aggregate.h):\\n\\n* AVG() complete coverage\\n* SUM() complete coverage\\n* MAX() complete coverage\\n* MIN() complete coverage\\n* COUNT() complete coverage\\n* STDDEV + friends complete coverage\\n* regr_*,covar_pop,covar_samp,corr not touched these.\\n* bool*() complete coverage\\n* bitwise aggs. complete coverage\\n* Remaining are not touched. I see diminishing returns with making\\nthese parallel for now. I think I might not be worth pushing myself\\nany harder to make these ones work.\\n\\nDoes what I have done + floating point aggs from Haribabu seem\\nreasonable for 9.6?\\n\\n> I think it would be good to actually document which aggregates support\\n> parallel execution, and which don't. Currently the docs don't mention this\\n> at all, and it's tricky for regular users to deduce that as it's related to\\n> the type of the internal state and not to the input types. An example of\\n> that is the SUM(bigint) example mentioned above.\\n\\nI agree. I will look for a suitable place.\\n\\n-- \\n David Rowley 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, 16 Mar 2016 10:34:37 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 16, 2016 at 8:34 AM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> On 16 March 2016 at 06:39, Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote:\\n>> After looking at the parallel aggregate patch, I also looked at this one, as\\n>> it's naturally related. Sadly I haven't found any issue that I could nag\\n>> about ;-) The patch seems well baked, as it was in the oven for quite a long\\n>> time already.\\n>\\n> Thanks for looking at this.\\n>\\n>> The one concern I do have is that it only adds (de)serialize functions for\\n>> SUM(numeric) and AVG(numeric). I think it's reasonable not to include that\\n>> into the patch, but it will look a bit silly if that's all that gets into\\n>> 9.6.\\n>\\n> Yes me too, so I spent several hours yesterday writing all of the\\n> combine functions and serialisation/deserialisation that are required\\n> for all of SUM(), AVG() STDDEV*(). I also noticed that I had missed\\n> using some existing functions for bool_and() and bool_or() so I added\\n> those to pg_aggregate.h. I'm just chasing down a crash bug on\\n> HAVE_INT128 enabled builds, so should be posting a patch quite soon. I\\n> didn't touch the FLOAT4 and FLOAT8 aggregates as I believe Haribabu\\n> has a patch for that over on the parallel aggregate thread. I've not\\n> looked at it in detail yet.\\n\\nThe additional combine function patch that I posted handles all float4 and\\nfloat8 aggregates. There is an OID conflict with the latest source code,\\nI will update the patch and post it in that thread.\\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, 16 Mar 2016 21:54:44 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 16 March 2016 at 10:34, David Rowley <david.rowley@2ndquadrant.com> wrote:\\n> If Haribabu's patch does all that's required for the numerical\\n> aggregates for floating point types then the status of covered\\n> aggregates is (in order of pg_aggregate.h):\\n>\\n> * AVG() complete coverage\\n> * SUM() complete coverage\\n> * MAX() complete coverage\\n> * MIN() complete coverage\\n> * COUNT() complete coverage\\n> * STDDEV + friends complete coverage\\n> * regr_*,covar_pop,covar_samp,corr not touched these.\\n> * bool*() complete coverage\\n> * bitwise aggs. complete coverage\\n> * Remaining are not touched. I see diminishing returns with making\\n> these parallel for now. I think I might not be worth pushing myself\\n> any harder to make these ones work.\\n>\\n> Does what I have done + floating point aggs from Haribabu seem\\n> reasonable for 9.6?\\n\\nI've attached a series of patches.\\n\\nPatch 1:\\nThis is the parallel aggregate patch, not intended for review here.\\nHowever, all further patches are based on this, and this adds the\\nrequired planner changes to make it possible to test patches 2 and 3.\\n\\nPatch 2:\\nThis adds the serial/deserial aggregate infrastructure, pg_dump\\nsupport, CREATE AGGREGATE changes, and nodeAgg.c changes to have it\\nserialise and deserialise aggregate states when instructed to do so.\\n\\nPatch 3:\\nThis adds a boat load of serial/deserial functions, and combine\\nfunctions for most of the built-in numerical aggregate functions. It\\nalso contains some regression tests which should really be in patch 2,\\nbut I with patch 2 there's no suitable serialisation or\\nde-serialisation functions to test CREATE AGGREGATE with. I think\\nhaving them here is ok, as patch 2 is quite useless without patch 3\\nanyway.\\n\\nAnother thing to note about this patch is that I've gone and created\\nserial/de-serial functions for when PolyNumAggState both require\\nsumX2, and don't require sumX2. I had thought about perhaps putting an\\nextra byte in the serial format to indicate if a sumX2 is included,\\nbut I ended up not doing it this way. I don't really want these serial\\nformats getting too complex as we might like to do fun things like\\npass them along to sharded servers one day, so it might be nice to\\nkeep them simple.\\n\\nPatch 4:\\nAdds a bunch of opr_sanity regression tests. This could be part of\\npatch 3, but 3 was quite big already.\\n\\nPatch 5:\\nAdds some documentation to indicate which aggregates allow partial mode.\\n\\nComments welcome.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\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, 17 Mar 2016 00:03:58 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 16 March 2016 at 23:54, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:\\n> On Wed, Mar 16, 2016 at 8:34 AM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n>> Yes me too, so I spent several hours yesterday writing all of the\\n>> combine functions and serialisation/deserialisation that are required\\n>> for all of SUM(), AVG() STDDEV*(). I also noticed that I had missed\\n>> using some existing functions for bool_and() and bool_or() so I added\\n>> those to pg_aggregate.h. I'm just chasing down a crash bug on\\n>> HAVE_INT128 enabled builds, so should be posting a patch quite soon. I\\n>> didn't touch the FLOAT4 and FLOAT8 aggregates as I believe Haribabu\\n>> has a patch for that over on the parallel aggregate thread. I've not\\n>> looked at it in detail yet.\\n>\\n> The additional combine function patch that I posted handles all float4 and\\n> float8 aggregates. There is an OID conflict with the latest source code,\\n> I will update the patch and post it in that thread.\\n\\nThanks! I just send a series of patches which add a whole host of\\nserial/deserial functions, and a patch which adds some documentation.\\nMaybe you could base your patch on the 0005 patch, and update the\\ndocuments to remove the \\\"All types apart from floating-point types\\\"\\ntext and replace that with \\\"Yes\\\".\\n\\n-- \\n David Rowley 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, 17 Mar 2016 00:08:49 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 03/16/2016 12:03 PM, David Rowley wrote:\\n> On 16 March 2016 at 10:34, David Rowley <david.rowley@2ndquadrant.com> wrote:\\n>> If Haribabu's patch does all that's required for the numerical\\n>> aggregates for floating point types then the status of covered\\n>> aggregates is (in order of pg_aggregate.h):\\n>>\\n>> * AVG() complete coverage\\n>> * SUM() complete coverage\\n>> * MAX() complete coverage\\n>> * MIN() complete coverage\\n>> * COUNT() complete coverage\\n>> * STDDEV + friends complete coverage\\n>> * regr_*,covar_pop,covar_samp,corr not touched these.\\n>> * bool*() complete coverage\\n>> * bitwise aggs. complete coverage\\n>> * Remaining are not touched. I see diminishing returns with making\\n>> these parallel for now. I think I might not be worth pushing myself\\n>> any harder to make these ones work.\\n>>\\n>> Does what I have done + floating point aggs from Haribabu seem\\n>> reasonable for 9.6?\\n>\\n> I've attached a series of patches.\\n\\nThanks. Haven't really done much testing of this, aside from reading \\nthrough the patches and \\\"it compiles\\\".\\n>\\n> Patch 1:\\n> This is the parallel aggregate patch, not intended for review here.\\n> However, all further patches are based on this, and this adds the\\n> required planner changes to make it possible to test patches 2 and 3.\\n>\\n> Patch 2:\\n> This adds the serial/deserial aggregate infrastructure, pg_dump\\n> support, CREATE AGGREGATE changes, and nodeAgg.c changes to have it\\n> serialise and deserialise aggregate states when instructed to do so.\\n>\\n> Patch 3:\\n> This adds a boat load of serial/deserial functions, and combine\\n> functions for most of the built-in numerical aggregate functions. It\\n> also contains some regression tests which should really be in patch 2,\\n> but I with patch 2 there's no suitable serialisation or\\n> de-serialisation functions to test CREATE AGGREGATE with. I think\\n> having them here is ok, as patch 2 is quite useless without patch 3\\n> anyway.\\n\\nI don't see how you could move the tests into #2 when the functions are \\ndefined in #3? IMHO this is the right place for the regression tests.\\n\\n>\\n> Another thing to note about this patch is that I've gone and created\\n> serial/de-serial functions for when PolyNumAggState both require\\n> sumX2, and don't require sumX2. I had thought about perhaps putting an\\n> extra byte in the serial format to indicate if a sumX2 is included,\\n> but I ended up not doing it this way. I don't really want these serial\\n> formats getting too complex as we might like to do fun things like\\n> pass them along to sharded servers one day, so it might be nice to\\n> keep them simple.\\n\\nHmmm, I've noticed that while eyeballing the diffs, and I'm not sure if \\nit's worth the additional complexity at this point. I mean, one byte is \\nhardly going to make a measurable difference - we're probably wasting \\nmore than that due to alignment, for example.\\n\\nAs you've mentioned sharded servers - how stable should the serialized \\nformat be? I've assumed it to be transient, i.e. in the extreme case it \\nmight change after restarting a database - for the parallel aggregate \\nthat's clearly sufficient.\\n\\nBut if we expect this to eventually work in a sharded environment, \\nthat's going to be way more complicated. I guess in these cases we could \\nrely on implicit format versioning, via the major the version (doubt \\nwe'd tweak the format in a minor version anyway).\\n\\nI'm not sure the sharding is particularly useful argument at this point, \\nconsidering we don't really know if the current format is actually a \\ngood match for that.\\n\\nregards\\n\\n-- \\nTomas Vondra 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, 17 Mar 2016 02:25:02 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 16, 2016 at 10:08 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> On 16 March 2016 at 23:54, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:\\n>> On Wed, Mar 16, 2016 at 8:34 AM, David Rowley\\n>> <david.rowley@2ndquadrant.com> wrote:\\n>>> Yes me too, so I spent several hours yesterday writing all of the\\n>>> combine functions and serialisation/deserialisation that are required\\n>>> for all of SUM(), AVG() STDDEV*(). I also noticed that I had missed\\n>>> using some existing functions for bool_and() and bool_or() so I added\\n>>> those to pg_aggregate.h. I'm just chasing down a crash bug on\\n>>> HAVE_INT128 enabled builds, so should be posting a patch quite soon. I\\n>>> didn't touch the FLOAT4 and FLOAT8 aggregates as I believe Haribabu\\n>>> has a patch for that over on the parallel aggregate thread. I've not\\n>>> looked at it in detail yet.\\n>>\\n>> The additional combine function patch that I posted handles all float4 and\\n>> float8 aggregates. There is an OID conflict with the latest source code,\\n>> I will update the patch and post it in that thread.\\n>\\n> Thanks! I just send a series of patches which add a whole host of\\n> serial/deserial functions, and a patch which adds some documentation.\\n> Maybe you could base your patch on the 0005 patch, and update the\\n> documents to remove the \\\"All types apart from floating-point types\\\"\\n> text and replace that with \\\"Yes\\\".\\n\\nHere I attached updated float aggregates patch based on 0005 patch.\\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\": \"Thu, 17 Mar 2016 14:30:52 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 17 March 2016 at 16:30, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:\\n> On Wed, Mar 16, 2016 at 10:08 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n>> On 16 March 2016 at 23:54, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:\\n>>> On Wed, Mar 16, 2016 at 8:34 AM, David Rowley\\n>>> <david.rowley@2ndquadrant.com> wrote:\\n>>>> Yes me too, so I spent several hours yesterday writing all of the\\n>>>> combine functions and serialisation/deserialisation that are required\\n>>>> for all of SUM(), AVG() STDDEV*(). I also noticed that I had missed\\n>>>> using some existing functions for bool_and() and bool_or() so I added\\n>>>> those to pg_aggregate.h. I'm just chasing down a crash bug on\\n>>>> HAVE_INT128 enabled builds, so should be posting a patch quite soon. I\\n>>>> didn't touch the FLOAT4 and FLOAT8 aggregates as I believe Haribabu\\n>>>> has a patch for that over on the parallel aggregate thread. I've not\\n>>>> looked at it in detail yet.\\n>>>\\n>>> The additional combine function patch that I posted handles all float4 and\\n>>> float8 aggregates. There is an OID conflict with the latest source code,\\n>>> I will update the patch and post it in that thread.\\n>>\\n>> Thanks! I just send a series of patches which add a whole host of\\n>> serial/deserial functions, and a patch which adds some documentation.\\n>> Maybe you could base your patch on the 0005 patch, and update the\\n>> documents to remove the \\\"All types apart from floating-point types\\\"\\n>> text and replace that with \\\"Yes\\\".\\n>\\n> Here I attached updated float aggregates patch based on 0005 patch.\\n\\nGreat! Thanks for sending that.\\n\\nI just had a quick skim over the patch and noticed the naming\\nconvention you're using for the combine function is *_pl, and you have\\nfloat8_pl. There's already a function named float8pl() which is quite\\nclose to what you have. I've been sticking to *_combine() for these,\\nso maybe float8_combine() and float8_regr_combine() are better names.\\n\\n-- \\n David Rowley 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, 18 Mar 2016 00:53:04 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 03/17/2016 12:53 PM, David Rowley wrote:\\n >\\n...\\n >\\n> I just had a quick skim over the patch and noticed the naming\\n> convention you're using for the combine function is *_pl, and you have\\n> float8_pl. There's already a function named float8pl() which is quite\\n> close to what you have. I've been sticking to *_combine() for these,\\n> so maybe float8_combine() and float8_regr_combine() are better names.\\n\\n+1 to the _combine naming convention.\\n\\nregards\\n\\n-- \\nTomas Vondra 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, 17 Mar 2016 12:59:42 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 17, 2016 at 10:59 PM, Tomas Vondra\\n<tomas.vondra@2ndquadrant.com> wrote:\\n> Hi,\\n>\\n> On 03/17/2016 12:53 PM, David Rowley wrote:\\n>>\\n> ...\\n>>\\n>>\\n>> I just had a quick skim over the patch and noticed the naming\\n>> convention you're using for the combine function is *_pl, and you have\\n>> float8_pl. There's already a function named float8pl() which is quite\\n>> close to what you have. I've been sticking to *_combine() for these,\\n>> so maybe float8_combine() and float8_regr_combine() are better names.\\n>\\n>\\n> +1 to the _combine naming convention.\\n\\nThanks for the input. Makes sense, updated patch is attached with\\nthe changes.\\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\": \"Fri, 18 Mar 2016 11:39:09 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 18 March 2016 at 13:39, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:\\n> On Thu, Mar 17, 2016 at 10:59 PM, Tomas Vondra\\n> <tomas.vondra@2ndquadrant.com> wrote:\\n>> Hi,\\n>>\\n>> On 03/17/2016 12:53 PM, David Rowley wrote:\\n>>>\\n>> ...\\n>>>\\n>>>\\n>>> I just had a quick skim over the patch and noticed the naming\\n>>> convention you're using for the combine function is *_pl, and you have\\n>>> float8_pl. There's already a function named float8pl() which is quite\\n>>> close to what you have. I've been sticking to *_combine() for these,\\n>>> so maybe float8_combine() and float8_regr_combine() are better names.\\n>>\\n>>\\n>> +1 to the _combine naming convention.\\n>\\n> Thanks for the input. Makes sense, updated patch is attached with\\n> the changes.\\n\\nI've had a look over this. I had to first base it on the 0005 patch,\\nas it seemed like the pg_aggregate.h changes didn't include the\\nserialfn and deserialfn changes, and an OID was being consumed by\\nanother function I added in patch 0003.\\n\\nOn testing I also noticed some wrong results, which on investigation,\\nare due to the wrong array elements being added together.\\n\\nFor example:\\n\\npostgres=# select stddev(num) from f;\\n stddev\\n------------------\\n 28867.5149028984\\n(1 row)\\n\\n\\npostgres=# set max_parallel_degree=8;\\nSET\\npostgres=# select stddev(num) from f;\\n stddev\\n--------\\n 0\\n(1 row)\\n\\n+ N += transvalues2[0];\\n+ sumX += transvalues2[1];\\n+ CHECKFLOATVAL(sumX, isinf(transvalues1[1]) || isinf(transvalues2[1]), true);\\n+ sumX2 += transvalues1[2];\\n\\nThe last line should use transvalues2[2], not transvalues1[2].\\n\\nThere's also quite a few mistakes in float8_regr_combine()\\n\\n+ sumX2 += transvalues2[2];\\n+ CHECKFLOATVAL(sumX2, isinf(transvalues1[2]) || isinf(transvalues2[1]), true);\\n\\nWrong array element on isinf() check\\n\\n+ sumY2 += transvalues2[4];\\n+ CHECKFLOATVAL(sumY2, isinf(transvalues1[4]) || isinf(transvalues2[3]), true);\\n\\nWrong array element on isinf() check\\n\\n+ sumXY += transvalues2[5];\\n+ CHECKFLOATVAL(sumXY, isinf(transvalues1[5]) || isinf(transvalues2[1]) ||\\n+ isinf(transvalues2[3]), true);\\n\\nWrong array element on isinf() check, and the final\\nisinf(transvalues2[3]) check does not need to be there.\\n\\nI've re-based the patch and fixed these already, so will send updated\\npatches shortly.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sun, 20 Mar 2016 16:23:03 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 17 March 2016 at 14:25, Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote:\\n> On 03/16/2016 12:03 PM, David Rowley wrote:\\n>> Patch 2:\\n>> This adds the serial/deserial aggregate infrastructure, pg_dump\\n>> support, CREATE AGGREGATE changes, and nodeAgg.c changes to have it\\n>> serialise and deserialise aggregate states when instructed to do so.\\n>>\\n>> Patch 3:\\n>> This adds a boat load of serial/deserial functions, and combine\\n>> functions for most of the built-in numerical aggregate functions. It\\n>> also contains some regression tests which should really be in patch 2,\\n>> but I with patch 2 there's no suitable serialisation or\\n>> de-serialisation functions to test CREATE AGGREGATE with. I think\\n>> having them here is ok, as patch 2 is quite useless without patch 3\\n>> anyway.\\n>\\n> I don't see how you could move the tests into #2 when the functions are\\n> defined in #3? IMHO this is the right place for the regression tests.\\n\\nYeah, but the CREATE AGGREGATE changes which are being tested in 0003\\nwere actually added in 0002. I think 0002 is the right place to test\\nthe changes to CREATE AGGREGATE, but since there's a complete lack of\\nfunctions to use, then I've just delayed until 0003.\\n\\n>> Another thing to note about this patch is that I've gone and created\\n>> serial/de-serial functions for when PolyNumAggState both require\\n>> sumX2, and don't require sumX2. I had thought about perhaps putting an\\n>> extra byte in the serial format to indicate if a sumX2 is included,\\n>> but I ended up not doing it this way. I don't really want these serial\\n>> formats getting too complex as we might like to do fun things like\\n>> pass them along to sharded servers one day, so it might be nice to\\n>> keep them simple.\\n>\\n>\\n> Hmmm, I've noticed that while eyeballing the diffs, and I'm not sure if it's\\n> worth the additional complexity at this point. I mean, one byte is hardly\\n> going to make a measurable difference - we're probably wasting more than\\n> that due to alignment, for example.\\n\\nI don't think any alignment gets done here. Look at pq_sendint().\\nDid you mean the complexity of having extra functions, or having the\\nextra byte to check in the deserial func?\\n\\n> As you've mentioned sharded servers - how stable should the serialized\\n> format be? I've assumed it to be transient, i.e. in the extreme case it\\n> might change after restarting a database - for the parallel aggregate that's\\n> clearly sufficient.\\n>\\n> But if we expect this to eventually work in a sharded environment, that's\\n> going to be way more complicated. I guess in these cases we could rely on\\n> implicit format versioning, via the major the version (doubt we'd tweak the\\n> format in a minor version anyway).\\n>\\n> I'm not sure the sharding is particularly useful argument at this point,\\n> considering we don't really know if the current format is actually a good\\n> match for that.\\n\\nPerhaps you're right. At this stage I've no idea if we'd want to\\nsupport shards on varying major versions. I think probably not, since\\nthe node write functions might not be compatible with the node read\\nfunctions on the other server.\\n\\nI've attached another series of patches:\\n\\n0001: This is the latest Parallel Aggregate Patch, not intended for\\nreview here, but is required for the remaining patches. This patch has\\nchanged quite a bit from the previous one that I posted here, and the\\nremaining patches needed re-based due to those changes.\\n\\n0002: Adds serial/de-serial function support to CREATE AGGREGATE,\\ncontains minor fix-ups from last version.\\n\\n0003: Adds various combine/serial/de-serial functions for the standard\\nset of aggregates, apart from most float8 aggregates.\\n\\n0004: Adds regression tests for 0003 pg_aggregate.h changes.\\n\\n0005: Documents to mention which aggregate functions support partial mode.\\n\\n0006: Re-based version of Haribabu's floating point aggregate support,\\ncontaining some fixes by me.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\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, 20 Mar 2016 16:48:44 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 03/20/2016 04:48 AM, David Rowley wrote:\\n> On 17 March 2016 at 14:25, Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote:\\n>> On 03/16/2016 12:03 PM, David Rowley wrote:\\n>>> Patch 2:\\n>>> This adds the serial/deserial aggregate infrastructure, pg_dump\\n>>> support, CREATE AGGREGATE changes, and nodeAgg.c changes to have it\\n>>> serialise and deserialise aggregate states when instructed to do so.\\n>>>\\n>>> Patch 3:\\n>>> This adds a boat load of serial/deserial functions, and combine\\n>>> functions for most of the built-in numerical aggregate functions. It\\n>>> also contains some regression tests which should really be in patch 2,\\n>>> but I with patch 2 there's no suitable serialisation or\\n>>> de-serialisation functions to test CREATE AGGREGATE with. I think\\n>>> having them here is ok, as patch 2 is quite useless without patch 3\\n>>> anyway.\\n>>\\n>> I don't see how you could move the tests into #2 when the functions are\\n>> defined in #3? IMHO this is the right place for the regression tests.\\n>\\n> Yeah, but the CREATE AGGREGATE changes which are being tested in 0003\\n> were actually added in 0002. I think 0002 is the right place to test\\n> the changes to CREATE AGGREGATE, but since there's a complete lack of\\n> functions to use, then I've just delayed until 0003.\\n>\\n>>> Another thing to note about this patch is that I've gone and created\\n>>> serial/de-serial functions for when PolyNumAggState both require\\n>>> sumX2, and don't require sumX2. I had thought about perhaps putting an\\n>>> extra byte in the serial format to indicate if a sumX2 is included,\\n>>> but I ended up not doing it this way. I don't really want these serial\\n>>> formats getting too complex as we might like to do fun things like\\n>>> pass them along to sharded servers one day, so it might be nice to\\n>>> keep them simple.\\n>>\\n>>\\n>> Hmmm, I've noticed that while eyeballing the diffs, and I'm not\\n>> sure if it's worth the additional complexity at this point. I mean,\\n>> one byte is hardly going to make a measurable difference - we're\\n>> probably wasting more than that due to alignment, for example.\\n>\\n> I don't think any alignment gets done here. Look at pq_sendint(). Did\\n> you mean the complexity of having extra functions, or having the\\n> extra byte to check in the deserial func?\\n\\nI haven't realized alignment does not apply here, but I still think the \\nextra byte would be negligible in the bigger scheme of things. For \\nexample we're serializing the state into a bytea, which adds up to 4B \\nheader. So I don't think adding 1B would make any measurable difference.\\n\\nBut that assumes adding the 1B header would actually make the code \\nsimpler. Now that I think about it, that may not the case - in the end \\nyou'd probably get a function with two branches, with about the same \\nsize as the two functions combined. So not really an improvement.\\n\\n>> As you've mentioned sharded servers - how stable should the\\n>> serialized format be? I've assumed it to be transient, i.e. in the\\n>> extreme case it might change after restarting a database - for the\\n>> parallel aggregate that's clearly sufficient.\\n>>\\n>> But if we expect this to eventually work in a sharded environment,\\n>> that's going to be way more complicated. I guess in these cases we\\n>> could rely on implicit format versioning, via the major the\\n>> version (doubt we'd tweak the format in a minor version anyway).\\n>>\\n>> I'm not sure the sharding is particularly useful argument at this\\n>> point, considering we don't really know if the current format is\\n>> actually a good match for that.\\n>\\n> Perhaps you're right. At this stage I've no idea if we'd want to\\n> support shards on varying major versions. I think probably not,\\n> since the node write functions might not be compatible with the node\\n> read functions on the other server.\\n\\nOK, so let's ignore the sharded setup for now.\\n\\n> I've attached another series of patches:\\n>\\n> 0001: This is the latest Parallel Aggregate Patch, not intended for\\n> review here, but is required for the remaining patches. This patch has\\n> changed quite a bit from the previous one that I posted here, and the\\n> remaining patches needed re-based due to those changes.\\n>\\n> 0002: Adds serial/de-serial function support to CREATE AGGREGATE,\\n> contains minor fix-ups from last version.\\n>\\n> 0003: Adds various combine/serial/de-serial functions for the standard\\n> set of aggregates, apart from most float8 aggregates.\\n>\\n> 0004: Adds regression tests for 0003 pg_aggregate.h changes.\\n>\\n> 0005: Documents to mention which aggregate functions support partial mode.\\n>\\n> 0006: Re-based version of Haribabu's floating point aggregate support,\\n> containing some fixes by me.\\n\\nI went through the changes and found nothing suspicious, except maybe \\nfor the wording in one of the doc patches:\\n\\n All types apart from floating-point types\\n\\nIt may not be entirely clear for the readers, but this does not include \\n\\\"numeric\\\" data type, only float4 and float8. But I don't think this \\nreally matters unless we fail to commit the last patch adding functions \\neven for those data types.\\n\\nAlso, I think it's probably the right time to fix the failures in \\nopr_sanity regression tests. After all, we'll only whitelist a bunch of \\nserialize functions, so the tests will still detect any unexpected \\nfunctions dealing with 'internal' data type.\\n\\nregards\\n\\n-- \\nTomas Vondra 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\": \"Sun, 20 Mar 2016 22:55:19 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Sun, Mar 20, 2016 at 2:23 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n>\\n> I've had a look over this. I had to first base it on the 0005 patch,\\n> as it seemed like the pg_aggregate.h changes didn't include the\\n> serialfn and deserialfn changes, and an OID was being consumed by\\n> another function I added in patch 0003.\\n>\\n> On testing I also noticed some wrong results, which on investigation,\\n> are due to the wrong array elements being added together.\\n>\\n> For example:\\n>\\n> postgres=# select stddev(num) from f;\\n> stddev\\n> ------------------\\n> 28867.5149028984\\n> (1 row)\\n>\\n>\\n> postgres=# set max_parallel_degree=8;\\n> SET\\n> postgres=# select stddev(num) from f;\\n> stddev\\n> --------\\n> 0\\n> (1 row)\\n>\\n> + N += transvalues2[0];\\n> + sumX += transvalues2[1];\\n> + CHECKFLOATVAL(sumX, isinf(transvalues1[1]) || isinf(transvalues2[1]), true);\\n> + sumX2 += transvalues1[2];\\n>\\n> The last line should use transvalues2[2], not transvalues1[2].\\n\\nThanks.\\n\\n> There's also quite a few mistakes in float8_regr_combine()\\n>\\n> + sumX2 += transvalues2[2];\\n> + CHECKFLOATVAL(sumX2, isinf(transvalues1[2]) || isinf(transvalues2[1]), true);\\n>\\n> Wrong array element on isinf() check\\n>\\n> + sumY2 += transvalues2[4];\\n> + CHECKFLOATVAL(sumY2, isinf(transvalues1[4]) || isinf(transvalues2[3]), true);\\n>\\n> Wrong array element on isinf() check\\n>\\n> + sumXY += transvalues2[5];\\n> + CHECKFLOATVAL(sumXY, isinf(transvalues1[5]) || isinf(transvalues2[1]) ||\\n> + isinf(transvalues2[3]), true);\\n>\\n> Wrong array element on isinf() check, and the final\\n> isinf(transvalues2[3]) check does not need to be there.\\n\\nThanks for the changes, I just followed the float8_regr_accum function while\\nwriting float8_regr_combine function. Now I understood isinf proper usage.\\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\": \"Mon, 21 Mar 2016 12:44:58 +1100\",\n \"msg_from\": \"Haribabu Kommi <kommi.haribabu@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Sat, Mar 19, 2016 at 11:48 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> 0002: Adds serial/de-serial function support to CREATE AGGREGATE,\\n> contains minor fix-ups from last version.\\n\\nThis looks pretty good, but don't build_aggregate_serialfn_expr and\\nbuild_aggregate_deserialfn_expr compile down to identical machine\\ncode? Keeping two copies of the same code with different parameter\\nnames is a degree of neatnik-ism I'm not sure I can swallow.\\n\\nThe only caller to make_partialgroup_input_target() passes true for\\nthe additional argument. That doesn't seem right.\\n\\nMaybe error messages should refer to \\\"aggregate serialization\\nfunction\\\" and \\\"aggregate deserialization function\\\" instead of\\n\\\"aggregate serial function\\\" and \\\"aggregate de-serial function\\\".\\n\\n- * Other behavior is also supported and is controlled by the\\n'combineStates'\\n- * and 'finalizeAggs'. 'combineStates' controls whether the trans func or\\n- * the combine func is used during aggregation. When 'combineStates' is\\n- * true we expect other (previously) aggregated states as input\\nrather than\\n- * input tuples. This mode facilitates multiple aggregate stages which\\n- * allows us to support pushing aggregation down deeper into\\nthe plan rather\\n- * than leaving it for the final stage. For example with a query such as:\\n+ * Other behavior is also supported and is controlled by the\\n'combineStates',\\n+ * 'finalizeAggs' and 'serialStates' parameters. 'combineStates' controls\\n+ * whether the trans func or the combine func is used during aggregation.\\n+ * When 'combineStates' is true we expect other (previously) aggregated\\n+ * states as input rather than input tuples. This mode\\nfacilitates multiple\\n+ * aggregate stages which allows us to support pushing aggregation down\\n+ * deeper into the plan rather than leaving it for the final stage. For\\n+ * example with a query such as:\\n\\nI'd omit this hunk. The serialStates thing is really separate from\\nwhat's being talked about here, and you discuss it further down.\\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, 21 Mar 2016 12:27:19 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 20 March 2016 at 16:48, David Rowley <david.rowley@2ndquadrant.com> wrote:\\n> I've attached another series of patches:\\n>\\n> 0001: This is the latest Parallel Aggregate Patch, not intended for\\n> review here, but is required for the remaining patches. This patch has\\n> changed quite a bit from the previous one that I posted here, and the\\n> remaining patches needed re-based due to those changes.\\n>\\n> 0002: Adds serial/de-serial function support to CREATE AGGREGATE,\\n> contains minor fix-ups from last version.\\n>\\n> 0003: Adds various combine/serial/de-serial functions for the standard\\n> set of aggregates, apart from most float8 aggregates.\\n>\\n> 0004: Adds regression tests for 0003 pg_aggregate.h changes.\\n>\\n> 0005: Documents to mention which aggregate functions support partial mode.\\n>\\n> 0006: Re-based version of Haribabu's floating point aggregate support,\\n> containing some fixes by me.\\n\\nNow that parallel aggregate has been committed, 0002 no longer applies.\\nPlease find attached a new 0001 patch, intended to replace the\\nprevious 0001 and 0002.\\n\\nThe previous 0003, 0004, 0005 and 0006 should still apply onto the new 0001.\\n\\n-- \\n David Rowley 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, 22 Mar 2016 05:33:05 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 22 March 2016 at 05:27, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Sat, Mar 19, 2016 at 11:48 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n>> 0002: Adds serial/de-serial function support to CREATE AGGREGATE,\\n>> contains minor fix-ups from last version.\\n>\\n> This looks pretty good, but don't build_aggregate_serialfn_expr and\\n> build_aggregate_deserialfn_expr compile down to identical machine\\n> code? Keeping two copies of the same code with different parameter\\n> names is a degree of neatnik-ism I'm not sure I can swallow.\\n\\nGood point. I've removed the deserial one, and renamed the arguments\\nto arg_input_type and arg_output_type.\\n\\n> The only caller to make_partialgroup_input_target() passes true for\\n> the additional argument. That doesn't seem right.\\n\\nYeah, I coded that with non-parallel use cases in mind. I very much\\nthing we'll end up using that flag in 9.7, but perhaps that's not a\\ngood enough reason to be adding it in 9.6. I've pulled it out of the\\npatch for now, but I could also go further and pull the flag from the\\nwhole patch, as we could just serialize the states in nodeAgg.c when\\nfinalizeAggs == false and deserialize when combineStates == true.\\n\\n> Maybe error messages should refer to \\\"aggregate serialization\\n> function\\\" and \\\"aggregate deserialization function\\\" instead of\\n> \\\"aggregate serial function\\\" and \\\"aggregate de-serial function\\\".\\n\\nThat's probably a good idea.\\n\\n> - * Other behavior is also supported and is controlled by the\\n> 'combineStates'\\n> - * and 'finalizeAggs'. 'combineStates' controls whether the trans func or\\n> - * the combine func is used during aggregation. When 'combineStates' is\\n> - * true we expect other (previously) aggregated states as input\\n> rather than\\n> - * input tuples. This mode facilitates multiple aggregate stages which\\n> - * allows us to support pushing aggregation down deeper into\\n> the plan rather\\n> - * than leaving it for the final stage. For example with a query such as:\\n> + * Other behavior is also supported and is controlled by the\\n> 'combineStates',\\n> + * 'finalizeAggs' and 'serialStates' parameters. 'combineStates' controls\\n> + * whether the trans func or the combine func is used during aggregation.\\n> + * When 'combineStates' is true we expect other (previously) aggregated\\n> + * states as input rather than input tuples. This mode\\n> facilitates multiple\\n> + * aggregate stages which allows us to support pushing aggregation down\\n> + * deeper into the plan rather than leaving it for the final stage. For\\n> + * example with a query such as:\\n>\\n> I'd omit this hunk. The serialStates thing is really separate from\\n> what's being talked about here, and you discuss it further down.\\n\\nRemoved.\\n\\nI've attached 2 of the patches which are affected by the changes.\\n\\nThanks for looking over this.\\n\\n-- \\n David Rowley 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, 22 Mar 2016 07:18:35 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 21 January 2016 at 08:06, Robert Haas <robertmhaas@gmail.com> wrote:\\n> I re-reviewed this and have committed most of it with only minor\\n> kibitizing. A few notes:\\n\\nI realised today that the combinefunc is rather undocumented. I've\\nattached a patch which aims to fix this.\\n\\nComments are welcome.\\n\\n-- \\n David Rowley 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, 24 Mar 2016 22:22:15 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 24, 2016 at 5:22 AM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> On 21 January 2016 at 08:06, Robert Haas <robertmhaas@gmail.com> wrote:\\n>> I re-reviewed this and have committed most of it with only minor\\n>> kibitizing. A few notes:\\n>\\n> I realised today that the combinefunc is rather undocumented. I've\\n> attached a patch which aims to fix this.\\n>\\n> Comments are welcome.\\n\\nLooks good to me. 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, 24 Mar 2016 13:00:09 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 21, 2016 at 2:18 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> I've attached 2 of the patches which are affected by the changes.\\n\\nI think the documentation for 0001 needs some work yet. The\\nadditional paragraph that you've added...\\n\\n(1) doesn't seem to appear at a very logical place in the\\ndocumentation - I think it should be much further down, as it's a\\nminor detail. Maybe document this the same way as the documentation\\npatch you just sent for the combine-function stuff does it; and\\n\\n(2) isn't indented consistently with the surrounding paragraphs; and\\n\\n(3) is missing a closing </para> tag\\n\\nAlso, I'd just cut this:\\n\\n+ This is required due to\\n+ the process model being unable to pass references to <literal>INTERNAL\\n+ </literal> types between different <productname>PostgreSQL</productname>\\n+ processes.\\n\\nInstead, I'd change the earlier sentence in the paragraph, which\\ncurrently reads:\\n\\n+ These\\n+ functions are required in order to allow parallel aggregation for aggregates\\n+ with an <replaceable class=\\\"PARAMETER\\\">stype</replaceable> of <literal>\\n+ INTERNAL</>.\\n\\nI'd replace the period at end with a comma and add \\\"since\\n<literal>INTERNAL</> values represent arbitrary in-memory data\\nstructures which can't be passed between processes\\\". I think that's a\\nbit smoother.\\n\\nI'm going to read through the code again 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, 24 Mar 2016 13:17:06 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 24, 2016 at 1:17 PM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> I'm going to read through the code again now.\\n\\nOK, I noticed another documentation problem: you need to update\\ncatalogs.sgml for these new columns.\\n\\n+ * Validate the serial function, if present. We must ensure\\nthat the return\\n+ * Validate the de-serial function, if present. We must ensure that the\\n\\nI think that you should refer to these consistently in the comments as\\nthe \\\"serialization function\\\" and the \\\"deserialization function\\\", even\\nthough the SQL syntax is different. And unhyphenated.\\n\\n+ /* check that we also got a serial type */\\n+ if (!OidIsValid(aggSerialType))\\n+ ereport(ERROR,\\n+ (errcode(ERRCODE_INVALID_FUNCTION_DEFINITION),\\n+ errmsg(\\\"must specify serialization\\ntype when specifying serialization function\\\")));\\n\\nI think that in parallel cases, this check is in DefineAggregate(),\\nnot here. See, e.g. \\\"aggregate mfinalfunc must not be specified\\nwithout mstype\\\".\\n\\nExisting type parameters to CREATE AGGREGATE have IsPolymorphicType()\\nchecks to enforce sanity in various ways, but you seem not to have\\nadded that for the serial type.\\n\\n+ /* don't call a strict serial function with\\nNULL input */\\n+ if (pertrans->serialfn.fn_strict &&\\n+ pergroupstate->transValueIsNull)\\n+ continue;\\n\\nShouldn't this instead set aggnulls[aggno] = true? And doesn't the\\nhunk in combine_aggregates() have the same problem?\\n\\n+ /*\\n+ * serial and de-serial functions must match, if\\npresent. Remember that\\n+ * these will be InvalidOid if they're not required\\nfor this agg node\\n+ */\\n\\nExplain WHY they need to match. And maybe update the overall comment\\nfor the function.\\n\\n+ \\\"'-' AS\\naggdeserialfn,aggmtransfn, aggminvtransfn, \\\"\\n\\nWhitespace.\\n\\nIn your pg_aggregate.h changes, avg(numeric) sets aggserialtype but no\\naggserialfn or aggdeserialfn.\\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, 24 Mar 2016 14:26:50 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 25 March 2016 at 06:17, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Mon, Mar 21, 2016 at 2:18 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n>> I've attached 2 of the patches which are affected by the changes.\\n>\\n> I think the documentation for 0001 needs some work yet. The\\n> additional paragraph that you've added...\\n>\\n> (1) doesn't seem to appear at a very logical place in the\\n> documentation - I think it should be much further down, as it's a\\n> minor detail. Maybe document this the same way as the documentation\\n> patch you just sent for the combine-function stuff does it; and\\n\\nThanks. I also realised this when writing the combine documents fix.\\n\\n> (2) isn't indented consistently with the surrounding paragraphs; and\\n>\\n> (3) is missing a closing </para> tag\\n>\\n> Also, I'd just cut this:\\n>\\n> + This is required due to\\n> + the process model being unable to pass references to <literal>INTERNAL\\n> + </literal> types between different <productname>PostgreSQL</productname>\\n> + processes.\\n>\\n> Instead, I'd change the earlier sentence in the paragraph, which\\n> currently reads:\\n>\\n> + These\\n> + functions are required in order to allow parallel aggregation for aggregates\\n> + with an <replaceable class=\\\"PARAMETER\\\">stype</replaceable> of <literal>\\n> + INTERNAL</>.\\n>\\n> I'd replace the period at end with a comma and add \\\"since\\n> <literal>INTERNAL</> values represent arbitrary in-memory data\\n> structures which can't be passed between processes\\\". I think that's a\\n> bit smoother.\\n>\\n\\nIn my rewrite I've incorporated these words.\\n\\nThanks for checking over this. A patch will follow in my response to\\nthe next email.\\n\\n-- \\n David Rowley 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, 25 Mar 2016 16:53:30 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 25 March 2016 at 07:26, Robert Haas <robertmhaas@gmail.com> wrote:\\n> OK, I noticed another documentation problem: you need to update\\n> catalogs.sgml for these new columns.\\n\\nThanks. Done.\\n\\n>\\n> + * Validate the serial function, if present. We must ensure\\n> that the return\\n> + * Validate the de-serial function, if present. We must ensure that the\\n> I think that you should refer to these consistently in the comments as\\n> the \\\"serialization function\\\" and the \\\"deserialization function\\\", even\\n> though the SQL syntax is different. And unhyphenated.\\n\\nThat's probably better. Fixed.\\n\\n> + /* check that we also got a serial type */\\n> + if (!OidIsValid(aggSerialType))\\n> + ereport(ERROR,\\n> + (errcode(ERRCODE_INVALID_FUNCTION_DEFINITION),\\n> + errmsg(\\\"must specify serialization\\n> type when specifying serialization function\\\")));\\n>\\n> I think that in parallel cases, this check is in DefineAggregate(),\\n> not here. See, e.g. \\\"aggregate mfinalfunc must not be specified\\n> without mstype\\\".\\n\\nGood point. I've fixed this.\\n\\n> Existing type parameters to CREATE AGGREGATE have IsPolymorphicType()\\n> checks to enforce sanity in various ways, but you seem not to have\\n> added that for the serial type.\\n\\nAdded check for this to disallow pseudo types that are not polymorphic.\\n\\n\\n> + /* don't call a strict serial function with\\n> NULL input */\\n> + if (pertrans->serialfn.fn_strict &&\\n> + pergroupstate->transValueIsNull)\\n> + continue;\\n>\\n> Shouldn't this instead set aggnulls[aggno] = true? And doesn't the\\n> hunk in combine_aggregates() have the same problem?\\n\\nhmm, no, I can't see a problem there. aggvalues[aggno] is what's going\\nto be set after that condition.\\n\\n\\n> + /*\\n> + * serial and de-serial functions must match, if\\n> present. Remember that\\n> + * these will be InvalidOid if they're not required\\n> for this agg node\\n> + */\\n>\\n> Explain WHY they need to match. And maybe update the overall comment\\n> for the function.\\n\\nI've added:\\n\\n/*\\n* The serialization and deserialization functions must match, if\\n* present, as we're unable to share the trans state for aggregates\\n* which will serialize or deserialize into different formats. Remember\\n* that these will be InvalidOid if they're not required for this agg\\n* node.\\n*/\\n\\nThis is for the code which shares transition states between aggregate\\nfunctions which have the same transition function, for example\\nSUM(NUMERIC) and AVG(NUMERIC). The serialization functions must match\\ntoo. Perhaps, providing the serial/deserial function does its job\\nproperly, then it does not matter, but it does seem better to be safe\\nhere. I can't really imagine why anyone would want to have the same\\ntransition function but have different serialization functions.\\n\\n> + \\\"'-' AS\\n> aggdeserialfn,aggmtransfn, aggminvtransfn, \\\"\\n>\\n> Whitespace.\\n\\nFixed.\\n\\n> In your pg_aggregate.h changes, avg(numeric) sets aggserialtype but no\\n> aggserialfn or aggdeserialfn.\\n\\nFixed.\\n\\nI've also added some checks to ensure a combine function being set on\\nan aggregate with an INTERNAL state is not-strict. I realised when\\nwriting the documentation for combine functions that it needed to be\\nthis way, as, if the function was strict, then\\nadvance_combine_function() could incorrectly assign the memory address\\nof the 2nd state to the transValue. Actually, this seems a little\\ntricky to optimise well, as in the case when we're serialising states,\\nthe deserialisation function should allocate a new state in the\\ncorrect memory context, so technically the combine function could be\\nstrict in this case, as it would be fine to use state2 if state1 was\\nnull, but parallel aggregation is probably unique and could be the\\nonly time we'll need to use the serialisation/deserialisation\\nfunctions, so without serialStates = true, the memory address could\\nstill belong to some other aggregate node's memory context which is\\ncertainly not right, so the code as of this patch is playing it safe.\\nYou could say that this is technically a bug in master, but it should\\nbe impossible to hit it now due to lack of ability to combine internal\\naggregate states. The checks for the correct strict property in the\\ncombine function are 3-fold.\\n\\n1. CREATE AGGREGATE check.\\n2. executor startup for Agg nodes (must be here too, since someone\\ncould modify the strict property after CREATE AGGREGATE).\\n3. Regression test for built-in aggregates.\\n\\n... long pause...\\n\\nOk, so on further look at this I've decided to make changes and have\\nit so the serialisation function can be dumb about memory contexts in\\nthe same way as finalize_aggregate() allows the final function to be\\ndumb... notice at the end of the function it copies byref types into\\nthe correct memory context, if they're not already. So in the attached\\nthe serialisation function call code now does the same thing. In fact\\nI decided to move all that code off into a function called\\nfinalize_partialaggregate().\\n\\nAs for the deserialisation function... there's not much we can do\\nthere in regards to copying the value back to the correct memory\\ncontext, as we always expect the return value of that function to be\\nINTERNAL, which is byval (pointer). So I've just set the code to\\naggstate->tmpcontext->ecxt_per_tuple_memory, and leave it up to the\\ncombine function to do the right thing, and make sure the new state is\\nbuilt properly in the aggregate memory context, which is analogous to\\nhow INTERNAL state transition functions work.\\n\\nAs for the optimisation I mentioned above; what I have now is not\\nperfect, as when we Gather the first state per group from the first\\nworker to return it, we deserialise that serialised state into a per\\ntuple context, next the combine function pulls that state apart and\\nrebuilds it again into another newly built state in the correct memory\\ncontext. I can't really see a way around this.\\n\\nI'd really like it if someone could look over this and make sure its\\nall sane, as I'm not sure if I've fully wrapped my head around the\\nexpected life time of each context used here...\\n\\nStatus of other patched:\\n\\n0002:\\n\\n* Changed the new incorrectly defined combine functions so they're no\\nlonger strict.\\n* Removed all memory context switches from\\nserialisation/deserialisation functions.\\n\\n0003:\\n\\nI've added a few new regression tests to test for strictness of\\ncombine functions, and also threw in some tests to ensure the\\nserialisation/deserialisation functions are all strict, since the\\ncurrent ones can't handle NULLs, and it would be wasteful to make them\\ntrouble themselves with that ability.\\n\\n0004:\\nNo change from last time.\\n\\n0005:\\nHaribabu's patch; no change from last time.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\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\": \"Sat, 26 Mar 2016 15:07:28 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 26 March 2016 at 15:07, David Rowley <david.rowley@2ndquadrant.com> wrote:\\n> Ok, so on further look at this I've decided to make changes and have\\n> it so the serialisation function can be dumb about memory contexts in\\n> the same way as finalize_aggregate() allows the final function to be\\n> dumb... notice at the end of the function it copies byref types into\\n> the correct memory context, if they're not already. So in the attached\\n> the serialisation function call code now does the same thing. In fact\\n> I decided to move all that code off into a function called\\n> finalize_partialaggregate().\\n\\nPlease disregard the previous 0001 patch in favour of the attached one.\\n\\n-- \\n David Rowley 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\": \"Sat, 26 Mar 2016 17:24:52 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 26 March 2016 at 15:07, David Rowley <david.rowley@2ndquadrant.com> wrote:\\n\\nMany thanks Robert for committing the serialize states portion.\\n\\n> 0005:\\n> Haribabu's patch; no change from last time.\\n\\nJust in case you jump ahead. I just wanted to re-highlight something\\nHaribabu mentioned a while ago about combining floating point states\\n[1]. This discussion did not happen on this thread, so to make sure it\\ndoes not get lost...\\n\\nAs of today aggregate calculations for floating point types can vary\\ndepending on the order in which values are aggregated.\\n\\nFor example:\\n\\ncreate table f8 (num float8);\\ninsert into f8 select x/100.0 from generate_series(1,10000) x(x);\\nselect stddev(num order by num) from f8;\\n stddev\\n------------------\\n 28.8689567990717\\n(1 row)\\n\\nselect stddev(num order by num desc) from f8;\\n stddev\\n------------------\\n 28.8689567990716\\n(1 row)\\n\\nselect stddev(num order by random()) from f8;\\n stddev\\n------------------\\n 28.8689567990715\\n(1 row)\\n\\nAnd of course the execution plan can determine the order in which rows\\nare aggregated, even if the underlying data does not change.\\n\\nParallelising these aggregates increases the chances of seeing these\\nvariations as the number of rows aggregated in each worker is going to\\nvary on each run, so the numerical anomalies will also vary between\\neach run.\\n\\nI wrote in [1]:\\n> We do also warn about this in the manual: \\\"Inexact means that some\\n> values cannot be converted exactly to the internal format and are\\n> stored as approximations, so that storing and retrieving a value might\\n> show slight discrepancies. Managing these errors and how they\\n> propagate through calculations is the subject of an entire branch of\\n> mathematics and computer science and will not be discussed here,\\n> except for the following points:\\\" [1]\\n> [1] http://www.postgresql.org/docs/devel/static/datatype-numeric.html\\n\\nDoes this text in the documents stretch as far as the variable results\\nfrom parallel aggregate for floating point types? or should we be more\\ncareful and not parallelise these, similar to how we didn't end up\\nwith inverse aggregate transition functions for floating point types?\\n\\nI'm personally undecided, and would like to hear what others think.\\n\\n[1] http://www.postgresql.org/message-id/CAKJS1f_HpLFhkd2yLfrsrmUMBZWQkGvJCWX21B_xg1A-0pzwHw@mail.gmail.com\\n-- \\n David Rowley 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 Mar 2016 16:14:38 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 29, 2016 at 11:14 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> Many thanks Robert for committing the serialize states portion.\\n\\nyw, sorry I didn't get an email out about that.\\n\\n>> 0005:\\n>> Haribabu's patch; no change from last time.\\n\\nSo what's the distinction between 0002 and 0005? And what is the\\ncorrect author/reviewer attribution for each?\\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, 30 Mar 2016 07:48:59 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 31 March 2016 at 00:48, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Tue, Mar 29, 2016 at 11:14 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n>>> 0005:\\n>>> Haribabu's patch; no change from last time.\\n>\\n> So what's the distinction between 0002 and 0005? And what is the\\n> correct author/reviewer attribution for each?\\n\\nI wrote 0002 - 0004, these were reviewed by Tomas.\\n0005 is Haribabu's patch: Reviewed by Tomas and I.\\n\\nThe reason 0002 and 0005 are separate patches is just down to them\\nhaving different authors.\\n\\n-- \\n David Rowley 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, 31 Mar 2016 08:34:02 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 30, 2016 at 3:34 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> I wrote 0002 - 0004, these were reviewed by Tomas.\\n> 0005 is Haribabu's patch: Reviewed by Tomas and I.\\n\\nI think it might be a good idea if these patches made less use of\\nbytea and exposed the numeric transition values as, say, a 2-element\\narray of numeric. Maybe this doesn't really matter, but it's not\\nimpossible that these values could be exposed somewhere, and a numeric\\nis an awful lot more user-friendly than an essentially opaque bytea.\\nOne of the things I eventually want to figure out how to do with this\\nis distributed aggregation across multiple shards, and I think it\\nmight be better to have the value being sent over the wire be\\nsomething like {26.6,435.12} rather than \\\\x1234...\\n\\nThoughts?\\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, 5 Apr 2016 00:54:56 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 5 April 2016 at 05:54, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On Wed, Mar 30, 2016 at 3:34 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n> > I wrote 0002 - 0004, these were reviewed by Tomas.\\n> > 0005 is Haribabu's patch: Reviewed by Tomas and I.\\n>\\n> I think it might be a good idea if these patches made less use of\\n> bytea and exposed the numeric transition values as, say, a 2-element\\n> array of numeric. Maybe this doesn't really matter, but it's not\\n> impossible that these values could be exposed somewhere, and a numeric\\n> is an awful lot more user-friendly than an essentially opaque bytea.\\n> One of the things I eventually want to figure out how to do with this\\n> is distributed aggregation across multiple shards, and I think it\\n> might be better to have the value being sent over the wire be\\n> something like {26.6,435.12} rather than \\\\x1234...\\n>\\n> Thoughts?\\n\\n\\nRewriting something that works fine just before the deadline isn't a good\\nplan.\\n\\n\\\"Might be better\\\" isn't enough.\\n\\n-- \\nSimon Riggs http://www.2ndQuadrant.com/\\n<http://www.2ndquadrant.com/>\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\nOn 5 April 2016 at 05:54, Robert Haas <robertmhaas@gmail.com> wrote:On Wed, Mar 30, 2016 at 3:34 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> I wrote 0002 - 0004, these were reviewed by Tomas.\\n> 0005 is Haribabu's patch: Reviewed by Tomas and I.\\n\\nI think it might be a good idea if these patches made less use of\\nbytea and exposed the numeric transition values as, say, a 2-element\\narray of numeric. Maybe this doesn't really matter, but it's not\\nimpossible that these values could be exposed somewhere, and a numeric\\nis an awful lot more user-friendly than an essentially opaque bytea.\\nOne of the things I eventually want to figure out how to do with this\\nis distributed aggregation across multiple shards, and I think it\\nmight be better to have the value being sent over the wire be\\nsomething like {26.6,435.12} rather than \\\\x1234...\\n\\nThoughts?Rewriting something that works fine just before the deadline isn't a good plan. \\\"Might be better\\\" isn't enough.-- Simon Riggs http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Tue, 5 Apr 2016 08:18:19 +0100\",\n \"msg_from\": \"Simon Riggs <simon@2ndQuadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 5 April 2016 at 16:54, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Wed, Mar 30, 2016 at 3:34 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n>> I wrote 0002 - 0004, these were reviewed by Tomas.\\n>> 0005 is Haribabu's patch: Reviewed by Tomas and I.\\n>\\n> I think it might be a good idea if these patches made less use of\\n> bytea and exposed the numeric transition values as, say, a 2-element\\n> array of numeric.\\n\\n\\nWell, if you have a look at NumericAggState you can see it's not quite\\nas simple as an array of numeric, unless of course you'd be willing to\\nspend the extra cycles, use the extra memory, and bandwidth to convert\\nthose int64's to numeric too, then it could be made to work. To do are\\nyou describe properly, we'd need a composite type.\\n\\n> Maybe this doesn't really matter, but it's not\\n> impossible that these values could be exposed somewhere, and a numeric\\n> is an awful lot more user-friendly than an essentially opaque bytea.\\n\\nhmm, isn't that why we have a deserialisation functions? Do you see a\\nuse case where these won't be available?\\n\\n> One of the things I eventually want to figure out how to do with this\\n> is distributed aggregation across multiple shards, and I think it\\n> might be better to have the value being sent over the wire be\\n> something like {26.6,435.12} rather than \\\\x1234...\\n>\\n> Thoughts?\\n\\nI've not yet read the design spec for sharding in Postgres. If there's\\nsomething I can read over to get an idea of why you think this won't\\nwork, then maybe we can come to a good conclusion that way. But if I\\ntake a guess, then I'd have imagined that we'd not support sharding\\nover different major versions, and if we really needed to change the\\nserialisation format later, then we could do so. We could even put a\\nnote in the documents that the serialisation format may change between\\nmajor versions.\\n\\nTo be really honest, I'm quite worried that if I go and make this\\nchange then my time might be wasted as I really think making that\\nchange this late in the day is just setting this up for failure. I\\nreally don't think we can bat this patch over the Pacific Ocean too\\nmany times before we find ourselves inside the feature freeze. Of\\ncourse, if you really think it's no good, that's different, it can't\\nbe committed, but \\\"I think it might be better\\\" does not seem like a\\nsolid enough argument for me to want to risk trying this and delaying\\nfurther for that.\\n\\nBut either way, we should come try to come to a consensus quickly. I\\ndon't want to alienate you because you think I don't want to listen to\\nyou. We need to act quickly or we'll only have a handful of the useful\\naggregates working in parallel for 9.6. I'd rather we had them all. I\\nhope you do too.\\n\\nThanks for looking over the patch.\\n\\n-- \\n David Rowley 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, 5 Apr 2016 19:55:04 +1200\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 5, 2016 at 3:55 AM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n>> I think it might be a good idea if these patches made less use of\\n>> bytea and exposed the numeric transition values as, say, a 2-element\\n>> array of numeric.\\n>\\n> Well, if you have a look at NumericAggState you can see it's not quite\\n> as simple as an array of numeric, unless of course you'd be willing to\\n> spend the extra cycles, use the extra memory, and bandwidth to convert\\n> those int64's to numeric too, then it could be made to work. To do are\\n> you describe properly, we'd need a composite type.\\n\\nUggh, yeah. Yuck.\\n\\n> hmm, isn't that why we have a deserialisation functions? Do you see a\\n> use case where these won't be available?\\n...\\n> I've not yet read the design spec for sharding in Postgres. If there's\\n> something I can read over to get an idea of why you think this won't\\n> work, then maybe we can come to a good conclusion that way. But if I\\n> take a guess, then I'd have imagined that we'd not support sharding\\n> over different major versions, and if we really needed to change the\\n> serialisation format later, then we could do so. We could even put a\\n> note in the documents that the serialisation format may change between\\n> major versions.\\n\\nWell, OK, so here was my thought. For the sake of simplicity, let's\\nsuppose that creating a sharded table works more or less like what you\\ncan already do today: create a parent table with a non-inherited CHECK\\n(false) constraint and then create some inheritance children that are\\nforeign tables on various remote servers. Give those children CHECK\\nconstraints that explicate the partitioning scheme. This is probably\\nnot actually how we want this to work in detail (e.g. we probably want\\ndeclarative partitioning) but the details don't matter very much for\\npurposes of what I'm trying to explain here so let's just ignore them\\nfor the moment.\\n\\nNow, let's suppose that the user sets up a sharded table and then\\nsays: SELECT a, SUM(b), AVG(c) FROM sometab. At this point, what we'd\\nlike to have happen is that for each child foreign table, we go and\\nfetch partially aggregated results. Those children might be running\\nany version of PostgreSQL - I was not assuming that we'd insist on\\nmatching major versions, although of course that could be done - and\\nthere would probably need to be a minimum version of PostgreSQL\\nanyway. They could even be running some other database. As long as\\nthey can spit out partial aggregates in a format that we can\\nunderstand, we can deserialize those aggregates and run combine\\nfunctions on them. But if the remote side is, say, MariaDB, it's\\nprobably much easier to get it to spit out something that looks like a\\nPostgreSQL array than it is to make it spit out some bytea blob that's\\nin an entirely PostgreSQL-specific format.\\n\\nNow, maybe that doesn't matter. Even getting something like this\\nworking with PostgreSQL as the remote side is going to be hard.\\nMoreover, for this to have any chance of working with anything other\\nthan a compatible PostgreSQL server on the remote side, the FDW is\\ngoing to have to write bespoke code for each aggregate anyway, and\\nthat code can always construct the requisite bytea blobs internally.\\nSo who cares? I can't point to any really tangible advantage of\\nhaving serialized transition states be human-readable, so maybe there\\nisn't one. But I was thinking about it, for fear that there might be\\nsome advantage that I'm missing.\\n\\n> To be really honest, I'm quite worried that if I go and make this\\n> change then my time might be wasted as I really think making that\\n> change this late in the day is just setting this up for failure. I\\n> really don't think we can bat this patch over the Pacific Ocean too\\n> many times before we find ourselves inside the feature freeze. Of\\n> course, if you really think it's no good, that's different, it can't\\n> be committed, but \\\"I think it might be better\\\" does not seem like a\\n> solid enough argument for me to want to risk trying this and delaying\\n> further for that.\\n\\nI think I agree. Certainly, with what we've got here today, these are\\nnot user-exposed, so I think we could certainly change them next time\\naround if need be. If they ever become user-exposed, then maybe we\\nshould think a little harder.\\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, 5 Apr 2016 09:01:48 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 6 April 2016 at 01:01, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Tue, Apr 5, 2016 at 3:55 AM, David Rowley\\n>> To be really honest, I'm quite worried that if I go and make this\\n>> change then my time might be wasted as I really think making that\\n>> change this late in the day is just setting this up for failure. I\\n>> really don't think we can bat this patch over the Pacific Ocean too\\n>> many times before we find ourselves inside the feature freeze. Of\\n>> course, if you really think it's no good, that's different, it can't\\n>> be committed, but \\\"I think it might be better\\\" does not seem like a\\n>> solid enough argument for me to want to risk trying this and delaying\\n>> further for that.\\n>\\n> I think I agree. Certainly, with what we've got here today, these are\\n> not user-exposed, so I think we could certainly change them next time\\n> around if need be. If they ever become user-exposed, then maybe we\\n> should think a little harder.\\n\\nI understand your concern. Painting ourselves into a corner would be pretty bad.\\n\\nAlthough, I'm not so sure we'd want to go down the route of trying to\\nstabilise the format, even into something human readable. As far as I\\nknow we've never mentioned anything about what's stored in the\\ninternal aggregate states in the documentation, and we've made some\\nfairly hefty changes to these over the last few releases. For example;\\n959277a4, where internal int128 support was added to improve\\nperformance of various aggregates, like sum(bigint). I think if we're\\ngoing to head down the route of trying to keep this format stable,\\nthen we're going to struggle to make improvements to aggregates in the\\nfuture.\\n\\nIt's also not all that clear that all of the internal fields really\\nmake sense to other databases, for example NumericAggState's maxScale\\nand maxScaleCount. These really only are needed for moving aggregates.\\nWould we want some other database to have to magic up some numbers for\\nthese fields because our format requires it?\\n\\n-- \\n David Rowley 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, 6 Apr 2016 01:30:51 +1200\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 5, 2016 at 9:30 AM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n> On 6 April 2016 at 01:01, Robert Haas <robertmhaas@gmail.com> wrote:\\n>> On Tue, Apr 5, 2016 at 3:55 AM, David Rowley\\n>>> To be really honest, I'm quite worried that if I go and make this\\n>>> change then my time might be wasted as I really think making that\\n>>> change this late in the day is just setting this up for failure. I\\n>>> really don't think we can bat this patch over the Pacific Ocean too\\n>>> many times before we find ourselves inside the feature freeze. Of\\n>>> course, if you really think it's no good, that's different, it can't\\n>>> be committed, but \\\"I think it might be better\\\" does not seem like a\\n>>> solid enough argument for me to want to risk trying this and delaying\\n>>> further for that.\\n>>\\n>> I think I agree. Certainly, with what we've got here today, these are\\n>> not user-exposed, so I think we could certainly change them next time\\n>> around if need be. If they ever become user-exposed, then maybe we\\n>> should think a little harder.\\n>\\n> I understand your concern. Painting ourselves into a corner would be pretty bad.\\n>\\n> Although, I'm not so sure we'd want to go down the route of trying to\\n> stabilise the format, even into something human readable. As far as I\\n> know we've never mentioned anything about what's stored in the\\n> internal aggregate states in the documentation, and we've made some\\n> fairly hefty changes to these over the last few releases. For example;\\n> 959277a4, where internal int128 support was added to improve\\n> performance of various aggregates, like sum(bigint). I think if we're\\n> going to head down the route of trying to keep this format stable,\\n> then we're going to struggle to make improvements to aggregates in the\\n> future.\\n\\nThat might be true, but it's much more likely to be true if the\\ninternal format is an opaque blob. If the serialized transition state\\nfor average is {count,sum} then we can switch between having the\\ninternal representation being numeric and having it be int128 and\\nnothing breaks. With the opaque blob, maybe nothing breaks, or maybe\\nsomething does.\\n\\n> It's also not all that clear that all of the internal fields really\\n> make sense to other databases, for example NumericAggState's maxScale\\n> and maxScaleCount. These really only are needed for moving aggregates.\\n> Would we want some other database to have to magic up some numbers for\\n> these fields because our format requires it?\\n\\nWell, if we want to do cross-node aggregation with that database on\\nthe other side, those numbers are going to have to get magicked up\\nsomeplace along the line.\\n\\nI'm going to concede the point that this shouldn't really be a\\npriority for 9.6, but I might want to come back to it later.\\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, 5 Apr 2016 10:19:35 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 5, 2016 at 10:19 AM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> I'm going to concede the point that this shouldn't really be a\\n> priority for 9.6, but I might want to come back to it later.\\n\\nIt seems to me that if two aggregates are using the same transition\\nfunction, they ought to also be using the same combine, serialization,\\nand deserialization functions. I wrote a query to find cases where\\nthat wasn't so and found a few, which I changed before committing.\\n\\nDATA(insert ( 2100 n 0 int8_avg_accum numeric_poly_avg\\nint8_avg_combine int8_avg_serialize int8_avg_deserialize\\nint8_avg_accum int8_avg_accum_inv numeric_poly_avg f f 0 2281\\n 17 48 2281 48 _null_ _null_ ));\\nDATA(insert ( 2107 n 0 int8_avg_accum numeric_poly_sum\\nnumeric_poly_combine int8_avg_serialize int8_avg_deserialize\\n int8_avg_accum int8_avg_accum_inv numeric_poly_sum f f 0 2281\\n 17 48 2281 48 _null_ _null_ ));\\n\\nI changed the second of these from numeric_poly_combine to int8_avg_combine.\\n\\nDATA(insert ( 2103 n 0 numeric_avg_accum numeric_avg\\nnumeric_avg_combine numeric_avg_serialize numeric_avg_deserialize\\nnumeric_avg_accum numeric_accum_inv numeric_avg f f 0 2281\\n 17 128 2281 128 _null_ _null_ ));\\nDATA(insert ( 2114 n 0 numeric_avg_accum numeric_sum\\nnumeric_combine numeric_avg_serialize\\nnumeric_avg_deserialize numeric_avg_accum numeric_accum_inv\\nnumeric_sum f f 0 2281 17 128 2281 128 _null_ _null_\\n));\\n\\nI changed the second of these from numeric_combine to\\nnumeric_avg_combine. I also added a regression test for this. Please\\nlet me know if I'm off-base here.\\n\\nCommitted 0002+0003 with those changes, some minor cosmetic stuff, and\\nof course the obligatory catversion bump. Oh, and fixed an OID\\nconflict with the patch Magnus just 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\": \"Tue, 5 Apr 2016 14:33:50 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"Robert Haas wrote:\\n\\n> Now, let's suppose that the user sets up a sharded table and then\\n> says: SELECT a, SUM(b), AVG(c) FROM sometab. At this point, what we'd\\n> like to have happen is that for each child foreign table, we go and\\n> fetch partially aggregated results. Those children might be running\\n> any version of PostgreSQL - I was not assuming that we'd insist on\\n> matching major versions, although of course that could be done - and\\n> there would probably need to be a minimum version of PostgreSQL\\n> anyway. They could even be running some other database. As long as\\n> they can spit out partial aggregates in a format that we can\\n> understand, we can deserialize those aggregates and run combine\\n> functions on them. But if the remote side is, say, MariaDB, it's\\n> probably much easier to get it to spit out something that looks like a\\n> PostgreSQL array than it is to make it spit out some bytea blob that's\\n> in an entirely PostgreSQL-specific format.\\n\\nBasing parts of the Postgres sharding mechanism on FDWs sounds\\nacceptable. Trying to design things so that *any* FDW can be part of a\\nshard, so that you have some shards in Postgres and other shards in\\nMariaDB, seems ludicrous to me. Down that path lies madness.\\n\\nIn fact, trying to ensure cross-major-version compatibility already\\nsounds like asking for too much. Requiring matching major versions\\nsounds a perfectly acceptable restricting to me.\\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, 5 Apr 2016 15:52:51 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 5, 2016 at 2:52 PM, Alvaro Herrera <alvherre@2ndquadrant.com> wrote:\\n> Robert Haas wrote:\\n>> Now, let's suppose that the user sets up a sharded table and then\\n>> says: SELECT a, SUM(b), AVG(c) FROM sometab. At this point, what we'd\\n>> like to have happen is that for each child foreign table, we go and\\n>> fetch partially aggregated results. Those children might be running\\n>> any version of PostgreSQL - I was not assuming that we'd insist on\\n>> matching major versions, although of course that could be done - and\\n>> there would probably need to be a minimum version of PostgreSQL\\n>> anyway. They could even be running some other database. As long as\\n>> they can spit out partial aggregates in a format that we can\\n>> understand, we can deserialize those aggregates and run combine\\n>> functions on them. But if the remote side is, say, MariaDB, it's\\n>> probably much easier to get it to spit out something that looks like a\\n>> PostgreSQL array than it is to make it spit out some bytea blob that's\\n>> in an entirely PostgreSQL-specific format.\\n>\\n> Basing parts of the Postgres sharding mechanism on FDWs sounds\\n> acceptable. Trying to design things so that *any* FDW can be part of a\\n> shard, so that you have some shards in Postgres and other shards in\\n> MariaDB, seems ludicrous to me. Down that path lies madness.\\n\\nI'm doubtful that anyone wants to do the work to make that happen, but\\nI don't agree that we shouldn't care about whether it's possible.\\nExtensibility is a feature of the system that we've worked hard for,\\nand we shouldn't casually surrender it. For example, postgres_fdw now\\nimplements join pushdown, and I suspect few other FDW authors will\\ncare to do the work to add similar support to their implementations.\\nBut some may, and it's good that the code is structured in such a way\\nthat they have the option.\\n\\nActually, though, MariaDB is a bad example. What somebody's much more\\nlikely to want to do is have PostgreSQL as a frontend accessing data\\nthat's actually stored in Hadoop. There are lots of SQL interfaces to\\nHadoop already, so it's clearly a thing people want, and our SQL is\\nthe best SQL (right?) so if you could put that on top of Hadoop\\nsomebody'd probably like it. I'm not planning to try it myself, but I\\nthink it would be cool if somebody else did. I have been very pleased\\nto see that many of the bits and pieces of infrastructure that I\\ncreated for parallel query (parallel background workers, DSM, shm_mq)\\nhave attracted quite a bit of interest from other developers for\\ntotally unrelated purposes, and I think anything we do around\\nhorizontal scalability should be designed the same way: the goal\\nshould be to work with PostgreSQL on the other side, but the bits that\\ncan be made reusable for other purposes should be so constructed.\\n\\n> In fact, trying to ensure cross-major-version compatibility already\\n> sounds like asking for too much. Requiring matching major versions\\n> sounds a perfectly acceptable restricting to me.\\n\\nI disagree. One of the motivations (not the only one, by any means)\\nfor wanting logical replication in PostgreSQL is precisely to get\\naround the fact that physical replication requires matching major\\nversions. That restriction turns out to be fairly onerous, not least\\nbecause when you've got a cluster of several machines you'd rather\\nupgrade them one at a time rather than all at once. That's going to\\nbe even more true with a sharded cluster, which will probably tend to\\ninvolve more machines than a replication cluster, maybe a lot more.\\nIf you say that the user has got to shut the entire thing down,\\nupgrade all the machines, and turn it all back on again, and just hope\\nit works, that's going to be really painful. I think that we should\\ntreat this more like we do with libpq, where each major release can\\nadd new capabilities that new applications can use, but the old stuff\\nhas got to keep working essentially forever. Maybe the requirements\\nhere are not quite so tight, because it's probably reasonable to say,\\ne.g. that you must upgrade every machine in the cluster to at least\\nrelease 11.1 before upgrading any machine to 11.3 or higher, but the\\nfewer such requirements we have, the better. Getting people to\\nupgrade to new major releases is already fairly hard, and anything\\nthat makes it harder is an unwelcome development from my point of\\nview.\\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, 5 Apr 2016 15:25:51 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 5 April 2016 at 19:33, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n\\n> Committed 0002+0003 with those changes, some minor cosmetic stuff, and\\n> of course the obligatory catversion bump. Oh, and fixed an OID\\n> conflict with the patch Magnus just committed.\\n\\n\\nIs that everything now? I don't see anything about combining aggs in the\\ngit log and this is still showing as UnCommitted in the CF app.\\n\\n-- \\nSimon Riggs http://www.2ndQuadrant.com/\\n<http://www.2ndquadrant.com/>\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\nOn 5 April 2016 at 19:33, Robert Haas <robertmhaas@gmail.com> wrote: Committed 0002+0003 with those changes, some minor cosmetic stuff, and\\nof course the obligatory catversion bump. Oh, and fixed an OID\\nconflict with the patch Magnus just committed.Is that everything now? I don't see anything about combining aggs in the git log and this is still showing as UnCommitted in the CF app. -- Simon Riggs http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Wed, 6 Apr 2016 22:25:17 +0100\",\n \"msg_from\": \"Simon Riggs <simon@2ndQuadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 7 April 2016 at 09:25, Simon Riggs <simon@2ndquadrant.com> wrote:\\n> On 5 April 2016 at 19:33, Robert Haas <robertmhaas@gmail.com> wrote:\\n>\\n>>\\n>> Committed 0002+0003 with those changes, some minor cosmetic stuff, and\\n>> of course the obligatory catversion bump. Oh, and fixed an OID\\n>> conflict with the patch Magnus just committed.\\n>\\n>\\n> Is that everything now? I don't see anything about combining aggs in the git\\n> log and this is still showing as UnCommitted in the CF app.\\n\\nThere's just a documentation patch and two combine functions for\\nfloating point aggs left now (Haribabu's patch)\\n\\n\\n-- \\n David Rowley 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, 7 Apr 2016 09:28:46 +1200\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 6 April 2016 at 22:28, David Rowley <david.rowley@2ndquadrant.com> wrote:\\n\\n> On 7 April 2016 at 09:25, Simon Riggs <simon@2ndquadrant.com> wrote:\\n> > On 5 April 2016 at 19:33, Robert Haas <robertmhaas@gmail.com> wrote:\\n> >\\n> >>\\n> >> Committed 0002+0003 with those changes, some minor cosmetic stuff, and\\n> >> of course the obligatory catversion bump. Oh, and fixed an OID\\n> >> conflict with the patch Magnus just committed.\\n> >\\n> >\\n> > Is that everything now? I don't see anything about combining aggs in the\\n> git\\n> > log and this is still showing as UnCommitted in the CF app.\\n>\\n> There's just a documentation patch and two combine functions for\\n> floating point aggs left now (Haribabu's patch)\\n>\\n\\nAh, I see, it was the commit about \\\"multi-stage aggregation\\\".\\n\\nSo SELECT sum(x), avg(x) is now optimized?\\n\\n-- \\nSimon Riggs http://www.2ndQuadrant.com/\\n<http://www.2ndquadrant.com/>\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\nOn 6 April 2016 at 22:28, David Rowley <david.rowley@2ndquadrant.com> wrote:On 7 April 2016 at 09:25, Simon Riggs <simon@2ndquadrant.com> wrote:\\n> On 5 April 2016 at 19:33, Robert Haas <robertmhaas@gmail.com> wrote:\\n>\\n>>\\n>> Committed 0002+0003 with those changes, some minor cosmetic stuff, and\\n>> of course the obligatory catversion bump. Oh, and fixed an OID\\n>> conflict with the patch Magnus just committed.\\n>\\n>\\n> Is that everything now? I don't see anything about combining aggs in the git\\n> log and this is still showing as UnCommitted in the CF app.\\n\\nThere's just a documentation patch and two combine functions for\\nfloating point aggs left now (Haribabu's patch)Ah, I see, it was the commit about \\\"multi-stage aggregation\\\".So SELECT sum(x), avg(x) is now optimized? -- Simon Riggs http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Wed, 6 Apr 2016 22:59:12 +0100\",\n \"msg_from\": \"Simon Riggs <simon@2ndQuadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Wed, Apr 6, 2016 at 5:28 PM, David Rowley\\n<david.rowley@2ndquadrant.com> wrote:\\n>> Is that everything now? I don't see anything about combining aggs in the git\\n>> log and this is still showing as UnCommitted in the CF app.\\n>\\n> There's just a documentation patch and two combine functions for\\n> floating point aggs left now (Haribabu's patch)\\n\\nI'll go have a look at that now. Hopefully it will be in good shape\\nand committable; if not, it'll have to wait for 9.7.\\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, 8 Apr 2016 11:57:42 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On Fri, Apr 8, 2016 at 11:57 AM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Wed, Apr 6, 2016 at 5:28 PM, David Rowley\\n> <david.rowley@2ndquadrant.com> wrote:\\n>>> Is that everything now? I don't see anything about combining aggs in the git\\n>>> log and this is still showing as UnCommitted in the CF app.\\n>>\\n>> There's just a documentation patch and two combine functions for\\n>> floating point aggs left now (Haribabu's patch)\\n>\\n> I'll go have a look at that now. Hopefully it will be in good shape\\n> and committable; if not, it'll have to wait for 9.7.\\n\\nAnd... it's pretty hard to argue with any of this. Committed.\\n\\nI am *so* glad to be done with this patch series. Man, that was a lot of work.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 8 Apr 2016 13:49:30 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n },\n {\n \"msg_contents\": \"On 9 April 2016 at 05:49, Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Fri, Apr 8, 2016 at 11:57 AM, Robert Haas <robertmhaas@gmail.com> wrote:\\n>> On Wed, Apr 6, 2016 at 5:28 PM, David Rowley\\n>> <david.rowley@2ndquadrant.com> wrote:\\n>>>> Is that everything now? I don't see anything about combining aggs in the git\\n>>>> log and this is still showing as UnCommitted in the CF app.\\n>>>\\n>>> There's just a documentation patch and two combine functions for\\n>>> floating point aggs left now (Haribabu's patch)\\n>>\\n>> I'll go have a look at that now. Hopefully it will be in good shape\\n>> and committable; if not, it'll have to wait for 9.7.\\n>\\n> And... it's pretty hard to argue with any of this. Committed.\\n>\\n> I am *so* glad to be done with this patch series. Man, that was a lot of work.\\n\\nGreat!\\n\\nThanks for committing all these Robert. I really appreciate it, and I\\nknow I'm not the only one.\\n\\n-- \\n David Rowley 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, 9 Apr 2016 22:13:21 +1200\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Combining Aggregates\"\n }\n]","isConversational":false}}},{"rowIdx":39981,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi there,\n\nI was doing some translation on postgres.po and found a string which\nlooks mistaken.\n\n#: libpq/auth.c:1593\n#, fuzzy, c-format\nmsgid \"could not to look up local user ID %ld: %s\"\n\nIt looks like there is an extra *to* there , so the string should be:\n\n\"could not look up local user ID %ld: %s\"\n\nCheers,\n\n-- \nMartín Marqués http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 17 Dec 2014 09:58:24 -0300","msg_from":"=?UTF-8?B?TWFydMOtbiBNYXJxdcOpcw==?= <martin@2ndquadrant.com>","msg_from_op":true,"msg_subject":"postgres messages error"},{"msg_contents":"Martin,\n\n#: libpq/auth.c:1593\n> #, fuzzy, c-format\n> msgid \"could not to look up local user ID %ld: %s\"\n>\n> It looks like there is an extra *to* there , so the string should be:\n>\n> \"could not look up local user ID %ld: %s\"\n>\n\nI think you are right. FWIW, I have attached a patch that fixes it for\nconsideration if others concur.\n\n-Adam\n\n-- \nAdam Brightwell - adam.brightwell@crunchydatasolutions.com\nDatabase Engineer - www.crunchydatasolutions.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","msg_date":"Wed, 17 Dec 2014 12:07:00 -0500","msg_from":"Adam Brightwell <adam.brightwell@crunchydatasolutions.com>","msg_from_op":false,"msg_subject":"Re: postgres messages error"},{"msg_contents":"Adam Brightwell <adam.brightwell@crunchydatasolutions.com> writes:\n> Martin,\n> #: libpq/auth.c:1593\n>> #, fuzzy, c-format\n>> msgid \"could not to look up local user ID %ld: %s\"\n>> \n>> It looks like there is an extra *to* there , so the string should be:\n>> \n>> \"could not look up local user ID %ld: %s\"\n\n> I think you are right. FWIW, I have attached a patch that fixes it for\n> consideration if others concur.\n\nI agree, that good English is not. Will commit.\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, 17 Dec 2014 13:04:30 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: postgres messages error"},{"msg_contents":"Tom Lane wrote:\n> Adam Brightwell <adam.brightwell@crunchydatasolutions.com> writes:\n> > Martin,\n> > #: libpq/auth.c:1593\n> >> #, fuzzy, c-format\n> >> msgid \"could not to look up local user ID %ld: %s\"\n> >> \n> >> It looks like there is an extra *to* there , so the string should be:\n> >> \n> >> \"could not look up local user ID %ld: %s\"\n> \n> > I think you are right. FWIW, I have attached a patch that fixes it for\n> > consideration if others concur.\n> \n> I agree, that good English is not. Will commit.\n\nA quick grep reveals also\n\ninitdb.po:msgid \"%s: could not to allocate SIDs: error code %lu\\n\"\n\n-- \n�lvaro Herrera http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 17 Dec 2014 15:08:53 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: postgres messages error"},{"msg_contents":"Alvaro Herrera <alvherre@2ndquadrant.com> writes:\n> A quick grep reveals also\n> initdb.po:msgid \"%s: could not to allocate SIDs: error code %lu\\n\"\n\nGood catch, I'll get that one too.\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, 17 Dec 2014 13:18:17 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: postgres messages error"}],"string":"[\n {\n \"msg_contents\": \"Hi there,\\n\\nI was doing some translation on postgres.po and found a string which\\nlooks mistaken.\\n\\n#: libpq/auth.c:1593\\n#, fuzzy, c-format\\nmsgid \\\"could not to look up local user ID %ld: %s\\\"\\n\\nIt looks like there is an extra *to* there , so the string should be:\\n\\n\\\"could not look up local user ID %ld: %s\\\"\\n\\nCheers,\\n\\n-- \\nMartín Marqués http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 17 Dec 2014 09:58:24 -0300\",\n \"msg_from\": \"=?UTF-8?B?TWFydMOtbiBNYXJxdcOpcw==?= <martin@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"postgres messages error\"\n },\n {\n \"msg_contents\": \"Martin,\\n\\n#: libpq/auth.c:1593\\n> #, fuzzy, c-format\\n> msgid \\\"could not to look up local user ID %ld: %s\\\"\\n>\\n> It looks like there is an extra *to* there , so the string should be:\\n>\\n> \\\"could not look up local user ID %ld: %s\\\"\\n>\\n\\nI think you are right. FWIW, I have attached a patch that fixes it for\\nconsideration if others concur.\\n\\n-Adam\\n\\n-- \\nAdam Brightwell - adam.brightwell@crunchydatasolutions.com\\nDatabase Engineer - www.crunchydatasolutions.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Wed, 17 Dec 2014 12:07:00 -0500\",\n \"msg_from\": \"Adam Brightwell <adam.brightwell@crunchydatasolutions.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: postgres messages error\"\n },\n {\n \"msg_contents\": \"Adam Brightwell <adam.brightwell@crunchydatasolutions.com> writes:\\n> Martin,\\n> #: libpq/auth.c:1593\\n>> #, fuzzy, c-format\\n>> msgid \\\"could not to look up local user ID %ld: %s\\\"\\n>> \\n>> It looks like there is an extra *to* there , so the string should be:\\n>> \\n>> \\\"could not look up local user ID %ld: %s\\\"\\n\\n> I think you are right. FWIW, I have attached a patch that fixes it for\\n> consideration if others concur.\\n\\nI agree, that good English is not. Will commit.\\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, 17 Dec 2014 13:04:30 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: postgres messages error\"\n },\n {\n \"msg_contents\": \"Tom Lane wrote:\\n> Adam Brightwell <adam.brightwell@crunchydatasolutions.com> writes:\\n> > Martin,\\n> > #: libpq/auth.c:1593\\n> >> #, fuzzy, c-format\\n> >> msgid \\\"could not to look up local user ID %ld: %s\\\"\\n> >> \\n> >> It looks like there is an extra *to* there , so the string should be:\\n> >> \\n> >> \\\"could not look up local user ID %ld: %s\\\"\\n> \\n> > I think you are right. FWIW, I have attached a patch that fixes it for\\n> > consideration if others concur.\\n> \\n> I agree, that good English is not. Will commit.\\n\\nA quick grep reveals also\\n\\ninitdb.po:msgid \\\"%s: could not to allocate SIDs: error code %lu\\\\n\\\"\\n\\n-- \\n�lvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 17 Dec 2014 15:08:53 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: postgres messages error\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n> A quick grep reveals also\\n> initdb.po:msgid \\\"%s: could not to allocate SIDs: error code %lu\\\\n\\\"\\n\\nGood catch, I'll get that one too.\\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, 17 Dec 2014 13:18:17 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: postgres messages error\"\n }\n]","isConversational":false}}},{"rowIdx":39982,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"At the end of archive recovery, we copy the last segment from the old \ntimeline, to initialize the first segment on the new timeline. For \nexample, if the timeline switch happens in the middle of WAL segment \n000000010000000000000005, the whole 000000010000000000000005 segment is \ncopied to become 000000020000000000000005. The copying is necessary, so \nthat the new segment contains valid data up to the switch point.\n\nHowever, we wouldn't really need to copy the whole segment, copying up \nto the switch point would be enough. In fact, copying the whole segment \nis a bad idea, because the copied WAL looks valid on the new timeline \ntoo. When we read the WAL at crash recovery, we rely on a number of \nthings to determine if the next WAL record is valid. Most importantly, \nthe checksum, and the prev-pointer. The checksum protects any random \ndata from appearing valid, and the prev-pointer makes sure that a WAL \nrecord copied from another location in the WAL is not mistaken as valid. \nThe prev-pointer is particularly important when we recycle old WAL \nsegments as new, because the old segment contains valid WAL records with \nchecksums and all. When we copy a WAL segment with the same segment \nnumber, the prev pointer doesn't protect us, as there can be WAL records \nat the exact same locations in both segments. There is a timeline ID on \nthe page header, but we could still be mistaken within the page. Also, \nwe are lenient with the TLI at start of WAL recovery, when we read the \nfirst WAL record after the checkpoint. There are further safeguards, \nlike the fact that when writing WAL, we always write full blocks. But \nthe write could still be torn at the OS or disk level, if you crash \nafter writing the WAL, but before fsyncing it.\n\nThis is largely academic, but I was able to craft a test case where WAL \nrecovery mistakenly starts to replay the WAL copied from the old \ntimeline, as if it was on the new timeline. Attached is a shell script I \nused. It's very sensitive to the lengths of the WAL records, so probably \nonly works on a similar platform as mine (x86_64 Linux). Running \npitr-test.sh ends with this:\n\nS LOG: database system was interrupted; last known up at 2014-12-17 \n15:15:42 EET\nS LOG: database system was not properly shut down; automatic recovery \nin progress\nS LOG: redo starts at 0/50A2018\nS PANIC: heap_insert_redo: invalid max offset number\nS CONTEXT: xlog redo Heap/INSERT: off 28\nS LOG: startup process (PID 10640) was terminated by signal 6: Aborted\nS LOG: aborting startup due to startup process failure\n\nThat PANIC happens because it tries to apply WAL from different \ntimeline, and it doesn't work because it missed an earlier change to the \nsame page it modifies. (If you were unlucky, you could get silent \ncorruption instead, if the WAL record happens to apply without an error)\n\nA simple way to avoid this is to copy the old WAL segment only up to the \npoint of the timeline switch, and zero the rest.\n\n\nAnother thing I noticed is that we copy the last old WAL segment on the \nnew timeline, even if the timeline switch happens at a segment boundary. \nIn that case, the copied WAL segment is 100% identical to the old \nsegment; it contains no records belonging to the new timeline. I guess \nthat's not wrong per se, but it seems pointless and confusing.\n\nAttached is a patch that addresses both of those issues. This doesn't \nseem worth the risk to back-patch, but let's fix these in master.\n\n\nPS. The \"if (endTLI != ThisTimeLineID)\" test in exitArchiveRecovery was \nalways true, because we always switch to a new timeline after archive \nrecovery. I turned that into an Assert.\n\n- Heikki\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, 17 Dec 2014 15:40:32 +0200","msg_from":"Heikki Linnakangas <hlinnakangas@vmware.com>","msg_from_op":true,"msg_subject":"exitArchiveRecovery woes"},{"msg_contents":"On Wed, Dec 17, 2014 at 8:40 AM, Heikki Linnakangas\n<hlinnakangas@vmware.com> wrote:\n> At the end of archive recovery, we copy the last segment from the old\n> timeline, to initialize the first segment on the new timeline. For example,\n> if the timeline switch happens in the middle of WAL segment\n> 000000010000000000000005, the whole 000000010000000000000005 segment is\n> copied to become 000000020000000000000005. The copying is necessary, so that\n> the new segment contains valid data up to the switch point.\n>\n> However, we wouldn't really need to copy the whole segment, copying up to\n> the switch point would be enough. In fact, copying the whole segment is a\n> bad idea, because the copied WAL looks valid on the new timeline too.\n\nYour proposed change makes sense to me, but why do we need the segment\nto contain valid data up to the switch point? It seems like the\nswitch between timelines should be \"crisper\": replay WAL on the old\ntimeline only from the old segment, and from the new timeline only on\nthe new segment. Anything else seems like an invitation to unending\ncorner-case bugs.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 18 Dec 2014 08:53:33 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: exitArchiveRecovery woes"},{"msg_contents":"On 12/18/2014 03:53 PM, Robert Haas wrote:\n> On Wed, Dec 17, 2014 at 8:40 AM, Heikki Linnakangas\n> <hlinnakangas@vmware.com> wrote:\n>> At the end of archive recovery, we copy the last segment from the old\n>> timeline, to initialize the first segment on the new timeline. For example,\n>> if the timeline switch happens in the middle of WAL segment\n>> 000000010000000000000005, the whole 000000010000000000000005 segment is\n>> copied to become 000000020000000000000005. The copying is necessary, so that\n>> the new segment contains valid data up to the switch point.\n>>\n>> However, we wouldn't really need to copy the whole segment, copying up to\n>> the switch point would be enough. In fact, copying the whole segment is a\n>> bad idea, because the copied WAL looks valid on the new timeline too.\n>\n> Your proposed change makes sense to me, but why do we need the segment\n> to contain valid data up to the switch point? It seems like the\n> switch between timelines should be \"crisper\": replay WAL on the old\n> timeline only from the old segment, and from the new timeline only on\n> the new segment. Anything else seems like an invitation to unending\n> corner-case bugs.\n\nTrue. That would require some changes to the way archive recovery works, \nthough. Currently, when our recovery target timeline is, for example, 5, \nwhose parents are 4 and 3, and we're currently on timeline 3, we will \ntry to restore each segment first with timeline ID 5, then 4, then 3. \nIt's a bit silly, because we know the timeline history and the exact \npoints where the timelines changed, so we could just fetch the correct \none. That would be a good idea, but I'm going to go ahead with just this \nsmaller change now.\n\n- Heikki\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 18 Dec 2014 17:49:28 +0200","msg_from":"Heikki Linnakangas <hlinnakangas@vmware.com>","msg_from_op":true,"msg_subject":"Re: exitArchiveRecovery woes"},{"msg_contents":"On Thu, Dec 18, 2014 at 10:49 AM, Heikki Linnakangas\n<hlinnakangas@vmware.com> wrote:\n> On 12/18/2014 03:53 PM, Robert Haas wrote:\n>> On Wed, Dec 17, 2014 at 8:40 AM, Heikki Linnakangas\n>> <hlinnakangas@vmware.com> wrote:\n>>>\n>>> At the end of archive recovery, we copy the last segment from the old\n>>> timeline, to initialize the first segment on the new timeline. For\n>>> example,\n>>> if the timeline switch happens in the middle of WAL segment\n>>> 000000010000000000000005, the whole 000000010000000000000005 segment is\n>>> copied to become 000000020000000000000005. The copying is necessary, so\n>>> that\n>>> the new segment contains valid data up to the switch point.\n>>>\n>>> However, we wouldn't really need to copy the whole segment, copying up to\n>>> the switch point would be enough. In fact, copying the whole segment is a\n>>> bad idea, because the copied WAL looks valid on the new timeline too.\n>>\n>>\n>> Your proposed change makes sense to me, but why do we need the segment\n>> to contain valid data up to the switch point? It seems like the\n>> switch between timelines should be \"crisper\": replay WAL on the old\n>> timeline only from the old segment, and from the new timeline only on\n>> the new segment. Anything else seems like an invitation to unending\n>> corner-case bugs.\n>\n> True. That would require some changes to the way archive recovery works,\n> though. Currently, when our recovery target timeline is, for example, 5,\n> whose parents are 4 and 3, and we're currently on timeline 3, we will try to\n> restore each segment first with timeline ID 5, then 4, then 3. It's a bit\n> silly, because we know the timeline history and the exact points where the\n> timelines changed, so we could just fetch the correct one. That would be a\n> good idea, but I'm going to go ahead with just this smaller change now.\n\nYeah, it's a separate issue.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 18 Dec 2014 12:27:15 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: exitArchiveRecovery woes"}],"string":"[\n {\n \"msg_contents\": \"At the end of archive recovery, we copy the last segment from the old \\ntimeline, to initialize the first segment on the new timeline. For \\nexample, if the timeline switch happens in the middle of WAL segment \\n000000010000000000000005, the whole 000000010000000000000005 segment is \\ncopied to become 000000020000000000000005. The copying is necessary, so \\nthat the new segment contains valid data up to the switch point.\\n\\nHowever, we wouldn't really need to copy the whole segment, copying up \\nto the switch point would be enough. In fact, copying the whole segment \\nis a bad idea, because the copied WAL looks valid on the new timeline \\ntoo. When we read the WAL at crash recovery, we rely on a number of \\nthings to determine if the next WAL record is valid. Most importantly, \\nthe checksum, and the prev-pointer. The checksum protects any random \\ndata from appearing valid, and the prev-pointer makes sure that a WAL \\nrecord copied from another location in the WAL is not mistaken as valid. \\nThe prev-pointer is particularly important when we recycle old WAL \\nsegments as new, because the old segment contains valid WAL records with \\nchecksums and all. When we copy a WAL segment with the same segment \\nnumber, the prev pointer doesn't protect us, as there can be WAL records \\nat the exact same locations in both segments. There is a timeline ID on \\nthe page header, but we could still be mistaken within the page. Also, \\nwe are lenient with the TLI at start of WAL recovery, when we read the \\nfirst WAL record after the checkpoint. There are further safeguards, \\nlike the fact that when writing WAL, we always write full blocks. But \\nthe write could still be torn at the OS or disk level, if you crash \\nafter writing the WAL, but before fsyncing it.\\n\\nThis is largely academic, but I was able to craft a test case where WAL \\nrecovery mistakenly starts to replay the WAL copied from the old \\ntimeline, as if it was on the new timeline. Attached is a shell script I \\nused. It's very sensitive to the lengths of the WAL records, so probably \\nonly works on a similar platform as mine (x86_64 Linux). Running \\npitr-test.sh ends with this:\\n\\nS LOG: database system was interrupted; last known up at 2014-12-17 \\n15:15:42 EET\\nS LOG: database system was not properly shut down; automatic recovery \\nin progress\\nS LOG: redo starts at 0/50A2018\\nS PANIC: heap_insert_redo: invalid max offset number\\nS CONTEXT: xlog redo Heap/INSERT: off 28\\nS LOG: startup process (PID 10640) was terminated by signal 6: Aborted\\nS LOG: aborting startup due to startup process failure\\n\\nThat PANIC happens because it tries to apply WAL from different \\ntimeline, and it doesn't work because it missed an earlier change to the \\nsame page it modifies. (If you were unlucky, you could get silent \\ncorruption instead, if the WAL record happens to apply without an error)\\n\\nA simple way to avoid this is to copy the old WAL segment only up to the \\npoint of the timeline switch, and zero the rest.\\n\\n\\nAnother thing I noticed is that we copy the last old WAL segment on the \\nnew timeline, even if the timeline switch happens at a segment boundary. \\nIn that case, the copied WAL segment is 100% identical to the old \\nsegment; it contains no records belonging to the new timeline. I guess \\nthat's not wrong per se, but it seems pointless and confusing.\\n\\nAttached is a patch that addresses both of those issues. This doesn't \\nseem worth the risk to back-patch, but let's fix these in master.\\n\\n\\nPS. The \\\"if (endTLI != ThisTimeLineID)\\\" test in exitArchiveRecovery was \\nalways true, because we always switch to a new timeline after archive \\nrecovery. I turned that into an Assert.\\n\\n- Heikki\\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, 17 Dec 2014 15:40:32 +0200\",\n \"msg_from\": \"Heikki Linnakangas <hlinnakangas@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"exitArchiveRecovery woes\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 17, 2014 at 8:40 AM, Heikki Linnakangas\\n<hlinnakangas@vmware.com> wrote:\\n> At the end of archive recovery, we copy the last segment from the old\\n> timeline, to initialize the first segment on the new timeline. For example,\\n> if the timeline switch happens in the middle of WAL segment\\n> 000000010000000000000005, the whole 000000010000000000000005 segment is\\n> copied to become 000000020000000000000005. The copying is necessary, so that\\n> the new segment contains valid data up to the switch point.\\n>\\n> However, we wouldn't really need to copy the whole segment, copying up to\\n> the switch point would be enough. In fact, copying the whole segment is a\\n> bad idea, because the copied WAL looks valid on the new timeline too.\\n\\nYour proposed change makes sense to me, but why do we need the segment\\nto contain valid data up to the switch point? It seems like the\\nswitch between timelines should be \\\"crisper\\\": replay WAL on the old\\ntimeline only from the old segment, and from the new timeline only on\\nthe new segment. Anything else seems like an invitation to unending\\ncorner-case bugs.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 18 Dec 2014 08:53:33 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: exitArchiveRecovery woes\"\n },\n {\n \"msg_contents\": \"On 12/18/2014 03:53 PM, Robert Haas wrote:\\n> On Wed, Dec 17, 2014 at 8:40 AM, Heikki Linnakangas\\n> <hlinnakangas@vmware.com> wrote:\\n>> At the end of archive recovery, we copy the last segment from the old\\n>> timeline, to initialize the first segment on the new timeline. For example,\\n>> if the timeline switch happens in the middle of WAL segment\\n>> 000000010000000000000005, the whole 000000010000000000000005 segment is\\n>> copied to become 000000020000000000000005. The copying is necessary, so that\\n>> the new segment contains valid data up to the switch point.\\n>>\\n>> However, we wouldn't really need to copy the whole segment, copying up to\\n>> the switch point would be enough. In fact, copying the whole segment is a\\n>> bad idea, because the copied WAL looks valid on the new timeline too.\\n>\\n> Your proposed change makes sense to me, but why do we need the segment\\n> to contain valid data up to the switch point? It seems like the\\n> switch between timelines should be \\\"crisper\\\": replay WAL on the old\\n> timeline only from the old segment, and from the new timeline only on\\n> the new segment. Anything else seems like an invitation to unending\\n> corner-case bugs.\\n\\nTrue. That would require some changes to the way archive recovery works, \\nthough. Currently, when our recovery target timeline is, for example, 5, \\nwhose parents are 4 and 3, and we're currently on timeline 3, we will \\ntry to restore each segment first with timeline ID 5, then 4, then 3. \\nIt's a bit silly, because we know the timeline history and the exact \\npoints where the timelines changed, so we could just fetch the correct \\none. That would be a good idea, but I'm going to go ahead with just this \\nsmaller change now.\\n\\n- Heikki\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 18 Dec 2014 17:49:28 +0200\",\n \"msg_from\": \"Heikki Linnakangas <hlinnakangas@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: exitArchiveRecovery woes\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 18, 2014 at 10:49 AM, Heikki Linnakangas\\n<hlinnakangas@vmware.com> wrote:\\n> On 12/18/2014 03:53 PM, Robert Haas wrote:\\n>> On Wed, Dec 17, 2014 at 8:40 AM, Heikki Linnakangas\\n>> <hlinnakangas@vmware.com> wrote:\\n>>>\\n>>> At the end of archive recovery, we copy the last segment from the old\\n>>> timeline, to initialize the first segment on the new timeline. For\\n>>> example,\\n>>> if the timeline switch happens in the middle of WAL segment\\n>>> 000000010000000000000005, the whole 000000010000000000000005 segment is\\n>>> copied to become 000000020000000000000005. The copying is necessary, so\\n>>> that\\n>>> the new segment contains valid data up to the switch point.\\n>>>\\n>>> However, we wouldn't really need to copy the whole segment, copying up to\\n>>> the switch point would be enough. In fact, copying the whole segment is a\\n>>> bad idea, because the copied WAL looks valid on the new timeline too.\\n>>\\n>>\\n>> Your proposed change makes sense to me, but why do we need the segment\\n>> to contain valid data up to the switch point? It seems like the\\n>> switch between timelines should be \\\"crisper\\\": replay WAL on the old\\n>> timeline only from the old segment, and from the new timeline only on\\n>> the new segment. Anything else seems like an invitation to unending\\n>> corner-case bugs.\\n>\\n> True. That would require some changes to the way archive recovery works,\\n> though. Currently, when our recovery target timeline is, for example, 5,\\n> whose parents are 4 and 3, and we're currently on timeline 3, we will try to\\n> restore each segment first with timeline ID 5, then 4, then 3. It's a bit\\n> silly, because we know the timeline history and the exact points where the\\n> timelines changed, so we could just fetch the correct one. That would be a\\n> good idea, but I'm going to go ahead with just this smaller change now.\\n\\nYeah, it's a separate issue.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 18 Dec 2014 12:27:15 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: exitArchiveRecovery woes\"\n }\n]","isConversational":false}}},{"rowIdx":39983,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nThe attached patch just removes one bad-looking blank line in the\ncomments at the top of a function in explain.c.\n\nThanks,\n\nBest regards,\nEtsuro Fujita\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Thu, 18 Dec 2014 12:52:10 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Minor improvement to explain.c"},{"msg_contents":"On Thu, Dec 18, 2014 at 12:52 PM, Etsuro Fujita\n<fujita.etsuro@lab.ntt.co.jp> wrote:\n> Hi,\n>\n> The attached patch just removes one bad-looking blank line in the\n> comments at the top of a function in explain.c.\n\nApplied.\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":"Thu, 18 Dec 2014 17:34:59 +0900","msg_from":"Fujii Masao <masao.fujii@gmail.com>","msg_from_op":false,"msg_subject":"Re: Minor improvement to explain.c"},{"msg_contents":"(2014/12/18 17:34), Fujii Masao wrote:\n> On Thu, Dec 18, 2014 at 12:52 PM, Etsuro Fujita\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\n>> The attached patch just removes one bad-looking blank line in the\n>> comments at the top of a function in explain.c.\n>\n> Applied.\n\nThanks!\n\nBest regards,\nEtsuro Fujita\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 18 Dec 2014 17:57:11 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Minor improvement to explain.c"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nThe attached patch just removes one bad-looking blank line in the\\ncomments at the top of a function in explain.c.\\n\\nThanks,\\n\\nBest regards,\\nEtsuro Fujita\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Thu, 18 Dec 2014 12:52:10 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Minor improvement to explain.c\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 18, 2014 at 12:52 PM, Etsuro Fujita\\n<fujita.etsuro@lab.ntt.co.jp> wrote:\\n> Hi,\\n>\\n> The attached patch just removes one bad-looking blank line in the\\n> comments at the top of a function in explain.c.\\n\\nApplied.\\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\": \"Thu, 18 Dec 2014 17:34:59 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Minor improvement to explain.c\"\n },\n {\n \"msg_contents\": \"(2014/12/18 17:34), Fujii Masao wrote:\\n> On Thu, Dec 18, 2014 at 12:52 PM, Etsuro Fujita\\n> <fujita.etsuro@lab.ntt.co.jp> wrote:\\n>> The attached patch just removes one bad-looking blank line in the\\n>> comments at the top of a function in explain.c.\\n>\\n> Applied.\\n\\nThanks!\\n\\nBest regards,\\nEtsuro Fujita\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 18 Dec 2014 17:57:11 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Minor improvement to explain.c\"\n }\n]","isConversational":false}}},{"rowIdx":39984,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi all,\n\nToday I spent a bit of time looking at the activity of autovacuum for\none table particularly bloated. log_autovacuum_min_duration was\nenabled and set to a high value but even with that I had to deal with\nsome spam from the jobs of other tables. It would be cool to have the\npossibility to do some filtering IMO. So, what about having a relopt\nto control log_autovacuum_min_duration? RelationData->rd_options is\nlargely accessible for a given relation in the code paths of vacuum\nand analyze.\nThoughts?\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, 18 Dec 2014 22:23:30 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Table-level log_autovacuum_min_duration"},{"msg_contents":"Michael Paquier wrote:\n> Hi all,\n> \n> Today I spent a bit of time looking at the activity of autovacuum for\n> one table particularly bloated. log_autovacuum_min_duration was\n> enabled and set to a high value but even with that I had to deal with\n> some spam from the jobs of other tables. It would be cool to have the\n> possibility to do some filtering IMO. So, what about having a relopt\n> to control log_autovacuum_min_duration? RelationData->rd_options is\n> largely accessible for a given relation in the code paths of vacuum\n> and analyze.\n\n+1\n\n-- \n�lvaro Herrera http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 18 Dec 2014 11:15:40 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Thu, Dec 18, 2014 at 11:15 PM, Alvaro Herrera\n<alvherre@2ndquadrant.com> wrote:\n> Michael Paquier wrote:\n>> Hi all,\n>>\n>> Today I spent a bit of time looking at the activity of autovacuum for\n>> one table particularly bloated. log_autovacuum_min_duration was\n>> enabled and set to a high value but even with that I had to deal with\n>> some spam from the jobs of other tables. It would be cool to have the\n>> possibility to do some filtering IMO. So, what about having a relopt\n>> to control log_autovacuum_min_duration? RelationData->rd_options is\n>> largely accessible for a given relation in the code paths of vacuum\n>> and analyze.\n>\n> +1\n\nAs long as I got this idea in mind, I looked at the code to see how\nmuch it would be possible to tune log_autovacuum_min_duration in the\ncode paths of analyze and vacuum. First, I think that it would be good\nto have parameters for both parent relations and their toast relation\nsimilarly to the other parameters...\n\nBut now, here are some things to consider if we use directly the\nreloptions available with RelationData:\n- If the parameters toast.autovacuum_* are not set, toast relations\ninherit values from their parent relation. Looking at autovacuum.c\nwhich is the only place where autovacuum options are located, we keep\na hash table to save the mapping toast -> parent relation. Using that\nit is easy to fetch for a given toast relation the relopts of its\nparent. Note this is strictly located in the autovacuum code path, so\nto let vacuum and analyze now the custom value of\nlog_autovacuum_min_duration, with the inheritance property kept, we\nwould need to pass some extra values from autovacuum to the calls of\nvacuum().\n- It would not be possible to log autovacuum and analyze being skipped\nwhen lock is not available because in this case Relation cannot be\nsafely fetched, so there are no reltoptions directly available. This\nis for example this kind of message:\nskipping analyze of \"foo\" --- lock not available\n\nBoth those things could be solved by passing a value through\nVacuumStmt, like what we do when passing a value for\nmultixact_freeze_min_age, or with an extra argument in vacuum() for\nexample. Now I am not sure if it is worth it, and we could live as\nwell with a parameter that do not support the inheritance parent\nrelation -> toast, so log value set for a toast relation and its\nparent share no dependency. In short that's a trade between code\nsimplicity and usability.\n\nThoughts?\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 Dec 2014 23:17:21 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Sat, Dec 20, 2014 at 9:17 AM, Michael Paquier\n<michael.paquier@gmail.com> wrote:\n> But now, here are some things to consider if we use directly the\n> reloptions available with RelationData:\n> - If the parameters toast.autovacuum_* are not set, toast relations\n> inherit values from their parent relation. Looking at autovacuum.c\n> which is the only place where autovacuum options are located, we keep\n> a hash table to save the mapping toast -> parent relation. Using that\n> it is easy to fetch for a given toast relation the relopts of its\n> parent. Note this is strictly located in the autovacuum code path, so\n> to let vacuum and analyze now the custom value of\n> log_autovacuum_min_duration, with the inheritance property kept, we\n> would need to pass some extra values from autovacuum to the calls of\n> vacuum().\n> - It would not be possible to log autovacuum and analyze being skipped\n> when lock is not available because in this case Relation cannot be\n> safely fetched, so there are no reltoptions directly available. This\n> is for example this kind of message:\n> skipping analyze of \"foo\" --- lock not available\n>\n> Both those things could be solved by passing a value through\n> VacuumStmt, like what we do when passing a value for\n> multixact_freeze_min_age, or with an extra argument in vacuum() for\n> example. Now I am not sure if it is worth it, and we could live as\n> well with a parameter that do not support the inheritance parent\n> relation -> toast, so log value set for a toast relation and its\n> parent share no dependency. In short that's a trade between code\n> simplicity and usability.\n\nI'm not sure I follow all of the particulars of what you are asking\nhere, but in general I would say that you shouldn't hesitate to pass\nmore information down the call stack when that helps to obtain correct\nbehavior.\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 Dec 2014 08:31:14 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Sat, Dec 20, 2014 at 11:17 PM, Michael Paquier\n<michael.paquier@gmail.com> wrote:\n>> Michael Paquier wrote:\n>>> Today I spent a bit of time looking at the activity of autovacuum for\n>>> one table particularly bloated. log_autovacuum_min_duration was\n>>> enabled and set to a high value but even with that I had to deal with\n>>> some spam from the jobs of other tables. It would be cool to have the\n>>> possibility to do some filtering IMO. So, what about having a relopt\n>>> to control log_autovacuum_min_duration? RelationData->rd_options is\n>>> largely accessible for a given relation in the code paths of vacuum\n>>> and analyze.\n\nOK, instead of only words, attached is a patch adding relopts called\nlog_autovacuum_min_duration and toast.log_autovacuum_min_duration to\ncontrol the logging of autovacuum at relation-level. The default value\nof those parameters is -1, meaning that in this case the global\nlog_autovacuum_min_duration is used to control the logs. The patch has\nfinished by being simpler than I though first by using VacuumStmt to\npass the relopts from autovacuum to the code paths of analyze and\nvacuum. Note that this uses the same mechanisms as the other relopts,\nmeaning that the toast relations use the values of their parent tables\nif it is defined.\n\nI am adding it to the next CF.\n-- \nMichael\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Tue, 13 Jan 2015 14:26:11 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"I wrote:\n> I am adding it to the next CF.\nPatch 2 attached. I forgot the update of copyfuncs.c and equalfuncs.c.\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":"Thu, 15 Jan 2015 14:53:31 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"The following review has been posted through the commitfest application:\nmake installcheck-world: tested, failed\nImplements feature: tested, failed\nSpec compliant: tested, failed\nDocumentation: tested, failed\n\nHi,\n\nI'm Naoya Anzai.\nI've been working as a PostgreSQL Support Engineer for 6 years.\nI am a newcomer of reviewing, and My programming skill is not so high.\nBut the following members also participate in this review. (We say \n\"Two heads are better than one.\" :))\n\nAkira Kurosawa <kurosawa-akira@mxc.nes.nec.co.jp>\nTaiki Kondo <kondo-taiki@mxt.nes.nec.co.jp>\nHuong Dangminh <dangminh-huong@mxm.nes.nec.co.jp>\n\nSo I believe reviewing patches is not difficult for us.\n\nThis is a review of Table-level log_autovacuum_min_duration patch:\nhttp://www.postgresql.org/message-id/CAB7nPqTBQsbEgvb8cOH01K7ARGYs9KBuV8dr+aqGonFVb8koAQ@mail.gmail.com\n\nSubmission review\n====================\nThe patch applies cleanly to HEAD, and it works fine on Fedora \nrelease 20.\nThere is no regression test,but I think it is no problem \nbecause other parameter also is not tested.\n\n\nUsability review\n====================\npg_dump/pg_restore support is OK.\nI think this feature is a nice idea and I also want this feature.\n\n\nFeature test\n====================\nI executed following commands after setting \n\"log_autovacuum_min_duration\" to 1000 in the GUC. (\"bar\" table is \nalready created with no options.)\n\n CREATE TABLE foo ( ... ) WITH ( log_autovacuum_min_duration = 0 );\n ALTER TABLE bar SET (log_autovacuum_min_duration = 0 );\n\nThen, only in \"foo\" and \"bar\" table, autovacuum log was printed out \neven if elapsed time of autovacuum is lesser than 1000ms. This \nbehavior was expected and there was no crash or failed assertion. \nSo it looked good for me. But, I executed following command, in \n\"buzz\" table, autovacuum log was printed out if elapsed time is \nmore than 1000ms.\n\n CREATE TABLE buzz ( ... ) WITH ( log_autovacuum_min_duration = -1 );\n ^^\n\nI expect autovacuum log is NOT printed out even if elapsed time is \nmore than 1000ms in this case. My thought is wrong, isn't it? In my \nopinion, there is an use case that autovacuum log is always printed \nout in all tables excepting specified tables. I think there is a \nperson who wants to use it like this case, but I (or he) can NOT use \nin this situation.\n\nHow about your opinion?\n\n\nPerformance review\n====================\nNot reviewed from this point of view.\n\n\nCoding review\n====================\nI think description of log_autovacuum_min_duration in reloptions.c\n(line:215) should be like other parameters (match with guc.c). So \nit should be \"Sets the minimum execution time above which autovacuum \nactions will be logged.\" but not \"Log autovacuum execution for \ngiven threshold\".\n\nThere is no new function which is defined in this patch, and \nmodified contents are not related to OS-dependent contents, so I \nthink it will work fine on Windows/BSD etc.\n\n\nArchitecture review\n====================\nAbout the modification of gram.y.\n\nI think it is not good that log_min_duration is initialized to \nzeros in gram.y when \"FREEZE\" option is set. Because any VACUUM \nqueries never use this parameter. I think log_min_duration always \nshould be initialized to -1.\n\n\nRegards,\nNaoya Anzai (NEC Solution Innovators,Ltd.)\n\n\nThe new status of this patch is: Waiting on Author\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 06 Feb 2015 00:39:56 +0000","msg_from":"Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"> The following review has been posted through the commitfest application:\r\n> make installcheck-world: tested, failed\r\n> Implements feature: tested, failed\r\n> Spec compliant: tested, failed\r\n> Documentation: tested, failed\r\nI'm sorry, I just sent it by mistake.\r\nAll of them have passed.\r\n\r\n---\r\nNaoya Anzai\r\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 6 Feb 2015 00:50:08 +0000","msg_from":"Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Fri, Feb 6, 2015 at 4:50 AM, Naoya Anzai\n<anzai-naoya@mxu.nes.nec.co.jp> wrote:\n>> The following review has been posted through the commitfest application:\n>> make installcheck-world: tested, failed\n>> Implements feature: tested, failed\n>> Spec compliant: tested, failed\n>> Documentation: tested, failed\n> I'm sorry, I just sent it by mistake.\n> All of them have passed.\nThat's fine. I think you used the UI on the commit fest app, and it is\nnot intuitive that you need to check those boxes at first sight when\nusing it for the first time.\n-- \nMichael\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 6 Feb 2015 07:06:12 +0400","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Fri, Feb 6, 2015 at 4:39 AM, Naoya Anzai wrote:\n> I've been working as a PostgreSQL Support Engineer for 6 years.\n\nThanks for your input Anzai-san.\n\n> I am a newcomer of reviewing, and My programming skill is not so high.\n> But the following members also participate in this review. (We say\n> \"Two heads are better than one.\" :))\n\nFWIW, any review for any kind of patches are always welcome, either\nknowing or not the code you are looking at for a given patch. It is\neven encouraged to look at new areas when possible.\n\n> Feature test\n> ====================\n> [blah]\n> CREATE TABLE buzz ( ... ) WITH ( log_autovacuum_min_duration = -1 );\n> I expect autovacuum log is NOT printed out even if elapsed time is\n> more than 1000ms in this case. My thought is wrong, isn't it? In my\n> opinion, there is an use case that autovacuum log is always printed\n> out in all tables excepting specified tables. I think there is a\n> person who wants to use it like this case, but I (or he) can NOT use\n> in this situation.\n>\n> How about your opinion?\n\nThe patch is working as expected. Similarly to the other parameters\nlike vacuum_cost_delay, a value of -1 is the default behavior, meaning\nthat the underlying GUC parameter is used. I thought about using a\nspecial value like -2 to define the behavior you are mentioning here,\naka with \"-2\" disable custom value and GUC parameter but I don't think\nit is worth adding as that's an ugly 3 line of code of this type:\nif (log_min_duration == -2)\n enforce_log_min = -1;\nAnd you can actually get what you are looking for by setting\nmin_duration through ALTER TABLE to a very high value, like say 2e9 to\nsuppress the autovacuum output of a set of tables out of the rest.\n\n> Coding review\n> ====================\n> I think description of log_autovacuum_min_duration in reloptions.c\n> (line:215) should be like other parameters (match with guc.c). So\n> it should be \"Sets the minimum execution time above which autovacuum\n> actions will be logged.\" but not \"Log autovacuum execution for\n> given threshold\".\n\nWhat about that then?\n\"Minimum execution time above which autovacuum actions will be logged\"\n\n> Architecture review\n> ====================\n> About the modification of gram.y.\n>\n> I think it is not good that log_min_duration is initialized to\n> zeros in gram.y when \"FREEZE\" option is set. Because any VACUUM\n> queries never use this parameter. I think log_min_duration always\n> should be initialized to -1.\n\nLooking at this patch this morning, actually I think that my last\npatch as well as your suggestion are both wrong. To put it in short\nwords, log_min_duration should be set only if VACOPT_VERBOSE is\ndefined in query declaration. So I changed this portion of the patch\nthis way.\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":"Fri, 6 Feb 2015 11:14:45 +0400","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Fri, Feb 6, 2015 at 11:14 AM, Michael Paquier wrote:\n> Looking at this patch this morning, actually I think that my last\n> patch as well as your suggestion are both wrong. To put it in short\n> words, log_min_duration should be set only if VACOPT_VERBOSE is\n> defined in query declaration. So I changed this portion of the patch\n> this way.\nAnd I forgot to change VacuumStmt for the ANALYZE portion in gram.y...\nPatch updated attached.\n-- \nMichael\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Fri, 6 Feb 2015 12:14:14 +0400","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"Thanks for your reply.\n\n>> Feature test\n>> ====================\n(snip)\n> I thought about using a\n> special value like -2 to define the behavior you are mentioning here, \n> aka with \"-2\" disable custom value and GUC parameter but I don't think\n> it is worth adding as that's an ugly 3 line of code of this type:\n> if (log_min_duration == -2)\n> enforce_log_min = -1;\n\nI disscussed about this use case with my co-workers, who said \n\"that code is not good\", then we concluded that it was actually \na rare case. If such a case sometimes happens in fact, I (or someone) \nwill suggest a different way from this patch to handle this case.\n\nWe know there is a practical workaround. :)\n\n>> Coding review\n>> ====================\n>> I think description of log_autovacuum_min_duration in reloptions.c\n>> (line:215) should be like other parameters (match with guc.c). So\n>> it should be \"Sets the minimum execution time above which autovacuum\n>> actions will be logged.\" but not \"Log autovacuum execution for\n>> given threshold\".\n>\n>What about that then?\n>\"Minimum execution time above which autovacuum actions will be logged\"\n\nThat's roughly match. For matching with guc.c, you might be better to \nadd \"Sets the\" to that discription.\n\n>> Architecture review\n>> ====================\n>> About the modification of gram.y.\n>>\n>> I think it is not good that log_min_duration is initialized to\n>> zeros in gram.y when \"FREEZE\" option is set. Because any VACUUM\n>> queries never use this parameter. I think log_min_duration always\n>> should be initialized to -1.\n>\n>Looking at this patch this morning, actually I think that my last\n>patch as well as your suggestion are both wrong. To put it in short\n>words, log_min_duration should be set only if VACOPT_VERBOSE is\n>defined in query declaration. So I changed this portion of the patch\n>this way.\n\n>And I forgot to change VacuumStmt for the ANALYZE portion in gram.y...\n>Patch updated attached.\n\nSorry, I could not correctly express my opinion to you. I mean \n\"log_autovacuum_min_duration\" is used only by AutoVacuum, Any VACUUM \nqueries (including VACUUM VERBOSE) never use this parameter. So, \nwhen someone executes Manual Vacuum, \"log_min_duration\" always should \nbe initialized to -1.\n\nANALYZE should also be the same.\n\nIn other words, it is not necessary to initialize \"log_min_duration\" \nto zero when perform the VACUUM(or ANALYZE) VERBOSE. \"VERBOSE\" is \nprovided only for changing the log level of Manual VACUUM from \nDEBUG2 to INFO.\n\nRegards,\n-----\nNaoya.\n\n\n-- \nSent via 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, 9 Feb 2015 06:47:12 +0000","msg_from":"Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Mon, Feb 9, 2015 at 3:47 PM, Naoya Anzai\n<anzai-naoya@mxu.nes.nec.co.jp> wrote:\n>>> Feature test\n>>> ====================\n> (snip)\n>> I thought about using a\n>> special value like -2 to define the behavior you are mentioning here,\n>> aka with \"-2\" disable custom value and GUC parameter but I don't think\n>> it is worth adding as that's an ugly 3 line of code of this type:\n>> if (log_min_duration == -2)\n>> enforce_log_min = -1;\n>\n> I discussed about this use case with my co-workers, who said\n> \"that code is not good\", then we concluded that it was actually\n> a rare case. If such a case sometimes happens in fact, I (or someone)\n> will suggest a different way from this patch to handle this case.\n>\n> We know there is a practical workaround. :)\n\nOK, done.\n\n>>> Coding review\n>>> ====================\n>>> I think description of log_autovacuum_min_duration in reloptions.c\n>>> (line:215) should be like other parameters (match with guc.c). So\n>>> it should be \"Sets the minimum execution time above which autovacuum\n>>> actions will be logged.\" but not \"Log autovacuum execution for\n>>> given threshold\".\n>>\n>>What about that then?\n>>\"Minimum execution time above which autovacuum actions will be logged\"\n>\n> That's roughly match. For matching with guc.c, you might be better to\n> add \"Sets the\" to that discription.\n\nOK, done this way...\n\n>>And I forgot to change VacuumStmt for the ANALYZE portion in gram.y...\n>>Patch updated attached.\n>\n> Sorry, I could not correctly express my opinion to you. I mean\n> \"log_autovacuum_min_duration\" is used only by AutoVacuum, Any VACUUM\n> queries (including VACUUM VERBOSE) never use this parameter. So,\n> when someone executes Manual Vacuum, \"log_min_duration\" always should\n> be initialized to -1.\n>\n> ANALYZE should also be the same.\n>\n> In other words, it is not necessary to initialize \"log_min_duration\"\n> to zero when perform the VACUUM(or ANALYZE) VERBOSE. \"VERBOSE\" is\n> provided only for changing the log level of Manual VACUUM from\n> DEBUG2 to INFO.\n\nWell, I see your point but this is not completely true: we could as\nwell rely entirely on this parameter instead of VACOPT_VERBOSE to\ndetermine if autovacuum, a vacuum or an analyze are in verbose mode,\nand remove VACOPT_VERBOSE, but I can imagine people complaining if\nVACOPT_VERBOSE is removed. So let's set it up unconditionally to -1 in\ngram.y for now. However if people think that it is fine to remove\nVACOPT_VERBOSE, we could use exclusively this parameter in VacuumStmt.\nOr even add sanity checks at the top of vacuum() to ensure that\nVACOPT_VERBOSE is set only when log_min_duration is positive.\nAdditional opinions on this matter are welcome.\n-- \nMichael\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Tue, 10 Feb 2015 17:31:40 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"An updated patch is attached, I think that previous version broke value\nassignment of log_min_duration in table_recheck_autovac: the value in\nreltoptions should be used only if it is positive. At the same time I wrote\nsome more documentation about the fact that a toast table will use the\nvalue of its parent table if nothing is set.\n-- \nMichael\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Tue, 10 Feb 2015 22:11:48 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"Hi, Michael-san\n\n> An updated patch is attached, \n\nI'm sorry for confusing you.\n\nI think you don't have to implement this code to disable this\nfeature with using value \"-2\".Because this use case is a rare case,\nand there is a practical workaround using huge value like \"2e9\".\n(You suggested \"2e9\" to me, didn't you? :) ) So, please remove this code.\n\n\n> Well, I see your point but this is not completely true: we could as\n> well rely entirely on this parameter instead of VACOPT_VERBOSE to\n> determine if autovacuum, a vacuum or an analyze are in verbose mode,\n> and remove VACOPT_VERBOSE, but I can imagine people complaining if\n> VACOPT_VERBOSE is removed. So let's set it up unconditionally to -1 in\n> gram.y for now. However if people think that it is fine to remove\n> VACOPT_VERBOSE, we could use exclusively this parameter in VacuumStmt.\n> Or even add sanity checks at the top of vacuum() to ensure that\n> VACOPT_VERBOSE is set only when log_min_duration is positive.\n> Additional opinions on this matter are welcome.\n\nI understand your point at last. :)\n\nYou mean that ...\nLog_autovacuum_min_duration assumes a role of VACOPT_VERBOSE.\nEven if this parameter never use currently for manual vacuum,\nlog_autovacuum_min_duration should be set zero(anytime output)\nwhen we executes \"VACUUM(or ANALYZE) VERBOSE\".\nIs my understanding correct? If so,it sounds logical.\n\nIf we can abolish VERBOSE option,\nI think it's ideal that we will prepare a new parameter like \na log_min_duration_vacuum(and log_min_duration_analyze) which \nintegrating \"VERBOSE feature\" and \"log_autovacuum_min_duration\".\n\nRegards,\n---\nNaoya Anzai\n\n\n-- \nSent via 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 Feb 2015 08:44:46 +0000","msg_from":"Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Thu, Feb 12, 2015 at 5:44 PM, Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp>\nwrote:\n\n> Hi, Michael-san\n>\n> > An updated patch is attached,\n>\n> I'm sorry for confusing you.\n>\n> I think you don't have to implement this code to disable this\n> feature with using value \"-2\".Because this use case is a rare case,\n> and there is a practical workaround using huge value like \"2e9\".\n> (You suggested \"2e9\" to me, didn't you? :) ) So, please remove this code.\n>\n\nI will clean up the code.\n\n\n> > Well, I see your point but this is not completely true: we could as\n> > well rely entirely on this parameter instead of VACOPT_VERBOSE to\n> > determine if autovacuum, a vacuum or an analyze are in verbose mode,\n> > and remove VACOPT_VERBOSE, but I can imagine people complaining if\n> > VACOPT_VERBOSE is removed. So let's set it up unconditionally to -1 in\n> > gram.y for now. However if people think that it is fine to remove\n> > VACOPT_VERBOSE, we could use exclusively this parameter in VacuumStmt.\n> > Or even add sanity checks at the top of vacuum() to ensure that\n> > VACOPT_VERBOSE is set only when log_min_duration is positive.\n> > Additional opinions on this matter are welcome.\n>\n> I understand your point at last. :)\n>\n> You mean that ...\n> Log_autovacuum_min_duration assumes a role of VACOPT_VERBOSE.\n> Even if this parameter never use currently for manual vacuum,\n> log_autovacuum_min_duration should be set zero(anytime output)\n> when we executes \"VACUUM(or ANALYZE) VERBOSE\".\n> Is my understanding correct? If so,it sounds logical.\n>\n\nYup, that's my opinion. Now I don't know if people would mind to remove\nVACOPT_VERBOSE and replace the control it does by log_min_duration in\nVacuumStmt. At least both things are overlapping, and log_min_duration\noffers more options than the plain VACOPT_VERBOSE.\n\n\n> If we can abolish VERBOSE option,\n> I think it's ideal that we will prepare a new parameter like\n> a log_min_duration_vacuum(and log_min_duration_analyze) which\n> integrating \"VERBOSE feature\" and \"log_autovacuum_min_duration\".\n>\n\nWhat I think you are proposing here is a VERBOSE option that hypothetically\ngets activated if a manual VACUUM takes more than a certain amount\nspecified by those parameters. I doubt this would be useful. In any case\nthis is unrelated to this patch.\n-- \nMichael\n\nOn Thu, Feb 12, 2015 at 5:44 PM, Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp> wrote:Hi, Michael-san\n\n> An updated patch is attached,\n\nI'm sorry for confusing you.\n\nI think you don't have to implement this code to disable this\nfeature with using value \"-2\".Because this use case is a rare case,\nand there is a practical workaround using huge value like \"2e9\".\n(You suggested \"2e9\" to me, didn't you? :) ) So, please remove this code.I will clean up the code. \n> Well, I see your point but this is not completely true: we could as\n> well rely entirely on this parameter instead of VACOPT_VERBOSE to\n> determine if autovacuum, a vacuum or an analyze are in verbose mode,\n> and remove VACOPT_VERBOSE, but I can imagine people complaining if\n> VACOPT_VERBOSE is removed. So let's set it up unconditionally to -1 in\n> gram.y for now. However if people think that it is fine to remove\n> VACOPT_VERBOSE, we could use exclusively this parameter in VacuumStmt.\n> Or even add sanity checks at the top of vacuum() to ensure that\n> VACOPT_VERBOSE is set only when log_min_duration is positive.\n> Additional opinions on this matter are welcome.\n\nI understand your point at last. :)\n\nYou mean that ...\nLog_autovacuum_min_duration assumes a role of VACOPT_VERBOSE.\nEven if this parameter never use currently for manual vacuum,\nlog_autovacuum_min_duration should be set zero(anytime output)\nwhen we executes \"VACUUM(or ANALYZE) VERBOSE\".\nIs my understanding correct? If so,it sounds logical.Yup, that's my opinion. Now I don't know if people would mind to remove VACOPT_VERBOSE and replace the control it does by log_min_duration in VacuumStmt. At least both things are overlapping, and log_min_duration offers more options than the plain VACOPT_VERBOSE. \nIf we can abolish VERBOSE option,\nI think it's ideal that we will prepare a new parameter like\na log_min_duration_vacuum(and log_min_duration_analyze) which\nintegrating \"VERBOSE feature\" and \"log_autovacuum_min_duration\".\nWhat I think you are proposing here is a VERBOSE option that hypothetically gets activated if a manual VACUUM takes more than a certain amount specified by those parameters. I doubt this would be useful. In any case this is unrelated to this patch.-- Michael","msg_date":"Thu, 12 Feb 2015 20:26:21 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":">> You mean that ...\n>> Log_autovacuum_min_duration assumes a role of VACOPT_VERBOSE.\n>> Even if this parameter never use currently for manual vacuum,\n>> log_autovacuum_min_duration should be set zero(anytime output)\n>> when we executes \"VACUUM(or ANALYZE) VERBOSE\".\n>> Is my understanding correct? If so,it sounds logical.\n>>\n>\n>Yup, that's my opinion. Now I don't know if people would mind to remove\n>VACOPT_VERBOSE and replace the control it does by log_min_duration in\n>VacuumStmt. At least both things are overlapping, and log_min_duration\n>offers more options than the plain VACOPT_VERBOSE.\n\nOK. I agree with you.\nPlease re-implement as you are thinking.\n\n>> If we can abolish VERBOSE option,\n>> I think it's ideal that we will prepare a new parameter like\n>> a log_min_duration_vacuum(and log_min_duration_analyze) which\n>> integrating \"VERBOSE feature\" and \"log_autovacuum_min_duration\".\n>>\n>\n>What I think you are proposing here is a VERBOSE option that hypothetically\n>gets activated if a manual VACUUM takes more than a certain amount\n>specified by those parameters. I doubt this would be useful. In any case\n>this is unrelated to this patch.\n\nI suspect manual vacuum often executes as \"semi-auto vacuum\" \nby job-scheduler, cron, etc in actual maintenance cases. \nWhether auto or manual, I think that's important to output \ntheir logs in the same mechanism.\n\nSorry, I seem to have wandered from the subject.\n\nNaoya\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 13 Feb 2015 01:16:44 +0000","msg_from":"Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Fri, Feb 13, 2015 at 10:16 AM, Naoya Anzai\n<anzai-naoya@mxu.nes.nec.co.jp> wrote:\n>>> You mean that ...\n>>> Log_autovacuum_min_duration assumes a role of VACOPT_VERBOSE.\n>>> Even if this parameter never use currently for manual vacuum,\n>>> log_autovacuum_min_duration should be set zero(anytime output)\n>>> when we executes \"VACUUM(or ANALYZE) VERBOSE\".\n>>> Is my understanding correct? If so,it sounds logical.\n>>>\n>>\n>>Yup, that's my opinion. Now I don't know if people would mind to remove\n>>VACOPT_VERBOSE and replace the control it does by log_min_duration in\n>>VacuumStmt. At least both things are overlapping, and log_min_duration\n>>offers more options than the plain VACOPT_VERBOSE.\n>\n> OK. I agree with you.\n> Please re-implement as you are thinking.\n\nOK will do that today.\n\n>>> If we can abolish VERBOSE option,\n>>> I think it's ideal that we will prepare a new parameter like\n>>> a log_min_duration_vacuum(and log_min_duration_analyze) which\n>>> integrating \"VERBOSE feature\" and \"log_autovacuum_min_duration\".\n>>>\n>>\n>>What I think you are proposing here is a VERBOSE option that hypothetically\n>>gets activated if a manual VACUUM takes more than a certain amount\n>>specified by those parameters. I doubt this would be useful. In any case\n>>this is unrelated to this patch.\n>\n> I suspect manual vacuum often executes as \"semi-auto vacuum\"\n> by job-scheduler, cron, etc in actual maintenance cases.\n> Whether auto or manual, I think that's important to output\n> their logs in the same mechanism.\n>\n> Sorry, I seem to have wandered from the subject.\n\nNo problem. That's a constructive discussion :)\n-- \nMichael\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 13 Feb 2015 11:18:19 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Fri, Feb 13, 2015 at 10:16 AM, Naoya Anzai\n<anzai-naoya@mxu.nes.nec.co.jp> wrote:\n>>> You mean that ...\n>>> Log_autovacuum_min_duration assumes a role of VACOPT_VERBOSE.\n>>> Even if this parameter never use currently for manual vacuum,\n>>> log_autovacuum_min_duration should be set zero(anytime output)\n>>> when we executes \"VACUUM(or ANALYZE) VERBOSE\".\n>>> Is my understanding correct? If so,it sounds logical.\n>>>\n>>\n>>Yup, that's my opinion. Now I don't know if people would mind to remove\n>>VACOPT_VERBOSE and replace the control it does by log_min_duration in\n>>VacuumStmt. At least both things are overlapping, and log_min_duration\n>>offers more options than the plain VACOPT_VERBOSE.\n>\n> OK. I agree with you.\n> Please re-implement as you are thinking.\n\nThanks. Attached is an updated patch will all those things implemented.\n\n>>> If we can abolish VERBOSE option,\n>>> I think it's ideal that we will prepare a new parameter like\n>>> a log_min_duration_vacuum(and log_min_duration_analyze) which\n>>> integrating \"VERBOSE feature\" and \"log_autovacuum_min_duration\".\n>>>\n>>\n>>What I think you are proposing here is a VERBOSE option that hypothetically\n>>gets activated if a manual VACUUM takes more than a certain amount\n>>specified by those parameters. I doubt this would be useful. In any case\n>>this is unrelated to this patch.\n>\n> I suspect manual vacuum often executes as \"semi-auto vacuum\"\n> by job-scheduler, cron, etc in actual maintenance cases.\n> Whether auto or manual, I think that's important to output\n> their logs in the same mechanism.\n>\n> Sorry, I seem to have wandered from the subject.\n\nThis patch is a step in this direction as it enables any backend-side\ncode to set up VacuumStmt with a custom timestamp value to output logs\nafter a given timing, for example in background workers. For the\nclient-side of things, I am unsure if we'd actually want it, we should\nalways VERBOSE when this option is invoked through a query, and not\nadd any hypothetical condition on it...\n\nBtw, after hacking on this it happens that we cannot completely remove\nVACOPT_VERBOSE as gram.y needs to take into account options with\nparenthesis :)\nBut we can limit its use to the query parser only, similarly for VACOPT_FREEZE.\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":"Fri, 13 Feb 2015 13:21:34 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"Hi, Michael\n\nI found that definition of VERBOSE and log_autovacuum is not \npretty match. For example, \"VERBOSE\" can output logs of \nscanning indices and scanning detail of analyze, but \nlog_autovacuum can't output them.\n\nPlease see following sequences.\n\n1. execute these queries.\n\n DROP TABLE t1;\n CREATE TABLE t1(id integer primary key,name text);\n ALTER TABLE t1 SET (log_autovacuum_min_duration=0);\n ALTER TABLE t1 ALTER COLUMN name SET STORAGE EXTERNAL;\n INSERT INTO t1 SELECT GENERATE_SERIES(1,100),repeat('a',3000);\n UPDATE t1 SET name='update';\n\n2. after a while, output the following logs by autovacuum movement.\n(For your convenience, I put the \"### TYPE ###\" prefix on each logs.)\n\n ### VERBOSE ###INFO: vacuuming \"public.t1\"\n ### VERBOSE ###INFO: scanned index \"t1_pkey\" to remove 33 row versions\n ### VERBOSE ###DETAIL: CPU 0.00s/0.00u sec elapsed 0.00 sec.\n ### VERBOSE ###INFO: \"t1\": removed 33 row versions in 1 pages\n ### VERBOSE ###DETAIL: CPU 0.00s/0.00u sec elapsed 0.00 sec.\n ### VERBOSE ###INFO: index \"t1_pkey\" now contains 100 row versions in 2 pages\n ### VERBOSE ###DETAIL: 33 index row versions were removed.\n ### VERBOSE ### 0 index pages have been deleted, 0 are currently reusable.\n ### VERBOSE ### CPU 0.00s/0.00u sec elapsed 0.00 sec.\n ### VERBOSE ###INFO: \"t1\": found 100 removable, 100 nonremovable row versions in 2 out of 2 pages\n ### VERBOSE ###DETAIL: 0 dead row versions cannot be removed yet.\n ### VERBOSE ### There were 0 unused item pointers.\n ### VERBOSE ### Skipped 0 pages due to buffer pins.\n ### VERBOSE ### 0 pages are entirely empty.\n ### VERBOSE ### CPU 0.00s/0.00u sec elapsed 0.00 sec.\n ### LOG_AVAC ###LOG: automatic vacuum of table \"naoya.public.t1\": index scans: 1\n ### LOG_AVAC ### pages: 0 removed, 2 remain, 0 skipped due to pins\n ### LOG_AVAC ### tuples: 100 removed, 100 remain, 0 are dead but not yet removable\n ### LOG_AVAC ### buffer usage: 47 hits, 4 misses, 4 dirtied\n ### LOG_AVAC ### avg read rate: 14.192 MB/s, avg write rate: 14.192 MB/s\n ### LOG_AVAC ### system usage: CPU 0.00s/0.00u sec elapsed 0.00 sec\n ### VERBOSE ###INFO: analyzing \"public.t1\"\n ### VERBOSE ###INFO: \"t1\": scanned 2 of 2 pages, containing 100 live rows and 0 dead rows; 100 rows in sample, 100 estimated total rows\n ### LOG_AVAC ###LOG: automatic analyze of table \"naoya.public.t1\" system usage: CPU 0.00s/0.00u sec elapsed 0.04 sec\n ### VERBOSE ###INFO: vacuuming \"pg_toast.pg_toast_72882\"\n ### VERBOSE ###INFO: scanned index \"pg_toast_72882_index\" to remove 200 row versions\n ### VERBOSE ###DETAIL: CPU 0.00s/0.00u sec elapsed 0.00 sec.\n ### VERBOSE ###INFO: \"pg_toast_72882\": removed 200 row versions in 50 pages\n ### VERBOSE ###DETAIL: CPU 0.00s/0.00u sec elapsed 0.00 sec.\n ### VERBOSE ###INFO: index \"pg_toast_72882_index\" now contains 0 row versions in 2 pages\n ### VERBOSE ###DETAIL: 200 index row versions were removed.\n ### VERBOSE ### 0 index pages have been deleted, 0 are currently reusable.\n ### VERBOSE ### CPU 0.00s/0.00u sec elapsed 0.00 sec.\n ### VERBOSE ###INFO: \"pg_toast_72882\": found 200 removable, 0 nonremovable row versions in 50 out of 50 pages\n ### VERBOSE ###DETAIL: 0 dead row versions cannot be removed yet.\n ### VERBOSE ### There were 0 unused item pointers.\n ### VERBOSE ### Skipped 0 pages due to buffer pins.\n ### VERBOSE ### 0 pages are entirely empty.\n ### VERBOSE ### CPU 0.00s/0.00u sec elapsed 0.00 sec.\n ### VERBOSE ###INFO: \"pg_toast_72882\": truncated 50 to 0 pages\n ### VERBOSE ###DETAIL: CPU 0.00s/0.00u sec elapsed 0.03 sec.\n ### LOG_AVAC ###LOG: automatic vacuum of table \"naoya.pg_toast.pg_toast_72882\": index scans: 1\n ### LOG_AVAC ### pages: 50 removed, 0 remain, 0 skipped due to pins\n ### LOG_AVAC ### tuples: 200 removed, 0 remain, 0 are dead but not yet removable\n ### LOG_AVAC ### buffer usage: 223 hits, 2 misses, 1 dirtied\n ### LOG_AVAC ### avg read rate: 0.457 MB/s, avg write rate: 0.228 MB/s\n ### LOG_AVAC ### system usage: CPU 0.00s/0.00u sec elapsed 0.03 sec\n\nI think VACOPT_VERBOSE should not be easily replaced to log_min_duration>=0.\n\nAnd, in this v7 patch looks that VERBOSE log is always output \nin INFO-Level when log_autovacuum_min_duration is set to 0. \nIs this your point?\n\nRegards,\n---\nNaoya\n\n\n-- \nSent via 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 Feb 2015 08:16:26 +0000","msg_from":"Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Tue, Feb 17, 2015 at 5:16 PM, Naoya Anzai wrote:\n> I found that definition of VERBOSE and log_autovacuum is not\n> pretty match. For example, \"VERBOSE\" can output logs of\n> scanning indices and scanning detail of analyze, but\n> log_autovacuum can't output them.\n> [...]\n\nDoh. I forgot to add !IsAutoVacuumWorkerProcess() in the checks with\nVACOPT_VERBOSE that I previously removed...\n\n> And, in this v7 patch looks that VERBOSE log is always output\n> in INFO-Level when log_autovacuum_min_duration is set to 0.\n> Is this your point?\n\nThat's my bug, fixed in the attached.\n-- \nMichael\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Tue, 17 Feb 2015 23:01:33 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"Hi, Michael\n\nI thought about VACOPT_VERBOSE again.\n\nAs a result, I think you should not delete VACOPT_VERBOSE.\n\nAccording to the last mail I have posted, the difference of \nmanual-vacuum log and auto-vacuum log exists clearly.\nSo, at least you should not touch the mechanism of VACOPT_VERBOSE \nuntil both vacuum log formats are unified to a same format.\n\nIf you agree my think, please undo your removing VACOPT_VERBOSE work.\n\nRegards,\n---\nNaoya\n\n\n-- \nSent via 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 Feb 2015 05:13:00 +0000","msg_from":"Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Thu, Feb 19, 2015 at 2:13 PM, Naoya Anzai\n<anzai-naoya@mxu.nes.nec.co.jp> wrote:\n> As a result, I think you should not delete VACOPT_VERBOSE.\n\nIn v8 it is not deleted. It is still declared, and its use is isolated\nin gram.y, similarly to VACOPT_FREEZE.\n\n> According to the last mail I have posted, the difference of\n> manual-vacuum log and auto-vacuum log exists clearly.\n\nDid you test the latest patch v8? I have added checks in it to see if\na process is an autovacuum worker to control elevel and the extra logs\nof v7 do not show up.\n\n> So, at least you should not touch the mechanism of VACOPT_VERBOSE\n> until both vacuum log formats are unified to a same format.\n\nIf you mean that we should have the same kind of log outputs for\nautovacuum and manual vacuum, I think that this is not going to\nhappen. Autovacuum entries are kept less verbose on purpose, contract\nthat v7 clealy broke.\n\n> If you agree my think, please undo your removing VACOPT_VERBOSE work.\n\nWell, I don't agree :) And I am guessing that you did not look at v8\nas well. Centralizing the control of logs using log_min_duration is\nmore extensible than simply having VACOPT_VERBOSE.\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, 19 Feb 2015 15:32:17 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Thu, Feb 19, 2015 at 3:32 PM, Michael Paquier\n<michael.paquier@gmail.com> wrote:\n> On Thu, Feb 19, 2015 at 2:13 PM, Naoya Anzai\n> <anzai-naoya@mxu.nes.nec.co.jp> wrote:\n>> As a result, I think you should not delete VACOPT_VERBOSE.\n>\n> In v8 it is not deleted. It is still declared, and its use is isolated\n> in gram.y, similarly to VACOPT_FREEZE.\n>\n>> According to the last mail I have posted, the difference of\n>> manual-vacuum log and auto-vacuum log exists clearly.\n>\n> Did you test the latest patch v8? I have added checks in it to see if\n> a process is an autovacuum worker to control elevel and the extra logs\n> of v7 do not show up.\n>\n>> So, at least you should not touch the mechanism of VACOPT_VERBOSE\n>> until both vacuum log formats are unified to a same format.\n>\n> If you mean that we should have the same kind of log outputs for\n> autovacuum and manual vacuum, I think that this is not going to\n> happen. Autovacuum entries are kept less verbose on purpose, contract\n> that v7 clealy broke.\n>\n>> If you agree my think, please undo your removing VACOPT_VERBOSE work.\n>\n> Well, I don't agree :) And I am guessing that you did not look at v8\n> as well. Centralizing the control of logs using log_min_duration is\n> more extensible than simply having VACOPT_VERBOSE.\n\nWith the patch, VACUUM ANALYZE VERBOSE doesn't emit any verbose message.\nWhy did you remove that functionality?\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":"Thu, 5 Mar 2015 19:10:08 +0900","msg_from":"Fujii Masao <masao.fujii@gmail.com>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Thu, Mar 5, 2015 at 7:10 PM, Fujii Masao wrote:\n> With the patch, VACUUM ANALYZE VERBOSE doesn't emit any verbose message.\n> Why did you remove that functionality?\n\nOops. Sorry about that. In gram.y, the combination of VacuumStmt with\nAnalyzeStmt overwrote the value of log_min_duration incorrectly. I\nalso found another bug related to logging of ANALYZE not working\ncorrectly because of the use of IsAutoVacuumWorkerProcess() instead of\nVACOPT_VERBOSE (this is reducing the diffs of the patch btw). All\nthose things are fixed in the attached.\n-- \nMichael\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Thu, 5 Mar 2015 21:49:03 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Thu, Mar 5, 2015 at 9:49 PM, Michael Paquier\n<michael.paquier@gmail.com> wrote:\n> On Thu, Mar 5, 2015 at 7:10 PM, Fujii Masao wrote:\n>> With the patch, VACUUM ANALYZE VERBOSE doesn't emit any verbose message.\n>> Why did you remove that functionality?\n>\n> Oops. Sorry about that. In gram.y, the combination of VacuumStmt with\n> AnalyzeStmt overwrote the value of log_min_duration incorrectly. I\n> also found another bug related to logging of ANALYZE not working\n> correctly because of the use of IsAutoVacuumWorkerProcess() instead of\n> VACOPT_VERBOSE (this is reducing the diffs of the patch btw). All\n> those things are fixed in the attached.\n\nThanks for updating the patch!\n\nWhy does log_min_duration need to be set even when manual VACUUM command is\nexecuted? Per the latest version of the patch, log_min_duration is checked only\nwhen the process is autovacuum worker. So ISTM that log_min_duration doesn't\nneed to be set in gram.y. It's even confusing to me. Or if you're going to\nimplement something like \"VACUUM VERBOSE DURATION n\" (i.e., verbose message\nis output if n seconds have been elapsed), that might be necessary, though...\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":"Fri, 6 Mar 2015 12:44:11 +0900","msg_from":"Fujii Masao <masao.fujii@gmail.com>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Fri, Mar 6, 2015 at 12:44 PM, Fujii Masao <masao.fujii@gmail.com> wrote:\n> On Thu, Mar 5, 2015 at 9:49 PM, Michael Paquier\n> <michael.paquier@gmail.com> wrote:\n>> On Thu, Mar 5, 2015 at 7:10 PM, Fujii Masao wrote:\n>>> With the patch, VACUUM ANALYZE VERBOSE doesn't emit any verbose message.\n>>> Why did you remove that functionality?\n>>\n>> Oops. Sorry about that. In gram.y, the combination of VacuumStmt with\n>> AnalyzeStmt overwrote the value of log_min_duration incorrectly. I\n>> also found another bug related to logging of ANALYZE not working\n>> correctly because of the use of IsAutoVacuumWorkerProcess() instead of\n>> VACOPT_VERBOSE (this is reducing the diffs of the patch btw). All\n>> those things are fixed in the attached.\n>\n> Thanks for updating the patch!\n>\n> Why does log_min_duration need to be set even when manual VACUUM command is\n> executed? Per the latest version of the patch, log_min_duration is checked only\n> when the process is autovacuum worker. So ISTM that log_min_duration doesn't\n> need to be set in gram.y. It's even confusing to me. Or if you're going to\n> implement something like \"VACUUM VERBOSE DURATION n\" (i.e., verbose message\n> is output if n seconds have been elapsed), that might be necessary, though...\n\nThanks for reminding. The DURATION-like clause was exactly a point\nmentioned by Anzai-san upthread, and it made sense to me to be in-line\nwith the other parameters controlling the freeze (discussion somewhat\nrelated to that =>\nhttp://www.postgresql.org/message-id/CAB7nPqRZX7Pv2B-R7xHmAh52tfjAQGfy9btqwFstgQgXks=iSw@mail.gmail.com)\nbut we can live without it for this patch as VACOPT_VERBOSE is used\nonly by manual VACUUM and not by autovacuum to choose the log elevel.\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":"Fri, 6 Mar 2015 13:07:04 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Fri, Mar 6, 2015 at 1:07 PM, Michael Paquier\n<michael.paquier@gmail.com> wrote:\n> On Fri, Mar 6, 2015 at 12:44 PM, Fujii Masao <masao.fujii@gmail.com> wrote:\n>> On Thu, Mar 5, 2015 at 9:49 PM, Michael Paquier\n>> <michael.paquier@gmail.com> wrote:\n>>> On Thu, Mar 5, 2015 at 7:10 PM, Fujii Masao wrote:\n>>>> With the patch, VACUUM ANALYZE VERBOSE doesn't emit any verbose message.\n>>>> Why did you remove that functionality?\n>>>\n>>> Oops. Sorry about that. In gram.y, the combination of VacuumStmt with\n>>> AnalyzeStmt overwrote the value of log_min_duration incorrectly. I\n>>> also found another bug related to logging of ANALYZE not working\n>>> correctly because of the use of IsAutoVacuumWorkerProcess() instead of\n>>> VACOPT_VERBOSE (this is reducing the diffs of the patch btw). All\n>>> those things are fixed in the attached.\n>>\n>> Thanks for updating the patch!\n>>\n>> Why does log_min_duration need to be set even when manual VACUUM command is\n>> executed? Per the latest version of the patch, log_min_duration is checked only\n>> when the process is autovacuum worker. So ISTM that log_min_duration doesn't\n>> need to be set in gram.y. It's even confusing to me. Or if you're going to\n>> implement something like \"VACUUM VERBOSE DURATION n\" (i.e., verbose message\n>> is output if n seconds have been elapsed), that might be necessary, though...\n>\n> Thanks for reminding. The DURATION-like clause was exactly a point\n> mentioned by Anzai-san upthread, and it made sense to me to be in-line\n> with the other parameters controlling the freeze (discussion somewhat\n> related to that =>\n> http://www.postgresql.org/message-id/CAB7nPqRZX7Pv2B-R7xHmAh52tfjAQGfy9btqwFstgQgXks=iSw@mail.gmail.com)\n> but we can live without it for this patch as VACOPT_VERBOSE is used\n> only by manual VACUUM and not by autovacuum to choose the log elevel.\n\nAre you planning to update the patch so that it's based on the commit 0d83138?\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":"Thu, 19 Mar 2015 12:23:01 +0900","msg_from":"Fujii Masao <masao.fujii@gmail.com>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Thu, Mar 19, 2015 at 12:23 PM, Fujii Masao <masao.fujii@gmail.com> wrote:\n> Are you planning to update the patch so that it's based on the commit 0d83138?\n\nYes... Very soon.\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, 19 Mar 2015 12:40:40 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Thu, Mar 19, 2015 at 12:40 PM, Michael Paquier\n<michael.paquier@gmail.com> wrote:\n> On Thu, Mar 19, 2015 at 12:23 PM, Fujii Masao <masao.fujii@gmail.com> wrote:\n>> Are you planning to update the patch so that it's based on the commit 0d83138?\n>\n> Yes... Very soon.\n\nAnd here is the rebased patch.\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":"Thu, 19 Mar 2015 13:43:52 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Thu, Mar 19, 2015 at 1:43 PM, Michael Paquier\n<michael.paquier@gmail.com> wrote:\n> On Thu, Mar 19, 2015 at 12:40 PM, Michael Paquier\n> <michael.paquier@gmail.com> wrote:\n>> On Thu, Mar 19, 2015 at 12:23 PM, Fujii Masao <masao.fujii@gmail.com> wrote:\n>>> Are you planning to update the patch so that it's based on the commit 0d83138?\n>>\n>> Yes... Very soon.\n>\n> And here is the rebased patch.\n\nThanks for rebasing the patch! Looks good to me.\n\nOne concern about this patch is; currently log_autovacuum_min_duration can be\nchanged even while autovacuum worker is running. So, for example, when\nthe admin notices that autovacuum is taking very long time, he or she can\nenable logging of autovacuum activity on the fly. But this patch completely\nprevents us from doing that, because, with the patch, autovacuum worker always\npicks up the latest setting value at its starting time and then keeps using it\nto the end. Probably I can live with this. But does anyone has other thought?\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":"Mon, 23 Mar 2015 13:54:03 +0900","msg_from":"Fujii Masao <masao.fujii@gmail.com>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Mon, Mar 23, 2015 at 1:54 PM, Fujii Masao <masao.fujii@gmail.com> wrote:\n> On Thu, Mar 19, 2015 at 1:43 PM, Michael Paquier\n> <michael.paquier@gmail.com> wrote:\n>> On Thu, Mar 19, 2015 at 12:40 PM, Michael Paquier\n>> <michael.paquier@gmail.com> wrote:\n>>> On Thu, Mar 19, 2015 at 12:23 PM, Fujii Masao <masao.fujii@gmail.com> wrote:\n>>>> Are you planning to update the patch so that it's based on the commit 0d83138?\n>>>\n>>> Yes... Very soon.\n>>\n>> And here is the rebased patch.\n>\n> Thanks for rebasing the patch! Looks good to me.\n>\n> One concern about this patch is; currently log_autovacuum_min_duration can be\n> changed even while autovacuum worker is running. So, for example, when\n> the admin notices that autovacuum is taking very long time, he or she can\n> enable logging of autovacuum activity on the fly. But this patch completely\n> prevents us from doing that, because, with the patch, autovacuum worker always\n> picks up the latest setting value at its starting time and then keeps using it\n> to the end. Probably I can live with this. But does anyone has other thought?\n\nIn AutoVacWorkerMain, I am reading the following:\n\n * Currently, we don't pay attention to postgresql.conf changes that\n * happen during a single daemon iteration, so we can ignore SIGHUP.\n */\n pqsignal(SIGHUP, SIG_IGN);\n\nSo a worker does not see changes in postgresql.conf once it is run and\nprocesses a database, no? The launcher does run ProcessConfigFile()\nwhen SIGHUP shows up though.\nRegards,\n-- \nMichael\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 23 Mar 2015 14:23:36 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"Michael Paquier wrote:\n\n> In AutoVacWorkerMain, I am reading the following:\n> \n> * Currently, we don't pay attention to postgresql.conf changes that\n> * happen during a single daemon iteration, so we can ignore SIGHUP.\n> */\n> pqsignal(SIGHUP, SIG_IGN);\n> \n> So a worker does not see changes in postgresql.conf once it is run and\n> processes a database, no? The launcher does run ProcessConfigFile()\n> when SIGHUP shows up though.\n\nMaybe this is something that we should change. For example, I wonder if\nthis can also affect the cost-delay balancing heuristics; if two\nbackends run the rebalance with different GUC settings because\npostgresql.conf changed in between each of them starting, would the\nsettings bounce back and forth. I think it's worth reconsidering this.\n(Don't really remember in detail how it works; maybe it's fine now.)\n\nIn any case, for log_autovacuum_min_duration it also seems worth keeping\nreasonably close track of GUC changes. I think reading them just before\nstarting vacuum of a new relation should be enough.\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":"Mon, 23 Mar 2015 10:25:20 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"Alvaro Herrera <alvherre@2ndquadrant.com> writes:\n> Michael Paquier wrote:\n>> So a worker does not see changes in postgresql.conf once it is run and\n>> processes a database, no? The launcher does run ProcessConfigFile()\n>> when SIGHUP shows up though.\n\n> Maybe this is something that we should change.\n\nYeah, checking for SIGHUP in the worker outer loop (ie once per table)\nseems like a reasonable thing.\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, 23 Mar 2015 10:07:12 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Mon, Mar 23, 2015 at 7:07 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n\n> Alvaro Herrera <alvherre@2ndquadrant.com> writes:\n> > Michael Paquier wrote:\n> >> So a worker does not see changes in postgresql.conf once it is run and\n> >> processes a database, no? The launcher does run ProcessConfigFile()\n> >> when SIGHUP shows up though.\n>\n> > Maybe this is something that we should change.\n>\n> Yeah, checking for SIGHUP in the worker outer loop (ie once per table)\n> seems like a reasonable thing.\n>\n> regards, tom lane\n>\n\nCould it be done more often? Maybe every time it is about to do a\ncost_delay sleep?\n\nI've certainly regretted the inability to\nchange autovacuum_vacuum_cost_delay mid-table on more than one occasion.\n\nThis was mostly during doing testing work, but still I'm sure other people\nhave run into this problem, perhaps without knowing it.\n\nCheers,\n\nJeff\n\nOn Mon, Mar 23, 2015 at 7:07 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:Alvaro Herrera <alvherre@2ndquadrant.com> writes:\n> Michael Paquier wrote:\n>> So a worker does not see changes in postgresql.conf once it is run and\n>> processes a database, no? The launcher does run ProcessConfigFile()\n>> when SIGHUP shows up though.\n\n> Maybe this is something that we should change.\n\nYeah, checking for SIGHUP in the worker outer loop (ie once per table)\nseems like a reasonable thing.\n\n regards, tom laneCould it be done more often? Maybe every time it is about to do a cost_delay sleep?I've certainly regretted the inability to change autovacuum_vacuum_cost_delay mid-table on more than one occasion.This was mostly during doing testing work, but still I'm sure other people have run into this problem, perhaps without knowing it.Cheers,Jeff","msg_date":"Mon, 23 Mar 2015 08:23:00 -0700","msg_from":"Jeff Janes <jeff.janes@gmail.com>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"Jeff Janes <jeff.janes@gmail.com> writes:\n> On Mon, Mar 23, 2015 at 7:07 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> Yeah, checking for SIGHUP in the worker outer loop (ie once per table)\n>> seems like a reasonable thing.\n\n> Could it be done more often? Maybe every time it is about to do a\n> cost_delay sleep?\n\nThat sounds risky from here. Normal backends don't check it more often\nthan once per command.\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, 23 Mar 2015 11:34:21 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Mon, Mar 23, 2015 at 11:07 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Alvaro Herrera <alvherre@2ndquadrant.com> writes:\n>> Michael Paquier wrote:\n>>> So a worker does not see changes in postgresql.conf once it is run and\n>>> processes a database, no? The launcher does run ProcessConfigFile()\n>>> when SIGHUP shows up though.\n>\n>> Maybe this is something that we should change.\n>\n> Yeah, checking for SIGHUP in the worker outer loop (ie once per table)\n> seems like a reasonable thing.\n\nThat sounds fine to me as well. A patch would not be complicated, but\nis this portion really 9.5 material?\n\nAlso, this is a discussion wider than only\nlog_autovacuum_min_duration, as autovacuum cost parameters are also\navailable as reloptions.\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 Mar 2015 10:45:15 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"Michael Paquier wrote:\n> On Mon, Mar 23, 2015 at 11:07 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> > Alvaro Herrera <alvherre@2ndquadrant.com> writes:\n> >> Michael Paquier wrote:\n> >>> So a worker does not see changes in postgresql.conf once it is run and\n> >>> processes a database, no? The launcher does run ProcessConfigFile()\n> >>> when SIGHUP shows up though.\n> >\n> >> Maybe this is something that we should change.\n> >\n> > Yeah, checking for SIGHUP in the worker outer loop (ie once per table)\n> > seems like a reasonable thing.\n> \n> That sounds fine to me as well. A patch would not be complicated, but\n> is this portion really 9.5 material?\n\nIMO yes.\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":"Mon, 23 Mar 2015 23:46:21 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Mon, Mar 23, 2015 at 7:46 PM, Alvaro Herrera wrote:\n> Michael Paquier wrote:\n>> On Mon, Mar 23, 2015 at 11:07 PM, Tom Lane wrote:\n>> > Alvaro Herrera <alvherre@2ndquadrant.com> writes:\n>> >> Michael Paquier wrote:\n>> >>> So a worker does not see changes in postgresql.conf once it is run and\n>> >>> processes a database, no? The launcher does run ProcessConfigFile()\n>> >>> when SIGHUP shows up though.\n>> >\n>> >> Maybe this is something that we should change.\n>> >\n>> > Yeah, checking for SIGHUP in the worker outer loop (ie once per table)\n>> > seems like a reasonable thing.\n>>\n>> That sounds fine to me as well. A patch would not be complicated, but\n>> is this portion really 9.5 material?\n>\n> IMO yes.\n\nSo, attached are two patches:\n- 0001 enables SIGHUP tracking in do_autovacuum(), which is checked\nbefore processing one table. I reused avl_sighup_handler for the\nworker, renaming it av_sighup_handler..\n- 0002 is the patch to add log_autovacuum_min_duration as a reloption.\nThere is nothing really changed, the value of\nlog_autovacuum_min_duration being picked up at startup with\ntable_recheck_autovac() is used until the end for one table.\nRegards,\n-- \nMichael\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 Mar 2015 15:54:16 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"Michael Paquier wrote:\n\n> So, attached are two patches:\n> - 0001 enables SIGHUP tracking in do_autovacuum(), which is checked\n> before processing one table. I reused avl_sighup_handler for the\n> worker, renaming it av_sighup_handler..\n> - 0002 is the patch to add log_autovacuum_min_duration as a reloption.\n> There is nothing really changed, the value of\n> log_autovacuum_min_duration being picked up at startup with\n> table_recheck_autovac() is used until the end for one table.\n\nThanks.\n\nYou added this in the worker loop processing each table:\n\n\t\t/*\n\t\t * Check for config changes before processing each collected table.\n\t\t */\n\t\tif (got_SIGHUP)\n\t\t{\n\t\t\tgot_SIGHUP = false;\n\t\t\tProcessConfigFile(PGC_SIGHUP);\n\n\t\t\t/* shutdown requested in config file? */\n\t\t\tif (!AutoVacuumingActive())\n\t\t\t\tbreak;\n\t\t}\n\nI think bailing out if autovac is disabled is a good idea; for instance,\nif this is a for-wraparound vacuum, we should keep running. My first\nreaction is that this part of the test ought to be moved down a\nscreenful or so, until we've ran the recheck step and we have the\nfor-wraparound flag in hand; only bail out if that one is not set. But\non the other hand, maybe the worker will process some tables that are\nnot for-wraparound in between some other tables that are, so that might\nturn out to be a bad idea as well. (Though I'm not 100% that it works\nthat way; consider commit f51ead09df for instance.) Maybe the test to\nuse for this is something along the lines of \"if autovacuum was enabled\nbefore and is no longer enabled, bail out, otherwise keep running\".\nThis implies that we need to keep track of whether autovac was enabled\nat the start of the worker run.\n\n\nAnother thing, but not directly related to this patch: while looking at\nthe doc changes, I noticed that we don't have index entries for the\nreloptions we have; for instance, the word \"fillfactor\" doesn't appear\nin the bookindex.html page at all. Having these things all crammed in\nthe CREATE TABLE page seems somewhat poor. I think we could improve on\nthis by having a listing of settable parameters in a separate section,\nand have CREATE TABLE, ALTER TABLE, and the Server Configuration chapter\nlink to that; we could also have index entries for these items.\n\nOf course, the simpler fix is to just add a few <indexterm> tags.\n\n\nAs a note, there is no point to \"Assert(foo != NULL)\" tests when foo is\nlater dereferenced in the same routine: the crash is going to happen\nanyway at the dereference, so it's just a LOC uselessly wasted.\n\n-- \n�lvaro Herrera http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 2 Apr 2015 17:57:14 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"\n--- a/doc/src/sgml/ref/create_table.sgml\n+++ b/doc/src/sgml/ref/create_table.sgml\n@@ -881,9 +881,11 @@ CREATE [ [ GLOBAL | LOCAL ] { TEMPORARY | TEMP } | UNLOGGED ] TABLE [ IF NOT EXI\n <literal>toast.</literal>, which can be used to control the behavior of the\n table's secondary <acronym>TOAST</> table, if any\n (see <xref linkend=\"storage-toast\"> for more information about TOAST).\n- Note that the TOAST table inherits the\n- <literal>autovacuum_*</literal> values from its parent table, if there are\n- no <literal>toast.autovacuum_*</literal> settings set.\n+ Note that the TOAST table inherits values of <literal>autovacuum_*</literal>\n+ and <literal>log_autovacuum_min_duration</literal> from its parent table, if\n+ there are no values set for respectively\n+ <literal>toast.autovacuum_*</literal> and\n+ <literal>toast.log_autovacuum_min_duration</literal>.\n </para>\n \n\nI think this could use some wordsmithing; I didn't like listing the\nparameters explicitely, and Jaime Casanova suggested not using the terms\n\"inherit\" and \"parent table\" in this context. So I came up with this:\n\n Note that the TOAST table uses the parameter values defined for the main\n table, for each parameter applicable to TOAST tables and for which no\n value is set in the TOAST table itself.\n\nThoughts?\n\n-- \n�lvaro Herrera http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 2 Apr 2015 23:04:55 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Fri, Apr 3, 2015 at 5:57 AM, Alvaro Herrera wrote:\n> You added this in the worker loop processing each table:\n>\n> /*\n> * Check for config changes before processing each collected table.\n> */\n> if (got_SIGHUP)\n> {\n> got_SIGHUP = false;\n> ProcessConfigFile(PGC_SIGHUP);\n>\n> /* shutdown requested in config file? */\n> if (!AutoVacuumingActive())\n> break;\n> }\n>\n> I think bailing out if autovac is disabled is a good idea; for instance,\n> if this is a for-wraparound vacuum, we should keep running. My first\n> reaction is that this part of the test ought to be moved down a\n> screenful or so, until we've ran the recheck step and we have the\n> for-wraparound flag in hand; only bail out if that one is not set. But\n> on the other hand, maybe the worker will process some tables that are\n> not for-wraparound in between some other tables that are, so that might\n> turn out to be a bad idea as well. (Though I'm not 100% that it works\n> that way; consider commit f51ead09df for instance.) Maybe the test to\n> use for this is something along the lines of \"if autovacuum was enabled\n> before and is no longer enabled, bail out, otherwise keep running\".\n> This implies that we need to keep track of whether autovac was enabled\n> at the start of the worker run.\n\nI did consider the case of wraparound but came to the conclusion that\nbailing out as fast as possible was the idea. But well, I guess that\nwe could play it safe and not add this check after all. The main\nuse-case of this change is now to be able to do re-balancing of the\ncost parameters. We could still decide later if a worker could bail\nout earlier or not, depending on what.\n\n> Another thing, but not directly related to this patch: while looking at\n> the doc changes, I noticed that we don't have index entries for the\n> reloptions we have; for instance, the word \"fillfactor\" doesn't appear\n> in the bookindex.html page at all. Having these things all crammed in\n> the CREATE TABLE page seems somewhat poor. I think we could improve on\n> this by having a listing of settable parameters in a separate section,\n> and have CREATE TABLE, ALTER TABLE, and the Server Configuration chapter\n> link to that; we could also have index entries for these items.\n> Of course, the simpler fix is to just add a few <indexterm> tags.\n\nThis sounds like a good idea, and this refactoring would meritate a\ndifferent patch and a different thread. I guess that it should be a\nnew section in Server Configuration then, named \"Relation Options\",\nwith two different subsections for index options and table options.\n\n> As a note, there is no point to \"Assert(foo != NULL)\" tests when foo is\n> later dereferenced in the same routine: the crash is going to happen\n> anyway at the dereference, so it's just a LOC uselessly wasted.\n\nCheck. I still think that it is useful in this case though... But removed.\n\n> I think this could use some wordsmithing; I didn't like listing the\n> parameters explicitely, and Jaime Casanova suggested not using the terms\n> \"inherit\" and \"parent table\" in this context. So I came up with this:\n> Note that the TOAST table uses the parameter values defined for the main\n> table, for each parameter applicable to TOAST tables and for which no\n> value is set in the TOAST table itself.\n\nFine for me.\n-- \nMichael\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, 3 Apr 2015 15:26:56 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Fri, Apr 3, 2015 at 3:26 PM, Michael Paquier wrote:\n> [...]\n> Fine for me.\n\nAnd here are the correct patches. Sorry for that.\n-- \nMichael\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, 3 Apr 2015 15:28:33 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"Michael Paquier wrote:\n> On Fri, Apr 3, 2015 at 3:26 PM, Michael Paquier wrote:\n> > [...]\n> > Fine for me.\n> \n> And here are the correct patches. Sorry for that.\n\nThanks, pushed. I added one extra comment to the SIGHUP patch in the\nplace where you previously had the exit.\n\n-- \n�lvaro Herrera http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 3 Apr 2015 11:59:22 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Table-level log_autovacuum_min_duration"},{"msg_contents":"On Fri, Apr 3, 2015 at 11:59 PM, Alvaro Herrera\n<alvherre@2ndquadrant.com> wrote:\n> Michael Paquier wrote:\n>> On Fri, Apr 3, 2015 at 3:26 PM, Michael Paquier wrote:\n>> > [...]\n>> > Fine for me.\n>>\n>> And here are the correct patches. Sorry for that.\n>\n> Thanks, pushed. I added one extra comment to the SIGHUP patch in the\n> place where you previously had the exit.\n\nThanks!\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, 4 Apr 2015 06:35:39 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Re: Table-level log_autovacuum_min_duration"}],"string":"[\n {\n \"msg_contents\": \"Hi all,\\n\\nToday I spent a bit of time looking at the activity of autovacuum for\\none table particularly bloated. log_autovacuum_min_duration was\\nenabled and set to a high value but even with that I had to deal with\\nsome spam from the jobs of other tables. It would be cool to have the\\npossibility to do some filtering IMO. So, what about having a relopt\\nto control log_autovacuum_min_duration? RelationData->rd_options is\\nlargely accessible for a given relation in the code paths of vacuum\\nand analyze.\\nThoughts?\\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, 18 Dec 2014 22:23:30 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"Michael Paquier wrote:\\n> Hi all,\\n> \\n> Today I spent a bit of time looking at the activity of autovacuum for\\n> one table particularly bloated. log_autovacuum_min_duration was\\n> enabled and set to a high value but even with that I had to deal with\\n> some spam from the jobs of other tables. It would be cool to have the\\n> possibility to do some filtering IMO. So, what about having a relopt\\n> to control log_autovacuum_min_duration? RelationData->rd_options is\\n> largely accessible for a given relation in the code paths of vacuum\\n> and analyze.\\n\\n+1\\n\\n-- \\n�lvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 18 Dec 2014 11:15:40 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 18, 2014 at 11:15 PM, Alvaro Herrera\\n<alvherre@2ndquadrant.com> wrote:\\n> Michael Paquier wrote:\\n>> Hi all,\\n>>\\n>> Today I spent a bit of time looking at the activity of autovacuum for\\n>> one table particularly bloated. log_autovacuum_min_duration was\\n>> enabled and set to a high value but even with that I had to deal with\\n>> some spam from the jobs of other tables. It would be cool to have the\\n>> possibility to do some filtering IMO. So, what about having a relopt\\n>> to control log_autovacuum_min_duration? RelationData->rd_options is\\n>> largely accessible for a given relation in the code paths of vacuum\\n>> and analyze.\\n>\\n> +1\\n\\nAs long as I got this idea in mind, I looked at the code to see how\\nmuch it would be possible to tune log_autovacuum_min_duration in the\\ncode paths of analyze and vacuum. First, I think that it would be good\\nto have parameters for both parent relations and their toast relation\\nsimilarly to the other parameters...\\n\\nBut now, here are some things to consider if we use directly the\\nreloptions available with RelationData:\\n- If the parameters toast.autovacuum_* are not set, toast relations\\ninherit values from their parent relation. Looking at autovacuum.c\\nwhich is the only place where autovacuum options are located, we keep\\na hash table to save the mapping toast -> parent relation. Using that\\nit is easy to fetch for a given toast relation the relopts of its\\nparent. Note this is strictly located in the autovacuum code path, so\\nto let vacuum and analyze now the custom value of\\nlog_autovacuum_min_duration, with the inheritance property kept, we\\nwould need to pass some extra values from autovacuum to the calls of\\nvacuum().\\n- It would not be possible to log autovacuum and analyze being skipped\\nwhen lock is not available because in this case Relation cannot be\\nsafely fetched, so there are no reltoptions directly available. This\\nis for example this kind of message:\\nskipping analyze of \\\"foo\\\" --- lock not available\\n\\nBoth those things could be solved by passing a value through\\nVacuumStmt, like what we do when passing a value for\\nmultixact_freeze_min_age, or with an extra argument in vacuum() for\\nexample. Now I am not sure if it is worth it, and we could live as\\nwell with a parameter that do not support the inheritance parent\\nrelation -> toast, so log value set for a toast relation and its\\nparent share no dependency. In short that's a trade between code\\nsimplicity and usability.\\n\\nThoughts?\\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 Dec 2014 23:17:21 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Sat, Dec 20, 2014 at 9:17 AM, Michael Paquier\\n<michael.paquier@gmail.com> wrote:\\n> But now, here are some things to consider if we use directly the\\n> reloptions available with RelationData:\\n> - If the parameters toast.autovacuum_* are not set, toast relations\\n> inherit values from their parent relation. Looking at autovacuum.c\\n> which is the only place where autovacuum options are located, we keep\\n> a hash table to save the mapping toast -> parent relation. Using that\\n> it is easy to fetch for a given toast relation the relopts of its\\n> parent. Note this is strictly located in the autovacuum code path, so\\n> to let vacuum and analyze now the custom value of\\n> log_autovacuum_min_duration, with the inheritance property kept, we\\n> would need to pass some extra values from autovacuum to the calls of\\n> vacuum().\\n> - It would not be possible to log autovacuum and analyze being skipped\\n> when lock is not available because in this case Relation cannot be\\n> safely fetched, so there are no reltoptions directly available. This\\n> is for example this kind of message:\\n> skipping analyze of \\\"foo\\\" --- lock not available\\n>\\n> Both those things could be solved by passing a value through\\n> VacuumStmt, like what we do when passing a value for\\n> multixact_freeze_min_age, or with an extra argument in vacuum() for\\n> example. Now I am not sure if it is worth it, and we could live as\\n> well with a parameter that do not support the inheritance parent\\n> relation -> toast, so log value set for a toast relation and its\\n> parent share no dependency. In short that's a trade between code\\n> simplicity and usability.\\n\\nI'm not sure I follow all of the particulars of what you are asking\\nhere, but in general I would say that you shouldn't hesitate to pass\\nmore information down the call stack when that helps to obtain correct\\nbehavior.\\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 Dec 2014 08:31:14 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Sat, Dec 20, 2014 at 11:17 PM, Michael Paquier\\n<michael.paquier@gmail.com> wrote:\\n>> Michael Paquier wrote:\\n>>> Today I spent a bit of time looking at the activity of autovacuum for\\n>>> one table particularly bloated. log_autovacuum_min_duration was\\n>>> enabled and set to a high value but even with that I had to deal with\\n>>> some spam from the jobs of other tables. It would be cool to have the\\n>>> possibility to do some filtering IMO. So, what about having a relopt\\n>>> to control log_autovacuum_min_duration? RelationData->rd_options is\\n>>> largely accessible for a given relation in the code paths of vacuum\\n>>> and analyze.\\n\\nOK, instead of only words, attached is a patch adding relopts called\\nlog_autovacuum_min_duration and toast.log_autovacuum_min_duration to\\ncontrol the logging of autovacuum at relation-level. The default value\\nof those parameters is -1, meaning that in this case the global\\nlog_autovacuum_min_duration is used to control the logs. The patch has\\nfinished by being simpler than I though first by using VacuumStmt to\\npass the relopts from autovacuum to the code paths of analyze and\\nvacuum. Note that this uses the same mechanisms as the other relopts,\\nmeaning that the toast relations use the values of their parent tables\\nif it is defined.\\n\\nI am adding it to the next CF.\\n-- \\nMichael\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Tue, 13 Jan 2015 14:26:11 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> I am adding it to the next CF.\\nPatch 2 attached. I forgot the update of copyfuncs.c and equalfuncs.c.\\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\": \"Thu, 15 Jan 2015 14:53:31 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"The following review has been posted through the commitfest application:\\nmake installcheck-world: tested, failed\\nImplements feature: tested, failed\\nSpec compliant: tested, failed\\nDocumentation: tested, failed\\n\\nHi,\\n\\nI'm Naoya Anzai.\\nI've been working as a PostgreSQL Support Engineer for 6 years.\\nI am a newcomer of reviewing, and My programming skill is not so high.\\nBut the following members also participate in this review. (We say \\n\\\"Two heads are better than one.\\\" :))\\n\\nAkira Kurosawa <kurosawa-akira@mxc.nes.nec.co.jp>\\nTaiki Kondo <kondo-taiki@mxt.nes.nec.co.jp>\\nHuong Dangminh <dangminh-huong@mxm.nes.nec.co.jp>\\n\\nSo I believe reviewing patches is not difficult for us.\\n\\nThis is a review of Table-level log_autovacuum_min_duration patch:\\nhttp://www.postgresql.org/message-id/CAB7nPqTBQsbEgvb8cOH01K7ARGYs9KBuV8dr+aqGonFVb8koAQ@mail.gmail.com\\n\\nSubmission review\\n====================\\nThe patch applies cleanly to HEAD, and it works fine on Fedora \\nrelease 20.\\nThere is no regression test,but I think it is no problem \\nbecause other parameter also is not tested.\\n\\n\\nUsability review\\n====================\\npg_dump/pg_restore support is OK.\\nI think this feature is a nice idea and I also want this feature.\\n\\n\\nFeature test\\n====================\\nI executed following commands after setting \\n\\\"log_autovacuum_min_duration\\\" to 1000 in the GUC. (\\\"bar\\\" table is \\nalready created with no options.)\\n\\n CREATE TABLE foo ( ... ) WITH ( log_autovacuum_min_duration = 0 );\\n ALTER TABLE bar SET (log_autovacuum_min_duration = 0 );\\n\\nThen, only in \\\"foo\\\" and \\\"bar\\\" table, autovacuum log was printed out \\neven if elapsed time of autovacuum is lesser than 1000ms. This \\nbehavior was expected and there was no crash or failed assertion. \\nSo it looked good for me. But, I executed following command, in \\n\\\"buzz\\\" table, autovacuum log was printed out if elapsed time is \\nmore than 1000ms.\\n\\n CREATE TABLE buzz ( ... ) WITH ( log_autovacuum_min_duration = -1 );\\n ^^\\n\\nI expect autovacuum log is NOT printed out even if elapsed time is \\nmore than 1000ms in this case. My thought is wrong, isn't it? In my \\nopinion, there is an use case that autovacuum log is always printed \\nout in all tables excepting specified tables. I think there is a \\nperson who wants to use it like this case, but I (or he) can NOT use \\nin this situation.\\n\\nHow about your opinion?\\n\\n\\nPerformance review\\n====================\\nNot reviewed from this point of view.\\n\\n\\nCoding review\\n====================\\nI think description of log_autovacuum_min_duration in reloptions.c\\n(line:215) should be like other parameters (match with guc.c). So \\nit should be \\\"Sets the minimum execution time above which autovacuum \\nactions will be logged.\\\" but not \\\"Log autovacuum execution for \\ngiven threshold\\\".\\n\\nThere is no new function which is defined in this patch, and \\nmodified contents are not related to OS-dependent contents, so I \\nthink it will work fine on Windows/BSD etc.\\n\\n\\nArchitecture review\\n====================\\nAbout the modification of gram.y.\\n\\nI think it is not good that log_min_duration is initialized to \\nzeros in gram.y when \\\"FREEZE\\\" option is set. Because any VACUUM \\nqueries never use this parameter. I think log_min_duration always \\nshould be initialized to -1.\\n\\n\\nRegards,\\nNaoya Anzai (NEC Solution Innovators,Ltd.)\\n\\n\\nThe new status of this patch is: Waiting on Author\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 06 Feb 2015 00:39:56 +0000\",\n \"msg_from\": \"Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"> The following review has been posted through the commitfest application:\\r\\n> make installcheck-world: tested, failed\\r\\n> Implements feature: tested, failed\\r\\n> Spec compliant: tested, failed\\r\\n> Documentation: tested, failed\\r\\nI'm sorry, I just sent it by mistake.\\r\\nAll of them have passed.\\r\\n\\r\\n---\\r\\nNaoya Anzai\\r\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 6 Feb 2015 00:50:08 +0000\",\n \"msg_from\": \"Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 6, 2015 at 4:50 AM, Naoya Anzai\\n<anzai-naoya@mxu.nes.nec.co.jp> wrote:\\n>> The following review has been posted through the commitfest application:\\n>> make installcheck-world: tested, failed\\n>> Implements feature: tested, failed\\n>> Spec compliant: tested, failed\\n>> Documentation: tested, failed\\n> I'm sorry, I just sent it by mistake.\\n> All of them have passed.\\nThat's fine. I think you used the UI on the commit fest app, and it is\\nnot intuitive that you need to check those boxes at first sight when\\nusing it for the first time.\\n-- \\nMichael\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 6 Feb 2015 07:06:12 +0400\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 6, 2015 at 4:39 AM, Naoya Anzai wrote:\\n> I've been working as a PostgreSQL Support Engineer for 6 years.\\n\\nThanks for your input Anzai-san.\\n\\n> I am a newcomer of reviewing, and My programming skill is not so high.\\n> But the following members also participate in this review. (We say\\n> \\\"Two heads are better than one.\\\" :))\\n\\nFWIW, any review for any kind of patches are always welcome, either\\nknowing or not the code you are looking at for a given patch. It is\\neven encouraged to look at new areas when possible.\\n\\n> Feature test\\n> ====================\\n> [blah]\\n> CREATE TABLE buzz ( ... ) WITH ( log_autovacuum_min_duration = -1 );\\n> I expect autovacuum log is NOT printed out even if elapsed time is\\n> more than 1000ms in this case. My thought is wrong, isn't it? In my\\n> opinion, there is an use case that autovacuum log is always printed\\n> out in all tables excepting specified tables. I think there is a\\n> person who wants to use it like this case, but I (or he) can NOT use\\n> in this situation.\\n>\\n> How about your opinion?\\n\\nThe patch is working as expected. Similarly to the other parameters\\nlike vacuum_cost_delay, a value of -1 is the default behavior, meaning\\nthat the underlying GUC parameter is used. I thought about using a\\nspecial value like -2 to define the behavior you are mentioning here,\\naka with \\\"-2\\\" disable custom value and GUC parameter but I don't think\\nit is worth adding as that's an ugly 3 line of code of this type:\\nif (log_min_duration == -2)\\n enforce_log_min = -1;\\nAnd you can actually get what you are looking for by setting\\nmin_duration through ALTER TABLE to a very high value, like say 2e9 to\\nsuppress the autovacuum output of a set of tables out of the rest.\\n\\n> Coding review\\n> ====================\\n> I think description of log_autovacuum_min_duration in reloptions.c\\n> (line:215) should be like other parameters (match with guc.c). So\\n> it should be \\\"Sets the minimum execution time above which autovacuum\\n> actions will be logged.\\\" but not \\\"Log autovacuum execution for\\n> given threshold\\\".\\n\\nWhat about that then?\\n\\\"Minimum execution time above which autovacuum actions will be logged\\\"\\n\\n> Architecture review\\n> ====================\\n> About the modification of gram.y.\\n>\\n> I think it is not good that log_min_duration is initialized to\\n> zeros in gram.y when \\\"FREEZE\\\" option is set. Because any VACUUM\\n> queries never use this parameter. I think log_min_duration always\\n> should be initialized to -1.\\n\\nLooking at this patch this morning, actually I think that my last\\npatch as well as your suggestion are both wrong. To put it in short\\nwords, log_min_duration should be set only if VACOPT_VERBOSE is\\ndefined in query declaration. So I changed this portion of the patch\\nthis way.\\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\": \"Fri, 6 Feb 2015 11:14:45 +0400\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 6, 2015 at 11:14 AM, Michael Paquier wrote:\\n> Looking at this patch this morning, actually I think that my last\\n> patch as well as your suggestion are both wrong. To put it in short\\n> words, log_min_duration should be set only if VACOPT_VERBOSE is\\n> defined in query declaration. So I changed this portion of the patch\\n> this way.\\nAnd I forgot to change VacuumStmt for the ANALYZE portion in gram.y...\\nPatch updated attached.\\n-- \\nMichael\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Fri, 6 Feb 2015 12:14:14 +0400\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"Thanks for your reply.\\n\\n>> Feature test\\n>> ====================\\n(snip)\\n> I thought about using a\\n> special value like -2 to define the behavior you are mentioning here, \\n> aka with \\\"-2\\\" disable custom value and GUC parameter but I don't think\\n> it is worth adding as that's an ugly 3 line of code of this type:\\n> if (log_min_duration == -2)\\n> enforce_log_min = -1;\\n\\nI disscussed about this use case with my co-workers, who said \\n\\\"that code is not good\\\", then we concluded that it was actually \\na rare case. If such a case sometimes happens in fact, I (or someone) \\nwill suggest a different way from this patch to handle this case.\\n\\nWe know there is a practical workaround. :)\\n\\n>> Coding review\\n>> ====================\\n>> I think description of log_autovacuum_min_duration in reloptions.c\\n>> (line:215) should be like other parameters (match with guc.c). So\\n>> it should be \\\"Sets the minimum execution time above which autovacuum\\n>> actions will be logged.\\\" but not \\\"Log autovacuum execution for\\n>> given threshold\\\".\\n>\\n>What about that then?\\n>\\\"Minimum execution time above which autovacuum actions will be logged\\\"\\n\\nThat's roughly match. For matching with guc.c, you might be better to \\nadd \\\"Sets the\\\" to that discription.\\n\\n>> Architecture review\\n>> ====================\\n>> About the modification of gram.y.\\n>>\\n>> I think it is not good that log_min_duration is initialized to\\n>> zeros in gram.y when \\\"FREEZE\\\" option is set. Because any VACUUM\\n>> queries never use this parameter. I think log_min_duration always\\n>> should be initialized to -1.\\n>\\n>Looking at this patch this morning, actually I think that my last\\n>patch as well as your suggestion are both wrong. To put it in short\\n>words, log_min_duration should be set only if VACOPT_VERBOSE is\\n>defined in query declaration. So I changed this portion of the patch\\n>this way.\\n\\n>And I forgot to change VacuumStmt for the ANALYZE portion in gram.y...\\n>Patch updated attached.\\n\\nSorry, I could not correctly express my opinion to you. I mean \\n\\\"log_autovacuum_min_duration\\\" is used only by AutoVacuum, Any VACUUM \\nqueries (including VACUUM VERBOSE) never use this parameter. So, \\nwhen someone executes Manual Vacuum, \\\"log_min_duration\\\" always should \\nbe initialized to -1.\\n\\nANALYZE should also be the same.\\n\\nIn other words, it is not necessary to initialize \\\"log_min_duration\\\" \\nto zero when perform the VACUUM(or ANALYZE) VERBOSE. \\\"VERBOSE\\\" is \\nprovided only for changing the log level of Manual VACUUM from \\nDEBUG2 to INFO.\\n\\nRegards,\\n-----\\nNaoya.\\n\\n\\n-- \\nSent via pgsql-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, 9 Feb 2015 06:47:12 +0000\",\n \"msg_from\": \"Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 9, 2015 at 3:47 PM, Naoya Anzai\\n<anzai-naoya@mxu.nes.nec.co.jp> wrote:\\n>>> Feature test\\n>>> ====================\\n> (snip)\\n>> I thought about using a\\n>> special value like -2 to define the behavior you are mentioning here,\\n>> aka with \\\"-2\\\" disable custom value and GUC parameter but I don't think\\n>> it is worth adding as that's an ugly 3 line of code of this type:\\n>> if (log_min_duration == -2)\\n>> enforce_log_min = -1;\\n>\\n> I discussed about this use case with my co-workers, who said\\n> \\\"that code is not good\\\", then we concluded that it was actually\\n> a rare case. If such a case sometimes happens in fact, I (or someone)\\n> will suggest a different way from this patch to handle this case.\\n>\\n> We know there is a practical workaround. :)\\n\\nOK, done.\\n\\n>>> Coding review\\n>>> ====================\\n>>> I think description of log_autovacuum_min_duration in reloptions.c\\n>>> (line:215) should be like other parameters (match with guc.c). So\\n>>> it should be \\\"Sets the minimum execution time above which autovacuum\\n>>> actions will be logged.\\\" but not \\\"Log autovacuum execution for\\n>>> given threshold\\\".\\n>>\\n>>What about that then?\\n>>\\\"Minimum execution time above which autovacuum actions will be logged\\\"\\n>\\n> That's roughly match. For matching with guc.c, you might be better to\\n> add \\\"Sets the\\\" to that discription.\\n\\nOK, done this way...\\n\\n>>And I forgot to change VacuumStmt for the ANALYZE portion in gram.y...\\n>>Patch updated attached.\\n>\\n> Sorry, I could not correctly express my opinion to you. I mean\\n> \\\"log_autovacuum_min_duration\\\" is used only by AutoVacuum, Any VACUUM\\n> queries (including VACUUM VERBOSE) never use this parameter. So,\\n> when someone executes Manual Vacuum, \\\"log_min_duration\\\" always should\\n> be initialized to -1.\\n>\\n> ANALYZE should also be the same.\\n>\\n> In other words, it is not necessary to initialize \\\"log_min_duration\\\"\\n> to zero when perform the VACUUM(or ANALYZE) VERBOSE. \\\"VERBOSE\\\" is\\n> provided only for changing the log level of Manual VACUUM from\\n> DEBUG2 to INFO.\\n\\nWell, I see your point but this is not completely true: we could as\\nwell rely entirely on this parameter instead of VACOPT_VERBOSE to\\ndetermine if autovacuum, a vacuum or an analyze are in verbose mode,\\nand remove VACOPT_VERBOSE, but I can imagine people complaining if\\nVACOPT_VERBOSE is removed. So let's set it up unconditionally to -1 in\\ngram.y for now. However if people think that it is fine to remove\\nVACOPT_VERBOSE, we could use exclusively this parameter in VacuumStmt.\\nOr even add sanity checks at the top of vacuum() to ensure that\\nVACOPT_VERBOSE is set only when log_min_duration is positive.\\nAdditional opinions on this matter are welcome.\\n-- \\nMichael\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Tue, 10 Feb 2015 17:31:40 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"An updated patch is attached, I think that previous version broke value\\nassignment of log_min_duration in table_recheck_autovac: the value in\\nreltoptions should be used only if it is positive. At the same time I wrote\\nsome more documentation about the fact that a toast table will use the\\nvalue of its parent table if nothing is set.\\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, 10 Feb 2015 22:11:48 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"Hi, Michael-san\\n\\n> An updated patch is attached, \\n\\nI'm sorry for confusing you.\\n\\nI think you don't have to implement this code to disable this\\nfeature with using value \\\"-2\\\".Because this use case is a rare case,\\nand there is a practical workaround using huge value like \\\"2e9\\\".\\n(You suggested \\\"2e9\\\" to me, didn't you? :) ) So, please remove this code.\\n\\n\\n> Well, I see your point but this is not completely true: we could as\\n> well rely entirely on this parameter instead of VACOPT_VERBOSE to\\n> determine if autovacuum, a vacuum or an analyze are in verbose mode,\\n> and remove VACOPT_VERBOSE, but I can imagine people complaining if\\n> VACOPT_VERBOSE is removed. So let's set it up unconditionally to -1 in\\n> gram.y for now. However if people think that it is fine to remove\\n> VACOPT_VERBOSE, we could use exclusively this parameter in VacuumStmt.\\n> Or even add sanity checks at the top of vacuum() to ensure that\\n> VACOPT_VERBOSE is set only when log_min_duration is positive.\\n> Additional opinions on this matter are welcome.\\n\\nI understand your point at last. :)\\n\\nYou mean that ...\\nLog_autovacuum_min_duration assumes a role of VACOPT_VERBOSE.\\nEven if this parameter never use currently for manual vacuum,\\nlog_autovacuum_min_duration should be set zero(anytime output)\\nwhen we executes \\\"VACUUM(or ANALYZE) VERBOSE\\\".\\nIs my understanding correct? If so,it sounds logical.\\n\\nIf we can abolish VERBOSE option,\\nI think it's ideal that we will prepare a new parameter like \\na log_min_duration_vacuum(and log_min_duration_analyze) which \\nintegrating \\\"VERBOSE feature\\\" and \\\"log_autovacuum_min_duration\\\".\\n\\nRegards,\\n---\\nNaoya Anzai\\n\\n\\n-- \\nSent via pgsql-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 Feb 2015 08:44:46 +0000\",\n \"msg_from\": \"Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 12, 2015 at 5:44 PM, Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp>\\nwrote:\\n\\n> Hi, Michael-san\\n>\\n> > An updated patch is attached,\\n>\\n> I'm sorry for confusing you.\\n>\\n> I think you don't have to implement this code to disable this\\n> feature with using value \\\"-2\\\".Because this use case is a rare case,\\n> and there is a practical workaround using huge value like \\\"2e9\\\".\\n> (You suggested \\\"2e9\\\" to me, didn't you? :) ) So, please remove this code.\\n>\\n\\nI will clean up the code.\\n\\n\\n> > Well, I see your point but this is not completely true: we could as\\n> > well rely entirely on this parameter instead of VACOPT_VERBOSE to\\n> > determine if autovacuum, a vacuum or an analyze are in verbose mode,\\n> > and remove VACOPT_VERBOSE, but I can imagine people complaining if\\n> > VACOPT_VERBOSE is removed. So let's set it up unconditionally to -1 in\\n> > gram.y for now. However if people think that it is fine to remove\\n> > VACOPT_VERBOSE, we could use exclusively this parameter in VacuumStmt.\\n> > Or even add sanity checks at the top of vacuum() to ensure that\\n> > VACOPT_VERBOSE is set only when log_min_duration is positive.\\n> > Additional opinions on this matter are welcome.\\n>\\n> I understand your point at last. :)\\n>\\n> You mean that ...\\n> Log_autovacuum_min_duration assumes a role of VACOPT_VERBOSE.\\n> Even if this parameter never use currently for manual vacuum,\\n> log_autovacuum_min_duration should be set zero(anytime output)\\n> when we executes \\\"VACUUM(or ANALYZE) VERBOSE\\\".\\n> Is my understanding correct? If so,it sounds logical.\\n>\\n\\nYup, that's my opinion. Now I don't know if people would mind to remove\\nVACOPT_VERBOSE and replace the control it does by log_min_duration in\\nVacuumStmt. At least both things are overlapping, and log_min_duration\\noffers more options than the plain VACOPT_VERBOSE.\\n\\n\\n> If we can abolish VERBOSE option,\\n> I think it's ideal that we will prepare a new parameter like\\n> a log_min_duration_vacuum(and log_min_duration_analyze) which\\n> integrating \\\"VERBOSE feature\\\" and \\\"log_autovacuum_min_duration\\\".\\n>\\n\\nWhat I think you are proposing here is a VERBOSE option that hypothetically\\ngets activated if a manual VACUUM takes more than a certain amount\\nspecified by those parameters. I doubt this would be useful. In any case\\nthis is unrelated to this patch.\\n-- \\nMichael\\n\\nOn Thu, Feb 12, 2015 at 5:44 PM, Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp> wrote:Hi, Michael-san\\n\\n> An updated patch is attached,\\n\\nI'm sorry for confusing you.\\n\\nI think you don't have to implement this code to disable this\\nfeature with using value \\\"-2\\\".Because this use case is a rare case,\\nand there is a practical workaround using huge value like \\\"2e9\\\".\\n(You suggested \\\"2e9\\\" to me, didn't you? :) ) So, please remove this code.I will clean up the code. \\n> Well, I see your point but this is not completely true: we could as\\n> well rely entirely on this parameter instead of VACOPT_VERBOSE to\\n> determine if autovacuum, a vacuum or an analyze are in verbose mode,\\n> and remove VACOPT_VERBOSE, but I can imagine people complaining if\\n> VACOPT_VERBOSE is removed. So let's set it up unconditionally to -1 in\\n> gram.y for now. However if people think that it is fine to remove\\n> VACOPT_VERBOSE, we could use exclusively this parameter in VacuumStmt.\\n> Or even add sanity checks at the top of vacuum() to ensure that\\n> VACOPT_VERBOSE is set only when log_min_duration is positive.\\n> Additional opinions on this matter are welcome.\\n\\nI understand your point at last. :)\\n\\nYou mean that ...\\nLog_autovacuum_min_duration assumes a role of VACOPT_VERBOSE.\\nEven if this parameter never use currently for manual vacuum,\\nlog_autovacuum_min_duration should be set zero(anytime output)\\nwhen we executes \\\"VACUUM(or ANALYZE) VERBOSE\\\".\\nIs my understanding correct? If so,it sounds logical.Yup, that's my opinion. Now I don't know if people would mind to remove VACOPT_VERBOSE and replace the control it does by log_min_duration in VacuumStmt. At least both things are overlapping, and log_min_duration offers more options than the plain VACOPT_VERBOSE. \\nIf we can abolish VERBOSE option,\\nI think it's ideal that we will prepare a new parameter like\\na log_min_duration_vacuum(and log_min_duration_analyze) which\\nintegrating \\\"VERBOSE feature\\\" and \\\"log_autovacuum_min_duration\\\".\\nWhat I think you are proposing here is a VERBOSE option that hypothetically gets activated if a manual VACUUM takes more than a certain amount specified by those parameters. I doubt this would be useful. In any case this is unrelated to this patch.-- Michael\",\n \"msg_date\": \"Thu, 12 Feb 2015 20:26:21 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \">> You mean that ...\\n>> Log_autovacuum_min_duration assumes a role of VACOPT_VERBOSE.\\n>> Even if this parameter never use currently for manual vacuum,\\n>> log_autovacuum_min_duration should be set zero(anytime output)\\n>> when we executes \\\"VACUUM(or ANALYZE) VERBOSE\\\".\\n>> Is my understanding correct? If so,it sounds logical.\\n>>\\n>\\n>Yup, that's my opinion. Now I don't know if people would mind to remove\\n>VACOPT_VERBOSE and replace the control it does by log_min_duration in\\n>VacuumStmt. At least both things are overlapping, and log_min_duration\\n>offers more options than the plain VACOPT_VERBOSE.\\n\\nOK. I agree with you.\\nPlease re-implement as you are thinking.\\n\\n>> If we can abolish VERBOSE option,\\n>> I think it's ideal that we will prepare a new parameter like\\n>> a log_min_duration_vacuum(and log_min_duration_analyze) which\\n>> integrating \\\"VERBOSE feature\\\" and \\\"log_autovacuum_min_duration\\\".\\n>>\\n>\\n>What I think you are proposing here is a VERBOSE option that hypothetically\\n>gets activated if a manual VACUUM takes more than a certain amount\\n>specified by those parameters. I doubt this would be useful. In any case\\n>this is unrelated to this patch.\\n\\nI suspect manual vacuum often executes as \\\"semi-auto vacuum\\\" \\nby job-scheduler, cron, etc in actual maintenance cases. \\nWhether auto or manual, I think that's important to output \\ntheir logs in the same mechanism.\\n\\nSorry, I seem to have wandered from the subject.\\n\\nNaoya\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 13 Feb 2015 01:16:44 +0000\",\n \"msg_from\": \"Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 13, 2015 at 10:16 AM, Naoya Anzai\\n<anzai-naoya@mxu.nes.nec.co.jp> wrote:\\n>>> You mean that ...\\n>>> Log_autovacuum_min_duration assumes a role of VACOPT_VERBOSE.\\n>>> Even if this parameter never use currently for manual vacuum,\\n>>> log_autovacuum_min_duration should be set zero(anytime output)\\n>>> when we executes \\\"VACUUM(or ANALYZE) VERBOSE\\\".\\n>>> Is my understanding correct? If so,it sounds logical.\\n>>>\\n>>\\n>>Yup, that's my opinion. Now I don't know if people would mind to remove\\n>>VACOPT_VERBOSE and replace the control it does by log_min_duration in\\n>>VacuumStmt. At least both things are overlapping, and log_min_duration\\n>>offers more options than the plain VACOPT_VERBOSE.\\n>\\n> OK. I agree with you.\\n> Please re-implement as you are thinking.\\n\\nOK will do that today.\\n\\n>>> If we can abolish VERBOSE option,\\n>>> I think it's ideal that we will prepare a new parameter like\\n>>> a log_min_duration_vacuum(and log_min_duration_analyze) which\\n>>> integrating \\\"VERBOSE feature\\\" and \\\"log_autovacuum_min_duration\\\".\\n>>>\\n>>\\n>>What I think you are proposing here is a VERBOSE option that hypothetically\\n>>gets activated if a manual VACUUM takes more than a certain amount\\n>>specified by those parameters. I doubt this would be useful. In any case\\n>>this is unrelated to this patch.\\n>\\n> I suspect manual vacuum often executes as \\\"semi-auto vacuum\\\"\\n> by job-scheduler, cron, etc in actual maintenance cases.\\n> Whether auto or manual, I think that's important to output\\n> their logs in the same mechanism.\\n>\\n> Sorry, I seem to have wandered from the subject.\\n\\nNo problem. That's a constructive discussion :)\\n-- \\nMichael\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 13 Feb 2015 11:18:19 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 13, 2015 at 10:16 AM, Naoya Anzai\\n<anzai-naoya@mxu.nes.nec.co.jp> wrote:\\n>>> You mean that ...\\n>>> Log_autovacuum_min_duration assumes a role of VACOPT_VERBOSE.\\n>>> Even if this parameter never use currently for manual vacuum,\\n>>> log_autovacuum_min_duration should be set zero(anytime output)\\n>>> when we executes \\\"VACUUM(or ANALYZE) VERBOSE\\\".\\n>>> Is my understanding correct? If so,it sounds logical.\\n>>>\\n>>\\n>>Yup, that's my opinion. Now I don't know if people would mind to remove\\n>>VACOPT_VERBOSE and replace the control it does by log_min_duration in\\n>>VacuumStmt. At least both things are overlapping, and log_min_duration\\n>>offers more options than the plain VACOPT_VERBOSE.\\n>\\n> OK. I agree with you.\\n> Please re-implement as you are thinking.\\n\\nThanks. Attached is an updated patch will all those things implemented.\\n\\n>>> If we can abolish VERBOSE option,\\n>>> I think it's ideal that we will prepare a new parameter like\\n>>> a log_min_duration_vacuum(and log_min_duration_analyze) which\\n>>> integrating \\\"VERBOSE feature\\\" and \\\"log_autovacuum_min_duration\\\".\\n>>>\\n>>\\n>>What I think you are proposing here is a VERBOSE option that hypothetically\\n>>gets activated if a manual VACUUM takes more than a certain amount\\n>>specified by those parameters. I doubt this would be useful. In any case\\n>>this is unrelated to this patch.\\n>\\n> I suspect manual vacuum often executes as \\\"semi-auto vacuum\\\"\\n> by job-scheduler, cron, etc in actual maintenance cases.\\n> Whether auto or manual, I think that's important to output\\n> their logs in the same mechanism.\\n>\\n> Sorry, I seem to have wandered from the subject.\\n\\nThis patch is a step in this direction as it enables any backend-side\\ncode to set up VacuumStmt with a custom timestamp value to output logs\\nafter a given timing, for example in background workers. For the\\nclient-side of things, I am unsure if we'd actually want it, we should\\nalways VERBOSE when this option is invoked through a query, and not\\nadd any hypothetical condition on it...\\n\\nBtw, after hacking on this it happens that we cannot completely remove\\nVACOPT_VERBOSE as gram.y needs to take into account options with\\nparenthesis :)\\nBut we can limit its use to the query parser only, similarly for VACOPT_FREEZE.\\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\": \"Fri, 13 Feb 2015 13:21:34 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"Hi, Michael\\n\\nI found that definition of VERBOSE and log_autovacuum is not \\npretty match. For example, \\\"VERBOSE\\\" can output logs of \\nscanning indices and scanning detail of analyze, but \\nlog_autovacuum can't output them.\\n\\nPlease see following sequences.\\n\\n1. execute these queries.\\n\\n DROP TABLE t1;\\n CREATE TABLE t1(id integer primary key,name text);\\n ALTER TABLE t1 SET (log_autovacuum_min_duration=0);\\n ALTER TABLE t1 ALTER COLUMN name SET STORAGE EXTERNAL;\\n INSERT INTO t1 SELECT GENERATE_SERIES(1,100),repeat('a',3000);\\n UPDATE t1 SET name='update';\\n\\n2. after a while, output the following logs by autovacuum movement.\\n(For your convenience, I put the \\\"### TYPE ###\\\" prefix on each logs.)\\n\\n ### VERBOSE ###INFO: vacuuming \\\"public.t1\\\"\\n ### VERBOSE ###INFO: scanned index \\\"t1_pkey\\\" to remove 33 row versions\\n ### VERBOSE ###DETAIL: CPU 0.00s/0.00u sec elapsed 0.00 sec.\\n ### VERBOSE ###INFO: \\\"t1\\\": removed 33 row versions in 1 pages\\n ### VERBOSE ###DETAIL: CPU 0.00s/0.00u sec elapsed 0.00 sec.\\n ### VERBOSE ###INFO: index \\\"t1_pkey\\\" now contains 100 row versions in 2 pages\\n ### VERBOSE ###DETAIL: 33 index row versions were removed.\\n ### VERBOSE ### 0 index pages have been deleted, 0 are currently reusable.\\n ### VERBOSE ### CPU 0.00s/0.00u sec elapsed 0.00 sec.\\n ### VERBOSE ###INFO: \\\"t1\\\": found 100 removable, 100 nonremovable row versions in 2 out of 2 pages\\n ### VERBOSE ###DETAIL: 0 dead row versions cannot be removed yet.\\n ### VERBOSE ### There were 0 unused item pointers.\\n ### VERBOSE ### Skipped 0 pages due to buffer pins.\\n ### VERBOSE ### 0 pages are entirely empty.\\n ### VERBOSE ### CPU 0.00s/0.00u sec elapsed 0.00 sec.\\n ### LOG_AVAC ###LOG: automatic vacuum of table \\\"naoya.public.t1\\\": index scans: 1\\n ### LOG_AVAC ### pages: 0 removed, 2 remain, 0 skipped due to pins\\n ### LOG_AVAC ### tuples: 100 removed, 100 remain, 0 are dead but not yet removable\\n ### LOG_AVAC ### buffer usage: 47 hits, 4 misses, 4 dirtied\\n ### LOG_AVAC ### avg read rate: 14.192 MB/s, avg write rate: 14.192 MB/s\\n ### LOG_AVAC ### system usage: CPU 0.00s/0.00u sec elapsed 0.00 sec\\n ### VERBOSE ###INFO: analyzing \\\"public.t1\\\"\\n ### VERBOSE ###INFO: \\\"t1\\\": scanned 2 of 2 pages, containing 100 live rows and 0 dead rows; 100 rows in sample, 100 estimated total rows\\n ### LOG_AVAC ###LOG: automatic analyze of table \\\"naoya.public.t1\\\" system usage: CPU 0.00s/0.00u sec elapsed 0.04 sec\\n ### VERBOSE ###INFO: vacuuming \\\"pg_toast.pg_toast_72882\\\"\\n ### VERBOSE ###INFO: scanned index \\\"pg_toast_72882_index\\\" to remove 200 row versions\\n ### VERBOSE ###DETAIL: CPU 0.00s/0.00u sec elapsed 0.00 sec.\\n ### VERBOSE ###INFO: \\\"pg_toast_72882\\\": removed 200 row versions in 50 pages\\n ### VERBOSE ###DETAIL: CPU 0.00s/0.00u sec elapsed 0.00 sec.\\n ### VERBOSE ###INFO: index \\\"pg_toast_72882_index\\\" now contains 0 row versions in 2 pages\\n ### VERBOSE ###DETAIL: 200 index row versions were removed.\\n ### VERBOSE ### 0 index pages have been deleted, 0 are currently reusable.\\n ### VERBOSE ### CPU 0.00s/0.00u sec elapsed 0.00 sec.\\n ### VERBOSE ###INFO: \\\"pg_toast_72882\\\": found 200 removable, 0 nonremovable row versions in 50 out of 50 pages\\n ### VERBOSE ###DETAIL: 0 dead row versions cannot be removed yet.\\n ### VERBOSE ### There were 0 unused item pointers.\\n ### VERBOSE ### Skipped 0 pages due to buffer pins.\\n ### VERBOSE ### 0 pages are entirely empty.\\n ### VERBOSE ### CPU 0.00s/0.00u sec elapsed 0.00 sec.\\n ### VERBOSE ###INFO: \\\"pg_toast_72882\\\": truncated 50 to 0 pages\\n ### VERBOSE ###DETAIL: CPU 0.00s/0.00u sec elapsed 0.03 sec.\\n ### LOG_AVAC ###LOG: automatic vacuum of table \\\"naoya.pg_toast.pg_toast_72882\\\": index scans: 1\\n ### LOG_AVAC ### pages: 50 removed, 0 remain, 0 skipped due to pins\\n ### LOG_AVAC ### tuples: 200 removed, 0 remain, 0 are dead but not yet removable\\n ### LOG_AVAC ### buffer usage: 223 hits, 2 misses, 1 dirtied\\n ### LOG_AVAC ### avg read rate: 0.457 MB/s, avg write rate: 0.228 MB/s\\n ### LOG_AVAC ### system usage: CPU 0.00s/0.00u sec elapsed 0.03 sec\\n\\nI think VACOPT_VERBOSE should not be easily replaced to log_min_duration>=0.\\n\\nAnd, in this v7 patch looks that VERBOSE log is always output \\nin INFO-Level when log_autovacuum_min_duration is set to 0. \\nIs this your point?\\n\\nRegards,\\n---\\nNaoya\\n\\n\\n-- \\nSent via pgsql-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 Feb 2015 08:16:26 +0000\",\n \"msg_from\": \"Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Tue, Feb 17, 2015 at 5:16 PM, Naoya Anzai wrote:\\n> I found that definition of VERBOSE and log_autovacuum is not\\n> pretty match. For example, \\\"VERBOSE\\\" can output logs of\\n> scanning indices and scanning detail of analyze, but\\n> log_autovacuum can't output them.\\n> [...]\\n\\nDoh. I forgot to add !IsAutoVacuumWorkerProcess() in the checks with\\nVACOPT_VERBOSE that I previously removed...\\n\\n> And, in this v7 patch looks that VERBOSE log is always output\\n> in INFO-Level when log_autovacuum_min_duration is set to 0.\\n> Is this your point?\\n\\nThat's my bug, fixed in the attached.\\n-- \\nMichael\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Tue, 17 Feb 2015 23:01:33 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"Hi, Michael\\n\\nI thought about VACOPT_VERBOSE again.\\n\\nAs a result, I think you should not delete VACOPT_VERBOSE.\\n\\nAccording to the last mail I have posted, the difference of \\nmanual-vacuum log and auto-vacuum log exists clearly.\\nSo, at least you should not touch the mechanism of VACOPT_VERBOSE \\nuntil both vacuum log formats are unified to a same format.\\n\\nIf you agree my think, please undo your removing VACOPT_VERBOSE work.\\n\\nRegards,\\n---\\nNaoya\\n\\n\\n-- \\nSent via pgsql-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 Feb 2015 05:13:00 +0000\",\n \"msg_from\": \"Naoya Anzai <anzai-naoya@mxu.nes.nec.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 19, 2015 at 2:13 PM, Naoya Anzai\\n<anzai-naoya@mxu.nes.nec.co.jp> wrote:\\n> As a result, I think you should not delete VACOPT_VERBOSE.\\n\\nIn v8 it is not deleted. It is still declared, and its use is isolated\\nin gram.y, similarly to VACOPT_FREEZE.\\n\\n> According to the last mail I have posted, the difference of\\n> manual-vacuum log and auto-vacuum log exists clearly.\\n\\nDid you test the latest patch v8? I have added checks in it to see if\\na process is an autovacuum worker to control elevel and the extra logs\\nof v7 do not show up.\\n\\n> So, at least you should not touch the mechanism of VACOPT_VERBOSE\\n> until both vacuum log formats are unified to a same format.\\n\\nIf you mean that we should have the same kind of log outputs for\\nautovacuum and manual vacuum, I think that this is not going to\\nhappen. Autovacuum entries are kept less verbose on purpose, contract\\nthat v7 clealy broke.\\n\\n> If you agree my think, please undo your removing VACOPT_VERBOSE work.\\n\\nWell, I don't agree :) And I am guessing that you did not look at v8\\nas well. Centralizing the control of logs using log_min_duration is\\nmore extensible than simply having VACOPT_VERBOSE.\\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, 19 Feb 2015 15:32:17 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 19, 2015 at 3:32 PM, Michael Paquier\\n<michael.paquier@gmail.com> wrote:\\n> On Thu, Feb 19, 2015 at 2:13 PM, Naoya Anzai\\n> <anzai-naoya@mxu.nes.nec.co.jp> wrote:\\n>> As a result, I think you should not delete VACOPT_VERBOSE.\\n>\\n> In v8 it is not deleted. It is still declared, and its use is isolated\\n> in gram.y, similarly to VACOPT_FREEZE.\\n>\\n>> According to the last mail I have posted, the difference of\\n>> manual-vacuum log and auto-vacuum log exists clearly.\\n>\\n> Did you test the latest patch v8? I have added checks in it to see if\\n> a process is an autovacuum worker to control elevel and the extra logs\\n> of v7 do not show up.\\n>\\n>> So, at least you should not touch the mechanism of VACOPT_VERBOSE\\n>> until both vacuum log formats are unified to a same format.\\n>\\n> If you mean that we should have the same kind of log outputs for\\n> autovacuum and manual vacuum, I think that this is not going to\\n> happen. Autovacuum entries are kept less verbose on purpose, contract\\n> that v7 clealy broke.\\n>\\n>> If you agree my think, please undo your removing VACOPT_VERBOSE work.\\n>\\n> Well, I don't agree :) And I am guessing that you did not look at v8\\n> as well. Centralizing the control of logs using log_min_duration is\\n> more extensible than simply having VACOPT_VERBOSE.\\n\\nWith the patch, VACUUM ANALYZE VERBOSE doesn't emit any verbose message.\\nWhy did you remove that functionality?\\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\": \"Thu, 5 Mar 2015 19:10:08 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 5, 2015 at 7:10 PM, Fujii Masao wrote:\\n> With the patch, VACUUM ANALYZE VERBOSE doesn't emit any verbose message.\\n> Why did you remove that functionality?\\n\\nOops. Sorry about that. In gram.y, the combination of VacuumStmt with\\nAnalyzeStmt overwrote the value of log_min_duration incorrectly. I\\nalso found another bug related to logging of ANALYZE not working\\ncorrectly because of the use of IsAutoVacuumWorkerProcess() instead of\\nVACOPT_VERBOSE (this is reducing the diffs of the patch btw). All\\nthose things are fixed in the attached.\\n-- \\nMichael\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Thu, 5 Mar 2015 21:49:03 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 5, 2015 at 9:49 PM, Michael Paquier\\n<michael.paquier@gmail.com> wrote:\\n> On Thu, Mar 5, 2015 at 7:10 PM, Fujii Masao wrote:\\n>> With the patch, VACUUM ANALYZE VERBOSE doesn't emit any verbose message.\\n>> Why did you remove that functionality?\\n>\\n> Oops. Sorry about that. In gram.y, the combination of VacuumStmt with\\n> AnalyzeStmt overwrote the value of log_min_duration incorrectly. I\\n> also found another bug related to logging of ANALYZE not working\\n> correctly because of the use of IsAutoVacuumWorkerProcess() instead of\\n> VACOPT_VERBOSE (this is reducing the diffs of the patch btw). All\\n> those things are fixed in the attached.\\n\\nThanks for updating the patch!\\n\\nWhy does log_min_duration need to be set even when manual VACUUM command is\\nexecuted? Per the latest version of the patch, log_min_duration is checked only\\nwhen the process is autovacuum worker. So ISTM that log_min_duration doesn't\\nneed to be set in gram.y. It's even confusing to me. Or if you're going to\\nimplement something like \\\"VACUUM VERBOSE DURATION n\\\" (i.e., verbose message\\nis output if n seconds have been elapsed), that might be necessary, though...\\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\": \"Fri, 6 Mar 2015 12:44:11 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 6, 2015 at 12:44 PM, Fujii Masao <masao.fujii@gmail.com> wrote:\\n> On Thu, Mar 5, 2015 at 9:49 PM, Michael Paquier\\n> <michael.paquier@gmail.com> wrote:\\n>> On Thu, Mar 5, 2015 at 7:10 PM, Fujii Masao wrote:\\n>>> With the patch, VACUUM ANALYZE VERBOSE doesn't emit any verbose message.\\n>>> Why did you remove that functionality?\\n>>\\n>> Oops. Sorry about that. In gram.y, the combination of VacuumStmt with\\n>> AnalyzeStmt overwrote the value of log_min_duration incorrectly. I\\n>> also found another bug related to logging of ANALYZE not working\\n>> correctly because of the use of IsAutoVacuumWorkerProcess() instead of\\n>> VACOPT_VERBOSE (this is reducing the diffs of the patch btw). All\\n>> those things are fixed in the attached.\\n>\\n> Thanks for updating the patch!\\n>\\n> Why does log_min_duration need to be set even when manual VACUUM command is\\n> executed? Per the latest version of the patch, log_min_duration is checked only\\n> when the process is autovacuum worker. So ISTM that log_min_duration doesn't\\n> need to be set in gram.y. It's even confusing to me. Or if you're going to\\n> implement something like \\\"VACUUM VERBOSE DURATION n\\\" (i.e., verbose message\\n> is output if n seconds have been elapsed), that might be necessary, though...\\n\\nThanks for reminding. The DURATION-like clause was exactly a point\\nmentioned by Anzai-san upthread, and it made sense to me to be in-line\\nwith the other parameters controlling the freeze (discussion somewhat\\nrelated to that =>\\nhttp://www.postgresql.org/message-id/CAB7nPqRZX7Pv2B-R7xHmAh52tfjAQGfy9btqwFstgQgXks=iSw@mail.gmail.com)\\nbut we can live without it for this patch as VACOPT_VERBOSE is used\\nonly by manual VACUUM and not by autovacuum to choose the log elevel.\\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\": \"Fri, 6 Mar 2015 13:07:04 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 6, 2015 at 1:07 PM, Michael Paquier\\n<michael.paquier@gmail.com> wrote:\\n> On Fri, Mar 6, 2015 at 12:44 PM, Fujii Masao <masao.fujii@gmail.com> wrote:\\n>> On Thu, Mar 5, 2015 at 9:49 PM, Michael Paquier\\n>> <michael.paquier@gmail.com> wrote:\\n>>> On Thu, Mar 5, 2015 at 7:10 PM, Fujii Masao wrote:\\n>>>> With the patch, VACUUM ANALYZE VERBOSE doesn't emit any verbose message.\\n>>>> Why did you remove that functionality?\\n>>>\\n>>> Oops. Sorry about that. In gram.y, the combination of VacuumStmt with\\n>>> AnalyzeStmt overwrote the value of log_min_duration incorrectly. I\\n>>> also found another bug related to logging of ANALYZE not working\\n>>> correctly because of the use of IsAutoVacuumWorkerProcess() instead of\\n>>> VACOPT_VERBOSE (this is reducing the diffs of the patch btw). All\\n>>> those things are fixed in the attached.\\n>>\\n>> Thanks for updating the patch!\\n>>\\n>> Why does log_min_duration need to be set even when manual VACUUM command is\\n>> executed? Per the latest version of the patch, log_min_duration is checked only\\n>> when the process is autovacuum worker. So ISTM that log_min_duration doesn't\\n>> need to be set in gram.y. It's even confusing to me. Or if you're going to\\n>> implement something like \\\"VACUUM VERBOSE DURATION n\\\" (i.e., verbose message\\n>> is output if n seconds have been elapsed), that might be necessary, though...\\n>\\n> Thanks for reminding. The DURATION-like clause was exactly a point\\n> mentioned by Anzai-san upthread, and it made sense to me to be in-line\\n> with the other parameters controlling the freeze (discussion somewhat\\n> related to that =>\\n> http://www.postgresql.org/message-id/CAB7nPqRZX7Pv2B-R7xHmAh52tfjAQGfy9btqwFstgQgXks=iSw@mail.gmail.com)\\n> but we can live without it for this patch as VACOPT_VERBOSE is used\\n> only by manual VACUUM and not by autovacuum to choose the log elevel.\\n\\nAre you planning to update the patch so that it's based on the commit 0d83138?\\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\": \"Thu, 19 Mar 2015 12:23:01 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 19, 2015 at 12:23 PM, Fujii Masao <masao.fujii@gmail.com> wrote:\\n> Are you planning to update the patch so that it's based on the commit 0d83138?\\n\\nYes... Very soon.\\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, 19 Mar 2015 12:40:40 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 19, 2015 at 12:40 PM, Michael Paquier\\n<michael.paquier@gmail.com> wrote:\\n> On Thu, Mar 19, 2015 at 12:23 PM, Fujii Masao <masao.fujii@gmail.com> wrote:\\n>> Are you planning to update the patch so that it's based on the commit 0d83138?\\n>\\n> Yes... Very soon.\\n\\nAnd here is the rebased patch.\\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\": \"Thu, 19 Mar 2015 13:43:52 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 19, 2015 at 1:43 PM, Michael Paquier\\n<michael.paquier@gmail.com> wrote:\\n> On Thu, Mar 19, 2015 at 12:40 PM, Michael Paquier\\n> <michael.paquier@gmail.com> wrote:\\n>> On Thu, Mar 19, 2015 at 12:23 PM, Fujii Masao <masao.fujii@gmail.com> wrote:\\n>>> Are you planning to update the patch so that it's based on the commit 0d83138?\\n>>\\n>> Yes... Very soon.\\n>\\n> And here is the rebased patch.\\n\\nThanks for rebasing the patch! Looks good to me.\\n\\nOne concern about this patch is; currently log_autovacuum_min_duration can be\\nchanged even while autovacuum worker is running. So, for example, when\\nthe admin notices that autovacuum is taking very long time, he or she can\\nenable logging of autovacuum activity on the fly. But this patch completely\\nprevents us from doing that, because, with the patch, autovacuum worker always\\npicks up the latest setting value at its starting time and then keeps using it\\nto the end. Probably I can live with this. But does anyone has other thought?\\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\": \"Mon, 23 Mar 2015 13:54:03 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 23, 2015 at 1:54 PM, Fujii Masao <masao.fujii@gmail.com> wrote:\\n> On Thu, Mar 19, 2015 at 1:43 PM, Michael Paquier\\n> <michael.paquier@gmail.com> wrote:\\n>> On Thu, Mar 19, 2015 at 12:40 PM, Michael Paquier\\n>> <michael.paquier@gmail.com> wrote:\\n>>> On Thu, Mar 19, 2015 at 12:23 PM, Fujii Masao <masao.fujii@gmail.com> wrote:\\n>>>> Are you planning to update the patch so that it's based on the commit 0d83138?\\n>>>\\n>>> Yes... Very soon.\\n>>\\n>> And here is the rebased patch.\\n>\\n> Thanks for rebasing the patch! Looks good to me.\\n>\\n> One concern about this patch is; currently log_autovacuum_min_duration can be\\n> changed even while autovacuum worker is running. So, for example, when\\n> the admin notices that autovacuum is taking very long time, he or she can\\n> enable logging of autovacuum activity on the fly. But this patch completely\\n> prevents us from doing that, because, with the patch, autovacuum worker always\\n> picks up the latest setting value at its starting time and then keeps using it\\n> to the end. Probably I can live with this. But does anyone has other thought?\\n\\nIn AutoVacWorkerMain, I am reading the following:\\n\\n * Currently, we don't pay attention to postgresql.conf changes that\\n * happen during a single daemon iteration, so we can ignore SIGHUP.\\n */\\n pqsignal(SIGHUP, SIG_IGN);\\n\\nSo a worker does not see changes in postgresql.conf once it is run and\\nprocesses a database, no? The launcher does run ProcessConfigFile()\\nwhen SIGHUP shows up though.\\nRegards,\\n-- \\nMichael\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 23 Mar 2015 14:23:36 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"Michael Paquier wrote:\\n\\n> In AutoVacWorkerMain, I am reading the following:\\n> \\n> * Currently, we don't pay attention to postgresql.conf changes that\\n> * happen during a single daemon iteration, so we can ignore SIGHUP.\\n> */\\n> pqsignal(SIGHUP, SIG_IGN);\\n> \\n> So a worker does not see changes in postgresql.conf once it is run and\\n> processes a database, no? The launcher does run ProcessConfigFile()\\n> when SIGHUP shows up though.\\n\\nMaybe this is something that we should change. For example, I wonder if\\nthis can also affect the cost-delay balancing heuristics; if two\\nbackends run the rebalance with different GUC settings because\\npostgresql.conf changed in between each of them starting, would the\\nsettings bounce back and forth. I think it's worth reconsidering this.\\n(Don't really remember in detail how it works; maybe it's fine now.)\\n\\nIn any case, for log_autovacuum_min_duration it also seems worth keeping\\nreasonably close track of GUC changes. I think reading them just before\\nstarting vacuum of a new relation should be enough.\\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\": \"Mon, 23 Mar 2015 10:25:20 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n> Michael Paquier wrote:\\n>> So a worker does not see changes in postgresql.conf once it is run and\\n>> processes a database, no? The launcher does run ProcessConfigFile()\\n>> when SIGHUP shows up though.\\n\\n> Maybe this is something that we should change.\\n\\nYeah, checking for SIGHUP in the worker outer loop (ie once per table)\\nseems like a reasonable thing.\\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, 23 Mar 2015 10:07:12 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 23, 2015 at 7:07 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n\\n> Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n> > Michael Paquier wrote:\\n> >> So a worker does not see changes in postgresql.conf once it is run and\\n> >> processes a database, no? The launcher does run ProcessConfigFile()\\n> >> when SIGHUP shows up though.\\n>\\n> > Maybe this is something that we should change.\\n>\\n> Yeah, checking for SIGHUP in the worker outer loop (ie once per table)\\n> seems like a reasonable thing.\\n>\\n> regards, tom lane\\n>\\n\\nCould it be done more often? Maybe every time it is about to do a\\ncost_delay sleep?\\n\\nI've certainly regretted the inability to\\nchange autovacuum_vacuum_cost_delay mid-table on more than one occasion.\\n\\nThis was mostly during doing testing work, but still I'm sure other people\\nhave run into this problem, perhaps without knowing it.\\n\\nCheers,\\n\\nJeff\\n\\nOn Mon, Mar 23, 2015 at 7:07 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n> Michael Paquier wrote:\\n>> So a worker does not see changes in postgresql.conf once it is run and\\n>> processes a database, no? The launcher does run ProcessConfigFile()\\n>> when SIGHUP shows up though.\\n\\n> Maybe this is something that we should change.\\n\\nYeah, checking for SIGHUP in the worker outer loop (ie once per table)\\nseems like a reasonable thing.\\n\\n regards, tom laneCould it be done more often? Maybe every time it is about to do a cost_delay sleep?I've certainly regretted the inability to change autovacuum_vacuum_cost_delay mid-table on more than one occasion.This was mostly during doing testing work, but still I'm sure other people have run into this problem, perhaps without knowing it.Cheers,Jeff\",\n \"msg_date\": \"Mon, 23 Mar 2015 08:23:00 -0700\",\n \"msg_from\": \"Jeff Janes <jeff.janes@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"Jeff Janes <jeff.janes@gmail.com> writes:\\n> On Mon, Mar 23, 2015 at 7:07 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> Yeah, checking for SIGHUP in the worker outer loop (ie once per table)\\n>> seems like a reasonable thing.\\n\\n> Could it be done more often? Maybe every time it is about to do a\\n> cost_delay sleep?\\n\\nThat sounds risky from here. Normal backends don't check it more often\\nthan once per command.\\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, 23 Mar 2015 11:34:21 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 23, 2015 at 11:07 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n>> Michael Paquier wrote:\\n>>> So a worker does not see changes in postgresql.conf once it is run and\\n>>> processes a database, no? The launcher does run ProcessConfigFile()\\n>>> when SIGHUP shows up though.\\n>\\n>> Maybe this is something that we should change.\\n>\\n> Yeah, checking for SIGHUP in the worker outer loop (ie once per table)\\n> seems like a reasonable thing.\\n\\nThat sounds fine to me as well. A patch would not be complicated, but\\nis this portion really 9.5 material?\\n\\nAlso, this is a discussion wider than only\\nlog_autovacuum_min_duration, as autovacuum cost parameters are also\\navailable as reloptions.\\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 Mar 2015 10:45:15 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"Michael Paquier wrote:\\n> On Mon, Mar 23, 2015 at 11:07 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> > Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n> >> Michael Paquier wrote:\\n> >>> So a worker does not see changes in postgresql.conf once it is run and\\n> >>> processes a database, no? The launcher does run ProcessConfigFile()\\n> >>> when SIGHUP shows up though.\\n> >\\n> >> Maybe this is something that we should change.\\n> >\\n> > Yeah, checking for SIGHUP in the worker outer loop (ie once per table)\\n> > seems like a reasonable thing.\\n> \\n> That sounds fine to me as well. A patch would not be complicated, but\\n> is this portion really 9.5 material?\\n\\nIMO yes.\\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\": \"Mon, 23 Mar 2015 23:46:21 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 23, 2015 at 7:46 PM, Alvaro Herrera wrote:\\n> Michael Paquier wrote:\\n>> On Mon, Mar 23, 2015 at 11:07 PM, Tom Lane wrote:\\n>> > Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n>> >> Michael Paquier wrote:\\n>> >>> So a worker does not see changes in postgresql.conf once it is run and\\n>> >>> processes a database, no? The launcher does run ProcessConfigFile()\\n>> >>> when SIGHUP shows up though.\\n>> >\\n>> >> Maybe this is something that we should change.\\n>> >\\n>> > Yeah, checking for SIGHUP in the worker outer loop (ie once per table)\\n>> > seems like a reasonable thing.\\n>>\\n>> That sounds fine to me as well. A patch would not be complicated, but\\n>> is this portion really 9.5 material?\\n>\\n> IMO yes.\\n\\nSo, attached are two patches:\\n- 0001 enables SIGHUP tracking in do_autovacuum(), which is checked\\nbefore processing one table. I reused avl_sighup_handler for the\\nworker, renaming it av_sighup_handler..\\n- 0002 is the patch to add log_autovacuum_min_duration as a reloption.\\nThere is nothing really changed, the value of\\nlog_autovacuum_min_duration being picked up at startup with\\ntable_recheck_autovac() is used until the end for one table.\\nRegards,\\n-- \\nMichael\\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 Mar 2015 15:54:16 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"Michael Paquier wrote:\\n\\n> So, attached are two patches:\\n> - 0001 enables SIGHUP tracking in do_autovacuum(), which is checked\\n> before processing one table. I reused avl_sighup_handler for the\\n> worker, renaming it av_sighup_handler..\\n> - 0002 is the patch to add log_autovacuum_min_duration as a reloption.\\n> There is nothing really changed, the value of\\n> log_autovacuum_min_duration being picked up at startup with\\n> table_recheck_autovac() is used until the end for one table.\\n\\nThanks.\\n\\nYou added this in the worker loop processing each table:\\n\\n\\t\\t/*\\n\\t\\t * Check for config changes before processing each collected table.\\n\\t\\t */\\n\\t\\tif (got_SIGHUP)\\n\\t\\t{\\n\\t\\t\\tgot_SIGHUP = false;\\n\\t\\t\\tProcessConfigFile(PGC_SIGHUP);\\n\\n\\t\\t\\t/* shutdown requested in config file? */\\n\\t\\t\\tif (!AutoVacuumingActive())\\n\\t\\t\\t\\tbreak;\\n\\t\\t}\\n\\nI think bailing out if autovac is disabled is a good idea; for instance,\\nif this is a for-wraparound vacuum, we should keep running. My first\\nreaction is that this part of the test ought to be moved down a\\nscreenful or so, until we've ran the recheck step and we have the\\nfor-wraparound flag in hand; only bail out if that one is not set. But\\non the other hand, maybe the worker will process some tables that are\\nnot for-wraparound in between some other tables that are, so that might\\nturn out to be a bad idea as well. (Though I'm not 100% that it works\\nthat way; consider commit f51ead09df for instance.) Maybe the test to\\nuse for this is something along the lines of \\\"if autovacuum was enabled\\nbefore and is no longer enabled, bail out, otherwise keep running\\\".\\nThis implies that we need to keep track of whether autovac was enabled\\nat the start of the worker run.\\n\\n\\nAnother thing, but not directly related to this patch: while looking at\\nthe doc changes, I noticed that we don't have index entries for the\\nreloptions we have; for instance, the word \\\"fillfactor\\\" doesn't appear\\nin the bookindex.html page at all. Having these things all crammed in\\nthe CREATE TABLE page seems somewhat poor. I think we could improve on\\nthis by having a listing of settable parameters in a separate section,\\nand have CREATE TABLE, ALTER TABLE, and the Server Configuration chapter\\nlink to that; we could also have index entries for these items.\\n\\nOf course, the simpler fix is to just add a few <indexterm> tags.\\n\\n\\nAs a note, there is no point to \\\"Assert(foo != NULL)\\\" tests when foo is\\nlater dereferenced in the same routine: the crash is going to happen\\nanyway at the dereference, so it's just a LOC uselessly wasted.\\n\\n-- \\n�lvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 2 Apr 2015 17:57:14 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"\\n--- a/doc/src/sgml/ref/create_table.sgml\\n+++ b/doc/src/sgml/ref/create_table.sgml\\n@@ -881,9 +881,11 @@ CREATE [ [ GLOBAL | LOCAL ] { TEMPORARY | TEMP } | UNLOGGED ] TABLE [ IF NOT EXI\\n <literal>toast.</literal>, which can be used to control the behavior of the\\n table's secondary <acronym>TOAST</> table, if any\\n (see <xref linkend=\\\"storage-toast\\\"> for more information about TOAST).\\n- Note that the TOAST table inherits the\\n- <literal>autovacuum_*</literal> values from its parent table, if there are\\n- no <literal>toast.autovacuum_*</literal> settings set.\\n+ Note that the TOAST table inherits values of <literal>autovacuum_*</literal>\\n+ and <literal>log_autovacuum_min_duration</literal> from its parent table, if\\n+ there are no values set for respectively\\n+ <literal>toast.autovacuum_*</literal> and\\n+ <literal>toast.log_autovacuum_min_duration</literal>.\\n </para>\\n \\n\\nI think this could use some wordsmithing; I didn't like listing the\\nparameters explicitely, and Jaime Casanova suggested not using the terms\\n\\\"inherit\\\" and \\\"parent table\\\" in this context. So I came up with this:\\n\\n Note that the TOAST table uses the parameter values defined for the main\\n table, for each parameter applicable to TOAST tables and for which no\\n value is set in the TOAST table itself.\\n\\nThoughts?\\n\\n-- \\n�lvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 2 Apr 2015 23:04:55 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Fri, Apr 3, 2015 at 5:57 AM, Alvaro Herrera wrote:\\n> You added this in the worker loop processing each table:\\n>\\n> /*\\n> * Check for config changes before processing each collected table.\\n> */\\n> if (got_SIGHUP)\\n> {\\n> got_SIGHUP = false;\\n> ProcessConfigFile(PGC_SIGHUP);\\n>\\n> /* shutdown requested in config file? */\\n> if (!AutoVacuumingActive())\\n> break;\\n> }\\n>\\n> I think bailing out if autovac is disabled is a good idea; for instance,\\n> if this is a for-wraparound vacuum, we should keep running. My first\\n> reaction is that this part of the test ought to be moved down a\\n> screenful or so, until we've ran the recheck step and we have the\\n> for-wraparound flag in hand; only bail out if that one is not set. But\\n> on the other hand, maybe the worker will process some tables that are\\n> not for-wraparound in between some other tables that are, so that might\\n> turn out to be a bad idea as well. (Though I'm not 100% that it works\\n> that way; consider commit f51ead09df for instance.) Maybe the test to\\n> use for this is something along the lines of \\\"if autovacuum was enabled\\n> before and is no longer enabled, bail out, otherwise keep running\\\".\\n> This implies that we need to keep track of whether autovac was enabled\\n> at the start of the worker run.\\n\\nI did consider the case of wraparound but came to the conclusion that\\nbailing out as fast as possible was the idea. But well, I guess that\\nwe could play it safe and not add this check after all. The main\\nuse-case of this change is now to be able to do re-balancing of the\\ncost parameters. We could still decide later if a worker could bail\\nout earlier or not, depending on what.\\n\\n> Another thing, but not directly related to this patch: while looking at\\n> the doc changes, I noticed that we don't have index entries for the\\n> reloptions we have; for instance, the word \\\"fillfactor\\\" doesn't appear\\n> in the bookindex.html page at all. Having these things all crammed in\\n> the CREATE TABLE page seems somewhat poor. I think we could improve on\\n> this by having a listing of settable parameters in a separate section,\\n> and have CREATE TABLE, ALTER TABLE, and the Server Configuration chapter\\n> link to that; we could also have index entries for these items.\\n> Of course, the simpler fix is to just add a few <indexterm> tags.\\n\\nThis sounds like a good idea, and this refactoring would meritate a\\ndifferent patch and a different thread. I guess that it should be a\\nnew section in Server Configuration then, named \\\"Relation Options\\\",\\nwith two different subsections for index options and table options.\\n\\n> As a note, there is no point to \\\"Assert(foo != NULL)\\\" tests when foo is\\n> later dereferenced in the same routine: the crash is going to happen\\n> anyway at the dereference, so it's just a LOC uselessly wasted.\\n\\nCheck. I still think that it is useful in this case though... But removed.\\n\\n> I think this could use some wordsmithing; I didn't like listing the\\n> parameters explicitely, and Jaime Casanova suggested not using the terms\\n> \\\"inherit\\\" and \\\"parent table\\\" in this context. So I came up with this:\\n> Note that the TOAST table uses the parameter values defined for the main\\n> table, for each parameter applicable to TOAST tables and for which no\\n> value is set in the TOAST table itself.\\n\\nFine for me.\\n-- \\nMichael\\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, 3 Apr 2015 15:26:56 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Fri, Apr 3, 2015 at 3:26 PM, Michael Paquier wrote:\\n> [...]\\n> Fine for me.\\n\\nAnd here are the correct patches. Sorry for that.\\n-- \\nMichael\\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, 3 Apr 2015 15:28:33 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"Michael Paquier wrote:\\n> On Fri, Apr 3, 2015 at 3:26 PM, Michael Paquier wrote:\\n> > [...]\\n> > Fine for me.\\n> \\n> And here are the correct patches. Sorry for that.\\n\\nThanks, pushed. I added one extra comment to the SIGHUP patch in the\\nplace where you previously had the exit.\\n\\n-- \\n�lvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 3 Apr 2015 11:59:22 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n },\n {\n \"msg_contents\": \"On Fri, Apr 3, 2015 at 11:59 PM, Alvaro Herrera\\n<alvherre@2ndquadrant.com> wrote:\\n> Michael Paquier wrote:\\n>> On Fri, Apr 3, 2015 at 3:26 PM, Michael Paquier wrote:\\n>> > [...]\\n>> > Fine for me.\\n>>\\n>> And here are the correct patches. Sorry for that.\\n>\\n> Thanks, pushed. I added one extra comment to the SIGHUP patch in the\\n> place where you previously had the exit.\\n\\nThanks!\\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, 4 Apr 2015 06:35:39 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Table-level log_autovacuum_min_duration\"\n }\n]","isConversational":false}}},{"rowIdx":39985,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Change how first WAL segment on new timeline after promotion is created.\n\nTwo changes:\n\n1. When copying a WAL segment from old timeline to create the first segment\non the new timeline, only copy up to the point where the timeline switch\nhappens, and zero-fill the rest. This avoids corner cases where we might\nthink that the copied WAL from the previous timeline belong to the new\ntimeline.\n\n2. If the timeline switch happens at a segment boundary, don't copy the\nwhole old segment to the new timeline. It's pointless, because it's 100%\nidentical to the old segment.\n\nBranch\n------\nmaster\n\nDetails\n-------\nhttp://git.postgresql.org/pg/commitdiff/ba94518aad23beb800b657bd0cc8c4e7ea43ca33\n\nModified Files\n--------------\nsrc/backend/access/transam/xlog.c | 82 +++++++++++++++++++++++++++----------\n1 file changed, 61 insertions(+), 21 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":"Thu, 18 Dec 2014 18:25:19 +0000","msg_from":"Heikki Linnakangas <heikki.linnakangas@iki.fi>","msg_from_op":true,"msg_subject":"pgsql: Change how first WAL segment on new timeline after promotion\n is"},{"msg_contents":"Hi Heikki,\n\nWhile writing a test script for\nhttp://archives.postgresql.org/message-id/20141205002854.GE21964%40awork2.anarazel.de\nI noticed that this commit broke starting a pg_basebackup -X * without a\nrecovery.conf present. Which might not be the best idea, but imo is a\nperfectly valid thing to do.\n\nTo me the changes to StartupXLOG() in that commit look a bit bogus. The\nnew startLogSegNo is initialized to XLByteToSeg(EndOfLog)? Which points\nto the end of the record +1? Which thus isn't guaranteed to exist as a\nsegment (e.g. never if the last record was a XLOG_SWITCH). Did you\nperhaps intend to use XLogFileInit(use_existing = true) instead of\nXLogFileOpen()? That works for me.\n\nI've attached my preliminary testscript (note it's really not that\ninteresting at this point) that reliably reproduces the problem for me.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n\n-- \nSent via pgsql-committers mailing list (pgsql-committers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-committers","msg_date":"Sat, 3 Jan 2015 19:59:57 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: pgsql: Change how first WAL segment on new timeline\n after promotion is"},{"msg_contents":"On 01/03/2015 08:59 PM, Andres Freund wrote:\n> Hi Heikki,\n>\n> While writing a test script for\n> http://archives.postgresql.org/message-id/20141205002854.GE21964%40awork2.anarazel.de\n> I noticed that this commit broke starting a pg_basebackup -X * without a\n> recovery.conf present. Which might not be the best idea, but imo is a\n> perfectly valid thing to do.\n>\n> To me the changes to StartupXLOG() in that commit look a bit bogus. The\n> new startLogSegNo is initialized to XLByteToSeg(EndOfLog)? Which points\n> to the end of the record +1? Which thus isn't guaranteed to exist as a\n> segment (e.g. never if the last record was a XLOG_SWITCH).\n\nAh, good point.\n\n> Did you perhaps intend to use XLogFileInit(use_existing = true)\n> instead of XLogFileOpen()? That works for me.\n\nHmm, that doesn't sound right either. XLogFileInit is used when you \nswitch to a new segment, not to open an old segment for writing. It \nhappens to work, because with use_existing = true it will in fact always \nopen the old segment, instead of creating a new one, but I don't think \nthat's in the spirit of how that function's intended to be used.\n\nA very simple fix is to not try opening the segment at all. There isn't \nactually any requirement to have the segment open at that point, \nXLogWrite() will open it the first time the WAL is flushed. The WAL is \nflushed after writing the initial checkpoint or end-of-recovery record, \nwhich happens pretty soon anyway. Any objections to the attached?\n\n> I've attached my preliminary testscript (note it's really not that\n> interesting at this point) that reliably reproduces the problem for me.\n\nThanks!\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, 5 Jan 2015 18:45:22 +0200","msg_from":"Heikki Linnakangas <hlinnakangas@vmware.com>","msg_from_op":false,"msg_subject":"Re: [COMMITTERS] pgsql: Change how first WAL segment on new timeline\n after promotion is"},{"msg_contents":"On 2015-01-05 18:45:22 +0200, Heikki Linnakangas wrote:\n> On 01/03/2015 08:59 PM, Andres Freund wrote:\n> >Hi Heikki,\n> >\n> >While writing a test script for\n> >http://archives.postgresql.org/message-id/20141205002854.GE21964%40awork2.anarazel.de\n> >I noticed that this commit broke starting a pg_basebackup -X * without a\n> >recovery.conf present. Which might not be the best idea, but imo is a\n> >perfectly valid thing to do.\n> >\n> >To me the changes to StartupXLOG() in that commit look a bit bogus. The\n> >new startLogSegNo is initialized to XLByteToSeg(EndOfLog)? Which points\n> >to the end of the record +1? Which thus isn't guaranteed to exist as a\n> >segment (e.g. never if the last record was a XLOG_SWITCH).\n> \n> Ah, good point.\n> \n> >Did you perhaps intend to use XLogFileInit(use_existing = true)\n> >instead of XLogFileOpen()? That works for me.\n> \n> Hmm, that doesn't sound right either. XLogFileInit is used when you switch\n> to a new segment, not to open an old segment for writing. It happens to\n> work, because with use_existing = true it will in fact always open the old\n> segment, instead of creating a new one, but I don't think that's in the\n> spirit of how that function's intended to be used.\n\nWell, its docs say \"Create a new XLOG file segment, or open a\npre-existing one.\", so it's not that mismatched. We really don't know\nwhether the EndOfLog's segment already exist in this scenario. Also,\ndoesn't XLogWrite() essentially use it in the same way?\n\n> A very simple fix is to not try opening the segment at all. There isn't\n> actually any requirement to have the segment open at that point, XLogWrite()\n> will open it the first time the WAL is flushed. The WAL is flushed after\n> writing the initial checkpoint or end-of-recovery record, which happens\n> pretty soon anyway. Any objections to the attached?\n\nSounds like a good plan.\n\n> >I've attached my preliminary testscript (note it's really not that\n> >interesting at this point) that reliably reproduces the problem for me.\n> \n> Thanks!\n\nI've attached attached, for posterities sake, the last version of that\nscript. I think we should have at least some of these tests in the tap\ntests. I'd not used that framework because I wanted to test back to 9.1,\nbut ...\n\n> diff --git a/src/backend/access/transam/xlog.c b/src/backend/access/transam/xlog.c\n> index 5cc7e47..e623463 100644\n> --- a/src/backend/access/transam/xlog.c\n> +++ b/src/backend/access/transam/xlog.c\n> @@ -5646,7 +5646,6 @@ StartupXLOG(void)\n> \tXLogRecPtr\tRecPtr,\n> \t\t\t\tcheckPointLoc,\n> \t\t\t\tEndOfLog;\n> -\tXLogSegNo\tstartLogSegNo;\n> \tTimeLineID\tPrevTimeLineID;\n> \tXLogRecord *record;\n> \tTransactionId oldestActiveXID;\n> @@ -6633,7 +6632,6 @@ StartupXLOG(void)\n> \t */\n> \trecord = ReadRecord(xlogreader, LastRec, PANIC, false);\n> \tEndOfLog = EndRecPtr;\n> -\tXLByteToSeg(EndOfLog, startLogSegNo);\n> \n> \t/*\n> \t * Complain if we did not roll forward far enough to render the backup\n> @@ -6741,9 +6739,6 @@ StartupXLOG(void)\n> \t * buffer cache using the block containing the last record from the\n> \t * previous incarnation.\n> \t */\n> -\topenLogSegNo = startLogSegNo;\n> -\topenLogFile = XLogFileOpen(openLogSegNo);\n> -\topenLogOff = 0;\n> \tInsert = &XLogCtl->Insert;\n> \tInsert->PrevBytePos = XLogRecPtrToBytePos(LastRec);\n> \tInsert->CurrBytePos = XLogRecPtrToBytePos(EndOfLog);\n\nThe comment above could use some minor word smithing - the second part\nof it doesn't really seem to belong there.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 6 Jan 2015 14:22:57 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: [COMMITTERS] pgsql: Change how first WAL segment on new timeline\n after promotion is"},{"msg_contents":"On Tue, Jan 6, 2015 at 8:22 AM, Andres Freund <andres@2ndquadrant.com> wrote:\n> I've attached attached, for posterities sake, the last version of that\n> script. I think we should have at least some of these tests in the tap\n> tests. I'd not used that framework because I wanted to test back to 9.1,\n> but ...\n\nI don't see an attachment.\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 Jan 2015 13:46:57 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Change how first WAL segment on\n new timeline after promotion is"},{"msg_contents":"On 2015-01-06 13:46:57 -0500, Robert Haas wrote:\n> On Tue, Jan 6, 2015 at 8:22 AM, Andres Freund <andres@2ndquadrant.com> wrote:\n> > I've attached attached, for posterities sake, the last version of that\n> > script. I think we should have at least some of these tests in the tap\n> > tests. I'd not used that framework because I wanted to test back to 9.1,\n> > but ...\n> \n> I don't see an attachment.\n\nOops.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n\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, 6 Jan 2015 19:56:12 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Re: [COMMITTERS] pgsql: Change how first WAL segment\n on new timeline after promotion is"},{"msg_contents":"On 01/06/2015 03:22 PM, Andres Freund wrote:\n> On 2015-01-05 18:45:22 +0200, Heikki Linnakangas wrote:\n>> On 01/03/2015 08:59 PM, Andres Freund wrote:\n>>> Did you perhaps intend to use XLogFileInit(use_existing = true)\n>>> instead of XLogFileOpen()? That works for me.\n>>\n>> Hmm, that doesn't sound right either. XLogFileInit is used when you switch\n>> to a new segment, not to open an old segment for writing. It happens to\n>> work, because with use_existing = true it will in fact always open the old\n>> segment, instead of creating a new one, but I don't think that's in the\n>> spirit of how that function's intended to be used.\n>\n> Well, its docs say \"Create a new XLOG file segment, or open a\n> pre-existing one.\", so it's not that mismatched. We really don't know\n> whether the EndOfLog's segment already exist in this scenario. Also,\n> doesn't XLogWrite() essentially use it in the same way?\n\nXLogWrite() is smarter. It uses XLogFileInit() when switching to a new \nsegment, and XLogFileOpen when writing to the middle of a segment.\n\nCommitted the fix to not open the segment at all.\n\n- Heikki\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 7 Jan 2015 16:24:02 +0200","msg_from":"Heikki Linnakangas <hlinnakangas@vmware.com>","msg_from_op":false,"msg_subject":"Re: [COMMITTERS] pgsql: Change how first WAL segment on new timeline\n after promotion is"}],"string":"[\n {\n \"msg_contents\": \"Change how first WAL segment on new timeline after promotion is created.\\n\\nTwo changes:\\n\\n1. When copying a WAL segment from old timeline to create the first segment\\non the new timeline, only copy up to the point where the timeline switch\\nhappens, and zero-fill the rest. This avoids corner cases where we might\\nthink that the copied WAL from the previous timeline belong to the new\\ntimeline.\\n\\n2. If the timeline switch happens at a segment boundary, don't copy the\\nwhole old segment to the new timeline. It's pointless, because it's 100%\\nidentical to the old segment.\\n\\nBranch\\n------\\nmaster\\n\\nDetails\\n-------\\nhttp://git.postgresql.org/pg/commitdiff/ba94518aad23beb800b657bd0cc8c4e7ea43ca33\\n\\nModified Files\\n--------------\\nsrc/backend/access/transam/xlog.c | 82 +++++++++++++++++++++++++++----------\\n1 file changed, 61 insertions(+), 21 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\": \"Thu, 18 Dec 2014 18:25:19 +0000\",\n \"msg_from\": \"Heikki Linnakangas <heikki.linnakangas@iki.fi>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"pgsql: Change how first WAL segment on new timeline after promotion\\n is\"\n },\n {\n \"msg_contents\": \"Hi Heikki,\\n\\nWhile writing a test script for\\nhttp://archives.postgresql.org/message-id/20141205002854.GE21964%40awork2.anarazel.de\\nI noticed that this commit broke starting a pg_basebackup -X * without a\\nrecovery.conf present. Which might not be the best idea, but imo is a\\nperfectly valid thing to do.\\n\\nTo me the changes to StartupXLOG() in that commit look a bit bogus. The\\nnew startLogSegNo is initialized to XLByteToSeg(EndOfLog)? Which points\\nto the end of the record +1? Which thus isn't guaranteed to exist as a\\nsegment (e.g. never if the last record was a XLOG_SWITCH). Did you\\nperhaps intend to use XLogFileInit(use_existing = true) instead of\\nXLogFileOpen()? That works for me.\\n\\nI've attached my preliminary testscript (note it's really not that\\ninteresting at this point) that reliably reproduces the problem for me.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n\\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, 3 Jan 2015 19:59:57 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pgsql: Change how first WAL segment on new timeline\\n after promotion is\"\n },\n {\n \"msg_contents\": \"On 01/03/2015 08:59 PM, Andres Freund wrote:\\n> Hi Heikki,\\n>\\n> While writing a test script for\\n> http://archives.postgresql.org/message-id/20141205002854.GE21964%40awork2.anarazel.de\\n> I noticed that this commit broke starting a pg_basebackup -X * without a\\n> recovery.conf present. Which might not be the best idea, but imo is a\\n> perfectly valid thing to do.\\n>\\n> To me the changes to StartupXLOG() in that commit look a bit bogus. The\\n> new startLogSegNo is initialized to XLByteToSeg(EndOfLog)? Which points\\n> to the end of the record +1? Which thus isn't guaranteed to exist as a\\n> segment (e.g. never if the last record was a XLOG_SWITCH).\\n\\nAh, good point.\\n\\n> Did you perhaps intend to use XLogFileInit(use_existing = true)\\n> instead of XLogFileOpen()? That works for me.\\n\\nHmm, that doesn't sound right either. XLogFileInit is used when you \\nswitch to a new segment, not to open an old segment for writing. It \\nhappens to work, because with use_existing = true it will in fact always \\nopen the old segment, instead of creating a new one, but I don't think \\nthat's in the spirit of how that function's intended to be used.\\n\\nA very simple fix is to not try opening the segment at all. There isn't \\nactually any requirement to have the segment open at that point, \\nXLogWrite() will open it the first time the WAL is flushed. The WAL is \\nflushed after writing the initial checkpoint or end-of-recovery record, \\nwhich happens pretty soon anyway. Any objections to the attached?\\n\\n> I've attached my preliminary testscript (note it's really not that\\n> interesting at this point) that reliably reproduces the problem for me.\\n\\nThanks!\\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, 5 Jan 2015 18:45:22 +0200\",\n \"msg_from\": \"Heikki Linnakangas <hlinnakangas@vmware.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [COMMITTERS] pgsql: Change how first WAL segment on new timeline\\n after promotion is\"\n },\n {\n \"msg_contents\": \"On 2015-01-05 18:45:22 +0200, Heikki Linnakangas wrote:\\n> On 01/03/2015 08:59 PM, Andres Freund wrote:\\n> >Hi Heikki,\\n> >\\n> >While writing a test script for\\n> >http://archives.postgresql.org/message-id/20141205002854.GE21964%40awork2.anarazel.de\\n> >I noticed that this commit broke starting a pg_basebackup -X * without a\\n> >recovery.conf present. Which might not be the best idea, but imo is a\\n> >perfectly valid thing to do.\\n> >\\n> >To me the changes to StartupXLOG() in that commit look a bit bogus. The\\n> >new startLogSegNo is initialized to XLByteToSeg(EndOfLog)? Which points\\n> >to the end of the record +1? Which thus isn't guaranteed to exist as a\\n> >segment (e.g. never if the last record was a XLOG_SWITCH).\\n> \\n> Ah, good point.\\n> \\n> >Did you perhaps intend to use XLogFileInit(use_existing = true)\\n> >instead of XLogFileOpen()? That works for me.\\n> \\n> Hmm, that doesn't sound right either. XLogFileInit is used when you switch\\n> to a new segment, not to open an old segment for writing. It happens to\\n> work, because with use_existing = true it will in fact always open the old\\n> segment, instead of creating a new one, but I don't think that's in the\\n> spirit of how that function's intended to be used.\\n\\nWell, its docs say \\\"Create a new XLOG file segment, or open a\\npre-existing one.\\\", so it's not that mismatched. We really don't know\\nwhether the EndOfLog's segment already exist in this scenario. Also,\\ndoesn't XLogWrite() essentially use it in the same way?\\n\\n> A very simple fix is to not try opening the segment at all. There isn't\\n> actually any requirement to have the segment open at that point, XLogWrite()\\n> will open it the first time the WAL is flushed. The WAL is flushed after\\n> writing the initial checkpoint or end-of-recovery record, which happens\\n> pretty soon anyway. Any objections to the attached?\\n\\nSounds like a good plan.\\n\\n> >I've attached my preliminary testscript (note it's really not that\\n> >interesting at this point) that reliably reproduces the problem for me.\\n> \\n> Thanks!\\n\\nI've attached attached, for posterities sake, the last version of that\\nscript. I think we should have at least some of these tests in the tap\\ntests. I'd not used that framework because I wanted to test back to 9.1,\\nbut ...\\n\\n> diff --git a/src/backend/access/transam/xlog.c b/src/backend/access/transam/xlog.c\\n> index 5cc7e47..e623463 100644\\n> --- a/src/backend/access/transam/xlog.c\\n> +++ b/src/backend/access/transam/xlog.c\\n> @@ -5646,7 +5646,6 @@ StartupXLOG(void)\\n> \\tXLogRecPtr\\tRecPtr,\\n> \\t\\t\\t\\tcheckPointLoc,\\n> \\t\\t\\t\\tEndOfLog;\\n> -\\tXLogSegNo\\tstartLogSegNo;\\n> \\tTimeLineID\\tPrevTimeLineID;\\n> \\tXLogRecord *record;\\n> \\tTransactionId oldestActiveXID;\\n> @@ -6633,7 +6632,6 @@ StartupXLOG(void)\\n> \\t */\\n> \\trecord = ReadRecord(xlogreader, LastRec, PANIC, false);\\n> \\tEndOfLog = EndRecPtr;\\n> -\\tXLByteToSeg(EndOfLog, startLogSegNo);\\n> \\n> \\t/*\\n> \\t * Complain if we did not roll forward far enough to render the backup\\n> @@ -6741,9 +6739,6 @@ StartupXLOG(void)\\n> \\t * buffer cache using the block containing the last record from the\\n> \\t * previous incarnation.\\n> \\t */\\n> -\\topenLogSegNo = startLogSegNo;\\n> -\\topenLogFile = XLogFileOpen(openLogSegNo);\\n> -\\topenLogOff = 0;\\n> \\tInsert = &XLogCtl->Insert;\\n> \\tInsert->PrevBytePos = XLogRecPtrToBytePos(LastRec);\\n> \\tInsert->CurrBytePos = XLogRecPtrToBytePos(EndOfLog);\\n\\nThe comment above could use some minor word smithing - the second part\\nof it doesn't really seem to belong there.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 6 Jan 2015 14:22:57 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [COMMITTERS] pgsql: Change how first WAL segment on new timeline\\n after promotion is\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 6, 2015 at 8:22 AM, Andres Freund <andres@2ndquadrant.com> wrote:\\n> I've attached attached, for posterities sake, the last version of that\\n> script. I think we should have at least some of these tests in the tap\\n> tests. I'd not used that framework because I wanted to test back to 9.1,\\n> but ...\\n\\nI don't see an attachment.\\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 Jan 2015 13:46:57 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Change how first WAL segment on\\n new timeline after promotion is\"\n },\n {\n \"msg_contents\": \"On 2015-01-06 13:46:57 -0500, Robert Haas wrote:\\n> On Tue, Jan 6, 2015 at 8:22 AM, Andres Freund <andres@2ndquadrant.com> wrote:\\n> > I've attached attached, for posterities sake, the last version of that\\n> > script. I think we should have at least some of these tests in the tap\\n> > tests. I'd not used that framework because I wanted to test back to 9.1,\\n> > but ...\\n> \\n> I don't see an attachment.\\n\\nOops.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n\\n-- \\nSent via 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 Jan 2015 19:56:12 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: [COMMITTERS] pgsql: Change how first WAL segment\\n on new timeline after promotion is\"\n },\n {\n \"msg_contents\": \"On 01/06/2015 03:22 PM, Andres Freund wrote:\\n> On 2015-01-05 18:45:22 +0200, Heikki Linnakangas wrote:\\n>> On 01/03/2015 08:59 PM, Andres Freund wrote:\\n>>> Did you perhaps intend to use XLogFileInit(use_existing = true)\\n>>> instead of XLogFileOpen()? That works for me.\\n>>\\n>> Hmm, that doesn't sound right either. XLogFileInit is used when you switch\\n>> to a new segment, not to open an old segment for writing. It happens to\\n>> work, because with use_existing = true it will in fact always open the old\\n>> segment, instead of creating a new one, but I don't think that's in the\\n>> spirit of how that function's intended to be used.\\n>\\n> Well, its docs say \\\"Create a new XLOG file segment, or open a\\n> pre-existing one.\\\", so it's not that mismatched. We really don't know\\n> whether the EndOfLog's segment already exist in this scenario. Also,\\n> doesn't XLogWrite() essentially use it in the same way?\\n\\nXLogWrite() is smarter. It uses XLogFileInit() when switching to a new \\nsegment, and XLogFileOpen when writing to the middle of a segment.\\n\\nCommitted the fix to not open the segment at all.\\n\\n- Heikki\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 7 Jan 2015 16:24:02 +0200\",\n \"msg_from\": \"Heikki Linnakangas <hlinnakangas@vmware.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [COMMITTERS] pgsql: Change how first WAL segment on new timeline\\n after promotion is\"\n }\n]","isConversational":false}}},{"rowIdx":39986,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"One of our servers is currently running on postgres 9.2 using the 9.2.9-1.pgdg70+1 packages from pgdg.\n\nAfter an apt update this morning which brought in the libpq5 package version 9.4.0-1.pgdg70+1, connections to the database started failing with SSL errors logged on the server:\n\n [unknown] [unknown] LOG: could not accept SSL connection: digest too big for rsa key\n\nRolling back the server and client to libpq5 version 9.3.5-2.pgdg70+1 fixed it.\n\nThis is running on an otherwise up-to-date Debian Wheezy. The SSL certificate is locally issued using an internal CA which has been added to the local trust store. SSL-related config options are left set to the defaults.\n\n-- \nChris Butler\nZedcore Systems Ltd\n\nTelephone: 0114 303 0666\nDirect dial: 0114 303 0572\n\n\n-- \nSent via pgsql-pkg-debian mailing list (pgsql-pkg-debian@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-pkg-debian\n","msg_date":"Fri, 19 Dec 2014 07:49:36 +0000 (GMT)","msg_from":"Chris Butler <cbutler@zedcore.com>","msg_from_op":true,"msg_subject":"Updated libpq5 packages cause connection errors on postgresql 9.2"},{"msg_contents":"Re: Chris Butler 2014-12-19 <1155204201.65430.1418975376728.JavaMail.zimbra@zedcore.com>\n> One of our servers is currently running on postgres 9.2 using the 9.2.9-1.pgdg70+1 packages from pgdg.\n> \n> After an apt update this morning which brought in the libpq5 package version 9.4.0-1.pgdg70+1, connections to the database started failing with SSL errors logged on the server:\n> \n> [unknown] [unknown] LOG: could not accept SSL connection: digest too big for rsa key\n> \n> Rolling back the server and client to libpq5 version 9.3.5-2.pgdg70+1 fixed it.\n> \n> This is running on an otherwise up-to-date Debian Wheezy. The SSL certificate is locally issued using an internal CA which has been added to the local trust store. SSL-related config options are left set to the defaults.\n\nHi Chris,\n\nthanks for the report.\n\nGoogling for \"digest too big for rsa key\" seems to indicate that this\nproblem occurs when you are using (client?) certificates with short\nRSA keys. 512 bits is most often cited in the problem reports,\nsomething like 768 is around the minimum size that works, and of\ncourse, anything smaller than 1024 or really 1536 (or 2048) bits is\ntoo small for today's crypto standards.\n\nSo the question here is if this is also the problem you saw - are you\nusing client or server certificates with short keys?\n\nWhat this explanation doesn't explain is why the problem occurs with\n9.4's libpq5 while it works with 9.3's. The libssl version used for\nbuilding these packages should really be the same, 9.3.5-2.pgdg70+1\nwas built just two days ago as well.\n\nI'm CCing -hackers, maybe someone there has an idea.\n\nChristoph\n-- \ncb@df7cb.de | http://www.df7cb.de/\n\n\n-- \nSent via pgsql-pkg-debian mailing list (pgsql-pkg-debian@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-pkg-debian\n","msg_date":"Fri, 19 Dec 2014 11:52:43 +0100","msg_from":"Christoph Berg <cb@df7cb.de>","msg_from_op":false,"msg_subject":"Re: Updated libpq5 packages cause connection errors on postgresql 9.2"},{"msg_contents":"Hi Christoph,\n\n----- Original Message -----\n> From: \"Christoph Berg\" <cb@df7cb.de>\n> To: \"Chris Butler\" <cbutler@zedcore.com>\n>\n> Googling for \"digest too big for rsa key\" seems to indicate that this\n> problem occurs when you are using (client?) certificates with short\n> RSA keys. 512 bits is most often cited in the problem reports,\n> something like 768 is around the minimum size that works, and of\n> course, anything smaller than 1024 or really 1536 (or 2048) bits is\n> too small for today's crypto standards.\n> \n> So the question here is if this is also the problem you saw - are you\n> using client or server certificates with short keys?\n\nYes, that would appear to be the case - the key we're using is only 512 bits. I'll make sure we replace the certificate before re-applying the update (which will probably be after the holidays now).\n \n> What this explanation doesn't explain is why the problem occurs with\n> 9.4's libpq5 while it works with 9.3's. The libssl version used for\n> building these packages should really be the same, 9.3.5-2.pgdg70+1\n> was built just two days ago as well.\n\nFor info, I can confirm that both libraries are loading the same libssl:\n\nzedcore@web2:/tmp/usr/lib/x86_64-linux-gnu$ ldd /usr/lib/x86_64-linux-gnu/libpq.so.5 | grep libssl\n\tlibssl.so.1.0.0 => /usr/lib/x86_64-linux-gnu/libssl.so.1.0.0 (0x00007f3e8d898000)\nzedcore@web2:/tmp/usr/lib/x86_64-linux-gnu$ ldd ./libpq.so.5 | grep libssl\n\tlibssl.so.1.0.0 => /usr/lib/x86_64-linux-gnu/libssl.so.1.0.0 (0x00007f5d76176000)\n\n\nI can see a few changes are listed in the 9.4 changelog relating to SSL, so my guess would be one of those changes has altered the behaviour of libssl when presented with a small key.\n\n-- \nChris Butler\nZedcore Systems Ltd\n\nTelephone: 0114 303 0666\nDirect dial: 0114 303 0572\n\n\n-- \nSent via pgsql-pkg-debian mailing list (pgsql-pkg-debian@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-pkg-debian\n","msg_date":"Fri, 19 Dec 2014 11:16:14 +0000 (GMT)","msg_from":"Chris Butler <cbutler@zedcore.com>","msg_from_op":true,"msg_subject":"Re: Updated libpq5 packages cause connection errors on postgresql 9.2"},{"msg_contents":"On Fri, Dec 19, 2014 at 11:52 AM, Christoph Berg <cb@df7cb.de> wrote:\n>\n> Re: Chris Butler 2014-12-19 <\n> 1155204201.65430.1418975376728.JavaMail.zimbra@zedcore.com>\n> > One of our servers is currently running on postgres 9.2 using the\n> 9.2.9-1.pgdg70+1 packages from pgdg.\n> >\n> > After an apt update this morning which brought in the libpq5 package\n> version 9.4.0-1.pgdg70+1, connections to the database started failing with\n> SSL errors logged on the server:\n> >\n> > [unknown] [unknown] LOG: could not accept SSL connection: digest too\n> big for rsa key\n> >\n> > Rolling back the server and client to libpq5 version 9.3.5-2.pgdg70+1\n> fixed it.\n> >\n> > This is running on an otherwise up-to-date Debian Wheezy. The SSL\n> certificate is locally issued using an internal CA which has been added to\n> the local trust store. SSL-related config options are left set to the\n> defaults.\n>\n> Hi Chris,\n>\n> thanks for the report.\n>\n> Googling for \"digest too big for rsa key\" seems to indicate that this\n> problem occurs when you are using (client?) certificates with short\n> RSA keys. 512 bits is most often cited in the problem reports,\n> something like 768 is around the minimum size that works, and of\n> course, anything smaller than 1024 or really 1536 (or 2048) bits is\n> too small for today's crypto standards.\n>\n> So the question here is if this is also the problem you saw - are you\n> using client or server certificates with short keys?\n>\n> What this explanation doesn't explain is why the problem occurs with\n> 9.4's libpq5 while it works with 9.3's. The libssl version used for\n> building these packages should really be the same, 9.3.5-2.pgdg70+1\n> was built just two days ago as well.\n>\n> I'm CCing -hackers, maybe someone there has an idea.\n>\n\nSome googling shows that this could be because it's negotiating TLS 1.2\nwhich the key is just too small for. And we did change that in 9.4 - commit\n326e1d73c476a0b5061ef00134bdf57aed70d5e7 disabled SSL in favor of always\nusing TLS for security reasons.\n\n-- \n Magnus Hagander\n Me: http://www.hagander.net/\n Work: http://www.redpill-linpro.com/\n\nOn Fri, Dec 19, 2014 at 11:52 AM, Christoph Berg <cb@df7cb.de> wrote:Re: Chris Butler 2014-12-19 <1155204201.65430.1418975376728.JavaMail.zimbra@zedcore.com>\n> One of our servers is currently running on postgres 9.2 using the 9.2.9-1.pgdg70+1 packages from pgdg.\n>\n> After an apt update this morning which brought in the libpq5 package version 9.4.0-1.pgdg70+1, connections to the database started failing with SSL errors logged on the server:\n>\n> [unknown] [unknown] LOG: could not accept SSL connection: digest too big for rsa key\n>\n> Rolling back the server and client to libpq5 version 9.3.5-2.pgdg70+1 fixed it.\n>\n> This is running on an otherwise up-to-date Debian Wheezy. The SSL certificate is locally issued using an internal CA which has been added to the local trust store. SSL-related config options are left set to the defaults.\n\nHi Chris,\n\nthanks for the report.\n\nGoogling for \"digest too big for rsa key\" seems to indicate that this\nproblem occurs when you are using (client?) certificates with short\nRSA keys. 512 bits is most often cited in the problem reports,\nsomething like 768 is around the minimum size that works, and of\ncourse, anything smaller than 1024 or really 1536 (or 2048) bits is\ntoo small for today's crypto standards.\n\nSo the question here is if this is also the problem you saw - are you\nusing client or server certificates with short keys?\n\nWhat this explanation doesn't explain is why the problem occurs with\n9.4's libpq5 while it works with 9.3's. The libssl version used for\nbuilding these packages should really be the same, 9.3.5-2.pgdg70+1\nwas built just two days ago as well.\n\nI'm CCing -hackers, maybe someone there has an idea.Some googling shows that this could be because it's negotiating TLS 1.2 which the key is just too small for. And we did change that in 9.4 - commit 326e1d73c476a0b5061ef00134bdf57aed70d5e7 disabled SSL in favor of always using TLS for security reasons. -- Magnus Hagander Me: http://www.hagander.net/ Work: http://www.redpill-linpro.com/","msg_date":"Fri, 19 Dec 2014 13:08:18 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":false,"msg_subject":"Re: Updated libpq5 packages cause connection errors on postgresql 9.2"},{"msg_contents":"Magnus Hagander <magnus@hagander.net> writes:\n> On Fri, Dec 19, 2014 at 11:52 AM, Christoph Berg <cb@df7cb.de> wrote:\n>> Googling for \"digest too big for rsa key\" seems to indicate that this\n>> problem occurs when you are using (client?) certificates with short\n>> RSA keys. 512 bits is most often cited in the problem reports,\n>> something like 768 is around the minimum size that works, and of\n>> course, anything smaller than 1024 or really 1536 (or 2048) bits is\n>> too small for today's crypto standards.\n>> \n>> So the question here is if this is also the problem you saw - are you\n>> using client or server certificates with short keys?\n>> \n>> What this explanation doesn't explain is why the problem occurs with\n>> 9.4's libpq5 while it works with 9.3's. The libssl version used for\n>> building these packages should really be the same, 9.3.5-2.pgdg70+1\n>> was built just two days ago as well.\n\n> Some googling shows that this could be because it's negotiating TLS 1.2\n> which the key is just too small for. And we did change that in 9.4 - commit\n> 326e1d73c476a0b5061ef00134bdf57aed70d5e7 disabled SSL in favor of always\n> using TLS for security reasons.\n\nHm ... the 9.4 release notes fail to describe that change adequately, and\ncertainly don't mention that it would have any compatibility implications.\nGuess that needs to be fixed. Does anyone know offhand what the change in\nthe minimum key length is across SSL/TLS versions, exactly?\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-pkg-debian mailing list (pgsql-pkg-debian@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-pkg-debian\n","msg_date":"Fri, 19 Dec 2014 09:57:55 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [HACKERS] Re: Updated libpq5 packages cause connection errors on\n postgresql 9.2"},{"msg_contents":"On Fri, Dec 19, 2014 at 3:57 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> Magnus Hagander <magnus@hagander.net> writes:\n> > On Fri, Dec 19, 2014 at 11:52 AM, Christoph Berg <cb@df7cb.de> wrote:\n> >> Googling for \"digest too big for rsa key\" seems to indicate that this\n> >> problem occurs when you are using (client?) certificates with short\n> >> RSA keys. 512 bits is most often cited in the problem reports,\n> >> something like 768 is around the minimum size that works, and of\n> >> course, anything smaller than 1024 or really 1536 (or 2048) bits is\n> >> too small for today's crypto standards.\n> >>\n> >> So the question here is if this is also the problem you saw - are you\n> >> using client or server certificates with short keys?\n> >>\n> >> What this explanation doesn't explain is why the problem occurs with\n> >> 9.4's libpq5 while it works with 9.3's. The libssl version used for\n> >> building these packages should really be the same, 9.3.5-2.pgdg70+1\n> >> was built just two days ago as well.\n>\n> > Some googling shows that this could be because it's negotiating TLS 1.2\n> > which the key is just too small for. And we did change that in 9.4 -\n> commit\n> > 326e1d73c476a0b5061ef00134bdf57aed70d5e7 disabled SSL in favor of always\n> > using TLS for security reasons.\n>\n> Hm ... the 9.4 release notes fail to describe that change adequately, and\n> certainly don't mention that it would have any compatibility implications.\n> Guess that needs to be fixed. Does anyone know offhand what the change in\n> the minimum key length is across SSL/TLS versions, exactly?\n>\n\nI haven't seen a specific number, it might depend on exactly which cipher\nis negotiated. See for example\nhttp://openssl.6102.n7.nabble.com/What-is-the-reason-for-error-quot-SSL-negotiation-failed-error-04075070-rsa-routines-RSA-sign-digest-td43953.html\n\nAll references I have foud say at least 2014 is safe and 512 is broken, but\nthere are points in betwee nthat apparently works in some cases only.\n\nI think if we say \"use 1024 bits or more\" we err on the safe side.\n\n-- \n Magnus Hagander\n Me: http://www.hagander.net/\n Work: http://www.redpill-linpro.com/\n\nOn Fri, Dec 19, 2014 at 3:57 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:Magnus Hagander <magnus@hagander.net> writes:\n> On Fri, Dec 19, 2014 at 11:52 AM, Christoph Berg <cb@df7cb.de> wrote:\n>> Googling for \"digest too big for rsa key\" seems to indicate that this\n>> problem occurs when you are using (client?) certificates with short\n>> RSA keys. 512 bits is most often cited in the problem reports,\n>> something like 768 is around the minimum size that works, and of\n>> course, anything smaller than 1024 or really 1536 (or 2048) bits is\n>> too small for today's crypto standards.\n>>\n>> So the question here is if this is also the problem you saw - are you\n>> using client or server certificates with short keys?\n>>\n>> What this explanation doesn't explain is why the problem occurs with\n>> 9.4's libpq5 while it works with 9.3's. The libssl version used for\n>> building these packages should really be the same, 9.3.5-2.pgdg70+1\n>> was built just two days ago as well.\n\n> Some googling shows that this could be because it's negotiating TLS 1.2\n> which the key is just too small for. And we did change that in 9.4 - commit\n> 326e1d73c476a0b5061ef00134bdf57aed70d5e7 disabled SSL in favor of always\n> using TLS for security reasons.\n\nHm ... the 9.4 release notes fail to describe that change adequately, and\ncertainly don't mention that it would have any compatibility implications.\nGuess that needs to be fixed. Does anyone know offhand what the change in\nthe minimum key length is across SSL/TLS versions, exactly?I haven't seen a specific number, it might depend on exactly which cipher is negotiated. See for example http://openssl.6102.n7.nabble.com/What-is-the-reason-for-error-quot-SSL-negotiation-failed-error-04075070-rsa-routines-RSA-sign-digest-td43953.htmlAll references I have foud say at least 2014 is safe and 512 is broken, but there are points in betwee nthat apparently works in some cases only.I think if we say \"use 1024 bits or more\" we err on the safe side. -- Magnus Hagander Me: http://www.hagander.net/ Work: http://www.redpill-linpro.com/","msg_date":"Sat, 20 Dec 2014 12:27:05 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":false,"msg_subject":"Re: [HACKERS] Re: Updated libpq5 packages cause\n connection errors on postgresql 9.2"},{"msg_contents":"On Sat, Dec 20, 2014 at 12:27:05PM +0100, Magnus Hagander wrote:\n> I haven't seen a specific number, it might depend on exactly which cipher is\n> negotiated. See for example http://openssl.6102.n7.nabble.com/\n> What-is-the-reason-for-error-quot-SSL-negotiation-failed-error-04075070-rsa-routines-RSA-sign-digest-td43953.html\n> \n> All references I have foud say at least 2014 is safe and 512 is broken, but\n> there are points in betwee nthat apparently works in some cases only.\n> \n> I think if we say \"use 1024 bits or more\" we err on the safe side.�\n\nDid we ever decide on this?\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n + Everyone has their own god. +\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 1 Apr 2015 12:09:07 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Re: [pgsql-pkg-debian] Updated libpq5 packages cause\n connection errors on postgresql 9.2"},{"msg_contents":"Re: Bruce Momjian 2015-04-01 <20150401160907.GJ4466@momjian.us>\n> On Sat, Dec 20, 2014 at 12:27:05PM +0100, Magnus Hagander wrote:\n> > I haven't seen a specific number, it might depend on exactly which cipher is\n> > negotiated. See for example http://openssl.6102.n7.nabble.com/\n> > What-is-the-reason-for-error-quot-SSL-negotiation-failed-error-04075070-rsa-routines-RSA-sign-digest-td43953.html\n> > \n> > All references I have foud say at least 2014 is safe and 512 is broken, but\n> > there are points in betwee nthat apparently works in some cases only.\n> > \n> > I think if we say \"use 1024 bits or more\" we err on the safe side. \n> \n> Did we ever decide on this?\n\nThe question seems to be if we want to recommend \"1024 or more\" or\nsomething more sophisticated like \"use something between 512 and 1024\nbut ymmv .... 1024 should work in any case\". Given that more bits\nshould be more secure, and 1024 shouldn't pose a performance problem\nfor anyone, going for the short version shouldn't do any harm.\n\nChristoph\n-- \ncb@df7cb.de | http://www.df7cb.de/","msg_date":"Wed, 1 Apr 2015 22:11:52 +0200","msg_from":"Christoph Berg <cb@df7cb.de>","msg_from_op":false,"msg_subject":"Re: Re: [pgsql-pkg-debian] Updated libpq5 packages cause\n connection errors on postgresql 9.2"}],"string":"[\n {\n \"msg_contents\": \"One of our servers is currently running on postgres 9.2 using the 9.2.9-1.pgdg70+1 packages from pgdg.\\n\\nAfter an apt update this morning which brought in the libpq5 package version 9.4.0-1.pgdg70+1, connections to the database started failing with SSL errors logged on the server:\\n\\n [unknown] [unknown] LOG: could not accept SSL connection: digest too big for rsa key\\n\\nRolling back the server and client to libpq5 version 9.3.5-2.pgdg70+1 fixed it.\\n\\nThis is running on an otherwise up-to-date Debian Wheezy. The SSL certificate is locally issued using an internal CA which has been added to the local trust store. SSL-related config options are left set to the defaults.\\n\\n-- \\nChris Butler\\nZedcore Systems Ltd\\n\\nTelephone: 0114 303 0666\\nDirect dial: 0114 303 0572\\n\\n\\n-- \\nSent via pgsql-pkg-debian mailing list (pgsql-pkg-debian@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-pkg-debian\\n\",\n \"msg_date\": \"Fri, 19 Dec 2014 07:49:36 +0000 (GMT)\",\n \"msg_from\": \"Chris Butler <cbutler@zedcore.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Updated libpq5 packages cause connection errors on postgresql 9.2\"\n },\n {\n \"msg_contents\": \"Re: Chris Butler 2014-12-19 <1155204201.65430.1418975376728.JavaMail.zimbra@zedcore.com>\\n> One of our servers is currently running on postgres 9.2 using the 9.2.9-1.pgdg70+1 packages from pgdg.\\n> \\n> After an apt update this morning which brought in the libpq5 package version 9.4.0-1.pgdg70+1, connections to the database started failing with SSL errors logged on the server:\\n> \\n> [unknown] [unknown] LOG: could not accept SSL connection: digest too big for rsa key\\n> \\n> Rolling back the server and client to libpq5 version 9.3.5-2.pgdg70+1 fixed it.\\n> \\n> This is running on an otherwise up-to-date Debian Wheezy. The SSL certificate is locally issued using an internal CA which has been added to the local trust store. SSL-related config options are left set to the defaults.\\n\\nHi Chris,\\n\\nthanks for the report.\\n\\nGoogling for \\\"digest too big for rsa key\\\" seems to indicate that this\\nproblem occurs when you are using (client?) certificates with short\\nRSA keys. 512 bits is most often cited in the problem reports,\\nsomething like 768 is around the minimum size that works, and of\\ncourse, anything smaller than 1024 or really 1536 (or 2048) bits is\\ntoo small for today's crypto standards.\\n\\nSo the question here is if this is also the problem you saw - are you\\nusing client or server certificates with short keys?\\n\\nWhat this explanation doesn't explain is why the problem occurs with\\n9.4's libpq5 while it works with 9.3's. The libssl version used for\\nbuilding these packages should really be the same, 9.3.5-2.pgdg70+1\\nwas built just two days ago as well.\\n\\nI'm CCing -hackers, maybe someone there has an idea.\\n\\nChristoph\\n-- \\ncb@df7cb.de | http://www.df7cb.de/\\n\\n\\n-- \\nSent via pgsql-pkg-debian mailing list (pgsql-pkg-debian@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-pkg-debian\\n\",\n \"msg_date\": \"Fri, 19 Dec 2014 11:52:43 +0100\",\n \"msg_from\": \"Christoph Berg <cb@df7cb.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Updated libpq5 packages cause connection errors on postgresql 9.2\"\n },\n {\n \"msg_contents\": \"Hi Christoph,\\n\\n----- Original Message -----\\n> From: \\\"Christoph Berg\\\" <cb@df7cb.de>\\n> To: \\\"Chris Butler\\\" <cbutler@zedcore.com>\\n>\\n> Googling for \\\"digest too big for rsa key\\\" seems to indicate that this\\n> problem occurs when you are using (client?) certificates with short\\n> RSA keys. 512 bits is most often cited in the problem reports,\\n> something like 768 is around the minimum size that works, and of\\n> course, anything smaller than 1024 or really 1536 (or 2048) bits is\\n> too small for today's crypto standards.\\n> \\n> So the question here is if this is also the problem you saw - are you\\n> using client or server certificates with short keys?\\n\\nYes, that would appear to be the case - the key we're using is only 512 bits. I'll make sure we replace the certificate before re-applying the update (which will probably be after the holidays now).\\n \\n> What this explanation doesn't explain is why the problem occurs with\\n> 9.4's libpq5 while it works with 9.3's. The libssl version used for\\n> building these packages should really be the same, 9.3.5-2.pgdg70+1\\n> was built just two days ago as well.\\n\\nFor info, I can confirm that both libraries are loading the same libssl:\\n\\nzedcore@web2:/tmp/usr/lib/x86_64-linux-gnu$ ldd /usr/lib/x86_64-linux-gnu/libpq.so.5 | grep libssl\\n\\tlibssl.so.1.0.0 => /usr/lib/x86_64-linux-gnu/libssl.so.1.0.0 (0x00007f3e8d898000)\\nzedcore@web2:/tmp/usr/lib/x86_64-linux-gnu$ ldd ./libpq.so.5 | grep libssl\\n\\tlibssl.so.1.0.0 => /usr/lib/x86_64-linux-gnu/libssl.so.1.0.0 (0x00007f5d76176000)\\n\\n\\nI can see a few changes are listed in the 9.4 changelog relating to SSL, so my guess would be one of those changes has altered the behaviour of libssl when presented with a small key.\\n\\n-- \\nChris Butler\\nZedcore Systems Ltd\\n\\nTelephone: 0114 303 0666\\nDirect dial: 0114 303 0572\\n\\n\\n-- \\nSent via pgsql-pkg-debian mailing list (pgsql-pkg-debian@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-pkg-debian\\n\",\n \"msg_date\": \"Fri, 19 Dec 2014 11:16:14 +0000 (GMT)\",\n \"msg_from\": \"Chris Butler <cbutler@zedcore.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Updated libpq5 packages cause connection errors on postgresql 9.2\"\n },\n {\n \"msg_contents\": \"On Fri, Dec 19, 2014 at 11:52 AM, Christoph Berg <cb@df7cb.de> wrote:\\n>\\n> Re: Chris Butler 2014-12-19 <\\n> 1155204201.65430.1418975376728.JavaMail.zimbra@zedcore.com>\\n> > One of our servers is currently running on postgres 9.2 using the\\n> 9.2.9-1.pgdg70+1 packages from pgdg.\\n> >\\n> > After an apt update this morning which brought in the libpq5 package\\n> version 9.4.0-1.pgdg70+1, connections to the database started failing with\\n> SSL errors logged on the server:\\n> >\\n> > [unknown] [unknown] LOG: could not accept SSL connection: digest too\\n> big for rsa key\\n> >\\n> > Rolling back the server and client to libpq5 version 9.3.5-2.pgdg70+1\\n> fixed it.\\n> >\\n> > This is running on an otherwise up-to-date Debian Wheezy. The SSL\\n> certificate is locally issued using an internal CA which has been added to\\n> the local trust store. SSL-related config options are left set to the\\n> defaults.\\n>\\n> Hi Chris,\\n>\\n> thanks for the report.\\n>\\n> Googling for \\\"digest too big for rsa key\\\" seems to indicate that this\\n> problem occurs when you are using (client?) certificates with short\\n> RSA keys. 512 bits is most often cited in the problem reports,\\n> something like 768 is around the minimum size that works, and of\\n> course, anything smaller than 1024 or really 1536 (or 2048) bits is\\n> too small for today's crypto standards.\\n>\\n> So the question here is if this is also the problem you saw - are you\\n> using client or server certificates with short keys?\\n>\\n> What this explanation doesn't explain is why the problem occurs with\\n> 9.4's libpq5 while it works with 9.3's. The libssl version used for\\n> building these packages should really be the same, 9.3.5-2.pgdg70+1\\n> was built just two days ago as well.\\n>\\n> I'm CCing -hackers, maybe someone there has an idea.\\n>\\n\\nSome googling shows that this could be because it's negotiating TLS 1.2\\nwhich the key is just too small for. And we did change that in 9.4 - commit\\n326e1d73c476a0b5061ef00134bdf57aed70d5e7 disabled SSL in favor of always\\nusing TLS for security reasons.\\n\\n-- \\n Magnus Hagander\\n Me: http://www.hagander.net/\\n Work: http://www.redpill-linpro.com/\\n\\nOn Fri, Dec 19, 2014 at 11:52 AM, Christoph Berg <cb@df7cb.de> wrote:Re: Chris Butler 2014-12-19 <1155204201.65430.1418975376728.JavaMail.zimbra@zedcore.com>\\n> One of our servers is currently running on postgres 9.2 using the 9.2.9-1.pgdg70+1 packages from pgdg.\\n>\\n> After an apt update this morning which brought in the libpq5 package version 9.4.0-1.pgdg70+1, connections to the database started failing with SSL errors logged on the server:\\n>\\n> [unknown] [unknown] LOG: could not accept SSL connection: digest too big for rsa key\\n>\\n> Rolling back the server and client to libpq5 version 9.3.5-2.pgdg70+1 fixed it.\\n>\\n> This is running on an otherwise up-to-date Debian Wheezy. The SSL certificate is locally issued using an internal CA which has been added to the local trust store. SSL-related config options are left set to the defaults.\\n\\nHi Chris,\\n\\nthanks for the report.\\n\\nGoogling for \\\"digest too big for rsa key\\\" seems to indicate that this\\nproblem occurs when you are using (client?) certificates with short\\nRSA keys. 512 bits is most often cited in the problem reports,\\nsomething like 768 is around the minimum size that works, and of\\ncourse, anything smaller than 1024 or really 1536 (or 2048) bits is\\ntoo small for today's crypto standards.\\n\\nSo the question here is if this is also the problem you saw - are you\\nusing client or server certificates with short keys?\\n\\nWhat this explanation doesn't explain is why the problem occurs with\\n9.4's libpq5 while it works with 9.3's. The libssl version used for\\nbuilding these packages should really be the same, 9.3.5-2.pgdg70+1\\nwas built just two days ago as well.\\n\\nI'm CCing -hackers, maybe someone there has an idea.Some googling shows that this could be because it's negotiating TLS 1.2 which the key is just too small for. And we did change that in 9.4 - commit 326e1d73c476a0b5061ef00134bdf57aed70d5e7 disabled SSL in favor of always using TLS for security reasons. -- Magnus Hagander Me: http://www.hagander.net/ Work: http://www.redpill-linpro.com/\",\n \"msg_date\": \"Fri, 19 Dec 2014 13:08:18 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Updated libpq5 packages cause connection errors on postgresql 9.2\"\n },\n {\n \"msg_contents\": \"Magnus Hagander <magnus@hagander.net> writes:\\n> On Fri, Dec 19, 2014 at 11:52 AM, Christoph Berg <cb@df7cb.de> wrote:\\n>> Googling for \\\"digest too big for rsa key\\\" seems to indicate that this\\n>> problem occurs when you are using (client?) certificates with short\\n>> RSA keys. 512 bits is most often cited in the problem reports,\\n>> something like 768 is around the minimum size that works, and of\\n>> course, anything smaller than 1024 or really 1536 (or 2048) bits is\\n>> too small for today's crypto standards.\\n>> \\n>> So the question here is if this is also the problem you saw - are you\\n>> using client or server certificates with short keys?\\n>> \\n>> What this explanation doesn't explain is why the problem occurs with\\n>> 9.4's libpq5 while it works with 9.3's. The libssl version used for\\n>> building these packages should really be the same, 9.3.5-2.pgdg70+1\\n>> was built just two days ago as well.\\n\\n> Some googling shows that this could be because it's negotiating TLS 1.2\\n> which the key is just too small for. And we did change that in 9.4 - commit\\n> 326e1d73c476a0b5061ef00134bdf57aed70d5e7 disabled SSL in favor of always\\n> using TLS for security reasons.\\n\\nHm ... the 9.4 release notes fail to describe that change adequately, and\\ncertainly don't mention that it would have any compatibility implications.\\nGuess that needs to be fixed. Does anyone know offhand what the change in\\nthe minimum key length is across SSL/TLS versions, exactly?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-pkg-debian mailing list (pgsql-pkg-debian@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-pkg-debian\\n\",\n \"msg_date\": \"Fri, 19 Dec 2014 09:57:55 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [HACKERS] Re: Updated libpq5 packages cause connection errors on\\n postgresql 9.2\"\n },\n {\n \"msg_contents\": \"On Fri, Dec 19, 2014 at 3:57 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> Magnus Hagander <magnus@hagander.net> writes:\\n> > On Fri, Dec 19, 2014 at 11:52 AM, Christoph Berg <cb@df7cb.de> wrote:\\n> >> Googling for \\\"digest too big for rsa key\\\" seems to indicate that this\\n> >> problem occurs when you are using (client?) certificates with short\\n> >> RSA keys. 512 bits is most often cited in the problem reports,\\n> >> something like 768 is around the minimum size that works, and of\\n> >> course, anything smaller than 1024 or really 1536 (or 2048) bits is\\n> >> too small for today's crypto standards.\\n> >>\\n> >> So the question here is if this is also the problem you saw - are you\\n> >> using client or server certificates with short keys?\\n> >>\\n> >> What this explanation doesn't explain is why the problem occurs with\\n> >> 9.4's libpq5 while it works with 9.3's. The libssl version used for\\n> >> building these packages should really be the same, 9.3.5-2.pgdg70+1\\n> >> was built just two days ago as well.\\n>\\n> > Some googling shows that this could be because it's negotiating TLS 1.2\\n> > which the key is just too small for. And we did change that in 9.4 -\\n> commit\\n> > 326e1d73c476a0b5061ef00134bdf57aed70d5e7 disabled SSL in favor of always\\n> > using TLS for security reasons.\\n>\\n> Hm ... the 9.4 release notes fail to describe that change adequately, and\\n> certainly don't mention that it would have any compatibility implications.\\n> Guess that needs to be fixed. Does anyone know offhand what the change in\\n> the minimum key length is across SSL/TLS versions, exactly?\\n>\\n\\nI haven't seen a specific number, it might depend on exactly which cipher\\nis negotiated. See for example\\nhttp://openssl.6102.n7.nabble.com/What-is-the-reason-for-error-quot-SSL-negotiation-failed-error-04075070-rsa-routines-RSA-sign-digest-td43953.html\\n\\nAll references I have foud say at least 2014 is safe and 512 is broken, but\\nthere are points in betwee nthat apparently works in some cases only.\\n\\nI think if we say \\\"use 1024 bits or more\\\" we err on the safe side.\\n\\n-- \\n Magnus Hagander\\n Me: http://www.hagander.net/\\n Work: http://www.redpill-linpro.com/\\n\\nOn Fri, Dec 19, 2014 at 3:57 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:Magnus Hagander <magnus@hagander.net> writes:\\n> On Fri, Dec 19, 2014 at 11:52 AM, Christoph Berg <cb@df7cb.de> wrote:\\n>> Googling for \\\"digest too big for rsa key\\\" seems to indicate that this\\n>> problem occurs when you are using (client?) certificates with short\\n>> RSA keys. 512 bits is most often cited in the problem reports,\\n>> something like 768 is around the minimum size that works, and of\\n>> course, anything smaller than 1024 or really 1536 (or 2048) bits is\\n>> too small for today's crypto standards.\\n>>\\n>> So the question here is if this is also the problem you saw - are you\\n>> using client or server certificates with short keys?\\n>>\\n>> What this explanation doesn't explain is why the problem occurs with\\n>> 9.4's libpq5 while it works with 9.3's. The libssl version used for\\n>> building these packages should really be the same, 9.3.5-2.pgdg70+1\\n>> was built just two days ago as well.\\n\\n> Some googling shows that this could be because it's negotiating TLS 1.2\\n> which the key is just too small for. And we did change that in 9.4 - commit\\n> 326e1d73c476a0b5061ef00134bdf57aed70d5e7 disabled SSL in favor of always\\n> using TLS for security reasons.\\n\\nHm ... the 9.4 release notes fail to describe that change adequately, and\\ncertainly don't mention that it would have any compatibility implications.\\nGuess that needs to be fixed. Does anyone know offhand what the change in\\nthe minimum key length is across SSL/TLS versions, exactly?I haven't seen a specific number, it might depend on exactly which cipher is negotiated. See for example http://openssl.6102.n7.nabble.com/What-is-the-reason-for-error-quot-SSL-negotiation-failed-error-04075070-rsa-routines-RSA-sign-digest-td43953.htmlAll references I have foud say at least 2014 is safe and 512 is broken, but there are points in betwee nthat apparently works in some cases only.I think if we say \\\"use 1024 bits or more\\\" we err on the safe side. -- Magnus Hagander Me: http://www.hagander.net/ Work: http://www.redpill-linpro.com/\",\n \"msg_date\": \"Sat, 20 Dec 2014 12:27:05 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [HACKERS] Re: Updated libpq5 packages cause\\n connection errors on postgresql 9.2\"\n },\n {\n \"msg_contents\": \"On Sat, Dec 20, 2014 at 12:27:05PM +0100, Magnus Hagander wrote:\\n> I haven't seen a specific number, it might depend on exactly which cipher is\\n> negotiated. See for example http://openssl.6102.n7.nabble.com/\\n> What-is-the-reason-for-error-quot-SSL-negotiation-failed-error-04075070-rsa-routines-RSA-sign-digest-td43953.html\\n> \\n> All references I have foud say at least 2014 is safe and 512 is broken, but\\n> there are points in betwee nthat apparently works in some cases only.\\n> \\n> I think if we say \\\"use 1024 bits or more\\\" we err on the safe side.�\\n\\nDid we ever decide on this?\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n + Everyone has their own god. +\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 1 Apr 2015 12:09:07 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: [pgsql-pkg-debian] Updated libpq5 packages cause\\n connection errors on postgresql 9.2\"\n },\n {\n \"msg_contents\": \"Re: Bruce Momjian 2015-04-01 <20150401160907.GJ4466@momjian.us>\\n> On Sat, Dec 20, 2014 at 12:27:05PM +0100, Magnus Hagander wrote:\\n> > I haven't seen a specific number, it might depend on exactly which cipher is\\n> > negotiated. See for example http://openssl.6102.n7.nabble.com/\\n> > What-is-the-reason-for-error-quot-SSL-negotiation-failed-error-04075070-rsa-routines-RSA-sign-digest-td43953.html\\n> > \\n> > All references I have foud say at least 2014 is safe and 512 is broken, but\\n> > there are points in betwee nthat apparently works in some cases only.\\n> > \\n> > I think if we say \\\"use 1024 bits or more\\\" we err on the safe side. \\n> \\n> Did we ever decide on this?\\n\\nThe question seems to be if we want to recommend \\\"1024 or more\\\" or\\nsomething more sophisticated like \\\"use something between 512 and 1024\\nbut ymmv .... 1024 should work in any case\\\". Given that more bits\\nshould be more secure, and 1024 shouldn't pose a performance problem\\nfor anyone, going for the short version shouldn't do any harm.\\n\\nChristoph\\n-- \\ncb@df7cb.de | http://www.df7cb.de/\",\n \"msg_date\": \"Wed, 1 Apr 2015 22:11:52 +0200\",\n \"msg_from\": \"Christoph Berg <cb@df7cb.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: [pgsql-pkg-debian] Updated libpq5 packages cause\\n connection errors on postgresql 9.2\"\n }\n]","isConversational":false}}},{"rowIdx":39987,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"When streaming replication was introduced in 9.0, we started to recycle \nold WAL segments in archive recovery, like we do during normal \noperation. The WAL segments are recycled on the current timeline. There \nis no guarantee that they are useful, if the current timeline changes, \nbecause we step to recover another timeline after that, or the standby \nis promoted, but that was thought to be harmless.\n\nHowever, consider what happens after a server is promoted, and WAL \narchiving is enabled. The server's pg_xlog directory will look something \nlike this:\n\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:22 000000010000000000000005\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000006\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000007\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000008\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000009\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 00000001000000000000000A\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 00000001000000000000000B\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 00000001000000000000000C\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 00000001000000000000000D\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 00000001000000000000000E\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 00000001000000000000000F\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000010\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000011\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000012\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000013\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000014\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000015\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000016\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000017\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000018\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:24 000000010000000000000019\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:22 00000001000000000000001A\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:22 00000001000000000000001B\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:22 00000001000000000000001C\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:24 000000020000000000000019\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:24 00000002000000000000001A\n> -rw------- 1 heikki heikki 42 Dec 19 14:24 00000002.history\n\nThe files on timeline 1, up to 000000010000000000000019, are valid \nsegments, streamed from the primary or restored from the WAL archive. \nThe segments 00000001000000000000001A and 00000001000000000000001B are \nrecycled segments that haven't been reused yet. Their contents are not \nvalid (they contain records from some earlier point in WAL, but it might \nas well be garbage).\n\nThe server was promoted within the segment 19, and a new timeline was \nstarted. Segments 000000020000000000000019 and 00000002000000000000001A \ncontain valid WAL on the new timeline.\n\nNow, after enough time passes that the bogus 00000001000000000000001A \nand 00000001000000000000001B segments become old enough to be recycled, \nthe system will see that there is no .ready or .done file for them, and \nwill create .ready files so that they are archived. And they are \narchived. That's bogus, because the files are bogus. Worse, if the \nprimary server where this server was forked off from continues running, \nand creates the genuine 00000001000000000000001A and \n00000001000000000000001B segments, it can fail to archive them if the \nstandby had already archived the bogus segments with the same names.\n\nWe must somehow prevent the recycled, but not yet used, segments from \nbeing archived. One idea is to not create them in the first place, i.e. \ndon't recycle old segments during recovery, just delete them and have \nnew ones be created on demand. That's simple, but would hurt performance.\n\nI'm thinking that we should add a step to promotion, where we scan \npg_xlog for any segments higher than the timeline switch point, and \nremove them, or mark them with .done so that they are not archived. \nThere might be some real WAL that was streamed from the primary, but not \nyet applied, but such WAL is of no interest to that server anyway, after \nit's been promoted. It's a bit disconcerting to zap WAL that's valid, \neven if doesn't belong to the current server's timeline history, because \nas a general rule it's good to avoid destroying evidence that might be \nuseful in debugging. There isn't much difference between removing them \nimmediately and marking them as .done, though, because they will \neventually be removed/recycled anyway if they're marked as .done.\n\nThe archival behaviour at promotion is a bit inconsistent and weird \nanyway; even valid, streamed WAL is marked as .done and not archived \nanyway, except for the last partial segment. We're discussing that in \nthe other thread (Streaming replication and WAL archive interactions, \nhttp://www.postgresql.org/message-id/689EB259-44C2-4820-B901-4F6B1C55A1E4@simply.name), \nbut it would be good have a small, back-patchable fix to prevent bogus \nsegments from being archived.\n\n- Heikki\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 19 Dec 2014 14:55:16 +0200","msg_from":"Heikki Linnakangas <hlinnakangas@vmware.com>","msg_from_op":true,"msg_subject":"Bogus WAL segments archived after promotion"},{"msg_contents":"On 12/19/2014 02:55 PM, Heikki Linnakangas wrote:\n> I'm thinking that we should add a step to promotion, where we scan\n> pg_xlog for any segments higher than the timeline switch point, and\n> remove them, or mark them with .done so that they are not archived.\n> There might be some real WAL that was streamed from the primary, but not\n> yet applied, but such WAL is of no interest to that server anyway, after\n> it's been promoted. It's a bit disconcerting to zap WAL that's valid,\n> even if doesn't belong to the current server's timeline history, because\n> as a general rule it's good to avoid destroying evidence that might be\n> useful in debugging. There isn't much difference between removing them\n> immediately and marking them as .done, though, because they will\n> eventually be removed/recycled anyway if they're marked as .done.\n\nThis is what I came up with. This patch removes the suspect segments at \ntimeline switch. The alternative of creating .done files for them would \npreserve more evidence for debugging, but OTOH it would also be very \nconfusing to have valid-looking WAL segments in pg_xlog, with .done \nfiles, that in fact contain garbage.\n\nThe patch is a bit longer than it otherwise would be, because I moved \nthe code to remove a single file from RemoveOldXlogFiles() to a new \nfunction. I think that makes it more readable in any case, simply \nbecause it was so deeply indented in RemoveOldXlogFiles.\n\nThoughts?\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":"Fri, 19 Dec 2014 22:26:34 +0200","msg_from":"Heikki Linnakangas <hlinnakangas@vmware.com>","msg_from_op":true,"msg_subject":"Re: Bogus WAL segments archived after promotion"},{"msg_contents":"On Fri, Dec 19, 2014 at 10:26:34PM +0200, Heikki Linnakangas wrote:\n> On 12/19/2014 02:55 PM, Heikki Linnakangas wrote:\n> >I'm thinking that we should add a step to promotion, where we scan\n> >pg_xlog for any segments higher than the timeline switch point, and\n> >remove them, or mark them with .done so that they are not archived.\n> >There might be some real WAL that was streamed from the primary, but not\n> >yet applied, but such WAL is of no interest to that server anyway, after\n> >it's been promoted. It's a bit disconcerting to zap WAL that's valid,\n> >even if doesn't belong to the current server's timeline history, because\n> >as a general rule it's good to avoid destroying evidence that might be\n> >useful in debugging. There isn't much difference between removing them\n> >immediately and marking them as .done, though, because they will\n> >eventually be removed/recycled anyway if they're marked as .done.\n> \n> This is what I came up with. This patch removes the suspect segments\n> at timeline switch. The alternative of creating .done files for them\n> would preserve more evidence for debugging, but OTOH it would also\n> be very confusing to have valid-looking WAL segments in pg_xlog,\n> with .done files, that in fact contain garbage.\n> \n> The patch is a bit longer than it otherwise would be, because I\n> moved the code to remove a single file from RemoveOldXlogFiles() to\n> a new function. I think that makes it more readable in any case,\n> simply because it was so deeply indented in RemoveOldXlogFiles.\n\nWhere are we on this?\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n + Everyone has their own god. +\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 1 Apr 2015 12:12:57 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Bogus WAL segments archived after promotion"},{"msg_contents":"On 04/01/2015 07:12 PM, Bruce Momjian wrote:\n> On Fri, Dec 19, 2014 at 10:26:34PM +0200, Heikki Linnakangas wrote:\n>> On 12/19/2014 02:55 PM, Heikki Linnakangas wrote:\n>>> I'm thinking that we should add a step to promotion, where we scan\n>>> pg_xlog for any segments higher than the timeline switch point, and\n>>> remove them, or mark them with .done so that they are not archived.\n>>> There might be some real WAL that was streamed from the primary, but not\n>>> yet applied, but such WAL is of no interest to that server anyway, after\n>>> it's been promoted. It's a bit disconcerting to zap WAL that's valid,\n>>> even if doesn't belong to the current server's timeline history, because\n>>> as a general rule it's good to avoid destroying evidence that might be\n>>> useful in debugging. There isn't much difference between removing them\n>>> immediately and marking them as .done, though, because they will\n>>> eventually be removed/recycled anyway if they're marked as .done.\n>>\n>> This is what I came up with. This patch removes the suspect segments\n>> at timeline switch. The alternative of creating .done files for them\n>> would preserve more evidence for debugging, but OTOH it would also\n>> be very confusing to have valid-looking WAL segments in pg_xlog,\n>> with .done files, that in fact contain garbage.\n>>\n>> The patch is a bit longer than it otherwise would be, because I\n>> moved the code to remove a single file from RemoveOldXlogFiles() to\n>> a new function. I think that makes it more readable in any case,\n>> simply because it was so deeply indented in RemoveOldXlogFiles.\n>\n> Where are we on this?\n\nI didn't hear any better ideas, so committed this now.\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, 13 Apr 2015 17:57:31 +0300","msg_from":"Heikki Linnakangas <hlinnaka@iki.fi>","msg_from_op":false,"msg_subject":"Re: Bogus WAL segments archived after promotion"},{"msg_contents":"On Mon, Apr 13, 2015 at 11:57 PM, Heikki Linnakangas <hlinnaka@iki.fi> wrote:\n> On 04/01/2015 07:12 PM, Bruce Momjian wrote:\n>>\n>> On Fri, Dec 19, 2014 at 10:26:34PM +0200, Heikki Linnakangas wrote:\n>>>\n>>> On 12/19/2014 02:55 PM, Heikki Linnakangas wrote:\n>>>>\n>>>> I'm thinking that we should add a step to promotion, where we scan\n>>>> pg_xlog for any segments higher than the timeline switch point, and\n>>>> remove them, or mark them with .done so that they are not archived.\n>>>> There might be some real WAL that was streamed from the primary, but not\n>>>> yet applied, but such WAL is of no interest to that server anyway, after\n>>>> it's been promoted. It's a bit disconcerting to zap WAL that's valid,\n>>>> even if doesn't belong to the current server's timeline history, because\n>>>> as a general rule it's good to avoid destroying evidence that might be\n>>>> useful in debugging. There isn't much difference between removing them\n>>>> immediately and marking them as .done, though, because they will\n>>>> eventually be removed/recycled anyway if they're marked as .done.\n>>>\n>>>\n>>> This is what I came up with. This patch removes the suspect segments\n>>> at timeline switch. The alternative of creating .done files for them\n>>> would preserve more evidence for debugging, but OTOH it would also\n>>> be very confusing to have valid-looking WAL segments in pg_xlog,\n>>> with .done files, that in fact contain garbage.\n>>>\n>>> The patch is a bit longer than it otherwise would be, because I\n>>> moved the code to remove a single file from RemoveOldXlogFiles() to\n>>> a new function. I think that makes it more readable in any case,\n>>> simply because it was so deeply indented in RemoveOldXlogFiles.\n>>\n>>\n>> Where are we on this?\n>\n>\n> I didn't hear any better ideas, so committed this now.\n\nFinally looking at that... The commit log of b2a5545 is a bit\nmisleading. Segment files that were recycled during archive recovery\nare not necessarily removed, they could be recycled as well during\npromotion on the new timeline in line with what RemoveOldXlogFiles()\ndoes. Hence I think that the comment on top of\nRemoveNonParentXlogFiles() should be updated to reflect that like in\nthe patch attached.\n\nSomething minor: perhaps we could refactor xlogarchive.c to have\nXLogArchiveCheckDone() and XLogArchiveIsBusy() use the new\nXLogArchiveIsReady().\nRegards,\n-- \nMichael\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Thu, 23 Apr 2015 14:57:59 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: Bogus WAL segments archived after promotion"},{"msg_contents":"On Thu, Apr 23, 2015 at 02:57:59PM +0900, Michael Paquier wrote:\n> Finally looking at that... The commit log of b2a5545 is a bit\n> misleading. Segment files that were recycled during archive recovery\n> are not necessarily removed, they could be recycled as well during\n> promotion on the new timeline in line with what RemoveOldXlogFiles()\n> does. Hence I think that the comment on top of\n> RemoveNonParentXlogFiles() should be updated to reflect that like in\n> the patch attached.\n> \n> Something minor: perhaps we could refactor xlogarchive.c to have\n> XLogArchiveCheckDone() and XLogArchiveIsBusy() use the new\n> XLogArchiveIsReady().\n> Regards,\n\nOld patch, but still valid, so applied to master, thanks.\n\n-- \n Bruce Momjian <bruce@momjian.us> https://momjian.us\n EDB https://enterprisedb.com\n\n Only you can decide what is important to you.\n\n\n","msg_date":"Sat, 25 Nov 2023 10:48:53 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: [HACKERS] Bogus WAL segments archived after promotion"}],"string":"[\n {\n \"msg_contents\": \"When streaming replication was introduced in 9.0, we started to recycle \\nold WAL segments in archive recovery, like we do during normal \\noperation. The WAL segments are recycled on the current timeline. There \\nis no guarantee that they are useful, if the current timeline changes, \\nbecause we step to recover another timeline after that, or the standby \\nis promoted, but that was thought to be harmless.\\n\\nHowever, consider what happens after a server is promoted, and WAL \\narchiving is enabled. The server's pg_xlog directory will look something \\nlike this:\\n\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:22 000000010000000000000005\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000006\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000007\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000008\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000009\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 00000001000000000000000A\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 00000001000000000000000B\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 00000001000000000000000C\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 00000001000000000000000D\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 00000001000000000000000E\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 00000001000000000000000F\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000010\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000011\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000012\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000013\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000014\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000015\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000016\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000017\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:23 000000010000000000000018\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:24 000000010000000000000019\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:22 00000001000000000000001A\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:22 00000001000000000000001B\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:22 00000001000000000000001C\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:24 000000020000000000000019\\n> -rw------- 1 heikki heikki 16777216 Dec 19 14:24 00000002000000000000001A\\n> -rw------- 1 heikki heikki 42 Dec 19 14:24 00000002.history\\n\\nThe files on timeline 1, up to 000000010000000000000019, are valid \\nsegments, streamed from the primary or restored from the WAL archive. \\nThe segments 00000001000000000000001A and 00000001000000000000001B are \\nrecycled segments that haven't been reused yet. Their contents are not \\nvalid (they contain records from some earlier point in WAL, but it might \\nas well be garbage).\\n\\nThe server was promoted within the segment 19, and a new timeline was \\nstarted. Segments 000000020000000000000019 and 00000002000000000000001A \\ncontain valid WAL on the new timeline.\\n\\nNow, after enough time passes that the bogus 00000001000000000000001A \\nand 00000001000000000000001B segments become old enough to be recycled, \\nthe system will see that there is no .ready or .done file for them, and \\nwill create .ready files so that they are archived. And they are \\narchived. That's bogus, because the files are bogus. Worse, if the \\nprimary server where this server was forked off from continues running, \\nand creates the genuine 00000001000000000000001A and \\n00000001000000000000001B segments, it can fail to archive them if the \\nstandby had already archived the bogus segments with the same names.\\n\\nWe must somehow prevent the recycled, but not yet used, segments from \\nbeing archived. One idea is to not create them in the first place, i.e. \\ndon't recycle old segments during recovery, just delete them and have \\nnew ones be created on demand. That's simple, but would hurt performance.\\n\\nI'm thinking that we should add a step to promotion, where we scan \\npg_xlog for any segments higher than the timeline switch point, and \\nremove them, or mark them with .done so that they are not archived. \\nThere might be some real WAL that was streamed from the primary, but not \\nyet applied, but such WAL is of no interest to that server anyway, after \\nit's been promoted. It's a bit disconcerting to zap WAL that's valid, \\neven if doesn't belong to the current server's timeline history, because \\nas a general rule it's good to avoid destroying evidence that might be \\nuseful in debugging. There isn't much difference between removing them \\nimmediately and marking them as .done, though, because they will \\neventually be removed/recycled anyway if they're marked as .done.\\n\\nThe archival behaviour at promotion is a bit inconsistent and weird \\nanyway; even valid, streamed WAL is marked as .done and not archived \\nanyway, except for the last partial segment. We're discussing that in \\nthe other thread (Streaming replication and WAL archive interactions, \\nhttp://www.postgresql.org/message-id/689EB259-44C2-4820-B901-4F6B1C55A1E4@simply.name), \\nbut it would be good have a small, back-patchable fix to prevent bogus \\nsegments from being archived.\\n\\n- Heikki\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 19 Dec 2014 14:55:16 +0200\",\n \"msg_from\": \"Heikki Linnakangas <hlinnakangas@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Bogus WAL segments archived after promotion\"\n },\n {\n \"msg_contents\": \"On 12/19/2014 02:55 PM, Heikki Linnakangas wrote:\\n> I'm thinking that we should add a step to promotion, where we scan\\n> pg_xlog for any segments higher than the timeline switch point, and\\n> remove them, or mark them with .done so that they are not archived.\\n> There might be some real WAL that was streamed from the primary, but not\\n> yet applied, but such WAL is of no interest to that server anyway, after\\n> it's been promoted. It's a bit disconcerting to zap WAL that's valid,\\n> even if doesn't belong to the current server's timeline history, because\\n> as a general rule it's good to avoid destroying evidence that might be\\n> useful in debugging. There isn't much difference between removing them\\n> immediately and marking them as .done, though, because they will\\n> eventually be removed/recycled anyway if they're marked as .done.\\n\\nThis is what I came up with. This patch removes the suspect segments at \\ntimeline switch. The alternative of creating .done files for them would \\npreserve more evidence for debugging, but OTOH it would also be very \\nconfusing to have valid-looking WAL segments in pg_xlog, with .done \\nfiles, that in fact contain garbage.\\n\\nThe patch is a bit longer than it otherwise would be, because I moved \\nthe code to remove a single file from RemoveOldXlogFiles() to a new \\nfunction. I think that makes it more readable in any case, simply \\nbecause it was so deeply indented in RemoveOldXlogFiles.\\n\\nThoughts?\\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, 19 Dec 2014 22:26:34 +0200\",\n \"msg_from\": \"Heikki Linnakangas <hlinnakangas@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Bogus WAL segments archived after promotion\"\n },\n {\n \"msg_contents\": \"On Fri, Dec 19, 2014 at 10:26:34PM +0200, Heikki Linnakangas wrote:\\n> On 12/19/2014 02:55 PM, Heikki Linnakangas wrote:\\n> >I'm thinking that we should add a step to promotion, where we scan\\n> >pg_xlog for any segments higher than the timeline switch point, and\\n> >remove them, or mark them with .done so that they are not archived.\\n> >There might be some real WAL that was streamed from the primary, but not\\n> >yet applied, but such WAL is of no interest to that server anyway, after\\n> >it's been promoted. It's a bit disconcerting to zap WAL that's valid,\\n> >even if doesn't belong to the current server's timeline history, because\\n> >as a general rule it's good to avoid destroying evidence that might be\\n> >useful in debugging. There isn't much difference between removing them\\n> >immediately and marking them as .done, though, because they will\\n> >eventually be removed/recycled anyway if they're marked as .done.\\n> \\n> This is what I came up with. This patch removes the suspect segments\\n> at timeline switch. The alternative of creating .done files for them\\n> would preserve more evidence for debugging, but OTOH it would also\\n> be very confusing to have valid-looking WAL segments in pg_xlog,\\n> with .done files, that in fact contain garbage.\\n> \\n> The patch is a bit longer than it otherwise would be, because I\\n> moved the code to remove a single file from RemoveOldXlogFiles() to\\n> a new function. I think that makes it more readable in any case,\\n> simply because it was so deeply indented in RemoveOldXlogFiles.\\n\\nWhere are we on this?\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n + Everyone has their own god. +\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 1 Apr 2015 12:12:57 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Bogus WAL segments archived after promotion\"\n },\n {\n \"msg_contents\": \"On 04/01/2015 07:12 PM, Bruce Momjian wrote:\\n> On Fri, Dec 19, 2014 at 10:26:34PM +0200, Heikki Linnakangas wrote:\\n>> On 12/19/2014 02:55 PM, Heikki Linnakangas wrote:\\n>>> I'm thinking that we should add a step to promotion, where we scan\\n>>> pg_xlog for any segments higher than the timeline switch point, and\\n>>> remove them, or mark them with .done so that they are not archived.\\n>>> There might be some real WAL that was streamed from the primary, but not\\n>>> yet applied, but such WAL is of no interest to that server anyway, after\\n>>> it's been promoted. It's a bit disconcerting to zap WAL that's valid,\\n>>> even if doesn't belong to the current server's timeline history, because\\n>>> as a general rule it's good to avoid destroying evidence that might be\\n>>> useful in debugging. There isn't much difference between removing them\\n>>> immediately and marking them as .done, though, because they will\\n>>> eventually be removed/recycled anyway if they're marked as .done.\\n>>\\n>> This is what I came up with. This patch removes the suspect segments\\n>> at timeline switch. The alternative of creating .done files for them\\n>> would preserve more evidence for debugging, but OTOH it would also\\n>> be very confusing to have valid-looking WAL segments in pg_xlog,\\n>> with .done files, that in fact contain garbage.\\n>>\\n>> The patch is a bit longer than it otherwise would be, because I\\n>> moved the code to remove a single file from RemoveOldXlogFiles() to\\n>> a new function. I think that makes it more readable in any case,\\n>> simply because it was so deeply indented in RemoveOldXlogFiles.\\n>\\n> Where are we on this?\\n\\nI didn't hear any better ideas, so committed this now.\\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\": \"Mon, 13 Apr 2015 17:57:31 +0300\",\n \"msg_from\": \"Heikki Linnakangas <hlinnaka@iki.fi>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Bogus WAL segments archived after promotion\"\n },\n {\n \"msg_contents\": \"On Mon, Apr 13, 2015 at 11:57 PM, Heikki Linnakangas <hlinnaka@iki.fi> wrote:\\n> On 04/01/2015 07:12 PM, Bruce Momjian wrote:\\n>>\\n>> On Fri, Dec 19, 2014 at 10:26:34PM +0200, Heikki Linnakangas wrote:\\n>>>\\n>>> On 12/19/2014 02:55 PM, Heikki Linnakangas wrote:\\n>>>>\\n>>>> I'm thinking that we should add a step to promotion, where we scan\\n>>>> pg_xlog for any segments higher than the timeline switch point, and\\n>>>> remove them, or mark them with .done so that they are not archived.\\n>>>> There might be some real WAL that was streamed from the primary, but not\\n>>>> yet applied, but such WAL is of no interest to that server anyway, after\\n>>>> it's been promoted. It's a bit disconcerting to zap WAL that's valid,\\n>>>> even if doesn't belong to the current server's timeline history, because\\n>>>> as a general rule it's good to avoid destroying evidence that might be\\n>>>> useful in debugging. There isn't much difference between removing them\\n>>>> immediately and marking them as .done, though, because they will\\n>>>> eventually be removed/recycled anyway if they're marked as .done.\\n>>>\\n>>>\\n>>> This is what I came up with. This patch removes the suspect segments\\n>>> at timeline switch. The alternative of creating .done files for them\\n>>> would preserve more evidence for debugging, but OTOH it would also\\n>>> be very confusing to have valid-looking WAL segments in pg_xlog,\\n>>> with .done files, that in fact contain garbage.\\n>>>\\n>>> The patch is a bit longer than it otherwise would be, because I\\n>>> moved the code to remove a single file from RemoveOldXlogFiles() to\\n>>> a new function. I think that makes it more readable in any case,\\n>>> simply because it was so deeply indented in RemoveOldXlogFiles.\\n>>\\n>>\\n>> Where are we on this?\\n>\\n>\\n> I didn't hear any better ideas, so committed this now.\\n\\nFinally looking at that... The commit log of b2a5545 is a bit\\nmisleading. Segment files that were recycled during archive recovery\\nare not necessarily removed, they could be recycled as well during\\npromotion on the new timeline in line with what RemoveOldXlogFiles()\\ndoes. Hence I think that the comment on top of\\nRemoveNonParentXlogFiles() should be updated to reflect that like in\\nthe patch attached.\\n\\nSomething minor: perhaps we could refactor xlogarchive.c to have\\nXLogArchiveCheckDone() and XLogArchiveIsBusy() use the new\\nXLogArchiveIsReady().\\nRegards,\\n-- \\nMichael\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Thu, 23 Apr 2015 14:57:59 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Bogus WAL segments archived after promotion\"\n },\n {\n \"msg_contents\": \"On Thu, Apr 23, 2015 at 02:57:59PM +0900, Michael Paquier wrote:\\n> Finally looking at that... The commit log of b2a5545 is a bit\\n> misleading. Segment files that were recycled during archive recovery\\n> are not necessarily removed, they could be recycled as well during\\n> promotion on the new timeline in line with what RemoveOldXlogFiles()\\n> does. Hence I think that the comment on top of\\n> RemoveNonParentXlogFiles() should be updated to reflect that like in\\n> the patch attached.\\n> \\n> Something minor: perhaps we could refactor xlogarchive.c to have\\n> XLogArchiveCheckDone() and XLogArchiveIsBusy() use the new\\n> XLogArchiveIsReady().\\n> Regards,\\n\\nOld patch, but still valid, so applied to master, thanks.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> https://momjian.us\\n EDB https://enterprisedb.com\\n\\n Only you can decide what is important to you.\\n\\n\\n\",\n \"msg_date\": \"Sat, 25 Nov 2023 10:48:53 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [HACKERS] Bogus WAL segments archived after promotion\"\n }\n]","isConversational":false}}},{"rowIdx":39988,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nWhen debugging lwlock issues I found PRINT_LWDEBUG/LOG_LWDEBUG rather\npainful to use because of the amount of elog contexts/statements\nemitted. Given the number of lwlock acquirations that's just not doable.\n\nTo solve that during development I've solved that by basically\nreplacing:\n if (Trace_lwlocks)\n elog(LOG, \"%s(%s %d): %s\", where, name, index, msg);\n\nwith something like\n\n if (Trace_lwlocks)\n {\n ErrorContextCallback *old_error_context_stack;\n...\n old_error_context_stack = error_context_stack;\n error_context_stack = NULL;\n ereport(LOG,\n (errhidestmt(true),\n errmsg(\"%s(%s %d): %s\", where, T_NAME(lock),\nT_ID(lock), msg)));\n\nI think it'd generally be useful to have something like errhidecontext()\nakin to errhidestatement() to avoid things like the above.\n\nThe usecases wher eI see this as being useful is high volume debug\nlogging, not normal messages...\n\nGreetings,\n\nAndres Freund\n\n--\n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 19 Dec 2014 17:06:04 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Suppressing elog.c context messages (was Re: Wait free\n LW_SHARED acquisition)"},{"msg_contents":"On Fri, Dec 19, 2014 at 9:36 PM, Andres Freund <andres@2ndquadrant.com>\nwrote:\n>\n> Hi,\n>\n> When debugging lwlock issues I found PRINT_LWDEBUG/LOG_LWDEBUG rather\n> painful to use because of the amount of elog contexts/statements\n> emitted. Given the number of lwlock acquirations that's just not doable.\n>\n> To solve that during development I've solved that by basically\n> replacing:\n> if (Trace_lwlocks)\n> elog(LOG, \"%s(%s %d): %s\", where, name, index, msg);\n>\n> with something like\n>\n> if (Trace_lwlocks)\n> {\n> ErrorContextCallback *old_error_context_stack;\n> ...\n> old_error_context_stack = error_context_stack;\n> error_context_stack = NULL;\n> ereport(LOG,\n> (errhidestmt(true),\n> errmsg(\"%s(%s %d): %s\", where, T_NAME(lock),\n> T_ID(lock), msg)));\n>\n> I think it'd generally be useful to have something like errhidecontext()\n> akin to errhidestatement() to avoid things like the above.\n>\n\nUnder this proposal, do you want to suppress the context/statement\nunconditionally or via some hook/variable, because it might be useful to\nprint the contexts/statements in certain cases where there is complex\napplication and we don't know which part of application code causes\nproblem.\n\n> The usecases wher eI see this as being useful is high volume debug\n> logging, not normal messages...\n>\n\nI think usecase is valid, it is really difficult to dig such a log\nespecially\nwhen statement size is big. Also I think even with above, the number\nof logs generated are high for any statement which could still make\ndebugging difficult, do you think it would be helpful if PRINT_LWDEBUG\nand LOG_LWDEBUG are used with separate defines (LOCK_DEBUG and\nLOCK_BLOCK_DEBUG) as in certain cases we might want to print info\nabout locks which are acquired after waiting or in other words that gets\nblocked.\n\n\nWith Regards,\nAmit Kapila.\nEnterpriseDB: http://www.enterprisedb.com\n\nOn Fri, Dec 19, 2014 at 9:36 PM, Andres Freund <andres@2ndquadrant.com> wrote:>> Hi,>> When debugging lwlock issues I found PRINT_LWDEBUG/LOG_LWDEBUG rather> painful to use because of the amount of elog contexts/statements> emitted. Given the number of lwlock acquirations that's just not doable.>> To solve that during development I've solved that by basically> replacing:> if (Trace_lwlocks)> elog(LOG, \"%s(%s %d): %s\", where, name, index, msg);>> with something like>> if (Trace_lwlocks)> {> ErrorContextCallback *old_error_context_stack;> ...> old_error_context_stack = error_context_stack;> error_context_stack = NULL;> ereport(LOG,> (errhidestmt(true),> errmsg(\"%s(%s %d): %s\", where, T_NAME(lock),> T_ID(lock), msg)));>> I think it'd generally be useful to have something like errhidecontext()> akin to errhidestatement() to avoid things like the above.>Under this proposal, do you want to suppress the context/statementunconditionally or via some hook/variable, because it might be useful toprint the contexts/statements in certain cases where there is complexapplication and we don't know which part of application code causesproblem. > The usecases wher eI see this as being useful is high volume debug> logging, not normal messages...>I think usecase is valid, it is really difficult to dig such a log especiallywhen statement size is big. Also I think even with above, the numberof logs generated are high for any statement which could still makedebugging difficult, do you think it would be helpful if PRINT_LWDEBUGand LOG_LWDEBUG are used with separate defines (LOCK_DEBUG andLOCK_BLOCK_DEBUG) as in certain cases we might want to print infoabout locks which are acquired after waiting or in other words that getsblocked.With Regards,Amit Kapila.EnterpriseDB: http://www.enterprisedb.com","msg_date":"Mon, 22 Dec 2014 10:35:35 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Suppressing elog.c context messages (was Re: Wait free\n LW_SHARED acquisition)"},{"msg_contents":"On 2014-12-22 10:35:35 +0530, Amit Kapila wrote:\n> On Fri, Dec 19, 2014 at 9:36 PM, Andres Freund <andres@2ndquadrant.com>\n> wrote:\n> >\n> > Hi,\n> >\n> > When debugging lwlock issues I found PRINT_LWDEBUG/LOG_LWDEBUG rather\n> > painful to use because of the amount of elog contexts/statements\n> > emitted. Given the number of lwlock acquirations that's just not doable.\n> >\n> > To solve that during development I've solved that by basically\n> > replacing:\n> > if (Trace_lwlocks)\n> > elog(LOG, \"%s(%s %d): %s\", where, name, index, msg);\n> >\n> > with something like\n> >\n> > if (Trace_lwlocks)\n> > {\n> > ErrorContextCallback *old_error_context_stack;\n> > ...\n> > old_error_context_stack = error_context_stack;\n> > error_context_stack = NULL;\n> > ereport(LOG,\n> > (errhidestmt(true),\n> > errmsg(\"%s(%s %d): %s\", where, T_NAME(lock),\n> > T_ID(lock), msg)));\n> >\n> > I think it'd generally be useful to have something like errhidecontext()\n> > akin to errhidestatement() to avoid things like the above.\n> >\n> \n> Under this proposal, do you want to suppress the context/statement\n> unconditionally or via some hook/variable, because it might be useful to\n> print the contexts/statements in certain cases where there is complex\n> application and we don't know which part of application code causes\n> problem.\n\nI'm proposing to do model it after errhidestatement(). I.e. add\nerrhidecontext().\n\nI've attached what I was tinkering with when I wrote this message.\n\n> > The usecases wher eI see this as being useful is high volume debug\n> > logging, not normal messages...\n> >\n> \n> I think usecase is valid, it is really difficult to dig such a log\n> especially when statement size is big.\n\nRight, that was what triggered to look me into it. I'd cases where the\nsame context was printed thousands of times.\n\n> Also I think even with above, the number\n> of logs generated are high for any statement which could still make\n> debugging difficult, do you think it would be helpful if PRINT_LWDEBUG\n> and LOG_LWDEBUG are used with separate defines (LOCK_DEBUG and\n> LOCK_BLOCK_DEBUG) as in certain cases we might want to print info\n> about locks which are acquired after waiting or in other words that gets\n> blocked.\n\nHm, that seems like a separate thing. Personally I don't find it\ninteresting enough to write a patch for it, although I could see it\nbeing interesting for somebody.\n\nIf you're looking at that level it's easy enough to just add a early\nreturn in either routine...\n\nGreetings,\n\nAndres Freund\n\n\n\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 Dec 2014 18:41:29 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Suppressing elog.c context messages (was Re: Wait free\n LW_SHARED acquisition)"},{"msg_contents":"On 12/23/14 11:41 AM, Andres Freund wrote:\n>>> > >I think it'd generally be useful to have something like errhidecontext()\n>>> > >akin to errhidestatement() to avoid things like the above.\n>>> > >\n>> >\n>> >Under this proposal, do you want to suppress the context/statement\n>> >unconditionally or via some hook/variable, because it might be useful to\n>> >print the contexts/statements in certain cases where there is complex\n>> >application and we don't know which part of application code causes\n>> >problem.\n> I'm proposing to do model it after errhidestatement(). I.e. add\n> errhidecontext().\n>\n> I've attached what I was tinkering with when I wrote this message.\n>\n>>> > >The usecases wher eI see this as being useful is high volume debug\n>>> > >logging, not normal messages...\n>>> > >\n>> >\n>> >I think usecase is valid, it is really difficult to dig such a log\n>> >especially when statement size is big.\n> Right, that was what triggered to look me into it. I'd cases where the\n> same context was printed thousands of times.\n\nIn case anyone picks this up... this problem exists in other places too, such as RAISE DEBUG in plpgsql. I think it'd be useful to have clien_min_context and log_min_context that operate akin to *_min_messages but control context output.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 26 Jan 2015 18:30:13 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":false,"msg_subject":"Re: Suppressing elog.c context messages (was Re: Wait free\n LW_SHARED acquisition)"},{"msg_contents":"On 2015-01-26 18:30:13 -0600, Jim Nasby wrote:\n> On 12/23/14 11:41 AM, Andres Freund wrote:\n> >>>> >I think it'd generally be useful to have something like errhidecontext()\n> >>>> >akin to errhidestatement() to avoid things like the above.\n> >>>> >\n> >>>\n> >>>Under this proposal, do you want to suppress the context/statement\n> >>>unconditionally or via some hook/variable, because it might be useful to\n> >>>print the contexts/statements in certain cases where there is complex\n> >>>application and we don't know which part of application code causes\n> >>>problem.\n> >I'm proposing to do model it after errhidestatement(). I.e. add\n> >errhidecontext().\n> >\n> >I've attached what I was tinkering with when I wrote this message.\n> >\n> >>>> >The usecases wher eI see this as being useful is high volume debug\n> >>>> >logging, not normal messages...\n> >>>> >\n> >>>\n> >>>I think usecase is valid, it is really difficult to dig such a log\n> >>>especially when statement size is big.\n> >Right, that was what triggered to look me into it. I'd cases where the\n> >same context was printed thousands of times.\n> \n> In case anyone picks this up... this problem exists in other places too, such as RAISE DEBUG in plpgsql. I think it'd be useful to have clien_min_context and log_min_context that operate akin to *_min_messages but control context output.\n\nI've pushed it already IIRC. I don't think we can just apply it without\nregard for possible users to that many cases - I think the RAISE DEBUG\ncase is better addressed by a plpgsql option to *optionall* suppress\ncontext, than do it unconditionally. That's discussed somewhere nearby.\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 27 Jan 2015 01:32:49 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Suppressing elog.c context messages (was Re: Wait free\n LW_SHARED acquisition)"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nWhen debugging lwlock issues I found PRINT_LWDEBUG/LOG_LWDEBUG rather\\npainful to use because of the amount of elog contexts/statements\\nemitted. Given the number of lwlock acquirations that's just not doable.\\n\\nTo solve that during development I've solved that by basically\\nreplacing:\\n if (Trace_lwlocks)\\n elog(LOG, \\\"%s(%s %d): %s\\\", where, name, index, msg);\\n\\nwith something like\\n\\n if (Trace_lwlocks)\\n {\\n ErrorContextCallback *old_error_context_stack;\\n...\\n old_error_context_stack = error_context_stack;\\n error_context_stack = NULL;\\n ereport(LOG,\\n (errhidestmt(true),\\n errmsg(\\\"%s(%s %d): %s\\\", where, T_NAME(lock),\\nT_ID(lock), msg)));\\n\\nI think it'd generally be useful to have something like errhidecontext()\\nakin to errhidestatement() to avoid things like the above.\\n\\nThe usecases wher eI see this as being useful is high volume debug\\nlogging, not normal messages...\\n\\nGreetings,\\n\\nAndres Freund\\n\\n--\\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 19 Dec 2014 17:06:04 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Suppressing elog.c context messages (was Re: Wait free\\n LW_SHARED acquisition)\"\n },\n {\n \"msg_contents\": \"On Fri, Dec 19, 2014 at 9:36 PM, Andres Freund <andres@2ndquadrant.com>\\nwrote:\\n>\\n> Hi,\\n>\\n> When debugging lwlock issues I found PRINT_LWDEBUG/LOG_LWDEBUG rather\\n> painful to use because of the amount of elog contexts/statements\\n> emitted. Given the number of lwlock acquirations that's just not doable.\\n>\\n> To solve that during development I've solved that by basically\\n> replacing:\\n> if (Trace_lwlocks)\\n> elog(LOG, \\\"%s(%s %d): %s\\\", where, name, index, msg);\\n>\\n> with something like\\n>\\n> if (Trace_lwlocks)\\n> {\\n> ErrorContextCallback *old_error_context_stack;\\n> ...\\n> old_error_context_stack = error_context_stack;\\n> error_context_stack = NULL;\\n> ereport(LOG,\\n> (errhidestmt(true),\\n> errmsg(\\\"%s(%s %d): %s\\\", where, T_NAME(lock),\\n> T_ID(lock), msg)));\\n>\\n> I think it'd generally be useful to have something like errhidecontext()\\n> akin to errhidestatement() to avoid things like the above.\\n>\\n\\nUnder this proposal, do you want to suppress the context/statement\\nunconditionally or via some hook/variable, because it might be useful to\\nprint the contexts/statements in certain cases where there is complex\\napplication and we don't know which part of application code causes\\nproblem.\\n\\n> The usecases wher eI see this as being useful is high volume debug\\n> logging, not normal messages...\\n>\\n\\nI think usecase is valid, it is really difficult to dig such a log\\nespecially\\nwhen statement size is big. Also I think even with above, the number\\nof logs generated are high for any statement which could still make\\ndebugging difficult, do you think it would be helpful if PRINT_LWDEBUG\\nand LOG_LWDEBUG are used with separate defines (LOCK_DEBUG and\\nLOCK_BLOCK_DEBUG) as in certain cases we might want to print info\\nabout locks which are acquired after waiting or in other words that gets\\nblocked.\\n\\n\\nWith Regards,\\nAmit Kapila.\\nEnterpriseDB: http://www.enterprisedb.com\\n\\nOn Fri, Dec 19, 2014 at 9:36 PM, Andres Freund <andres@2ndquadrant.com> wrote:>> Hi,>> When debugging lwlock issues I found PRINT_LWDEBUG/LOG_LWDEBUG rather> painful to use because of the amount of elog contexts/statements> emitted. Given the number of lwlock acquirations that's just not doable.>> To solve that during development I've solved that by basically> replacing:> if (Trace_lwlocks)> elog(LOG, \\\"%s(%s %d): %s\\\", where, name, index, msg);>> with something like>> if (Trace_lwlocks)> {> ErrorContextCallback *old_error_context_stack;> ...> old_error_context_stack = error_context_stack;> error_context_stack = NULL;> ereport(LOG,> (errhidestmt(true),> errmsg(\\\"%s(%s %d): %s\\\", where, T_NAME(lock),> T_ID(lock), msg)));>> I think it'd generally be useful to have something like errhidecontext()> akin to errhidestatement() to avoid things like the above.>Under this proposal, do you want to suppress the context/statementunconditionally or via some hook/variable, because it might be useful toprint the contexts/statements in certain cases where there is complexapplication and we don't know which part of application code causesproblem. > The usecases wher eI see this as being useful is high volume debug> logging, not normal messages...>I think usecase is valid, it is really difficult to dig such a log especiallywhen statement size is big. Also I think even with above, the numberof logs generated are high for any statement which could still makedebugging difficult, do you think it would be helpful if PRINT_LWDEBUGand LOG_LWDEBUG are used with separate defines (LOCK_DEBUG andLOCK_BLOCK_DEBUG) as in certain cases we might want to print infoabout locks which are acquired after waiting or in other words that getsblocked.With Regards,Amit Kapila.EnterpriseDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Mon, 22 Dec 2014 10:35:35 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Suppressing elog.c context messages (was Re: Wait free\\n LW_SHARED acquisition)\"\n },\n {\n \"msg_contents\": \"On 2014-12-22 10:35:35 +0530, Amit Kapila wrote:\\n> On Fri, Dec 19, 2014 at 9:36 PM, Andres Freund <andres@2ndquadrant.com>\\n> wrote:\\n> >\\n> > Hi,\\n> >\\n> > When debugging lwlock issues I found PRINT_LWDEBUG/LOG_LWDEBUG rather\\n> > painful to use because of the amount of elog contexts/statements\\n> > emitted. Given the number of lwlock acquirations that's just not doable.\\n> >\\n> > To solve that during development I've solved that by basically\\n> > replacing:\\n> > if (Trace_lwlocks)\\n> > elog(LOG, \\\"%s(%s %d): %s\\\", where, name, index, msg);\\n> >\\n> > with something like\\n> >\\n> > if (Trace_lwlocks)\\n> > {\\n> > ErrorContextCallback *old_error_context_stack;\\n> > ...\\n> > old_error_context_stack = error_context_stack;\\n> > error_context_stack = NULL;\\n> > ereport(LOG,\\n> > (errhidestmt(true),\\n> > errmsg(\\\"%s(%s %d): %s\\\", where, T_NAME(lock),\\n> > T_ID(lock), msg)));\\n> >\\n> > I think it'd generally be useful to have something like errhidecontext()\\n> > akin to errhidestatement() to avoid things like the above.\\n> >\\n> \\n> Under this proposal, do you want to suppress the context/statement\\n> unconditionally or via some hook/variable, because it might be useful to\\n> print the contexts/statements in certain cases where there is complex\\n> application and we don't know which part of application code causes\\n> problem.\\n\\nI'm proposing to do model it after errhidestatement(). I.e. add\\nerrhidecontext().\\n\\nI've attached what I was tinkering with when I wrote this message.\\n\\n> > The usecases wher eI see this as being useful is high volume debug\\n> > logging, not normal messages...\\n> >\\n> \\n> I think usecase is valid, it is really difficult to dig such a log\\n> especially when statement size is big.\\n\\nRight, that was what triggered to look me into it. I'd cases where the\\nsame context was printed thousands of times.\\n\\n> Also I think even with above, the number\\n> of logs generated are high for any statement which could still make\\n> debugging difficult, do you think it would be helpful if PRINT_LWDEBUG\\n> and LOG_LWDEBUG are used with separate defines (LOCK_DEBUG and\\n> LOCK_BLOCK_DEBUG) as in certain cases we might want to print info\\n> about locks which are acquired after waiting or in other words that gets\\n> blocked.\\n\\nHm, that seems like a separate thing. Personally I don't find it\\ninteresting enough to write a patch for it, although I could see it\\nbeing interesting for somebody.\\n\\nIf you're looking at that level it's easy enough to just add a early\\nreturn in either routine...\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Tue, 23 Dec 2014 18:41:29 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Suppressing elog.c context messages (was Re: Wait free\\n LW_SHARED acquisition)\"\n },\n {\n \"msg_contents\": \"On 12/23/14 11:41 AM, Andres Freund wrote:\\n>>> > >I think it'd generally be useful to have something like errhidecontext()\\n>>> > >akin to errhidestatement() to avoid things like the above.\\n>>> > >\\n>> >\\n>> >Under this proposal, do you want to suppress the context/statement\\n>> >unconditionally or via some hook/variable, because it might be useful to\\n>> >print the contexts/statements in certain cases where there is complex\\n>> >application and we don't know which part of application code causes\\n>> >problem.\\n> I'm proposing to do model it after errhidestatement(). I.e. add\\n> errhidecontext().\\n>\\n> I've attached what I was tinkering with when I wrote this message.\\n>\\n>>> > >The usecases wher eI see this as being useful is high volume debug\\n>>> > >logging, not normal messages...\\n>>> > >\\n>> >\\n>> >I think usecase is valid, it is really difficult to dig such a log\\n>> >especially when statement size is big.\\n> Right, that was what triggered to look me into it. I'd cases where the\\n> same context was printed thousands of times.\\n\\nIn case anyone picks this up... this problem exists in other places too, such as RAISE DEBUG in plpgsql. I think it'd be useful to have clien_min_context and log_min_context that operate akin to *_min_messages but control context output.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 26 Jan 2015 18:30:13 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Suppressing elog.c context messages (was Re: Wait free\\n LW_SHARED acquisition)\"\n },\n {\n \"msg_contents\": \"On 2015-01-26 18:30:13 -0600, Jim Nasby wrote:\\n> On 12/23/14 11:41 AM, Andres Freund wrote:\\n> >>>> >I think it'd generally be useful to have something like errhidecontext()\\n> >>>> >akin to errhidestatement() to avoid things like the above.\\n> >>>> >\\n> >>>\\n> >>>Under this proposal, do you want to suppress the context/statement\\n> >>>unconditionally or via some hook/variable, because it might be useful to\\n> >>>print the contexts/statements in certain cases where there is complex\\n> >>>application and we don't know which part of application code causes\\n> >>>problem.\\n> >I'm proposing to do model it after errhidestatement(). I.e. add\\n> >errhidecontext().\\n> >\\n> >I've attached what I was tinkering with when I wrote this message.\\n> >\\n> >>>> >The usecases wher eI see this as being useful is high volume debug\\n> >>>> >logging, not normal messages...\\n> >>>> >\\n> >>>\\n> >>>I think usecase is valid, it is really difficult to dig such a log\\n> >>>especially when statement size is big.\\n> >Right, that was what triggered to look me into it. I'd cases where the\\n> >same context was printed thousands of times.\\n> \\n> In case anyone picks this up... this problem exists in other places too, such as RAISE DEBUG in plpgsql. I think it'd be useful to have clien_min_context and log_min_context that operate akin to *_min_messages but control context output.\\n\\nI've pushed it already IIRC. I don't think we can just apply it without\\nregard for possible users to that many cases - I think the RAISE DEBUG\\ncase is better addressed by a plpgsql option to *optionall* suppress\\ncontext, than do it unconditionally. That's discussed somewhere nearby.\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 27 Jan 2015 01:32:49 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Suppressing elog.c context messages (was Re: Wait free\\n LW_SHARED acquisition)\"\n }\n]","isConversational":false}}},{"rowIdx":39989,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Right now, both table and domain constraints share the same\nclassification in the ObjType enum, OBJECT_CONSTRAINT. For my patch to\nimplement improved object drop reporting in event triggers, it is\nnecessary to separate them, as per in the attach patch.\n\nOne emerging side-effect is that it is now possible to execute COMMENT\nON domain constraints, which wasn't previously possible. (I added \\dd\nsupport to psql as well.)\n\nThis is part of the patch for \"replicating DROP commands across\nservers\",\nhttp://archives.postgresql.org/message-id/20140826135459.GD6343@eldon.alvh.no-ip.org\nwhich I just split up.\n\n-- \n�lvaro Herrera http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Fri, 19 Dec 2014 23:23:08 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":true,"msg_subject":"table vs domain constraints in ObjType"}],"string":"[\n {\n \"msg_contents\": \"Right now, both table and domain constraints share the same\\nclassification in the ObjType enum, OBJECT_CONSTRAINT. For my patch to\\nimplement improved object drop reporting in event triggers, it is\\nnecessary to separate them, as per in the attach patch.\\n\\nOne emerging side-effect is that it is now possible to execute COMMENT\\nON domain constraints, which wasn't previously possible. (I added \\\\dd\\nsupport to psql as well.)\\n\\nThis is part of the patch for \\\"replicating DROP commands across\\nservers\\\",\\nhttp://archives.postgresql.org/message-id/20140826135459.GD6343@eldon.alvh.no-ip.org\\nwhich I just split up.\\n\\n-- \\n�lvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Fri, 19 Dec 2014 23:23:08 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"table vs domain constraints in ObjType\"\n }\n]","isConversational":false}}},{"rowIdx":39990,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Buildfarm members magpie, treepie and fulmar went absent on 2014-10-29. Since\nreturning on 2014-11-16, they have consistently failed with 'initdb: invalid\nlocale name \"cs_CZ.WIN-1250\"'. No commits in that period readily explain a\nregression in this area. Did the underlying system configurations change?\n\nThanks,\nnm\n\n\n-- \nSent via 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 Dec 2014 01:39:23 -0500","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"On 20.12.2014 07:39, Noah Misch wrote:\n> Buildfarm members magpie, treepie and fulmar went absent on\n> 2014-10-29. Since returning on 2014-11-16, they have consistently\n> failed with 'initdb: invalid locale name \"cs_CZ.WIN-1250\"'. No\n> commits in that period readily explain a regression in this area. Did\n> the underlying system configurations change?\n\nI'm pretty sure it's because of broken locales at the system level. It was\nworking fine, and I haven't done any substantial changes to the system\nexcept for occasional \"yum update\", so I'm unaware of what went wong :(\n\nThe issue looks like this:\n\n# locale -a | grep en_US\nen_US\nen_US.iso88591\nen_US.iso885915\nen_US.utf8\n\n# locale en_US\nlocale: unknown name \"en_US\"\n\nThe only reasons I can think of is that some of the updates required\na reboot, and I haven't done that because that would kill all the VMs\nrunning on that HW, including the one with CLOBBER_CACHE_RECURSIVE\ntests. And that'd throw away tests running for ~3 months.\n\nI've disabled the three animals (magpie, fulmar, treepie) for now, because\nthere's no point in running the tests until the locale issues are fixed. If\nanyone has an idea of what might be wrong, let me know.\n\nTomas\n\n\n\n\n\n\n On 20.12.2014 07:39, Noah Misch wrote:\n> Buildfarm members magpie,\n treepie and fulmar went absent on\n > 2014-10-29. Since returning on 2014-11-16, they have\n consistently\n > failed with 'initdb: invalid locale name \"cs_CZ.WIN-1250\"'.\n No\n > commits in that period readily explain a regression in this\n area. Did\n > the underlying system configurations change?\n\n I'm pretty sure it's because of broken locales at the system level.\n It was\n working fine, and I haven't done any substantial changes to the\n system\n except for occasional \"yum update\", so I'm unaware of what went wong\n :(\n\n The issue looks like this:\n\n # locale -a | grep en_US\n en_US\n en_US.iso88591\n en_US.iso885915\n en_US.utf8\n\n # locale en_US\n locale: unknown name \"en_US\"\n\n The only reasons I can think of is that some of the updates required\n a reboot, and I haven't done that because that would kill all the\n VMs\n running on that HW, including the one with CLOBBER_CACHE_RECURSIVE \n tests. And that'd throw away tests running for ~3 months.\n\n I've disabled the three animals (magpie, fulmar, treepie) for now,\n because\n there's no point in running the tests until the locale issues are\n fixed. If\n anyone has an idea of what might be wrong, let me know.\n\n Tomas","msg_date":"Sat, 20 Dec 2014 17:14:03 +0100","msg_from":"CSPUG <info@cspug.cz>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"On Sat, Dec 20, 2014 at 05:14:03PM +0100, CSPUG wrote:\n> On 20.12.2014 07:39, Noah Misch wrote:\n> > Buildfarm members magpie, treepie and fulmar went absent on\n> > 2014-10-29. Since returning on 2014-11-16, they have consistently\n> > failed with 'initdb: invalid locale name \"cs_CZ.WIN-1250\"'. No\n> > commits in that period readily explain a regression in this area. Did\n> > the underlying system configurations change?\n> \n> I'm pretty sure it's because of broken locales at the system level. It was\n> working fine, and I haven't done any substantial changes to the system\n> except for occasional \"yum update\", so I'm unaware of what went wong :(\n> \n> The issue looks like this:\n> \n> # locale -a | grep en_US\n> en_US\n> en_US.iso88591\n> en_US.iso885915\n> en_US.utf8\n> \n> # locale en_US\n> locale: unknown name \"en_US\"\n\nThe NAME argument to the \"locale\" tool is something like LC_PAPER, not a\nlocale name. Use \"LANG=en_US locale LC_NUMERIC\" to test locale loading.\n\n> The only reasons I can think of is that some of the updates required\n> a reboot, and I haven't done that because that would kill all the VMs\n> running on that HW, including the one with CLOBBER_CACHE_RECURSIVE\n> tests. And that'd throw away tests running for ~3 months.\n> \n> I've disabled the three animals (magpie, fulmar, treepie) for now, because\n> there's no point in running the tests until the locale issues are fixed. If\n> anyone has an idea of what might be wrong, let me know.\n\nThose animals have been successfully completing initdb for several locales,\nincluding en_US, before failing at cs_CZ.WIN-1250. You could disable just the\ncs_CZ.WIN-1250 steps. A CentOS 6.6 system here also lacks such a locale:\n\n$ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\nlocale: Cannot set LC_CTYPE to default locale: No such file or directory\nlocale: Cannot set LC_MESSAGES to default locale: No such file or directory\nlocale: Cannot set LC_ALL to default locale: No such file or directory\n.\n\n-1\n46\n0\nANSI_X3.4-1968\n$ locale -a|grep cs_\ncs_CZ\ncs_CZ.iso88592\ncs_CZ.utf8\n\nPerhaps an update made the system stricter about rejecting unknown locales.\n\n\n-- \nSent via 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 Dec 2014 11:48:43 -0500","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"Noah Misch <noah@leadboat.com> writes:\n> On Sat, Dec 20, 2014 at 05:14:03PM +0100, CSPUG wrote:\n>> On 20.12.2014 07:39, Noah Misch wrote:\n>>> Buildfarm members magpie, treepie and fulmar went absent on\n>>> 2014-10-29. Since returning on 2014-11-16, they have consistently\n>>> failed with 'initdb: invalid locale name \"cs_CZ.WIN-1250\"'. No\n>>> commits in that period readily explain a regression in this area. Did\n>>> the underlying system configurations change?\n\nThat's what I'd been assuming, but I had not got round to inquiring.\n\n>> I'm pretty sure it's because of broken locales at the system level. It was\n>> working fine, and I haven't done any substantial changes to the system\n>> except for occasional \"yum update\", so I'm unaware of what went wong :(\n>> \n>> The only reasons I can think of is that some of the updates required\n>> a reboot, and I haven't done that because that would kill all the VMs\n>> running on that HW, including the one with CLOBBER_CACHE_RECURSIVE\n>> tests. And that'd throw away tests running for ~3 months.\n>> \n>> I've disabled the three animals (magpie, fulmar, treepie) for now, because\n>> there's no point in running the tests until the locale issues are fixed. If\n>> anyone has an idea of what might be wrong, let me know.\n\n> Those animals have been successfully completing initdb for several locales,\n> including en_US, before failing at cs_CZ.WIN-1250. You could disable just the\n> cs_CZ.WIN-1250 steps. A CentOS 6.6 system here also lacks such a locale:\n\n> $ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\n> locale: Cannot set LC_CTYPE to default locale: No such file or directory\n> locale: Cannot set LC_MESSAGES to default locale: No such file or directory\n> locale: Cannot set LC_ALL to default locale: No such file or directory\n\nFWIW, an actual RHEL 6.6 system doesn't like that either; at least not\nwith a fairly minimal set of locales installed.\n\n> Perhaps an update made the system stricter about rejecting unknown locales.\n\nRed Hat released RHEL 6.6 on Oct 14. I am not sure how long it took the\nCentOS crew to follow suit, but it's well within the realm of plausibility\nthat when these buildfarm critters went offline it was associated with a\nsystem update from 6.5 to 6.6 ... which was a *major* update. (On my\nmachine, something like 480 out of 1500 packages were updated.) It's easy\nto believe that either the glibc locale code is now stricter, or that Red\nHat reshuffled the packaging a bit and the missing locale is now in a\nnew package that didn't get installed.\n\nPersonally I'd not have tried to run such a system without rebooting ...\nit's unlikely that not rebooting has anything directly to do with this\nproblem, but I would not be surprised by issues elsewhere.\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 Dec 2014 12:12:04 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"2014-12-20 17:48 GMT+01:00 Noah Misch <noah@leadboat.com>:\n\n> On Sat, Dec 20, 2014 at 05:14:03PM +0100, CSPUG wrote:\n> > On 20.12.2014 07:39, Noah Misch wrote:\n> > > Buildfarm members magpie, treepie and fulmar went absent on\n> > > 2014-10-29. Since returning on 2014-11-16, they have consistently\n> > > failed with 'initdb: invalid locale name \"cs_CZ.WIN-1250\"'. No\n> > > commits in that period readily explain a regression in this area. Did\n> > > the underlying system configurations change?\n> >\n> > I'm pretty sure it's because of broken locales at the system level. It\n> was\n> > working fine, and I haven't done any substantial changes to the system\n> > except for occasional \"yum update\", so I'm unaware of what went wong :(\n> >\n> > The issue looks like this:\n> >\n> > # locale -a | grep en_US\n> > en_US\n> > en_US.iso88591\n> > en_US.iso885915\n> > en_US.utf8\n> >\n> > # locale en_US\n> > locale: unknown name \"en_US\"\n>\n> The NAME argument to the \"locale\" tool is something like LC_PAPER, not a\n> locale name. Use \"LANG=en_US locale LC_NUMERIC\" to test locale loading.\n>\n> > The only reasons I can think of is that some of the updates required\n> > a reboot, and I haven't done that because that would kill all the VMs\n> > running on that HW, including the one with CLOBBER_CACHE_RECURSIVE\n> > tests. And that'd throw away tests running for ~3 months.\n> >\n> > I've disabled the three animals (magpie, fulmar, treepie) for now,\n> because\n> > there's no point in running the tests until the locale issues are fixed.\n> If\n> > anyone has an idea of what might be wrong, let me know.\n>\n> Those animals have been successfully completing initdb for several locales,\n> including en_US, before failing at cs_CZ.WIN-1250. You could disable just\n> the\n> cs_CZ.WIN-1250 steps. A CentOS 6.6 system here also lacks such a locale:\n>\n> $ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\n>\n\nIt is Microsoft encoding, - it is not available on Linux\n\nRegards\n\nPavel\n\n\n> locale: Cannot set LC_CTYPE to default locale: No such file or directory\n> locale: Cannot set LC_MESSAGES to default locale: No such file or directory\n> locale: Cannot set LC_ALL to default locale: No such file or directory\n> .\n>\n> -1\n> 46\n> 0\n> ANSI_X3.4-1968\n> $ locale -a|grep cs_\n> cs_CZ\n> cs_CZ.iso88592\n> cs_CZ.utf8\n>\n> Perhaps an update made the system stricter about rejecting unknown locales.\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\n2014-12-20 17:48 GMT+01:00 Noah Misch <noah@leadboat.com>:On Sat, Dec 20, 2014 at 05:14:03PM +0100, CSPUG wrote:\n> On 20.12.2014 07:39, Noah Misch wrote:\n> > Buildfarm members magpie, treepie and fulmar went absent on\n> > 2014-10-29. Since returning on 2014-11-16, they have consistently\n> > failed with 'initdb: invalid locale name \"cs_CZ.WIN-1250\"'. No\n> > commits in that period readily explain a regression in this area. Did\n> > the underlying system configurations change?\n>\n> I'm pretty sure it's because of broken locales at the system level. It was\n> working fine, and I haven't done any substantial changes to the system\n> except for occasional \"yum update\", so I'm unaware of what went wong :(\n>\n> The issue looks like this:\n>\n> # locale -a | grep en_US\n> en_US\n> en_US.iso88591\n> en_US.iso885915\n> en_US.utf8\n>\n> # locale en_US\n> locale: unknown name \"en_US\"\n\nThe NAME argument to the \"locale\" tool is something like LC_PAPER, not a\nlocale name. Use \"LANG=en_US locale LC_NUMERIC\" to test locale loading.\n\n> The only reasons I can think of is that some of the updates required\n> a reboot, and I haven't done that because that would kill all the VMs\n> running on that HW, including the one with CLOBBER_CACHE_RECURSIVE\n> tests. And that'd throw away tests running for ~3 months.\n>\n> I've disabled the three animals (magpie, fulmar, treepie) for now, because\n> there's no point in running the tests until the locale issues are fixed. If\n> anyone has an idea of what might be wrong, let me know.\n\nThose animals have been successfully completing initdb for several locales,\nincluding en_US, before failing at cs_CZ.WIN-1250. You could disable just the\ncs_CZ.WIN-1250 steps. A CentOS 6.6 system here also lacks such a locale:\n\n$ LANG=cs_CZ.WIN-1250 locale LC_NUMERICIt is Microsoft encoding, - it is not available on LinuxRegardsPavel \nlocale: Cannot set LC_CTYPE to default locale: No such file or directory\nlocale: Cannot set LC_MESSAGES to default locale: No such file or directory\nlocale: Cannot set LC_ALL to default locale: No such file or directory\n.\n\n-1\n46\n0\nANSI_X3.4-1968\n$ locale -a|grep cs_\ncs_CZ\ncs_CZ.iso88592\ncs_CZ.utf8\n\nPerhaps an update made the system stricter about rejecting unknown locales.\n\n\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 Dec 2014 18:13:05 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"On 20.12.2014 18:13, Pavel Stehule wrote:\n> \n> 2014-12-20 17:48 GMT+01:00 Noah Misch <noah@leadboat.com\n>\n> $ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\n> \n> \n> It is Microsoft encoding, - it is not available on Linux\n\nNot true. It is available on Linux, and the regression tests were\nrunning with it for a long time (essentially from the moment magpie was\nadded to the buildfarm).\n\nTomas\n\n\n-- \nSent via 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 Dec 2014 18:17:11 +0100","msg_from":"Tomas Vondra <tv@fuzzy.cz>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"2014-12-20 18:17 GMT+01:00 Tomas Vondra <tv@fuzzy.cz>:\n\n> On 20.12.2014 18:13, Pavel Stehule wrote:\n> >\n> > 2014-12-20 17:48 GMT+01:00 Noah Misch <noah@leadboat.com\n> >\n> > $ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\n> >\n> >\n> > It is Microsoft encoding, - it is not available on Linux\n>\n> Not true. It is available on Linux, and the regression tests were\n> running with it for a long time (essentially from the moment magpie was\n> added to the buildfarm).\n>\n\naha\n\nI am sorry for my mistake\n\nPavel\n\n\n>\n> Tomas\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\n2014-12-20 18:17 GMT+01:00 Tomas Vondra <tv@fuzzy.cz>:On 20.12.2014 18:13, Pavel Stehule wrote:\n>\n> 2014-12-20 17:48 GMT+01:00 Noah Misch <noah@leadboat.com\n>\n> $ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\n>\n>\n> It is Microsoft encoding, - it is not available on Linux\n\nNot true. It is available on Linux, and the regression tests were\nrunning with it for a long time (essentially from the moment magpie was\nadded to the buildfarm).ahaI am sorry for my mistakePavel \n\nTomas\n\n\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 Dec 2014 18:25:17 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"Tomas Vondra <tv@fuzzy.cz> writes:\n> On 20.12.2014 18:13, Pavel Stehule wrote:\n>> It is Microsoft encoding, - it is not available on Linux\n\n> Not true. It is available on Linux, and the regression tests were\n> running with it for a long time (essentially from the moment magpie was\n> added to the buildfarm).\n\nWell, you had it at one time, but it sure doesn't seem to be there now.\n\nI poked around in the RHEL 6.6 release notes, and found this possibly\nrelevant tidbit:\n\n\t mingw component, BZ#1063396\n\n\t Following the deprecation of Matahari packages in Red Hat\n\t Enterprise Linux 6.3, at which time the mingw packages were noted\n\t as deprecated, and the subsequent removal of Matahari packages\n\t from Red Hat Enterprise Linux 6.4, the mingw packages are now\n\t being removed from Red Hat Enterprise Linux 6.6.\n\n\t The mingw packages will no longer be shipped in future Red Hat\n\t Enterprise Linux 6 minor releases, nor will they receive\n\t security-related updates. Consequently, users are advised to\n\t uninstall any earlier releases of the mingw packages from their\n\t Red Hat Enterprise Linux 6 systems.\n\nIt seems plausible that a WIN-1250 locale would have been something\nthat would've been supplied by mingw rather than being part of either\nglibc-common or any official locale package. I'm not sure how that\ntranslates to it actively disappearing from your machine --- as the note\nsays, you're \"advised to uninstall\" mingw, but I don't think the 6.6\nupdate would have removed it automatically. Still, the outlines of a\ntheory are starting to come into focus.\n\nImmediate recommendation is to restart the buildfarm critters with the\nmissing locale removed from their configurations. If you can find the\nlocale again somewhere, by all means re-enable it, but better less testing\nthan no testing in the meantime.\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 Dec 2014 12:32:33 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"\nOn 12/20/2014 12:32 PM, Tom Lane wrote:\n\n>\n> Immediate recommendation is to restart the buildfarm critters with the\n> missing locale removed from their configurations. If you can find the\n> locale again somewhere, by all means re-enable it, but better less testing\n> than no testing in the meantime.\n>\n> \t\t\t\n\n\nYeah, all they need to do is alter the locales setting in the config \nfile. If they want to force new runs, they can use \n<https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Tips_and_Tricks>\n\n\ncheers\n\nandrew\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 20 Dec 2014 12:50:34 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"On 20.12.2014 17:48, Noah Misch wrote:\n> On Sat, Dec 20, 2014 at 05:14:03PM +0100, CSPUG wrote:\n>> On 20.12.2014 07:39, Noah Misch wrote:\n>>> Buildfarm members magpie, treepie and fulmar went absent on\n>>> 2014-10-29. Since returning on 2014-11-16, they have consistently\n>>> failed with 'initdb: invalid locale name \"cs_CZ.WIN-1250\"'. No\n>>> commits in that period readily explain a regression in this area. Did\n>>> the underlying system configurations change?\n>>\n>> I'm pretty sure it's because of broken locales at the system level. It was\n>> working fine, and I haven't done any substantial changes to the system\n>> except for occasional \"yum update\", so I'm unaware of what went wong :(\n>>\n>> The issue looks like this:\n>>\n>> # locale -a | grep en_US\n>> en_US\n>> en_US.iso88591\n>> en_US.iso885915\n>> en_US.utf8\n>>\n>> # locale en_US\n>> locale: unknown name \"en_US\"\n> \n> The NAME argument to the \"locale\" tool is something like LC_PAPER, not a\n> locale name. Use \"LANG=en_US locale LC_NUMERIC\" to test locale loading.\n\nD'oh! That kinda explains the strange `locale` failures, of course ...\n\nBut still, the three animals (magpie, fulmar, treepie) are all running\nwith these locales:\n\n C en_US\n cs_CZ.UTF-8 cs_CZ.ISO-8859-2 cs_CZ.WIN-1250\n sk_SK.UTF-8 sk_SK.ISO-8859-2 sk_SK.WIN-1250\n\nand the only one missing (so the 'locale LC_NUMERIC' failed) was\nsk_SK.WIN-1250. I've fixed that, and now all the locales work fine.\n\nThat however does not explain why the tests are failing with\ncs_CZ.WIN-1250 because that locale was working for some time.\n\n> \n>> The only reasons I can think of is that some of the updates required\n>> a reboot, and I haven't done that because that would kill all the VMs\n>> running on that HW, including the one with CLOBBER_CACHE_RECURSIVE\n>> tests. And that'd throw away tests running for ~3 months.\n>>\n>> I've disabled the three animals (magpie, fulmar, treepie) for now, because\n>> there's no point in running the tests until the locale issues are fixed. If\n>> anyone has an idea of what might be wrong, let me know.\n> \n> Those animals have been successfully completing initdb for several locales,\n> including en_US, before failing at cs_CZ.WIN-1250. You could disable just the\n> cs_CZ.WIN-1250 steps. A CentOS 6.6 system here also lacks such a locale:\n> \n> $ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\n> locale: Cannot set LC_CTYPE to default locale: No such file or directory\n> locale: Cannot set LC_MESSAGES to default locale: No such file or directory\n> locale: Cannot set LC_ALL to default locale: No such file or directory\n> .\n> \n> -1\n> 46\n> 0\n> ANSI_X3.4-1968\n> $ locale -a|grep cs_\n> cs_CZ\n> cs_CZ.iso88592\n> cs_CZ.utf8\n> \n> Perhaps an update made the system stricter about rejecting unknown locales.\n\nI believe the locale system (at the OS level) works just like before. I\nremember I had to manually create the locales while initially setting up\nthe animals. Then, ~2 months ago something happened (I asssume a yum\nupdate) and some of the locales disappeared. But I have recreated them,\nexcept for sk_SK.WIN-1250. But the tests fail because of cs_CZ.WIN-1250\nwhich does exist.\n\nAttached is a log of\n\nfor l in cs_CZ.UTF-8 sk_SK.UTF-8 cs_CZ.ISO-8859-2 sk_SK.ISO-8859-2\ncs_CZ.WIN-1250 sk_SK.WIN-1250; do\n echo $l; LANG=$l locale LC_NUMERIC;\ndone;\n\nshowing that all the locales exist. However when I tried to initialize a\ncluster with cs_CZ.WIN-1250, I got an error like this:\n\n[pgbuild@regular-builds ~]$ locale\nLANG=cs_CZ.WIN-1250\nLC_CTYPE=\"cs_CZ.WIN-1250\"\nLC_NUMERIC=\"cs_CZ.WIN-1250\"\nLC_TIME=\"cs_CZ.WIN-1250\"\nLC_COLLATE=\"cs_CZ.WIN-1250\"\nLC_MONETARY=\"cs_CZ.WIN-1250\"\nLC_MESSAGES=\"cs_CZ.WIN-1250\"\nLC_PAPER=\"cs_CZ.WIN-1250\"\nLC_NAME=\"cs_CZ.WIN-1250\"\nLC_ADDRESS=\"cs_CZ.WIN-1250\"\nLC_TELEPHONE=\"cs_CZ.WIN-1250\"\nLC_MEASUREMENT=\"cs_CZ.WIN-1250\"\nLC_IDENTIFICATION=\"cs_CZ.WIN-1250\"\nLC_ALL=\n[pgbuild@regular-builds ~]$ pg_ctl -D tmp-data init\nThe files belonging to this database system will be owned by user \"pgbuild\".\nThis user must also own the server process.\n\nThe database cluster will be initialized with locale \"cs_CZ.WIN-1250\".\ncould not determine encoding for locale \"cs_CZ.WIN-1250\": codeset is\n\"ANSI_X3.4-1968\"\ninitdb: could not find suitable encoding for locale \"cs_CZ.WIN-1250\"\nRerun initdb with the -E option.\nTry \"initdb --help\" for more information.\npg_ctl: database system initialization failed\n\nSo apparently the locale does exist, but we're unable to cope with it\nfor some reason ...\n\nTomas\n\n\n\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 Dec 2014 18:55:30 +0100","msg_from":"Tomas Vondra <tv@fuzzy.cz>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"On 20.12.2014 18:32, Tom Lane wrote:\n> Tomas Vondra <tv@fuzzy.cz> writes:\n>> On 20.12.2014 18:13, Pavel Stehule wrote:\n>>> It is Microsoft encoding, - it is not available on Linux\n> \n>> Not true. It is available on Linux, and the regression tests were \n>> running with it for a long time (essentially from the moment magpie\n>> was added to the buildfarm).\n> \n> Well, you had it at one time, but it sure doesn't seem to be there now.\n>\n> I poked around in the RHEL 6.6 release notes, and found this possibly\n> relevant tidbit:\n\nAh, apparently I upgraded this VM to 6.6 - that wasn't exactly planned.\n\n> \n> \t mingw component, BZ#1063396\n> \n> \t Following the deprecation of Matahari packages in Red Hat\n> \t Enterprise Linux 6.3, at which time the mingw packages were noted\n> \t as deprecated, and the subsequent removal of Matahari packages\n> \t from Red Hat Enterprise Linux 6.4, the mingw packages are now\n> \t being removed from Red Hat Enterprise Linux 6.6.\n> \n> \t The mingw packages will no longer be shipped in future Red Hat\n> \t Enterprise Linux 6 minor releases, nor will they receive\n> \t security-related updates. Consequently, users are advised to\n> \t uninstall any earlier releases of the mingw packages from their\n> \t Red Hat Enterprise Linux 6 systems.\n> \n> It seems plausible that a WIN-1250 locale would have been something \n> that would've been supplied by mingw rather than being part of\n> either glibc-common or any official locale package. I'm not sure how\n> that translates to it actively disappearing from your machine --- as\n> the note says, you're \"advised to uninstall\" mingw, but I don't think\n> the 6.6 update would have removed it automatically. Still, the\n> outlines of a theory are starting to come into focus.\n\nI don't think so. As I mentioned in the previous post, I still can\ncreate the locale, but initdb still fails with an error about\nANSI_X3.4-1968 encoding.\n\nBut clearly, something changed between RH 6.5 and 6.6, because on 6.5 I\nget this:\n\n$ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\n,\n�\n3;3\n44\n160\nCP1250\n\nwhile on 6.6 I get this:\n\n$ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\n,\n\n3;3\n44\n160\nANSI_X3.4-1968\n\nSo yeah, the locale definition did change a bit - the encoding is\ndifferent, and we can't cope with it.\n\n> Immediate recommendation is to restart the buildfarm critters with\n> the missing locale removed from their configurations. If you can find\n> the locale again somewhere, by all means re-enable it, but better\n> less testing than no testing in the meantime.\n\nI've removed cs_CZ.WIN-1250 and sk_SK.WIN-1250 from the config. Let's\nsee how that works.\n\n\nTomas\n\n\n-- \nSent via 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 Dec 2014 19:02:17 +0100","msg_from":"Tomas Vondra <tv@fuzzy.cz>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"Tomas Vondra <tv@fuzzy.cz> writes:\n> I believe the locale system (at the OS level) works just like before. I\n> remember I had to manually create the locales while initially setting up\n> the animals. Then, ~2 months ago something happened (I asssume a yum\n> update) and some of the locales disappeared. But I have recreated them,\n> except for sk_SK.WIN-1250. But the tests fail because of cs_CZ.WIN-1250\n> which does exist.\n\nI am betting that you recreated them differently from before.\n\n> However when I tried to initialize a\n> cluster with cs_CZ.WIN-1250, I got an error like this:\n\n> [pgbuild@regular-builds ~]$ pg_ctl -D tmp-data init\n> The files belonging to this database system will be owned by user \"pgbuild\".\n> This user must also own the server process.\n\n> The database cluster will be initialized with locale \"cs_CZ.WIN-1250\".\n> could not determine encoding for locale \"cs_CZ.WIN-1250\": codeset is\n> \"ANSI_X3.4-1968\"\n> initdb: could not find suitable encoding for locale \"cs_CZ.WIN-1250\"\n\nLocale cs_CZ.WIN-1250 is evidently marked with a codeset property of\n\"ANSI_X3.4-1968\" (which means old-school US-ASCII). That's certainly\nwrong. I believe the correct thing would be \"CP1250\".\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 Dec 2014 13:05:33 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"Tomas Vondra <tv@fuzzy.cz> writes:\n> But clearly, something changed between RH 6.5 and 6.6, because on 6.5 I\n> get this:\n\n> $ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\n> ,\n> �\n> 3;3\n> 44\n> 160\n> CP1250\n\n> while on 6.6 I get this:\n\n> $ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\n> ,\n\n> 3;3\n> 44\n> 160\n> ANSI_X3.4-1968\n\nThat's certainly broken. The entire point of having a cs_CZ.WIN-1250\nlocale (as opposed to cs_CZ.something-else) would be to specify a\ncodeset corresponding to WIN-1250. Our code recognizes the spelling\n\"CP1250\" for that. It's possible there are other spellings we should\nrecognize, but \"ANSI_X3.4-1968\" is certainly not one. As I said, that\njust means ASCII, so it's completely useless for determining which\nASCII-superset encoding is wanted.\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 Dec 2014 13:22:17 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"On 20.12.2014 19:05, Tom Lane wrote:\n> Tomas Vondra <tv@fuzzy.cz> writes:\n>> I believe the locale system (at the OS level) works just like before. I\n>> remember I had to manually create the locales while initially setting up\n>> the animals. Then, ~2 months ago something happened (I asssume a yum\n>> update) and some of the locales disappeared. But I have recreated them,\n>> except for sk_SK.WIN-1250. But the tests fail because of cs_CZ.WIN-1250\n>> which does exist.\n> \n> I am betting that you recreated them differently from before.\n\nAaaaand you're probably right. Apparently, I recreated them like this:\n\n $ localedef -v -c -i cs_CZ -f WIN-1250 cs_CZ.WIN-1250\n\nbut the correct way seems to be this:\n\n $ localedef -v -c -i cs_CZ -f CP1250 cs_CZ.WIN-1250\n\n>> However when I tried to initialize a\n>> cluster with cs_CZ.WIN-1250, I got an error like this:\n> \n>> [pgbuild@regular-builds ~]$ pg_ctl -D tmp-data init\n>> The files belonging to this database system will be owned by user \"pgbuild\".\n>> This user must also own the server process.\n> \n>> The database cluster will be initialized with locale \"cs_CZ.WIN-1250\".\n>> could not determine encoding for locale \"cs_CZ.WIN-1250\": codeset is\n>> \"ANSI_X3.4-1968\"\n>> initdb: could not find suitable encoding for locale \"cs_CZ.WIN-1250\"\n> \n> Locale cs_CZ.WIN-1250 is evidently marked with a codeset property of\n> \"ANSI_X3.4-1968\" (which means old-school US-ASCII). That's certainly\n> wrong. I believe the correct thing would be \"CP1250\".\n\nYes. I fixed the locales and added the locales back to the client\nconfiguration.\n\nregards\nTomas Vondra\n\n\n-- \nSent via 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 Dec 2014 19:28:33 +0100","msg_from":"Tomas Vondra <tv@fuzzy.cz>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"Tomas Vondra <tv@fuzzy.cz> writes:\n> On 20.12.2014 19:05, Tom Lane wrote:\n>> I am betting that you recreated them differently from before.\n\n> Aaaaand you're probably right. Apparently, I recreated them like this:\n\n> $ localedef -v -c -i cs_CZ -f WIN-1250 cs_CZ.WIN-1250\n\n> but the correct way seems to be this:\n\n> $ localedef -v -c -i cs_CZ -f CP1250 cs_CZ.WIN-1250\n\nInteresting. Apparently, instead of failing outright on an unrecognized\ncharmap name, localedef just substituted ASCII and plowed ahead. Bad dog.\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 Dec 2014 13:35:28 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"On 20.12.2014 19:35, Tom Lane wrote:\n> Tomas Vondra <tv@fuzzy.cz> writes:\n>> On 20.12.2014 19:05, Tom Lane wrote:\n>>> I am betting that you recreated them differently from before.\n> \n>> Aaaaand you're probably right. Apparently, I recreated them like this:\n> \n>> $ localedef -v -c -i cs_CZ -f WIN-1250 cs_CZ.WIN-1250\n> \n>> but the correct way seems to be this:\n> \n>> $ localedef -v -c -i cs_CZ -f CP1250 cs_CZ.WIN-1250\n> \n> Interesting. Apparently, instead of failing outright on an unrecognized\n> charmap name, localedef just substituted ASCII and plowed ahead. Bad dog.\n\nNot really. It's rather about abusive owner of the dog, using '-c' to\nforce the dog to create the locale even when there are warings:\n\n # localedef -i cs_CZ -f WIN-1250 cs_CZ.WIN-1250\n character map file `WIN-1250' not found: No such file or directory\n no output file produced because warnings were issued\n\nIn my defense, I've been using verbose mode, and that produces a lot of\nwarnings like 'non-symbolic character value should not be used' (which\ngets ignored in non-verbose mode) and thus missed the one important one.\n\nregards\nTomas\n\n\n-- \nSent via 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 Dec 2014 19:48:23 +0100","msg_from":"Tomas Vondra <tv@fuzzy.cz>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"On Sat, Dec 20, 2014 at 07:28:33PM +0100, Tomas Vondra wrote:\n> On 20.12.2014 19:05, Tom Lane wrote:\n> > Locale cs_CZ.WIN-1250 is evidently marked with a codeset property of\n> > \"ANSI_X3.4-1968\" (which means old-school US-ASCII). That's certainly\n> > wrong. I believe the correct thing would be \"CP1250\".\n> \n> Yes. I fixed the locales and added the locales back to the client\n> configuration.\n\nThanks. These animals are now OK except on REL9_0_STABLE. The log message of\ncommit 2dfa15d explains their ongoing 9.0 failures. I recommend adding\n--skip-steps=pl-install-check to their 9.0 invocations. Dropping the affected\nlocales is another option, but we benefit from the rare encoding coverage more\nthan we benefit from the src/bin/pl coverage.\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 23 Dec 2014 03:19:12 -0500","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"On 23.12.2014 09:19, Noah Misch wrote:\n> On Sat, Dec 20, 2014 at 07:28:33PM +0100, Tomas Vondra wrote:\n>> On 20.12.2014 19:05, Tom Lane wrote:\n>>> Locale cs_CZ.WIN-1250 is evidently marked with a codeset property of\n>>> \"ANSI_X3.4-1968\" (which means old-school US-ASCII). That's certainly\n>>> wrong. I believe the correct thing would be \"CP1250\".\n>>\n>> Yes. I fixed the locales and added the locales back to the client\n>> configuration.\n> \n> Thanks. These animals are now OK except on REL9_0_STABLE. The log message of\n> commit 2dfa15d explains their ongoing 9.0 failures. I recommend adding\n> --skip-steps=pl-install-check to their 9.0 invocations. Dropping the affected\n> locales is another option, but we benefit from the rare encoding coverage more\n> than we benefit from the src/bin/pl coverage.\n\nOK, can do. Something like this in build-farm.conf should work, right?\n\nif ($branch eq 'REL9_0_STABLE')\n{\n push(@{$conf{config_opts}},\"--skip-steps=pl-install-check\");\n}\n\n\nregards\nTomas\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 23 Dec 2014 13:14:08 +0100","msg_from":"Tomas Vondra <tv@fuzzy.cz>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"\nOn 12/23/2014 07:14 AM, Tomas Vondra wrote:\n> On 23.12.2014 09:19, Noah Misch wrote:\n>> On Sat, Dec 20, 2014 at 07:28:33PM +0100, Tomas Vondra wrote:\n>>> On 20.12.2014 19:05, Tom Lane wrote:\n>>>> Locale cs_CZ.WIN-1250 is evidently marked with a codeset property of\n>>>> \"ANSI_X3.4-1968\" (which means old-school US-ASCII). That's certainly\n>>>> wrong. I believe the correct thing would be \"CP1250\".\n>>> Yes. I fixed the locales and added the locales back to the client\n>>> configuration.\n>> Thanks. These animals are now OK except on REL9_0_STABLE. The log message of\n>> commit 2dfa15d explains their ongoing 9.0 failures. I recommend adding\n>> --skip-steps=pl-install-check to their 9.0 invocations. Dropping the affected\n>> locales is another option, but we benefit from the rare encoding coverage more\n>> than we benefit from the src/bin/pl coverage.\n> OK, can do. Something like this in build-farm.conf should work, right?\n>\n> if ($branch eq 'REL9_0_STABLE')\n> {\n> push(@{$conf{config_opts}},\"--skip-steps=pl-install-check\");\n> }\n>\n\n\nNo, config_opts is what's passed to configure. Try something like:\n\n if ($branch eq 'REL9_0_STABLE')\n {\n $skip_steps{'pl-install-check'} = 1;\n }\n\ncheers\n\nandrew\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 23 Dec 2014 09:21:01 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"On 23.12.2014 15:21, Andrew Dunstan wrote:\n> \n> No, config_opts is what's passed to configure. Try something like:\n> \n> if ($branch eq 'REL9_0_STABLE')\n> {\n> $skip_steps{'pl-install-check'} = 1;\n> }\n\nApplied to all three animals.\n\nTomas\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 23 Dec 2014 15:32:59 +0100","msg_from":"Tomas Vondra <tv@fuzzy.cz>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"On Tue, Dec 23, 2014 at 03:32:59PM +0100, Tomas Vondra wrote:\n> On 23.12.2014 15:21, Andrew Dunstan wrote:\n> > \n> > No, config_opts is what's passed to configure. Try something like:\n> > \n> > if ($branch eq 'REL9_0_STABLE')\n> > {\n> > $skip_steps{'pl-install-check'} = 1;\n> > }\n> \n> Applied to all three animals.\n\nAs of the time of this change, the animals ceased to report in.\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 27 Dec 2014 18:46:51 -0500","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"On 28.12.2014 00:46, Noah Misch wrote:\n> On Tue, Dec 23, 2014 at 03:32:59PM +0100, Tomas Vondra wrote:\n>> On 23.12.2014 15:21, Andrew Dunstan wrote:\n>>>\n>>> No, config_opts is what's passed to configure. Try something like:\n>>>\n>>> if ($branch eq 'REL9_0_STABLE')\n>>> {\n>>> $skip_steps{'pl-install-check'} = 1;\n>>> }\n>>\n>> Applied to all three animals.\n> \n> As of the time of this change, the animals ceased to report in.\n\nThe strange thing is that while the first run worked fine (as\nillustrated by the results submitted to pgbuildfarm.org), all the\nfollowing runs fail like this:\n\nGlobal symbol \"%skip_steps\" requires explicit package name at\nbuild-farm.conf line 308.\nCompilation failed in require at ./run_branches.pl line 55.\n\nApparently, something is broken, but my Perl-fu is limited so I have no\nidea what/why :-(\n\nregards\nTomas\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 31 Dec 2014 13:49:20 +0100","msg_from":"Tomas Vondra <tv@fuzzy.cz>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"\nOn 12/31/2014 07:49 AM, Tomas Vondra wrote:\n> On 28.12.2014 00:46, Noah Misch wrote:\n>> On Tue, Dec 23, 2014 at 03:32:59PM +0100, Tomas Vondra wrote:\n>>> On 23.12.2014 15:21, Andrew Dunstan wrote:\n>>>> No, config_opts is what's passed to configure. Try something like:\n>>>>\n>>>> if ($branch eq 'REL9_0_STABLE')\n>>>> {\n>>>> $skip_steps{'pl-install-check'} = 1;\n>>>> }\n>>> Applied to all three animals.\n>> As of the time of this change, the animals ceased to report in.\n> The strange thing is that while the first run worked fine (as\n> illustrated by the results submitted to pgbuildfarm.org), all the\n> following runs fail like this:\n>\n> Global symbol \"%skip_steps\" requires explicit package name at\n> build-farm.conf line 308.\n> Compilation failed in require at ./run_branches.pl line 55.\n>\n> Apparently, something is broken, but my Perl-fu is limited so I have no\n> idea what/why :-(\n>\n\n\n\nSorry, I should have tested it. This seems to work:\n\n if ($branch eq 'REL9_0_STABLE')\n {\n $PGBuild::Options::skip_steps .= ' pl-install-check';\n $main::skip_steps{'pl-install-check'} = 1;\n }\n\ncheers\n\nandrew\n\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 31 Dec 2014 11:29:42 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"On 31.12.2014 17:29, Andrew Dunstan wrote:\n> \n> Sorry, I should have tested it. This seems to work:\n> \n> if ($branch eq 'REL9_0_STABLE')\n> {\n> $PGBuild::Options::skip_steps .= ' pl-install-check';\n> $main::skip_steps{'pl-install-check'} = 1;\n> }\n> \n> cheers\n\nMeh, no problem. We've fixed it in 2014, so it's OK. To quote one of my\ncolleagues \"I haven't tested it, I believe it works correctly.\" ;-)\n\nAny ideas why it worked for the first run?\n\nregards\nTomas\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 31 Dec 2014 18:26:46 +0100","msg_from":"Tomas Vondra <tv@fuzzy.cz>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"},{"msg_contents":"\nOn 12/31/2014 12:26 PM, Tomas Vondra wrote:\n> On 31.12.2014 17:29, Andrew Dunstan wrote:\n>> Sorry, I should have tested it. This seems to work:\n>>\n>> if ($branch eq 'REL9_0_STABLE')\n>> {\n>> $PGBuild::Options::skip_steps .= ' pl-install-check';\n>> $main::skip_steps{'pl-install-check'} = 1;\n>> }\n>>\n>> cheers\n> Meh, no problem. We've fixed it in 2014, so it's OK. To quote one of my\n> colleagues \"I haven't tested it, I believe it works correctly.\" ;-)\n>\n> Any ideas why it worked for the first run?\n>\n\nNo. It failed for me right off the bat.\n\ncheers\n\nandrew\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 31 Dec 2014 12:39:04 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures"}],"string":"[\n {\n \"msg_contents\": \"Buildfarm members magpie, treepie and fulmar went absent on 2014-10-29. Since\\nreturning on 2014-11-16, they have consistently failed with 'initdb: invalid\\nlocale name \\\"cs_CZ.WIN-1250\\\"'. No commits in that period readily explain a\\nregression in this area. Did the underlying system configurations change?\\n\\nThanks,\\nnm\\n\\n\\n-- \\nSent via pgsql-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 Dec 2014 01:39:23 -0500\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"On 20.12.2014 07:39, Noah Misch wrote:\\n> Buildfarm members magpie, treepie and fulmar went absent on\\n> 2014-10-29. Since returning on 2014-11-16, they have consistently\\n> failed with 'initdb: invalid locale name \\\"cs_CZ.WIN-1250\\\"'. No\\n> commits in that period readily explain a regression in this area. Did\\n> the underlying system configurations change?\\n\\nI'm pretty sure it's because of broken locales at the system level. It was\\nworking fine, and I haven't done any substantial changes to the system\\nexcept for occasional \\\"yum update\\\", so I'm unaware of what went wong :(\\n\\nThe issue looks like this:\\n\\n# locale -a | grep en_US\\nen_US\\nen_US.iso88591\\nen_US.iso885915\\nen_US.utf8\\n\\n# locale en_US\\nlocale: unknown name \\\"en_US\\\"\\n\\nThe only reasons I can think of is that some of the updates required\\na reboot, and I haven't done that because that would kill all the VMs\\nrunning on that HW, including the one with CLOBBER_CACHE_RECURSIVE\\ntests. And that'd throw away tests running for ~3 months.\\n\\nI've disabled the three animals (magpie, fulmar, treepie) for now, because\\nthere's no point in running the tests until the locale issues are fixed. If\\nanyone has an idea of what might be wrong, let me know.\\n\\nTomas\\n\\n\\n\\n\\n\\n\\n On 20.12.2014 07:39, Noah Misch wrote:\\n> Buildfarm members magpie,\\n treepie and fulmar went absent on\\n > 2014-10-29. Since returning on 2014-11-16, they have\\n consistently\\n > failed with 'initdb: invalid locale name \\\"cs_CZ.WIN-1250\\\"'.\\n No\\n > commits in that period readily explain a regression in this\\n area. Did\\n > the underlying system configurations change?\\n\\n I'm pretty sure it's because of broken locales at the system level.\\n It was\\n working fine, and I haven't done any substantial changes to the\\n system\\n except for occasional \\\"yum update\\\", so I'm unaware of what went wong\\n :(\\n\\n The issue looks like this:\\n\\n # locale -a | grep en_US\\n en_US\\n en_US.iso88591\\n en_US.iso885915\\n en_US.utf8\\n\\n # locale en_US\\n locale: unknown name \\\"en_US\\\"\\n\\n The only reasons I can think of is that some of the updates required\\n a reboot, and I haven't done that because that would kill all the\\n VMs\\n running on that HW, including the one with CLOBBER_CACHE_RECURSIVE \\n tests. And that'd throw away tests running for ~3 months.\\n\\n I've disabled the three animals (magpie, fulmar, treepie) for now,\\n because\\n there's no point in running the tests until the locale issues are\\n fixed. If\\n anyone has an idea of what might be wrong, let me know.\\n\\n Tomas\",\n \"msg_date\": \"Sat, 20 Dec 2014 17:14:03 +0100\",\n \"msg_from\": \"CSPUG <info@cspug.cz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"On Sat, Dec 20, 2014 at 05:14:03PM +0100, CSPUG wrote:\\n> On 20.12.2014 07:39, Noah Misch wrote:\\n> > Buildfarm members magpie, treepie and fulmar went absent on\\n> > 2014-10-29. Since returning on 2014-11-16, they have consistently\\n> > failed with 'initdb: invalid locale name \\\"cs_CZ.WIN-1250\\\"'. No\\n> > commits in that period readily explain a regression in this area. Did\\n> > the underlying system configurations change?\\n> \\n> I'm pretty sure it's because of broken locales at the system level. It was\\n> working fine, and I haven't done any substantial changes to the system\\n> except for occasional \\\"yum update\\\", so I'm unaware of what went wong :(\\n> \\n> The issue looks like this:\\n> \\n> # locale -a | grep en_US\\n> en_US\\n> en_US.iso88591\\n> en_US.iso885915\\n> en_US.utf8\\n> \\n> # locale en_US\\n> locale: unknown name \\\"en_US\\\"\\n\\nThe NAME argument to the \\\"locale\\\" tool is something like LC_PAPER, not a\\nlocale name. Use \\\"LANG=en_US locale LC_NUMERIC\\\" to test locale loading.\\n\\n> The only reasons I can think of is that some of the updates required\\n> a reboot, and I haven't done that because that would kill all the VMs\\n> running on that HW, including the one with CLOBBER_CACHE_RECURSIVE\\n> tests. And that'd throw away tests running for ~3 months.\\n> \\n> I've disabled the three animals (magpie, fulmar, treepie) for now, because\\n> there's no point in running the tests until the locale issues are fixed. If\\n> anyone has an idea of what might be wrong, let me know.\\n\\nThose animals have been successfully completing initdb for several locales,\\nincluding en_US, before failing at cs_CZ.WIN-1250. You could disable just the\\ncs_CZ.WIN-1250 steps. A CentOS 6.6 system here also lacks such a locale:\\n\\n$ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\\nlocale: Cannot set LC_CTYPE to default locale: No such file or directory\\nlocale: Cannot set LC_MESSAGES to default locale: No such file or directory\\nlocale: Cannot set LC_ALL to default locale: No such file or directory\\n.\\n\\n-1\\n46\\n0\\nANSI_X3.4-1968\\n$ locale -a|grep cs_\\ncs_CZ\\ncs_CZ.iso88592\\ncs_CZ.utf8\\n\\nPerhaps an update made the system stricter about rejecting unknown locales.\\n\\n\\n-- \\nSent via pgsql-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 Dec 2014 11:48:43 -0500\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"Noah Misch <noah@leadboat.com> writes:\\n> On Sat, Dec 20, 2014 at 05:14:03PM +0100, CSPUG wrote:\\n>> On 20.12.2014 07:39, Noah Misch wrote:\\n>>> Buildfarm members magpie, treepie and fulmar went absent on\\n>>> 2014-10-29. Since returning on 2014-11-16, they have consistently\\n>>> failed with 'initdb: invalid locale name \\\"cs_CZ.WIN-1250\\\"'. No\\n>>> commits in that period readily explain a regression in this area. Did\\n>>> the underlying system configurations change?\\n\\nThat's what I'd been assuming, but I had not got round to inquiring.\\n\\n>> I'm pretty sure it's because of broken locales at the system level. It was\\n>> working fine, and I haven't done any substantial changes to the system\\n>> except for occasional \\\"yum update\\\", so I'm unaware of what went wong :(\\n>> \\n>> The only reasons I can think of is that some of the updates required\\n>> a reboot, and I haven't done that because that would kill all the VMs\\n>> running on that HW, including the one with CLOBBER_CACHE_RECURSIVE\\n>> tests. And that'd throw away tests running for ~3 months.\\n>> \\n>> I've disabled the three animals (magpie, fulmar, treepie) for now, because\\n>> there's no point in running the tests until the locale issues are fixed. If\\n>> anyone has an idea of what might be wrong, let me know.\\n\\n> Those animals have been successfully completing initdb for several locales,\\n> including en_US, before failing at cs_CZ.WIN-1250. You could disable just the\\n> cs_CZ.WIN-1250 steps. A CentOS 6.6 system here also lacks such a locale:\\n\\n> $ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\\n> locale: Cannot set LC_CTYPE to default locale: No such file or directory\\n> locale: Cannot set LC_MESSAGES to default locale: No such file or directory\\n> locale: Cannot set LC_ALL to default locale: No such file or directory\\n\\nFWIW, an actual RHEL 6.6 system doesn't like that either; at least not\\nwith a fairly minimal set of locales installed.\\n\\n> Perhaps an update made the system stricter about rejecting unknown locales.\\n\\nRed Hat released RHEL 6.6 on Oct 14. I am not sure how long it took the\\nCentOS crew to follow suit, but it's well within the realm of plausibility\\nthat when these buildfarm critters went offline it was associated with a\\nsystem update from 6.5 to 6.6 ... which was a *major* update. (On my\\nmachine, something like 480 out of 1500 packages were updated.) It's easy\\nto believe that either the glibc locale code is now stricter, or that Red\\nHat reshuffled the packaging a bit and the missing locale is now in a\\nnew package that didn't get installed.\\n\\nPersonally I'd not have tried to run such a system without rebooting ...\\nit's unlikely that not rebooting has anything directly to do with this\\nproblem, but I would not be surprised by issues elsewhere.\\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 Dec 2014 12:12:04 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"2014-12-20 17:48 GMT+01:00 Noah Misch <noah@leadboat.com>:\\n\\n> On Sat, Dec 20, 2014 at 05:14:03PM +0100, CSPUG wrote:\\n> > On 20.12.2014 07:39, Noah Misch wrote:\\n> > > Buildfarm members magpie, treepie and fulmar went absent on\\n> > > 2014-10-29. Since returning on 2014-11-16, they have consistently\\n> > > failed with 'initdb: invalid locale name \\\"cs_CZ.WIN-1250\\\"'. No\\n> > > commits in that period readily explain a regression in this area. Did\\n> > > the underlying system configurations change?\\n> >\\n> > I'm pretty sure it's because of broken locales at the system level. It\\n> was\\n> > working fine, and I haven't done any substantial changes to the system\\n> > except for occasional \\\"yum update\\\", so I'm unaware of what went wong :(\\n> >\\n> > The issue looks like this:\\n> >\\n> > # locale -a | grep en_US\\n> > en_US\\n> > en_US.iso88591\\n> > en_US.iso885915\\n> > en_US.utf8\\n> >\\n> > # locale en_US\\n> > locale: unknown name \\\"en_US\\\"\\n>\\n> The NAME argument to the \\\"locale\\\" tool is something like LC_PAPER, not a\\n> locale name. Use \\\"LANG=en_US locale LC_NUMERIC\\\" to test locale loading.\\n>\\n> > The only reasons I can think of is that some of the updates required\\n> > a reboot, and I haven't done that because that would kill all the VMs\\n> > running on that HW, including the one with CLOBBER_CACHE_RECURSIVE\\n> > tests. And that'd throw away tests running for ~3 months.\\n> >\\n> > I've disabled the three animals (magpie, fulmar, treepie) for now,\\n> because\\n> > there's no point in running the tests until the locale issues are fixed.\\n> If\\n> > anyone has an idea of what might be wrong, let me know.\\n>\\n> Those animals have been successfully completing initdb for several locales,\\n> including en_US, before failing at cs_CZ.WIN-1250. You could disable just\\n> the\\n> cs_CZ.WIN-1250 steps. A CentOS 6.6 system here also lacks such a locale:\\n>\\n> $ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\\n>\\n\\nIt is Microsoft encoding, - it is not available on Linux\\n\\nRegards\\n\\nPavel\\n\\n\\n> locale: Cannot set LC_CTYPE to default locale: No such file or directory\\n> locale: Cannot set LC_MESSAGES to default locale: No such file or directory\\n> locale: Cannot set LC_ALL to default locale: No such file or directory\\n> .\\n>\\n> -1\\n> 46\\n> 0\\n> ANSI_X3.4-1968\\n> $ locale -a|grep cs_\\n> cs_CZ\\n> cs_CZ.iso88592\\n> cs_CZ.utf8\\n>\\n> Perhaps an update made the system stricter about rejecting unknown locales.\\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\\n2014-12-20 17:48 GMT+01:00 Noah Misch <noah@leadboat.com>:On Sat, Dec 20, 2014 at 05:14:03PM +0100, CSPUG wrote:\\n> On 20.12.2014 07:39, Noah Misch wrote:\\n> > Buildfarm members magpie, treepie and fulmar went absent on\\n> > 2014-10-29. Since returning on 2014-11-16, they have consistently\\n> > failed with 'initdb: invalid locale name \\\"cs_CZ.WIN-1250\\\"'. No\\n> > commits in that period readily explain a regression in this area. Did\\n> > the underlying system configurations change?\\n>\\n> I'm pretty sure it's because of broken locales at the system level. It was\\n> working fine, and I haven't done any substantial changes to the system\\n> except for occasional \\\"yum update\\\", so I'm unaware of what went wong :(\\n>\\n> The issue looks like this:\\n>\\n> # locale -a | grep en_US\\n> en_US\\n> en_US.iso88591\\n> en_US.iso885915\\n> en_US.utf8\\n>\\n> # locale en_US\\n> locale: unknown name \\\"en_US\\\"\\n\\nThe NAME argument to the \\\"locale\\\" tool is something like LC_PAPER, not a\\nlocale name. Use \\\"LANG=en_US locale LC_NUMERIC\\\" to test locale loading.\\n\\n> The only reasons I can think of is that some of the updates required\\n> a reboot, and I haven't done that because that would kill all the VMs\\n> running on that HW, including the one with CLOBBER_CACHE_RECURSIVE\\n> tests. And that'd throw away tests running for ~3 months.\\n>\\n> I've disabled the three animals (magpie, fulmar, treepie) for now, because\\n> there's no point in running the tests until the locale issues are fixed. If\\n> anyone has an idea of what might be wrong, let me know.\\n\\nThose animals have been successfully completing initdb for several locales,\\nincluding en_US, before failing at cs_CZ.WIN-1250. You could disable just the\\ncs_CZ.WIN-1250 steps. A CentOS 6.6 system here also lacks such a locale:\\n\\n$ LANG=cs_CZ.WIN-1250 locale LC_NUMERICIt is Microsoft encoding, - it is not available on LinuxRegardsPavel \\nlocale: Cannot set LC_CTYPE to default locale: No such file or directory\\nlocale: Cannot set LC_MESSAGES to default locale: No such file or directory\\nlocale: Cannot set LC_ALL to default locale: No such file or directory\\n.\\n\\n-1\\n46\\n0\\nANSI_X3.4-1968\\n$ locale -a|grep cs_\\ncs_CZ\\ncs_CZ.iso88592\\ncs_CZ.utf8\\n\\nPerhaps an update made the system stricter about rejecting unknown locales.\\n\\n\\n--\\nSent via 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 Dec 2014 18:13:05 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"On 20.12.2014 18:13, Pavel Stehule wrote:\\n> \\n> 2014-12-20 17:48 GMT+01:00 Noah Misch <noah@leadboat.com\\n>\\n> $ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\\n> \\n> \\n> It is Microsoft encoding, - it is not available on Linux\\n\\nNot true. It is available on Linux, and the regression tests were\\nrunning with it for a long time (essentially from the moment magpie was\\nadded to the buildfarm).\\n\\nTomas\\n\\n\\n-- \\nSent via pgsql-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 Dec 2014 18:17:11 +0100\",\n \"msg_from\": \"Tomas Vondra <tv@fuzzy.cz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"2014-12-20 18:17 GMT+01:00 Tomas Vondra <tv@fuzzy.cz>:\\n\\n> On 20.12.2014 18:13, Pavel Stehule wrote:\\n> >\\n> > 2014-12-20 17:48 GMT+01:00 Noah Misch <noah@leadboat.com\\n> >\\n> > $ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\\n> >\\n> >\\n> > It is Microsoft encoding, - it is not available on Linux\\n>\\n> Not true. It is available on Linux, and the regression tests were\\n> running with it for a long time (essentially from the moment magpie was\\n> added to the buildfarm).\\n>\\n\\naha\\n\\nI am sorry for my mistake\\n\\nPavel\\n\\n\\n>\\n> Tomas\\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\\n2014-12-20 18:17 GMT+01:00 Tomas Vondra <tv@fuzzy.cz>:On 20.12.2014 18:13, Pavel Stehule wrote:\\n>\\n> 2014-12-20 17:48 GMT+01:00 Noah Misch <noah@leadboat.com\\n>\\n> $ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\\n>\\n>\\n> It is Microsoft encoding, - it is not available on Linux\\n\\nNot true. It is available on Linux, and the regression tests were\\nrunning with it for a long time (essentially from the moment magpie was\\nadded to the buildfarm).ahaI am sorry for my mistakePavel \\n\\nTomas\\n\\n\\n--\\nSent via 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 Dec 2014 18:25:17 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"Tomas Vondra <tv@fuzzy.cz> writes:\\n> On 20.12.2014 18:13, Pavel Stehule wrote:\\n>> It is Microsoft encoding, - it is not available on Linux\\n\\n> Not true. It is available on Linux, and the regression tests were\\n> running with it for a long time (essentially from the moment magpie was\\n> added to the buildfarm).\\n\\nWell, you had it at one time, but it sure doesn't seem to be there now.\\n\\nI poked around in the RHEL 6.6 release notes, and found this possibly\\nrelevant tidbit:\\n\\n\\t mingw component, BZ#1063396\\n\\n\\t Following the deprecation of Matahari packages in Red Hat\\n\\t Enterprise Linux 6.3, at which time the mingw packages were noted\\n\\t as deprecated, and the subsequent removal of Matahari packages\\n\\t from Red Hat Enterprise Linux 6.4, the mingw packages are now\\n\\t being removed from Red Hat Enterprise Linux 6.6.\\n\\n\\t The mingw packages will no longer be shipped in future Red Hat\\n\\t Enterprise Linux 6 minor releases, nor will they receive\\n\\t security-related updates. Consequently, users are advised to\\n\\t uninstall any earlier releases of the mingw packages from their\\n\\t Red Hat Enterprise Linux 6 systems.\\n\\nIt seems plausible that a WIN-1250 locale would have been something\\nthat would've been supplied by mingw rather than being part of either\\nglibc-common or any official locale package. I'm not sure how that\\ntranslates to it actively disappearing from your machine --- as the note\\nsays, you're \\\"advised to uninstall\\\" mingw, but I don't think the 6.6\\nupdate would have removed it automatically. Still, the outlines of a\\ntheory are starting to come into focus.\\n\\nImmediate recommendation is to restart the buildfarm critters with the\\nmissing locale removed from their configurations. If you can find the\\nlocale again somewhere, by all means re-enable it, but better less testing\\nthan no testing in the meantime.\\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 Dec 2014 12:32:33 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"\\nOn 12/20/2014 12:32 PM, Tom Lane wrote:\\n\\n>\\n> Immediate recommendation is to restart the buildfarm critters with the\\n> missing locale removed from their configurations. If you can find the\\n> locale again somewhere, by all means re-enable it, but better less testing\\n> than no testing in the meantime.\\n>\\n> \\t\\t\\t\\n\\n\\nYeah, all they need to do is alter the locales setting in the config \\nfile. If they want to force new runs, they can use \\n<https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Tips_and_Tricks>\\n\\n\\ncheers\\n\\nandrew\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 20 Dec 2014 12:50:34 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"On 20.12.2014 17:48, Noah Misch wrote:\\n> On Sat, Dec 20, 2014 at 05:14:03PM +0100, CSPUG wrote:\\n>> On 20.12.2014 07:39, Noah Misch wrote:\\n>>> Buildfarm members magpie, treepie and fulmar went absent on\\n>>> 2014-10-29. Since returning on 2014-11-16, they have consistently\\n>>> failed with 'initdb: invalid locale name \\\"cs_CZ.WIN-1250\\\"'. No\\n>>> commits in that period readily explain a regression in this area. Did\\n>>> the underlying system configurations change?\\n>>\\n>> I'm pretty sure it's because of broken locales at the system level. It was\\n>> working fine, and I haven't done any substantial changes to the system\\n>> except for occasional \\\"yum update\\\", so I'm unaware of what went wong :(\\n>>\\n>> The issue looks like this:\\n>>\\n>> # locale -a | grep en_US\\n>> en_US\\n>> en_US.iso88591\\n>> en_US.iso885915\\n>> en_US.utf8\\n>>\\n>> # locale en_US\\n>> locale: unknown name \\\"en_US\\\"\\n> \\n> The NAME argument to the \\\"locale\\\" tool is something like LC_PAPER, not a\\n> locale name. Use \\\"LANG=en_US locale LC_NUMERIC\\\" to test locale loading.\\n\\nD'oh! That kinda explains the strange `locale` failures, of course ...\\n\\nBut still, the three animals (magpie, fulmar, treepie) are all running\\nwith these locales:\\n\\n C en_US\\n cs_CZ.UTF-8 cs_CZ.ISO-8859-2 cs_CZ.WIN-1250\\n sk_SK.UTF-8 sk_SK.ISO-8859-2 sk_SK.WIN-1250\\n\\nand the only one missing (so the 'locale LC_NUMERIC' failed) was\\nsk_SK.WIN-1250. I've fixed that, and now all the locales work fine.\\n\\nThat however does not explain why the tests are failing with\\ncs_CZ.WIN-1250 because that locale was working for some time.\\n\\n> \\n>> The only reasons I can think of is that some of the updates required\\n>> a reboot, and I haven't done that because that would kill all the VMs\\n>> running on that HW, including the one with CLOBBER_CACHE_RECURSIVE\\n>> tests. And that'd throw away tests running for ~3 months.\\n>>\\n>> I've disabled the three animals (magpie, fulmar, treepie) for now, because\\n>> there's no point in running the tests until the locale issues are fixed. If\\n>> anyone has an idea of what might be wrong, let me know.\\n> \\n> Those animals have been successfully completing initdb for several locales,\\n> including en_US, before failing at cs_CZ.WIN-1250. You could disable just the\\n> cs_CZ.WIN-1250 steps. A CentOS 6.6 system here also lacks such a locale:\\n> \\n> $ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\\n> locale: Cannot set LC_CTYPE to default locale: No such file or directory\\n> locale: Cannot set LC_MESSAGES to default locale: No such file or directory\\n> locale: Cannot set LC_ALL to default locale: No such file or directory\\n> .\\n> \\n> -1\\n> 46\\n> 0\\n> ANSI_X3.4-1968\\n> $ locale -a|grep cs_\\n> cs_CZ\\n> cs_CZ.iso88592\\n> cs_CZ.utf8\\n> \\n> Perhaps an update made the system stricter about rejecting unknown locales.\\n\\nI believe the locale system (at the OS level) works just like before. I\\nremember I had to manually create the locales while initially setting up\\nthe animals. Then, ~2 months ago something happened (I asssume a yum\\nupdate) and some of the locales disappeared. But I have recreated them,\\nexcept for sk_SK.WIN-1250. But the tests fail because of cs_CZ.WIN-1250\\nwhich does exist.\\n\\nAttached is a log of\\n\\nfor l in cs_CZ.UTF-8 sk_SK.UTF-8 cs_CZ.ISO-8859-2 sk_SK.ISO-8859-2\\ncs_CZ.WIN-1250 sk_SK.WIN-1250; do\\n echo $l; LANG=$l locale LC_NUMERIC;\\ndone;\\n\\nshowing that all the locales exist. However when I tried to initialize a\\ncluster with cs_CZ.WIN-1250, I got an error like this:\\n\\n[pgbuild@regular-builds ~]$ locale\\nLANG=cs_CZ.WIN-1250\\nLC_CTYPE=\\\"cs_CZ.WIN-1250\\\"\\nLC_NUMERIC=\\\"cs_CZ.WIN-1250\\\"\\nLC_TIME=\\\"cs_CZ.WIN-1250\\\"\\nLC_COLLATE=\\\"cs_CZ.WIN-1250\\\"\\nLC_MONETARY=\\\"cs_CZ.WIN-1250\\\"\\nLC_MESSAGES=\\\"cs_CZ.WIN-1250\\\"\\nLC_PAPER=\\\"cs_CZ.WIN-1250\\\"\\nLC_NAME=\\\"cs_CZ.WIN-1250\\\"\\nLC_ADDRESS=\\\"cs_CZ.WIN-1250\\\"\\nLC_TELEPHONE=\\\"cs_CZ.WIN-1250\\\"\\nLC_MEASUREMENT=\\\"cs_CZ.WIN-1250\\\"\\nLC_IDENTIFICATION=\\\"cs_CZ.WIN-1250\\\"\\nLC_ALL=\\n[pgbuild@regular-builds ~]$ pg_ctl -D tmp-data init\\nThe files belonging to this database system will be owned by user \\\"pgbuild\\\".\\nThis user must also own the server process.\\n\\nThe database cluster will be initialized with locale \\\"cs_CZ.WIN-1250\\\".\\ncould not determine encoding for locale \\\"cs_CZ.WIN-1250\\\": codeset is\\n\\\"ANSI_X3.4-1968\\\"\\ninitdb: could not find suitable encoding for locale \\\"cs_CZ.WIN-1250\\\"\\nRerun initdb with the -E option.\\nTry \\\"initdb --help\\\" for more information.\\npg_ctl: database system initialization failed\\n\\nSo apparently the locale does exist, but we're unable to cope with it\\nfor some reason ...\\n\\nTomas\\n\\n\\n\\n-- \\nSent via 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 Dec 2014 18:55:30 +0100\",\n \"msg_from\": \"Tomas Vondra <tv@fuzzy.cz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"On 20.12.2014 18:32, Tom Lane wrote:\\n> Tomas Vondra <tv@fuzzy.cz> writes:\\n>> On 20.12.2014 18:13, Pavel Stehule wrote:\\n>>> It is Microsoft encoding, - it is not available on Linux\\n> \\n>> Not true. It is available on Linux, and the regression tests were \\n>> running with it for a long time (essentially from the moment magpie\\n>> was added to the buildfarm).\\n> \\n> Well, you had it at one time, but it sure doesn't seem to be there now.\\n>\\n> I poked around in the RHEL 6.6 release notes, and found this possibly\\n> relevant tidbit:\\n\\nAh, apparently I upgraded this VM to 6.6 - that wasn't exactly planned.\\n\\n> \\n> \\t mingw component, BZ#1063396\\n> \\n> \\t Following the deprecation of Matahari packages in Red Hat\\n> \\t Enterprise Linux 6.3, at which time the mingw packages were noted\\n> \\t as deprecated, and the subsequent removal of Matahari packages\\n> \\t from Red Hat Enterprise Linux 6.4, the mingw packages are now\\n> \\t being removed from Red Hat Enterprise Linux 6.6.\\n> \\n> \\t The mingw packages will no longer be shipped in future Red Hat\\n> \\t Enterprise Linux 6 minor releases, nor will they receive\\n> \\t security-related updates. Consequently, users are advised to\\n> \\t uninstall any earlier releases of the mingw packages from their\\n> \\t Red Hat Enterprise Linux 6 systems.\\n> \\n> It seems plausible that a WIN-1250 locale would have been something \\n> that would've been supplied by mingw rather than being part of\\n> either glibc-common or any official locale package. I'm not sure how\\n> that translates to it actively disappearing from your machine --- as\\n> the note says, you're \\\"advised to uninstall\\\" mingw, but I don't think\\n> the 6.6 update would have removed it automatically. Still, the\\n> outlines of a theory are starting to come into focus.\\n\\nI don't think so. As I mentioned in the previous post, I still can\\ncreate the locale, but initdb still fails with an error about\\nANSI_X3.4-1968 encoding.\\n\\nBut clearly, something changed between RH 6.5 and 6.6, because on 6.5 I\\nget this:\\n\\n$ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\\n,\\n�\\n3;3\\n44\\n160\\nCP1250\\n\\nwhile on 6.6 I get this:\\n\\n$ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\\n,\\n\\n3;3\\n44\\n160\\nANSI_X3.4-1968\\n\\nSo yeah, the locale definition did change a bit - the encoding is\\ndifferent, and we can't cope with it.\\n\\n> Immediate recommendation is to restart the buildfarm critters with\\n> the missing locale removed from their configurations. If you can find\\n> the locale again somewhere, by all means re-enable it, but better\\n> less testing than no testing in the meantime.\\n\\nI've removed cs_CZ.WIN-1250 and sk_SK.WIN-1250 from the config. Let's\\nsee how that works.\\n\\n\\nTomas\\n\\n\\n-- \\nSent via pgsql-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 Dec 2014 19:02:17 +0100\",\n \"msg_from\": \"Tomas Vondra <tv@fuzzy.cz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"Tomas Vondra <tv@fuzzy.cz> writes:\\n> I believe the locale system (at the OS level) works just like before. I\\n> remember I had to manually create the locales while initially setting up\\n> the animals. Then, ~2 months ago something happened (I asssume a yum\\n> update) and some of the locales disappeared. But I have recreated them,\\n> except for sk_SK.WIN-1250. But the tests fail because of cs_CZ.WIN-1250\\n> which does exist.\\n\\nI am betting that you recreated them differently from before.\\n\\n> However when I tried to initialize a\\n> cluster with cs_CZ.WIN-1250, I got an error like this:\\n\\n> [pgbuild@regular-builds ~]$ pg_ctl -D tmp-data init\\n> The files belonging to this database system will be owned by user \\\"pgbuild\\\".\\n> This user must also own the server process.\\n\\n> The database cluster will be initialized with locale \\\"cs_CZ.WIN-1250\\\".\\n> could not determine encoding for locale \\\"cs_CZ.WIN-1250\\\": codeset is\\n> \\\"ANSI_X3.4-1968\\\"\\n> initdb: could not find suitable encoding for locale \\\"cs_CZ.WIN-1250\\\"\\n\\nLocale cs_CZ.WIN-1250 is evidently marked with a codeset property of\\n\\\"ANSI_X3.4-1968\\\" (which means old-school US-ASCII). That's certainly\\nwrong. I believe the correct thing would be \\\"CP1250\\\".\\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 Dec 2014 13:05:33 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"Tomas Vondra <tv@fuzzy.cz> writes:\\n> But clearly, something changed between RH 6.5 and 6.6, because on 6.5 I\\n> get this:\\n\\n> $ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\\n> ,\\n> �\\n> 3;3\\n> 44\\n> 160\\n> CP1250\\n\\n> while on 6.6 I get this:\\n\\n> $ LANG=cs_CZ.WIN-1250 locale LC_NUMERIC\\n> ,\\n\\n> 3;3\\n> 44\\n> 160\\n> ANSI_X3.4-1968\\n\\nThat's certainly broken. The entire point of having a cs_CZ.WIN-1250\\nlocale (as opposed to cs_CZ.something-else) would be to specify a\\ncodeset corresponding to WIN-1250. Our code recognizes the spelling\\n\\\"CP1250\\\" for that. It's possible there are other spellings we should\\nrecognize, but \\\"ANSI_X3.4-1968\\\" is certainly not one. As I said, that\\njust means ASCII, so it's completely useless for determining which\\nASCII-superset encoding is wanted.\\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 Dec 2014 13:22:17 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"On 20.12.2014 19:05, Tom Lane wrote:\\n> Tomas Vondra <tv@fuzzy.cz> writes:\\n>> I believe the locale system (at the OS level) works just like before. I\\n>> remember I had to manually create the locales while initially setting up\\n>> the animals. Then, ~2 months ago something happened (I asssume a yum\\n>> update) and some of the locales disappeared. But I have recreated them,\\n>> except for sk_SK.WIN-1250. But the tests fail because of cs_CZ.WIN-1250\\n>> which does exist.\\n> \\n> I am betting that you recreated them differently from before.\\n\\nAaaaand you're probably right. Apparently, I recreated them like this:\\n\\n $ localedef -v -c -i cs_CZ -f WIN-1250 cs_CZ.WIN-1250\\n\\nbut the correct way seems to be this:\\n\\n $ localedef -v -c -i cs_CZ -f CP1250 cs_CZ.WIN-1250\\n\\n>> However when I tried to initialize a\\n>> cluster with cs_CZ.WIN-1250, I got an error like this:\\n> \\n>> [pgbuild@regular-builds ~]$ pg_ctl -D tmp-data init\\n>> The files belonging to this database system will be owned by user \\\"pgbuild\\\".\\n>> This user must also own the server process.\\n> \\n>> The database cluster will be initialized with locale \\\"cs_CZ.WIN-1250\\\".\\n>> could not determine encoding for locale \\\"cs_CZ.WIN-1250\\\": codeset is\\n>> \\\"ANSI_X3.4-1968\\\"\\n>> initdb: could not find suitable encoding for locale \\\"cs_CZ.WIN-1250\\\"\\n> \\n> Locale cs_CZ.WIN-1250 is evidently marked with a codeset property of\\n> \\\"ANSI_X3.4-1968\\\" (which means old-school US-ASCII). That's certainly\\n> wrong. I believe the correct thing would be \\\"CP1250\\\".\\n\\nYes. I fixed the locales and added the locales back to the client\\nconfiguration.\\n\\nregards\\nTomas Vondra\\n\\n\\n-- \\nSent via pgsql-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 Dec 2014 19:28:33 +0100\",\n \"msg_from\": \"Tomas Vondra <tv@fuzzy.cz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"Tomas Vondra <tv@fuzzy.cz> writes:\\n> On 20.12.2014 19:05, Tom Lane wrote:\\n>> I am betting that you recreated them differently from before.\\n\\n> Aaaaand you're probably right. Apparently, I recreated them like this:\\n\\n> $ localedef -v -c -i cs_CZ -f WIN-1250 cs_CZ.WIN-1250\\n\\n> but the correct way seems to be this:\\n\\n> $ localedef -v -c -i cs_CZ -f CP1250 cs_CZ.WIN-1250\\n\\nInteresting. Apparently, instead of failing outright on an unrecognized\\ncharmap name, localedef just substituted ASCII and plowed ahead. Bad dog.\\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 Dec 2014 13:35:28 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"On 20.12.2014 19:35, Tom Lane wrote:\\n> Tomas Vondra <tv@fuzzy.cz> writes:\\n>> On 20.12.2014 19:05, Tom Lane wrote:\\n>>> I am betting that you recreated them differently from before.\\n> \\n>> Aaaaand you're probably right. Apparently, I recreated them like this:\\n> \\n>> $ localedef -v -c -i cs_CZ -f WIN-1250 cs_CZ.WIN-1250\\n> \\n>> but the correct way seems to be this:\\n> \\n>> $ localedef -v -c -i cs_CZ -f CP1250 cs_CZ.WIN-1250\\n> \\n> Interesting. Apparently, instead of failing outright on an unrecognized\\n> charmap name, localedef just substituted ASCII and plowed ahead. Bad dog.\\n\\nNot really. It's rather about abusive owner of the dog, using '-c' to\\nforce the dog to create the locale even when there are warings:\\n\\n # localedef -i cs_CZ -f WIN-1250 cs_CZ.WIN-1250\\n character map file `WIN-1250' not found: No such file or directory\\n no output file produced because warnings were issued\\n\\nIn my defense, I've been using verbose mode, and that produces a lot of\\nwarnings like 'non-symbolic character value should not be used' (which\\ngets ignored in non-verbose mode) and thus missed the one important one.\\n\\nregards\\nTomas\\n\\n\\n-- \\nSent via pgsql-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 Dec 2014 19:48:23 +0100\",\n \"msg_from\": \"Tomas Vondra <tv@fuzzy.cz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"On Sat, Dec 20, 2014 at 07:28:33PM +0100, Tomas Vondra wrote:\\n> On 20.12.2014 19:05, Tom Lane wrote:\\n> > Locale cs_CZ.WIN-1250 is evidently marked with a codeset property of\\n> > \\\"ANSI_X3.4-1968\\\" (which means old-school US-ASCII). That's certainly\\n> > wrong. I believe the correct thing would be \\\"CP1250\\\".\\n> \\n> Yes. I fixed the locales and added the locales back to the client\\n> configuration.\\n\\nThanks. These animals are now OK except on REL9_0_STABLE. The log message of\\ncommit 2dfa15d explains their ongoing 9.0 failures. I recommend adding\\n--skip-steps=pl-install-check to their 9.0 invocations. Dropping the affected\\nlocales is another option, but we benefit from the rare encoding coverage more\\nthan we benefit from the src/bin/pl coverage.\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 23 Dec 2014 03:19:12 -0500\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"On 23.12.2014 09:19, Noah Misch wrote:\\n> On Sat, Dec 20, 2014 at 07:28:33PM +0100, Tomas Vondra wrote:\\n>> On 20.12.2014 19:05, Tom Lane wrote:\\n>>> Locale cs_CZ.WIN-1250 is evidently marked with a codeset property of\\n>>> \\\"ANSI_X3.4-1968\\\" (which means old-school US-ASCII). That's certainly\\n>>> wrong. I believe the correct thing would be \\\"CP1250\\\".\\n>>\\n>> Yes. I fixed the locales and added the locales back to the client\\n>> configuration.\\n> \\n> Thanks. These animals are now OK except on REL9_0_STABLE. The log message of\\n> commit 2dfa15d explains their ongoing 9.0 failures. I recommend adding\\n> --skip-steps=pl-install-check to their 9.0 invocations. Dropping the affected\\n> locales is another option, but we benefit from the rare encoding coverage more\\n> than we benefit from the src/bin/pl coverage.\\n\\nOK, can do. Something like this in build-farm.conf should work, right?\\n\\nif ($branch eq 'REL9_0_STABLE')\\n{\\n push(@{$conf{config_opts}},\\\"--skip-steps=pl-install-check\\\");\\n}\\n\\n\\nregards\\nTomas\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 23 Dec 2014 13:14:08 +0100\",\n \"msg_from\": \"Tomas Vondra <tv@fuzzy.cz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"\\nOn 12/23/2014 07:14 AM, Tomas Vondra wrote:\\n> On 23.12.2014 09:19, Noah Misch wrote:\\n>> On Sat, Dec 20, 2014 at 07:28:33PM +0100, Tomas Vondra wrote:\\n>>> On 20.12.2014 19:05, Tom Lane wrote:\\n>>>> Locale cs_CZ.WIN-1250 is evidently marked with a codeset property of\\n>>>> \\\"ANSI_X3.4-1968\\\" (which means old-school US-ASCII). That's certainly\\n>>>> wrong. I believe the correct thing would be \\\"CP1250\\\".\\n>>> Yes. I fixed the locales and added the locales back to the client\\n>>> configuration.\\n>> Thanks. These animals are now OK except on REL9_0_STABLE. The log message of\\n>> commit 2dfa15d explains their ongoing 9.0 failures. I recommend adding\\n>> --skip-steps=pl-install-check to their 9.0 invocations. Dropping the affected\\n>> locales is another option, but we benefit from the rare encoding coverage more\\n>> than we benefit from the src/bin/pl coverage.\\n> OK, can do. Something like this in build-farm.conf should work, right?\\n>\\n> if ($branch eq 'REL9_0_STABLE')\\n> {\\n> push(@{$conf{config_opts}},\\\"--skip-steps=pl-install-check\\\");\\n> }\\n>\\n\\n\\nNo, config_opts is what's passed to configure. Try something like:\\n\\n if ($branch eq 'REL9_0_STABLE')\\n {\\n $skip_steps{'pl-install-check'} = 1;\\n }\\n\\ncheers\\n\\nandrew\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 23 Dec 2014 09:21:01 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"On 23.12.2014 15:21, Andrew Dunstan wrote:\\n> \\n> No, config_opts is what's passed to configure. Try something like:\\n> \\n> if ($branch eq 'REL9_0_STABLE')\\n> {\\n> $skip_steps{'pl-install-check'} = 1;\\n> }\\n\\nApplied to all three animals.\\n\\nTomas\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 23 Dec 2014 15:32:59 +0100\",\n \"msg_from\": \"Tomas Vondra <tv@fuzzy.cz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"On Tue, Dec 23, 2014 at 03:32:59PM +0100, Tomas Vondra wrote:\\n> On 23.12.2014 15:21, Andrew Dunstan wrote:\\n> > \\n> > No, config_opts is what's passed to configure. Try something like:\\n> > \\n> > if ($branch eq 'REL9_0_STABLE')\\n> > {\\n> > $skip_steps{'pl-install-check'} = 1;\\n> > }\\n> \\n> Applied to all three animals.\\n\\nAs of the time of this change, the animals ceased to report in.\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 27 Dec 2014 18:46:51 -0500\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"On 28.12.2014 00:46, Noah Misch wrote:\\n> On Tue, Dec 23, 2014 at 03:32:59PM +0100, Tomas Vondra wrote:\\n>> On 23.12.2014 15:21, Andrew Dunstan wrote:\\n>>>\\n>>> No, config_opts is what's passed to configure. Try something like:\\n>>>\\n>>> if ($branch eq 'REL9_0_STABLE')\\n>>> {\\n>>> $skip_steps{'pl-install-check'} = 1;\\n>>> }\\n>>\\n>> Applied to all three animals.\\n> \\n> As of the time of this change, the animals ceased to report in.\\n\\nThe strange thing is that while the first run worked fine (as\\nillustrated by the results submitted to pgbuildfarm.org), all the\\nfollowing runs fail like this:\\n\\nGlobal symbol \\\"%skip_steps\\\" requires explicit package name at\\nbuild-farm.conf line 308.\\nCompilation failed in require at ./run_branches.pl line 55.\\n\\nApparently, something is broken, but my Perl-fu is limited so I have no\\nidea what/why :-(\\n\\nregards\\nTomas\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 31 Dec 2014 13:49:20 +0100\",\n \"msg_from\": \"Tomas Vondra <tv@fuzzy.cz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"\\nOn 12/31/2014 07:49 AM, Tomas Vondra wrote:\\n> On 28.12.2014 00:46, Noah Misch wrote:\\n>> On Tue, Dec 23, 2014 at 03:32:59PM +0100, Tomas Vondra wrote:\\n>>> On 23.12.2014 15:21, Andrew Dunstan wrote:\\n>>>> No, config_opts is what's passed to configure. Try something like:\\n>>>>\\n>>>> if ($branch eq 'REL9_0_STABLE')\\n>>>> {\\n>>>> $skip_steps{'pl-install-check'} = 1;\\n>>>> }\\n>>> Applied to all three animals.\\n>> As of the time of this change, the animals ceased to report in.\\n> The strange thing is that while the first run worked fine (as\\n> illustrated by the results submitted to pgbuildfarm.org), all the\\n> following runs fail like this:\\n>\\n> Global symbol \\\"%skip_steps\\\" requires explicit package name at\\n> build-farm.conf line 308.\\n> Compilation failed in require at ./run_branches.pl line 55.\\n>\\n> Apparently, something is broken, but my Perl-fu is limited so I have no\\n> idea what/why :-(\\n>\\n\\n\\n\\nSorry, I should have tested it. This seems to work:\\n\\n if ($branch eq 'REL9_0_STABLE')\\n {\\n $PGBuild::Options::skip_steps .= ' pl-install-check';\\n $main::skip_steps{'pl-install-check'} = 1;\\n }\\n\\ncheers\\n\\nandrew\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 31 Dec 2014 11:29:42 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"On 31.12.2014 17:29, Andrew Dunstan wrote:\\n> \\n> Sorry, I should have tested it. This seems to work:\\n> \\n> if ($branch eq 'REL9_0_STABLE')\\n> {\\n> $PGBuild::Options::skip_steps .= ' pl-install-check';\\n> $main::skip_steps{'pl-install-check'} = 1;\\n> }\\n> \\n> cheers\\n\\nMeh, no problem. We've fixed it in 2014, so it's OK. To quote one of my\\ncolleagues \\\"I haven't tested it, I believe it works correctly.\\\" ;-)\\n\\nAny ideas why it worked for the first run?\\n\\nregards\\nTomas\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 31 Dec 2014 18:26:46 +0100\",\n \"msg_from\": \"Tomas Vondra <tv@fuzzy.cz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n },\n {\n \"msg_contents\": \"\\nOn 12/31/2014 12:26 PM, Tomas Vondra wrote:\\n> On 31.12.2014 17:29, Andrew Dunstan wrote:\\n>> Sorry, I should have tested it. This seems to work:\\n>>\\n>> if ($branch eq 'REL9_0_STABLE')\\n>> {\\n>> $PGBuild::Options::skip_steps .= ' pl-install-check';\\n>> $main::skip_steps{'pl-install-check'} = 1;\\n>> }\\n>>\\n>> cheers\\n> Meh, no problem. We've fixed it in 2014, so it's OK. To quote one of my\\n> colleagues \\\"I haven't tested it, I believe it works correctly.\\\" ;-)\\n>\\n> Any ideas why it worked for the first run?\\n>\\n\\nNo. It failed for me right off the bat.\\n\\ncheers\\n\\nandrew\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 31 Dec 2014 12:39:04 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Initdb-cs_CZ.WIN-1250 buildfarm failures\"\n }\n]","isConversational":false}}},{"rowIdx":39991,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"PostgreSQL 9.2 and later support Python 3.3 and Python 3.4; PostgreSQL 9.1 and\n9.0 do not. Buildfarm member raccoon upgraded from Python 2.7 to Python 3.3\nfor its 2013-08-18 run, and it has failed for REL9_1_STABLE and REL9_0_STABLE\never since. Please omit --with-python for those branches. (Pointing to an\nolder Python is also an option.) Incidentally, raccoon REL9_4_STABLE has\njammed for some time on a stale git lock file. As best I recall, you can just\nremove /usr/src/pgfarm/build/REL9_4_STABLE/pgsql and let the buildfarm client\nre-clone it.\n\nBuildfarm member jaguarundi, which has Python 3.4, activated --with-python for\nREL9_1_STABLE as of its 2014-12-15 run. Please remove --with-python or test\nagainst an older Python. It already omits --with-python for REL9_0_STABLE.\n\nThanks,\nnm\n\n\n-- \nSent via 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 Dec 2014 02:22:03 -0500","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"New Python vs. old PG on raccoon and jaguarundi"},{"msg_contents":"* From: Noah Misch [mailto:noah@leadboat.com]\n\n> Buildfarm member jaguarundi, which has Python 3.4, activated --with-\n> python for REL9_1_STABLE as of its 2014-12-15 run. Please remove --\n> with-python or test against an older Python. It already omits --with-\n> python for REL9_0_STABLE.\n\nDone; sorry for the noise.\n\n-- \nChristian\n\n\n\n-- \nSent via 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 Dec 2014 08:06:13 +0000","msg_from":"Christian Ullrich <chris@chrullrich.net>","msg_from_op":false,"msg_subject":"Re: New Python vs. old PG on raccoon and jaguarundi"}],"string":"[\n {\n \"msg_contents\": \"PostgreSQL 9.2 and later support Python 3.3 and Python 3.4; PostgreSQL 9.1 and\\n9.0 do not. Buildfarm member raccoon upgraded from Python 2.7 to Python 3.3\\nfor its 2013-08-18 run, and it has failed for REL9_1_STABLE and REL9_0_STABLE\\never since. Please omit --with-python for those branches. (Pointing to an\\nolder Python is also an option.) Incidentally, raccoon REL9_4_STABLE has\\njammed for some time on a stale git lock file. As best I recall, you can just\\nremove /usr/src/pgfarm/build/REL9_4_STABLE/pgsql and let the buildfarm client\\nre-clone it.\\n\\nBuildfarm member jaguarundi, which has Python 3.4, activated --with-python for\\nREL9_1_STABLE as of its 2014-12-15 run. Please remove --with-python or test\\nagainst an older Python. It already omits --with-python for REL9_0_STABLE.\\n\\nThanks,\\nnm\\n\\n\\n-- \\nSent via pgsql-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 Dec 2014 02:22:03 -0500\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"New Python vs. old PG on raccoon and jaguarundi\"\n },\n {\n \"msg_contents\": \"* From: Noah Misch [mailto:noah@leadboat.com]\\n\\n> Buildfarm member jaguarundi, which has Python 3.4, activated --with-\\n> python for REL9_1_STABLE as of its 2014-12-15 run. Please remove --\\n> with-python or test against an older Python. It already omits --with-\\n> python for REL9_0_STABLE.\\n\\nDone; sorry for the noise.\\n\\n-- \\nChristian\\n\\n\\n\\n-- \\nSent via pgsql-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 Dec 2014 08:06:13 +0000\",\n \"msg_from\": \"Christian Ullrich <chris@chrullrich.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: New Python vs. old PG on raccoon and jaguarundi\"\n }\n]","isConversational":false}}},{"rowIdx":39992,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi all,\n\nI work with some customer that have databases with a lot of schemas and\nsometimes we need to run manual VACUUM in one schema, and would be nice to\nhave a new option to run vacuum in relations from a specific schema.\n\nThe new syntax could be:\n\nVACUUM [ FULL ] [ FREEZE ] [ VERBOSE ] { [ table_name ] | SCHEMA\nschema_name }\n\nAlso I'll add a new option to \"vacuumdb\" client:\n\n-S, --schema=SCHEMA\n\nI can work on this feature to 2015/02 CF.\n\nThoughts?\n\n--\nFabrízio de Royes Mello\nConsultoria/Coaching PostgreSQL\n>> Timbira: http://www.timbira.com.br\n>> Blog: http://fabriziomello.github.io\n>> Linkedin: http://br.linkedin.com/in/fabriziomello\n>> Twitter: http://twitter.com/fabriziomello\n>> Github: http://github.com/fabriziomello\n\nHi all,I work with some customer that have databases with a lot of schemas and sometimes we need to run manual VACUUM in one schema, and would be nice to have a new option to run vacuum in relations from a specific schema.The new syntax could be:VACUUM [ FULL ] [ FREEZE ] [ VERBOSE ] { [ table_name ] | SCHEMA schema_name }Also I'll add a new option to \"vacuumdb\" client:-S, --schema=SCHEMAI can work on this feature to 2015/02 CF.Thoughts? --Fabrízio de Royes MelloConsultoria/Coaching PostgreSQL>> Timbira: http://www.timbira.com.br>> Blog: http://fabriziomello.github.io>> Linkedin: http://br.linkedin.com/in/fabriziomello>> Twitter: http://twitter.com/fabriziomello>> Github: http://github.com/fabriziomello","msg_date":"Sun, 21 Dec 2014 14:48:16 -0200","msg_from":"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com>","msg_from_op":true,"msg_subject":"Proposal \"VACUUM SCHEMA\""},{"msg_contents":"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com> writes:\n> I work with some customer that have databases with a lot of schemas and\n> sometimes we need to run manual VACUUM in one schema, and would be nice to\n> have a new option to run vacuum in relations from a specific schema.\n\nI'm pretty skeptical of this alleged use-case. Manual vacuuming ought\nto be mostly a thing of the past, and even if it's not, hitting\n*everything* in a schema should seldom be an appropriate thing to do.\n\nWhile the feature itself might be fairly innocuous, I'm just wondering\nwhy we need to encourage manual vacuuming. And why that, but not\nsay schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\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, 21 Dec 2014 14:18:33 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"On Sun, Dec 21, 2014 at 5:18 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> =?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com> writes:\n> > I work with some customer that have databases with a lot of schemas and\n> > sometimes we need to run manual VACUUM in one schema, and would be nice\nto\n> > have a new option to run vacuum in relations from a specific schema.\n>\n> I'm pretty skeptical of this alleged use-case. Manual vacuuming ought\n> to be mostly a thing of the past, and even if it's not, hitting\n> *everything* in a schema should seldom be an appropriate thing to do.\n>\n\nI agree manual vacuum is a thing of the past, but autovacuum doesn't solve\n100% of the cases, and sometimes we need to use it so my proposal is just\ndo help DBAs and/or Sysadmins to write simple maintenance scripts.\n\n\n> While the feature itself might be fairly innocuous, I'm just wondering\n> why we need to encourage manual vacuuming.\n\nIMHO we will not encourage manual vacuuming, just give more flexibility to\nusers.\n\n\n> And why that, but not\n> say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\n>\n\n+1. I can write patches for each of this maintenance statement too.\n\nRegards,\n\n--\nFabrízio de Royes Mello\nConsultoria/Coaching PostgreSQL\n>> Timbira: http://www.timbira.com.br\n>> Blog: http://fabriziomello.github.io\n>> Linkedin: http://br.linkedin.com/in/fabriziomello\n>> Twitter: http://twitter.com/fabriziomello\n>> Github: http://github.com/fabriziomello\n\nOn Sun, Dec 21, 2014 at 5:18 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:>> =?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com> writes:> > I work with some customer that have databases with a lot of schemas and> > sometimes we need to run manual VACUUM in one schema, and would be nice to> > have a new option to run vacuum in relations from a specific schema.>> I'm pretty skeptical of this alleged use-case. Manual vacuuming ought> to be mostly a thing of the past, and even if it's not, hitting> *everything* in a schema should seldom be an appropriate thing to do.>I agree manual vacuum is a thing of the past, but autovacuum doesn't solve 100% of the cases, and sometimes we need to use it so my proposal is just do help DBAs and/or Sysadmins to write simple maintenance scripts.> While the feature itself might be fairly innocuous, I'm just wondering> why we need to encourage manual vacuuming. IMHO we will not encourage manual vacuuming, just give more flexibility to users.> And why that, but not> say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...>+1. I can write patches for each of this maintenance statement too.Regards,--Fabrízio de Royes MelloConsultoria/Coaching PostgreSQL>> Timbira: http://www.timbira.com.br>> Blog: http://fabriziomello.github.io>> Linkedin: http://br.linkedin.com/in/fabriziomello>> Twitter: http://twitter.com/fabriziomello>> Github: http://github.com/fabriziomello","msg_date":"Sun, 21 Dec 2014 19:30:35 -0200","msg_from":"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com>","msg_from_op":true,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"On 12/21/14, 3:30 PM, Fabrízio de Royes Mello wrote:\n>\n> On Sun, Dec 21, 2014 at 5:18 PM, Tom Lane <tgl@sss.pgh.pa.us <mailto:tgl@sss.pgh.pa.us>> wrote:\n> >\n> > =?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com <mailto:fabriziomello@gmail.com>> writes:\n> > > I work with some customer that have databases with a lot of schemas and\n> > > sometimes we need to run manual VACUUM in one schema, and would be nice to\n> > > have a new option to run vacuum in relations from a specific schema.\n> >\n> > I'm pretty skeptical of this alleged use-case. Manual vacuuming ought\n> > to be mostly a thing of the past, and even if it's not, hitting\n> > *everything* in a schema should seldom be an appropriate thing to do.\n> >\n>\n> I agree manual vacuum is a thing of the past, but autovacuum doesn't solve 100% of the cases, and sometimes we need to use it so my proposal is just do help DBAs and/or Sysadmins to write simple maintenance scripts.\n\nJust one example of that is pre-emptively vacuuming during slower periods. Nothing spells \"fun\" like a freeze vacuum in the middle of a busy lunch period for a website.\n\nSimilarly, it's common to need to proactively vacuum after a data load, and since it's not unusual for there to be a schema dedicated to loading data, this makes that easier.\n\n> > And why that, but not\n> > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\n> >\n>\n> +1. I can write patches for each of this maintenance statement too.\n\nIf we're going to go that route, then perhaps it would make more sense to create a command that allows you to apply a second command to every object in a schema. We would have to be careful about PreventTransactionChain commands.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sun, 21 Dec 2014 20:32:32 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"Em segunda-feira, 22 de dezembro de 2014, Jim Nasby <\nJim.Nasby@bluetreble.com> escreveu:\n\n> On 12/21/14, 3:30 PM, Fabrízio de Royes Mello wrote:\n>\n>>\n>> On Sun, Dec 21, 2014 at 5:18 PM, Tom Lane <tgl@sss.pgh.pa.us <mailto:\n>> tgl@sss.pgh.pa.us>> wrote:\n>> >\n>> > =?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com\n>> <mailto:fabriziomello@gmail.com>> writes:\n>> > > I work with some customer that have databases with a lot of schemas\n>> and\n>> > > sometimes we need to run manual VACUUM in one schema, and would be\n>> nice to\n>> > > have a new option to run vacuum in relations from a specific schema.\n>> >\n>> > I'm pretty skeptical of this alleged use-case. Manual vacuuming ought\n>> > to be mostly a thing of the past, and even if it's not, hitting\n>> > *everything* in a schema should seldom be an appropriate thing to do.\n>> >\n>>\n>> I agree manual vacuum is a thing of the past, but autovacuum doesn't\n>> solve 100% of the cases, and sometimes we need to use it so my proposal is\n>> just do help DBAs and/or Sysadmins to write simple maintenance scripts.\n>>\n>\n> Just one example of that is pre-emptively vacuuming during slower periods.\n> Nothing spells \"fun\" like a freeze vacuum in the middle of a busy lunch\n> period for a website.\n>\n> Similarly, it's common to need to proactively vacuum after a data load,\n> and since it's not unusual for there to be a schema dedicated to loading\n> data, this makes that easier.\n\n\nGood example. Thanks.\n\n\n\n>\n> > And why that, but not\n>> > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\n>> >\n>>\n>> +1. I can write patches for each of this maintenance statement too.\n>>\n>\n> If we're going to go that route, then perhaps it would make more sense to\n> create a command that allows you to apply a second command to every object\n> in a schema. We would have to be careful about PreventTransactionChain\n> commands.\n\n\n Sorry but I don't understand what you meant. Can you explain more about\nyour idea?\n\nRegards,\n\nFabrízio Mello\n\n\n-- \nFabrízio de Royes Mello\nConsultoria/Coaching PostgreSQL\n>> Timbira: http://www.timbira.com.br\n>> Blog: http://fabriziomello.github.io\n>> Linkedin: http://br.linkedin.com/in/fabriziomello\n>> Twitter: http://twitter.com/fabriziomello\n>> Github: http://github.com/fabriziomello\n\nEm segunda-feira, 22 de dezembro de 2014, Jim Nasby <Jim.Nasby@bluetreble.com> escreveu:On 12/21/14, 3:30 PM, Fabrízio de Royes Mello wrote:\n\n\nOn Sun, Dec 21, 2014 at 5:18 PM, Tom Lane <tgl@sss.pgh.pa.us <mailto:tgl@sss.pgh.pa.us>> wrote:\n >\n > =?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com <mailto:fabriziomello@gmail.com>> writes:\n > > I work with some customer that have databases with a lot of schemas and\n > > sometimes we need to run manual VACUUM in one schema, and would be nice to\n > > have a new option to run vacuum in relations from a specific schema.\n >\n > I'm pretty skeptical of this alleged use-case. Manual vacuuming ought\n > to be mostly a thing of the past, and even if it's not, hitting\n > *everything* in a schema should seldom be an appropriate thing to do.\n >\n\nI agree manual vacuum is a thing of the past, but autovacuum doesn't solve 100% of the cases, and sometimes we need to use it so my proposal is just do help DBAs and/or Sysadmins to write simple maintenance scripts.\n\n\nJust one example of that is pre-emptively vacuuming during slower periods. Nothing spells \"fun\" like a freeze vacuum in the middle of a busy lunch period for a website.\n\nSimilarly, it's common to need to proactively vacuum after a data load, and since it's not unusual for there to be a schema dedicated to loading data, this makes that easier.Good example. Thanks. \n\n\n > And why that, but not\n > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\n >\n\n+1. I can write patches for each of this maintenance statement too.\n\n\nIf we're going to go that route, then perhaps it would make more sense to create a command that allows you to apply a second command to every object in a schema. We would have to be careful about PreventTransactionChain commands. Sorry but I don't understand what you meant. Can you explain more about your idea?Regards,Fabrízio Mello-- Fabrízio de Royes MelloConsultoria/Coaching PostgreSQL>> Timbira: http://www.timbira.com.br>> Blog: http://fabriziomello.github.io>> Linkedin: http://br.linkedin.com/in/fabriziomello>> Twitter: http://twitter.com/fabriziomello>> Github: http://github.com/fabriziomello","msg_date":"Mon, 22 Dec 2014 00:55:14 -0200","msg_from":"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com>","msg_from_op":true,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"On 12/21/2014 10:30 PM, Fabrízio de Royes Mello wrote:\n> [snip]\n\nI do agree that \"vacuum schema\" might very well be useful (I'll probably \nuse it myself from time to time, too).\nANALYZE SCHEMA (specially coupled with some transaction-wide \"SET \nstatistics_target\" could be beneficial)\n\n>\n> > And why that, but not\n> > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\n> >\n>\n> +1. I can write patches for each of this maintenance statement too.\n\nHmm... I think Tom might have been a bit rethorical (or even sarcastic \nwith that), but I can definitely be wrong.\n\nDo we really want to have some such operation potentially (and \ninadvertently) locking for *hours* at a time?\n\nCLUSTER SCHEMA somename;\n\n ... where schema \"somename\" contains \"myHugeTable\"\n\n Given that the cluster command exclusively locks and rewrites the \ntable, it might lock queries and overwhelm the I/O subsystem for quite a \nlong time.\n\n\nTRUNCATE SCHEMA whatever sounds quite dangerous, too.\n\n\n\nJust my .02€\n\n / J.L.\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, 22 Dec 2014 16:55:50 +0100","msg_from":"=?UTF-8?B?Sm9zw6kgTHVpcyBUYWxsw7Nu?= <jltallon@adv-solutions.net>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"On 2014-12-21 14:18:33 -0500, Tom Lane wrote:\n> =?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com> writes:\n> > I work with some customer that have databases with a lot of schemas and\n> > sometimes we need to run manual VACUUM in one schema, and would be nice to\n> > have a new option to run vacuum in relations from a specific schema.\n> \n> I'm pretty skeptical of this alleged use-case. Manual vacuuming ought\n> to be mostly a thing of the past, and even if it's not, hitting\n> *everything* in a schema should seldom be an appropriate thing to do.\n\nBased on my experience autovacuum isn't sufficient on bigger high\nthroughput databases. At the very least manual vacuuming with lower\nfreeze_table_age settings during low-load times is required lest\nanti-wraparound vacuums increase load too much during prime business\nhours.\nThat said, I don't see how this feature is actually helpful in those\ncases. In pretty much all of what I've seen you'd want to have more\ncomplex selection criteria than the schema.\n\n> While the feature itself might be fairly innocuous, I'm just wondering\n> why we need to encourage manual vacuuming. And why that, but not\n> say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\n\nThere's one argument for supporting more for VACUUM than the rest - it\ncan't be executed directly as the result of a query as the others\ncan... I wonder if that'd not better be answered by adding a feature to\nvacuumdb that allows selecting the to-be-vacuumed table by a user\ndefined query.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 22 Dec 2014 17:05:49 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"Jos� Luis Tall�n wrote:\n> On 12/21/2014 10:30 PM, Fabr�zio de Royes Mello wrote:\n> >[snip]\n> \n> I do agree that \"vacuum schema\" might very well be useful (I'll probably use\n> it myself from time to time, too).\n> ANALYZE SCHEMA (specially coupled with some transaction-wide \"SET\n> statistics_target\" could be beneficial)\n\nWe already have transanction-wide SET -- it's spelled SET LOCAL.\n\n> >\n> >> And why that, but not say schema-wide ANALYZE, CLUSTER, TRUNCATE,\n> >> ...\n> >\n> >+1. I can write patches for each of this maintenance statement too.\n> \n> Hmm... I think Tom might have been a bit rethorical (or even sarcastic with\n> that),\n\nThat was my impression too.\n\n> Do we really want to have some such operation potentially (and\n> inadvertently) locking for *hours* at a time?\n> \n> CLUSTER SCHEMA somename;\n> \n> ... where schema \"somename\" contains \"myHugeTable\"\n> \n> Given that the cluster command exclusively locks and rewrites the table,\n> it might lock queries and overwhelm the I/O subsystem for quite a long time.\n\nMulti-table CLUSTER uses multiple transactions, so this should not be an\nissue. That said, I don't think there's much point in CLUSTER SCHEMA,\nmuch less TRUNCATE SCHEMA. Do you normally organize your schemas so\nthat there are some that contain only tables that need to be truncated\ntogether? That would be a strange use case.\n\nOverall, this whole line of development seems like bloating the parse\ntables for little gain.\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":"Mon, 22 Dec 2014 13:51:57 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"Re: Alvaro Herrera 2014-12-22 <20141222165157.GD1768@alvh.no-ip.org>\n> Multi-table CLUSTER uses multiple transactions, so this should not be an\n> issue. That said, I don't think there's much point in CLUSTER SCHEMA,\n> much less TRUNCATE SCHEMA. Do you normally organize your schemas so\n> that there are some that contain only tables that need to be truncated\n> together? That would be a strange use case.\n\nHaving a schema that's only used for importing data in batch jobs\ndoesn't sound too unreasonable. It could then be cleaned in a simple\n\"TRUNCATE SCHEMA import_area\" command.\n\n> Overall, this whole line of development seems like bloating the parse\n> tables for little gain.\n\nReading the thread, my impression was that most people opposed the\nidea because there's ways to script \"vacuum schema\", or because of\n\"people shouldn't be invoking manual vacuums anyway\". I think the\npatch tries to solve a practical problem, and does have its merits.\n\nChristoph\n-- \ncb@df7cb.de | http://www.df7cb.de/\n\n\n-- \nSent via 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 Dec 2014 17:58:49 +0100","msg_from":"Christoph Berg <cb@df7cb.de>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"* Alvaro Herrera (alvherre@2ndquadrant.com) wrote:\n> Multi-table CLUSTER uses multiple transactions, so this should not be an\n> issue. That said, I don't think there's much point in CLUSTER SCHEMA,\n> much less TRUNCATE SCHEMA. Do you normally organize your schemas so\n> that there are some that contain only tables that need to be truncated\n> together? That would be a strange use case.\n\nI could see it happening in environments which use schemas when doing\npartitioning. eg: data_2014 contains all of the data_201401-201412\nmonthly (or perhaps weekly) tables.\n\n> Overall, this whole line of development seems like bloating the parse\n> tables for little gain.\n\nStill, I see this point also. I do think it'd be really great if we\ncould figure out a way to segregate these kinds of DDL / maintenance\ncommands from the normal select/insert/update/delete SQL parsing, such\nthat we could add more options, etc, to those longer running and less\nfrequent commands without impacting parse time for the high-volume\ncommands.\n\nI'm less concerned about the memory impact, except to the extent that it\nimpacts throughput and performance.\n\n\tThanks,\n\n\t\tStephen","msg_date":"Mon, 22 Dec 2014 12:05:42 -0500","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"Stephen Frost wrote:\n> * Alvaro Herrera (alvherre@2ndquadrant.com) wrote:\n\n> > Overall, this whole line of development seems like bloating the parse\n> > tables for little gain.\n> \n> Still, I see this point also. I do think it'd be really great if we\n> could figure out a way to segregate these kinds of DDL / maintenance\n> commands from the normal select/insert/update/delete SQL parsing, such\n> that we could add more options, etc, to those longer running and less\n> frequent commands without impacting parse time for the high-volume\n> commands.\n\nWe do have a parenthesized options clause in VACUUM. I think adding\nthis as a clause there would be pretty much free.\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":"Mon, 22 Dec 2014 14:11:07 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"* Andres Freund (andres@2ndquadrant.com) wrote:\n> On 2014-12-21 14:18:33 -0500, Tom Lane wrote:\n> > While the feature itself might be fairly innocuous, I'm just wondering\n> > why we need to encourage manual vacuuming. And why that, but not\n> > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\n> \n> There's one argument for supporting more for VACUUM than the rest - it\n> can't be executed directly as the result of a query as the others\n> can... I wonder if that'd not better be answered by adding a feature to\n> vacuumdb that allows selecting the to-be-vacuumed table by a user\n> defined query.\n\nWow. That's certainly an interesting idea.\n\nWe might end up turning the autovacuum process into a generalized\nscheduler/cron-like entity that way though. I'd rather we just build\nthat. Users would then be able to run a script periodically which\nwould add VACUUM commands to be run on whichever tables they want to\nthe jobs queue, either for immediate execution or at whatever time they\nwant (or possibly chronically :).\n\n\tThanks!\n\n\t\tStephen","msg_date":"Mon, 22 Dec 2014 12:12:12 -0500","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"On 2014-12-22 12:12:12 -0500, Stephen Frost wrote:\n> * Andres Freund (andres@2ndquadrant.com) wrote:\n> > On 2014-12-21 14:18:33 -0500, Tom Lane wrote:\n> > > While the feature itself might be fairly innocuous, I'm just wondering\n> > > why we need to encourage manual vacuuming. And why that, but not\n> > > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\n> > \n> > There's one argument for supporting more for VACUUM than the rest - it\n> > can't be executed directly as the result of a query as the others\n> > can... I wonder if that'd not better be answered by adding a feature to\n> > vacuumdb that allows selecting the to-be-vacuumed table by a user\n> > defined query.\n> \n> Wow. That's certainly an interesting idea.\n> \n> We might end up turning the autovacuum process into a generalized\n> scheduler/cron-like entity that way though.\n\nI'm not talking about autovacuum, just plain vacuumdb.\n\n> I'd rather we just build\n> that. Users would then be able to run a script periodically which\n> would add VACUUM commands to be run on whichever tables they want to\n> the jobs queue, either for immediate execution or at whatever time they\n> want (or possibly chronically :).\n\nAnd this discussion just feature creeped beyond anything realistic... :)\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 22 Dec 2014 18:15:51 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"Stephen Frost wrote:\n> * Andres Freund (andres@2ndquadrant.com) wrote:\n\n> > There's one argument for supporting more for VACUUM than the rest - it\n> > can't be executed directly as the result of a query as the others\n> > can... I wonder if that'd not better be answered by adding a feature to\n> > vacuumdb that allows selecting the to-be-vacuumed table by a user\n> > defined query.\n> \n> Wow. That's certainly an interesting idea.\n\n+1.\n\n> We might end up turning the autovacuum process into a generalized\n> scheduler/cron-like entity that way though. I'd rather we just build\n> that. Users would then be able to run a script periodically which\n> would add VACUUM commands to be run on whichever tables they want to\n> the jobs queue, either for immediate execution or at whatever time they\n> want (or possibly chronically :).\n\nThis too. I think there's one or two orders of magnitude of difference\nin implementation effort of these two ideas, however.\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":"Mon, 22 Dec 2014 14:17:15 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"* Andres Freund (andres@2ndquadrant.com) wrote:\n> On 2014-12-22 12:12:12 -0500, Stephen Frost wrote:\n> > We might end up turning the autovacuum process into a generalized\n> > scheduler/cron-like entity that way though.\n> \n> I'm not talking about autovacuum, just plain vacuumdb.\n\nOh, right, clearly I was thinking of autovacuum. Adding an option like\nthat to vacuumdb would certainly be a lot more straight-forward.\n\n> > I'd rather we just build\n> > that. Users would then be able to run a script periodically which\n> > would add VACUUM commands to be run on whichever tables they want to\n> > the jobs queue, either for immediate execution or at whatever time they\n> > want (or possibly chronically :).\n> \n> And this discussion just feature creeped beyond anything realistic... :)\n\nYeah, but I really *want* this... ;)\n\n\tThanks!\n\n\t\tStephen","msg_date":"Mon, 22 Dec 2014 12:17:33 -0500","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"On Mon, Dec 22, 2014 at 3:17 PM, Alvaro Herrera <alvherre@2ndquadrant.com>\nwrote:\n>\n> Stephen Frost wrote:\n> > * Andres Freund (andres@2ndquadrant.com) wrote:\n>\n> > > There's one argument for supporting more for VACUUM than the rest - it\n> > > can't be executed directly as the result of a query as the others\n> > > can... I wonder if that'd not better be answered by adding a feature\nto\n> > > vacuumdb that allows selecting the to-be-vacuumed table by a user\n> > > defined query.\n> >\n> > Wow. That's certainly an interesting idea.\n>\n> +1.\n>\n\nThen to simplify can we allow the \"--table\" option of vacuumdb act similar\nto the \"--table\" option of pg_dump??\n\nRegards,\n\n--\nFabrízio de Royes Mello\nConsultoria/Coaching PostgreSQL\n>> Timbira: http://www.timbira.com.br\n>> Blog: http://fabriziomello.github.io\n>> Linkedin: http://br.linkedin.com/in/fabriziomello\n>> Twitter: http://twitter.com/fabriziomello\n>> Github: http://github.com/fabriziomello\n\nOn Mon, Dec 22, 2014 at 3:17 PM, Alvaro Herrera <alvherre@2ndquadrant.com> wrote:>> Stephen Frost wrote:> > * Andres Freund (andres@2ndquadrant.com) wrote:>> > > There's one argument for supporting more for VACUUM than the rest - it> > > can't be executed directly as the result of a query as the others> > > can... I wonder if that'd not better be answered by adding a feature to> > > vacuumdb that allows selecting the to-be-vacuumed table by a user> > > defined query.> >> > Wow. That's certainly an interesting idea.>> +1.>Then to simplify can we allow the \"--table\" option of vacuumdb act similar to the \"--table\" option of pg_dump??Regards,--Fabrízio de Royes MelloConsultoria/Coaching PostgreSQL>> Timbira: http://www.timbira.com.br>> Blog: http://fabriziomello.github.io>> Linkedin: http://br.linkedin.com/in/fabriziomello>> Twitter: http://twitter.com/fabriziomello>> Github: http://github.com/fabriziomello","msg_date":"Mon, 22 Dec 2014 15:23:07 -0200","msg_from":"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com>","msg_from_op":true,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"\nOn 12/21/2014 02:18 PM, Tom Lane wrote:\n> =?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com> writes:\n>> I work with some customer that have databases with a lot of schemas and\n>> sometimes we need to run manual VACUUM in one schema, and would be nice to\n>> have a new option to run vacuum in relations from a specific schema.\n> I'm pretty skeptical of this alleged use-case. Manual vacuuming ought\n> to be mostly a thing of the past, and even if it's not, hitting\n> *everything* in a schema should seldom be an appropriate thing to do.\n>\n> While the feature itself might be fairly innocuous, I'm just wondering\n> why we need to encourage manual vacuuming. And why that, but not\n> say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\n>\n> \t\t\t\n\n\nSadly, manual vacuuming is very far from a thing of the past. Autovacuum \nsimply doesn't give us enough control in many cases.\n\nMaybe this gadget would be useful, but its application seems a bit \nlimited. Someone mentioned allowing multiple --table options to \nvacuumdb. That would be mopre flexible.\n\nBut really I think we need to work on how we can make autovacuum more \nuseful. For example, it would be nice not to have to do \"ALTER TABLE\" to \nchange the autovac settings. It would be nice to be able to specify \ntimes of day and days of week when autovacuum should be turned on or off \nfor a table. I'm sure there are plenty of other ideas.\n\ncheers\n\nandrew\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 22 Dec 2014 14:35:54 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"On 12/22/14, 10:05 AM, Andres Freund wrote:\n>> While the feature itself might be fairly innocuous, I'm just wondering\n>> >why we need to encourage manual vacuuming. And why that, but not\n>> >say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\n> There's one argument for supporting more for VACUUM than the rest - it\n> can't be executed directly as the result of a query as the others\n> can... I wonder if that'd not better be answered by adding a feature to\n> vacuumdb that allows selecting the to-be-vacuumed table by a user\n> defined query.\n\nI would MUCH rather that we find a way to special-case executing non-transactional commands dynamically, because VACUUM isn't the only one that suffers from this problem.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 22 Dec 2014 16:00:38 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"On 12/21/14, 8:55 PM, Fabrízio de Royes Mello wrote:\n> > And why that, but not\n> > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\n> >\n>\n> +1. I can write patches for each of this maintenance statement too.\n>\n>\n> If we're going to go that route, then perhaps it would make more sense to create a command that allows you to apply a second command to every object in a schema. We would have to be careful about PreventTransactionChain commands.\n>\n>\n> Sorry but I don't understand what you meant. Can you explain more about your idea?\n\nThere's a very large number of commands that could be useful to execute on every object in a schema. (RE)INDEX, CLUSTER, ALTER come to mind besides VACUUM.\n\nRight now a lot of people just work around this with things like DO blocks, but as mentioned elsewhere in the thread that fails for commands that can't be in a transaction.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 22 Dec 2014 16:02:57 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"On Mon, Dec 22, 2014 at 5:00 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\n> I would MUCH rather that we find a way to special-case executing\n> non-transactional commands dynamically, because VACUUM isn't the only one\n> that suffers from this problem.\n\nIs pg_background a solution to this problem?\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 Dec 2014 08:44:17 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"On Mon, Dec 22, 2014 at 11:51 AM, Alvaro Herrera\n<alvherre@2ndquadrant.com> wrote:\n> Multi-table CLUSTER uses multiple transactions, so this should not be an\n> issue. That said, I don't think there's much point in CLUSTER SCHEMA,\n> much less TRUNCATE SCHEMA. Do you normally organize your schemas so\n> that there are some that contain only tables that need to be truncated\n> together? That would be a strange use case.\n>\n> Overall, this whole line of development seems like bloating the parse\n> tables for little gain.\n\nWe added REINDEX SCHEMA less than three weeks ago; if we accept that\nthat was a good change, but think this is a bad one, it's not clear to\nme that there is any guiding principle here beyond who happened to\nweigh in on which threads.\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 Dec 2014 08:54:22 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"On Mon, Dec 22, 2014 at 8:02 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\n>\n> On 12/21/14, 8:55 PM, Fabrízio de Royes Mello wrote:\n>>\n>> > And why that, but not\n>> > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\n>> >\n>>\n>> +1. I can write patches for each of this maintenance statement\ntoo.\n>>\n>>\n>> If we're going to go that route, then perhaps it would make more\nsense to create a command that allows you to apply a second command to\nevery object in a schema. We would have to be careful about\nPreventTransactionChain commands.\n>>\n>>\n>> Sorry but I don't understand what you meant. Can you explain more\nabout your idea?\n>\n>\n> There's a very large number of commands that could be useful to execute\non every object in a schema. (RE)INDEX, CLUSTER, ALTER come to mind besides\nVACUUM.\n>\n\nANALYZE too...\n\n\n\n> Right now a lot of people just work around this with things like DO\nblocks, but as mentioned elsewhere in the thread that fails for commands\nthat can't be in a transaction.\n>\n\nI use \"dblink\" to solve it. :-)\n\n\nRegards,\n\n--\nFabrízio de Royes Mello\nConsultoria/Coaching PostgreSQL\n>> Timbira: http://www.timbira.com.br\n>> Blog: http://fabriziomello.github.io\n>> Linkedin: http://br.linkedin.com/in/fabriziomello\n>> Twitter: http://twitter.com/fabriziomello\n>> Github: http://github.com/fabriziomello\n\nOn Mon, Dec 22, 2014 at 8:02 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:>> On 12/21/14, 8:55 PM, Fabrízio de Royes Mello wrote:>>>> > And why that, but not>> > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...>> >>>>> +1. I can write patches for each of this maintenance statement too.>>>>>> If we're going to go that route, then perhaps it would make more sense to create a command that allows you to apply a second command to every object in a schema. We would have to be careful about PreventTransactionChain commands.>>>>>> Sorry but I don't understand what you meant. Can you explain more about your idea?>>> There's a very large number of commands that could be useful to execute on every object in a schema. (RE)INDEX, CLUSTER, ALTER come to mind besides VACUUM.>ANALYZE too... > Right now a lot of people just work around this with things like DO blocks, but as mentioned elsewhere in the thread that fails for commands that can't be in a transaction.>I use \"dblink\" to solve it. :-)Regards,--Fabrízio de Royes MelloConsultoria/Coaching PostgreSQL>> Timbira: http://www.timbira.com.br>> Blog: http://fabriziomello.github.io>> Linkedin: http://br.linkedin.com/in/fabriziomello>> Twitter: http://twitter.com/fabriziomello>> Github: http://github.com/fabriziomello","msg_date":"Tue, 23 Dec 2014 12:54:01 -0200","msg_from":"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com>","msg_from_op":true,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"Robert Haas wrote:\n> On Mon, Dec 22, 2014 at 11:51 AM, Alvaro Herrera\n> <alvherre@2ndquadrant.com> wrote:\n> > Multi-table CLUSTER uses multiple transactions, so this should not be an\n> > issue. That said, I don't think there's much point in CLUSTER SCHEMA,\n> > much less TRUNCATE SCHEMA. Do you normally organize your schemas so\n> > that there are some that contain only tables that need to be truncated\n> > together? That would be a strange use case.\n> >\n> > Overall, this whole line of development seems like bloating the parse\n> > tables for little gain.\n> \n> We added REINDEX SCHEMA less than three weeks ago; if we accept that\n> that was a good change, but think this is a bad one, it's not clear to\n> me that there is any guiding principle here beyond who happened to\n> weigh in on which threads.\n\nI didn't think much of REINDEX SCHEMA, TBH.\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, 23 Dec 2014 12:28:35 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"On 12/23/14, 8:54 AM, Fabrízio de Royes Mello wrote:\n> > Right now a lot of people just work around this with things like DO blocks, but as mentioned elsewhere in the thread that fails for commands that can't be in a transaction.\n> >\n>\n> I use \"dblink\" to solve it. :-)\n\nSo... how about instead of solving this only for vacuum we create something generic? :) Possibly using Robert's background worker work?\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 23 Dec 2014 19:06:25 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"On 12/23/14, 7:44 AM, Robert Haas wrote:\n> On Mon, Dec 22, 2014 at 5:00 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\n>> I would MUCH rather that we find a way to special-case executing\n>> non-transactional commands dynamically, because VACUUM isn't the only one\n>> that suffers from this problem.\n>\n> Is pg_background a solution to this problem?\n\nYes, since it allows you to do \"autonomous transactions\". It's probably not the most efficient way to solve this, but it should work.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 23 Dec 2014 19:09:27 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"Em terça-feira, 23 de dezembro de 2014, Jim Nasby <Jim.Nasby@bluetreble.com>\nescreveu:\n\n> On 12/23/14, 8:54 AM, Fabrízio de Royes Mello wrote:\n>\n>> > Right now a lot of people just work around this with things like DO\n>> blocks, but as mentioned elsewhere in the thread that fails for commands\n>> that can't be in a transaction.\n>> >\n>>\n>> I use \"dblink\" to solve it. :-)\n>>\n>\n> So... how about instead of solving this only for vacuum we create\n> something generic? :) Possibly using Robert's background worker work?\n\n\n Interesting idea.\n\nBut and what about the idea of improve the \"--table\" option from clients:\nvaccumdb and clusterdb?\n\nRegards,\n\nFabrízio Mello\n\n\n\n-- \nFabrízio de Royes Mello\nConsultoria/Coaching PostgreSQL\n>> Timbira: http://www.timbira.com.br\n>> Blog: http://fabriziomello.github.io\n>> Linkedin: http://br.linkedin.com/in/fabriziomello\n>> Twitter: http://twitter.com/fabriziomello\n>> Github: http://github.com/fabriziomello\n\nEm terça-feira, 23 de dezembro de 2014, Jim Nasby <Jim.Nasby@bluetreble.com> escreveu:On 12/23/14, 8:54 AM, Fabrízio de Royes Mello wrote:\n\n > Right now a lot of people just work around this with things like DO blocks, but as mentioned elsewhere in the thread that fails for commands that can't be in a transaction.\n >\n\nI use \"dblink\" to solve it. :-)\n\n\nSo... how about instead of solving this only for vacuum we create something generic? :) Possibly using Robert's background worker work? Interesting idea. But and what about the idea of improve the \"--table\" option from clients: vaccumdb and clusterdb?Regards,Fabrízio Mello -- Fabrízio de Royes MelloConsultoria/Coaching PostgreSQL>> Timbira: http://www.timbira.com.br>> Blog: http://fabriziomello.github.io>> Linkedin: http://br.linkedin.com/in/fabriziomello>> Twitter: http://twitter.com/fabriziomello>> Github: http://github.com/fabriziomello","msg_date":"Wed, 24 Dec 2014 00:49:40 -0200","msg_from":"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com>","msg_from_op":true,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"On 12/23/14, 8:49 PM, Fabrízio de Royes Mello wrote:\n> Em terça-feira, 23 de dezembro de 2014, Jim Nasby <Jim.Nasby@bluetreble.com <mailto:Jim.Nasby@bluetreble.com>> escreveu:\n>\n> On 12/23/14, 8:54 AM, Fabrízio de Royes Mello wrote:\n>\n> > Right now a lot of people just work around this with things like DO blocks, but as mentioned elsewhere in the thread that fails for commands that can't be in a transaction.\n> >\n>\n> I use \"dblink\" to solve it. :-)\n>\n>\n> So... how about instead of solving this only for vacuum we create something generic? :) Possibly using Robert's background worker work?\n>\n>\n> Interesting idea.\n>\n> But and what about the idea of improve the \"--table\" option from clients: vaccumdb and clusterdb?\n\nSeems reasonable.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 24 Dec 2014 12:53:06 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""},{"msg_contents":"21.12.2014, 18:48, Fabrízio de Royes Mello kirjoitti:\n> I work with some customer that have databases with a lot of schemas and\n> sometimes we need to run manual VACUUM in one schema, and would be nice\n> to have a new option to run vacuum in relations from a specific schema.\n> \n> The new syntax could be:\n> \n> VACUUM [ FULL ] [ FREEZE ] [ VERBOSE ] { [ table_name ] | SCHEMA\n> schema_name }\n> \n> Also I'll add a new option to \"vacuumdb\" client:\n> \n> -S, --schema=SCHEMA\n> \n> I can work on this feature to 2015/02 CF.\n> \n> Thoughts?\n\nThis would be useful for ANALYZE to make it easier to run analyze only\nfor the interesting schemas after a pg_upgrade. I have a database with\nmost of the actively used data in the \"public\" schema and a number of\nrarely accessed large logging and archive tables in other schemas. It'd\nbe useful to prioritize analyzing the main tables before doing anything\nabout the rarely used schemas to allow the database to be put back into\nproduction as soon as possible.\n\n/ Oskari\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sun, 28 Dec 2014 17:38:45 +0200","msg_from":"Oskari Saarenmaa <os@ohmu.fi>","msg_from_op":false,"msg_subject":"Re: Proposal \"VACUUM SCHEMA\""}],"string":"[\n {\n \"msg_contents\": \"Hi all,\\n\\nI work with some customer that have databases with a lot of schemas and\\nsometimes we need to run manual VACUUM in one schema, and would be nice to\\nhave a new option to run vacuum in relations from a specific schema.\\n\\nThe new syntax could be:\\n\\nVACUUM [ FULL ] [ FREEZE ] [ VERBOSE ] { [ table_name ] | SCHEMA\\nschema_name }\\n\\nAlso I'll add a new option to \\\"vacuumdb\\\" client:\\n\\n-S, --schema=SCHEMA\\n\\nI can work on this feature to 2015/02 CF.\\n\\nThoughts?\\n\\n--\\nFabrízio de Royes Mello\\nConsultoria/Coaching PostgreSQL\\n>> Timbira: http://www.timbira.com.br\\n>> Blog: http://fabriziomello.github.io\\n>> Linkedin: http://br.linkedin.com/in/fabriziomello\\n>> Twitter: http://twitter.com/fabriziomello\\n>> Github: http://github.com/fabriziomello\\n\\nHi all,I work with some customer that have databases with a lot of schemas and sometimes we need to run manual VACUUM in one schema, and would be nice to have a new option to run vacuum in relations from a specific schema.The new syntax could be:VACUUM [ FULL ] [ FREEZE ] [ VERBOSE ] { [ table_name ] | SCHEMA schema_name }Also I'll add a new option to \\\"vacuumdb\\\" client:-S, --schema=SCHEMAI can work on this feature to 2015/02 CF.Thoughts? --Fabrízio de Royes MelloConsultoria/Coaching PostgreSQL>> Timbira: http://www.timbira.com.br>> Blog: http://fabriziomello.github.io>> Linkedin: http://br.linkedin.com/in/fabriziomello>> Twitter: http://twitter.com/fabriziomello>> Github: http://github.com/fabriziomello\",\n \"msg_date\": \"Sun, 21 Dec 2014 14:48:16 -0200\",\n \"msg_from\": \"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com> writes:\\n> I work with some customer that have databases with a lot of schemas and\\n> sometimes we need to run manual VACUUM in one schema, and would be nice to\\n> have a new option to run vacuum in relations from a specific schema.\\n\\nI'm pretty skeptical of this alleged use-case. Manual vacuuming ought\\nto be mostly a thing of the past, and even if it's not, hitting\\n*everything* in a schema should seldom be an appropriate thing to do.\\n\\nWhile the feature itself might be fairly innocuous, I'm just wondering\\nwhy we need to encourage manual vacuuming. And why that, but not\\nsay schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\\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, 21 Dec 2014 14:18:33 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"On Sun, Dec 21, 2014 at 5:18 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> =?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com> writes:\\n> > I work with some customer that have databases with a lot of schemas and\\n> > sometimes we need to run manual VACUUM in one schema, and would be nice\\nto\\n> > have a new option to run vacuum in relations from a specific schema.\\n>\\n> I'm pretty skeptical of this alleged use-case. Manual vacuuming ought\\n> to be mostly a thing of the past, and even if it's not, hitting\\n> *everything* in a schema should seldom be an appropriate thing to do.\\n>\\n\\nI agree manual vacuum is a thing of the past, but autovacuum doesn't solve\\n100% of the cases, and sometimes we need to use it so my proposal is just\\ndo help DBAs and/or Sysadmins to write simple maintenance scripts.\\n\\n\\n> While the feature itself might be fairly innocuous, I'm just wondering\\n> why we need to encourage manual vacuuming.\\n\\nIMHO we will not encourage manual vacuuming, just give more flexibility to\\nusers.\\n\\n\\n> And why that, but not\\n> say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\\n>\\n\\n+1. I can write patches for each of this maintenance statement too.\\n\\nRegards,\\n\\n--\\nFabrízio de Royes Mello\\nConsultoria/Coaching PostgreSQL\\n>> Timbira: http://www.timbira.com.br\\n>> Blog: http://fabriziomello.github.io\\n>> Linkedin: http://br.linkedin.com/in/fabriziomello\\n>> Twitter: http://twitter.com/fabriziomello\\n>> Github: http://github.com/fabriziomello\\n\\nOn Sun, Dec 21, 2014 at 5:18 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:>> =?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com> writes:> > I work with some customer that have databases with a lot of schemas and> > sometimes we need to run manual VACUUM in one schema, and would be nice to> > have a new option to run vacuum in relations from a specific schema.>> I'm pretty skeptical of this alleged use-case. Manual vacuuming ought> to be mostly a thing of the past, and even if it's not, hitting> *everything* in a schema should seldom be an appropriate thing to do.>I agree manual vacuum is a thing of the past, but autovacuum doesn't solve 100% of the cases, and sometimes we need to use it so my proposal is just do help DBAs and/or Sysadmins to write simple maintenance scripts.> While the feature itself might be fairly innocuous, I'm just wondering> why we need to encourage manual vacuuming. IMHO we will not encourage manual vacuuming, just give more flexibility to users.> And why that, but not> say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...>+1. I can write patches for each of this maintenance statement too.Regards,--Fabrízio de Royes MelloConsultoria/Coaching PostgreSQL>> Timbira: http://www.timbira.com.br>> Blog: http://fabriziomello.github.io>> Linkedin: http://br.linkedin.com/in/fabriziomello>> Twitter: http://twitter.com/fabriziomello>> Github: http://github.com/fabriziomello\",\n \"msg_date\": \"Sun, 21 Dec 2014 19:30:35 -0200\",\n \"msg_from\": \"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"On 12/21/14, 3:30 PM, Fabrízio de Royes Mello wrote:\\n>\\n> On Sun, Dec 21, 2014 at 5:18 PM, Tom Lane <tgl@sss.pgh.pa.us <mailto:tgl@sss.pgh.pa.us>> wrote:\\n> >\\n> > =?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com <mailto:fabriziomello@gmail.com>> writes:\\n> > > I work with some customer that have databases with a lot of schemas and\\n> > > sometimes we need to run manual VACUUM in one schema, and would be nice to\\n> > > have a new option to run vacuum in relations from a specific schema.\\n> >\\n> > I'm pretty skeptical of this alleged use-case. Manual vacuuming ought\\n> > to be mostly a thing of the past, and even if it's not, hitting\\n> > *everything* in a schema should seldom be an appropriate thing to do.\\n> >\\n>\\n> I agree manual vacuum is a thing of the past, but autovacuum doesn't solve 100% of the cases, and sometimes we need to use it so my proposal is just do help DBAs and/or Sysadmins to write simple maintenance scripts.\\n\\nJust one example of that is pre-emptively vacuuming during slower periods. Nothing spells \\\"fun\\\" like a freeze vacuum in the middle of a busy lunch period for a website.\\n\\nSimilarly, it's common to need to proactively vacuum after a data load, and since it's not unusual for there to be a schema dedicated to loading data, this makes that easier.\\n\\n> > And why that, but not\\n> > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\\n> >\\n>\\n> +1. I can write patches for each of this maintenance statement too.\\n\\nIf we're going to go that route, then perhaps it would make more sense to create a command that allows you to apply a second command to every object in a schema. We would have to be careful about PreventTransactionChain commands.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sun, 21 Dec 2014 20:32:32 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"Em segunda-feira, 22 de dezembro de 2014, Jim Nasby <\\nJim.Nasby@bluetreble.com> escreveu:\\n\\n> On 12/21/14, 3:30 PM, Fabrízio de Royes Mello wrote:\\n>\\n>>\\n>> On Sun, Dec 21, 2014 at 5:18 PM, Tom Lane <tgl@sss.pgh.pa.us <mailto:\\n>> tgl@sss.pgh.pa.us>> wrote:\\n>> >\\n>> > =?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com\\n>> <mailto:fabriziomello@gmail.com>> writes:\\n>> > > I work with some customer that have databases with a lot of schemas\\n>> and\\n>> > > sometimes we need to run manual VACUUM in one schema, and would be\\n>> nice to\\n>> > > have a new option to run vacuum in relations from a specific schema.\\n>> >\\n>> > I'm pretty skeptical of this alleged use-case. Manual vacuuming ought\\n>> > to be mostly a thing of the past, and even if it's not, hitting\\n>> > *everything* in a schema should seldom be an appropriate thing to do.\\n>> >\\n>>\\n>> I agree manual vacuum is a thing of the past, but autovacuum doesn't\\n>> solve 100% of the cases, and sometimes we need to use it so my proposal is\\n>> just do help DBAs and/or Sysadmins to write simple maintenance scripts.\\n>>\\n>\\n> Just one example of that is pre-emptively vacuuming during slower periods.\\n> Nothing spells \\\"fun\\\" like a freeze vacuum in the middle of a busy lunch\\n> period for a website.\\n>\\n> Similarly, it's common to need to proactively vacuum after a data load,\\n> and since it's not unusual for there to be a schema dedicated to loading\\n> data, this makes that easier.\\n\\n\\nGood example. Thanks.\\n\\n\\n\\n>\\n> > And why that, but not\\n>> > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\\n>> >\\n>>\\n>> +1. I can write patches for each of this maintenance statement too.\\n>>\\n>\\n> If we're going to go that route, then perhaps it would make more sense to\\n> create a command that allows you to apply a second command to every object\\n> in a schema. We would have to be careful about PreventTransactionChain\\n> commands.\\n\\n\\n Sorry but I don't understand what you meant. Can you explain more about\\nyour idea?\\n\\nRegards,\\n\\nFabrízio Mello\\n\\n\\n-- \\nFabrízio de Royes Mello\\nConsultoria/Coaching PostgreSQL\\n>> Timbira: http://www.timbira.com.br\\n>> Blog: http://fabriziomello.github.io\\n>> Linkedin: http://br.linkedin.com/in/fabriziomello\\n>> Twitter: http://twitter.com/fabriziomello\\n>> Github: http://github.com/fabriziomello\\n\\nEm segunda-feira, 22 de dezembro de 2014, Jim Nasby <Jim.Nasby@bluetreble.com> escreveu:On 12/21/14, 3:30 PM, Fabrízio de Royes Mello wrote:\\n\\n\\nOn Sun, Dec 21, 2014 at 5:18 PM, Tom Lane <tgl@sss.pgh.pa.us <mailto:tgl@sss.pgh.pa.us>> wrote:\\n >\\n > =?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com <mailto:fabriziomello@gmail.com>> writes:\\n > > I work with some customer that have databases with a lot of schemas and\\n > > sometimes we need to run manual VACUUM in one schema, and would be nice to\\n > > have a new option to run vacuum in relations from a specific schema.\\n >\\n > I'm pretty skeptical of this alleged use-case. Manual vacuuming ought\\n > to be mostly a thing of the past, and even if it's not, hitting\\n > *everything* in a schema should seldom be an appropriate thing to do.\\n >\\n\\nI agree manual vacuum is a thing of the past, but autovacuum doesn't solve 100% of the cases, and sometimes we need to use it so my proposal is just do help DBAs and/or Sysadmins to write simple maintenance scripts.\\n\\n\\nJust one example of that is pre-emptively vacuuming during slower periods. Nothing spells \\\"fun\\\" like a freeze vacuum in the middle of a busy lunch period for a website.\\n\\nSimilarly, it's common to need to proactively vacuum after a data load, and since it's not unusual for there to be a schema dedicated to loading data, this makes that easier.Good example. Thanks. \\n\\n\\n > And why that, but not\\n > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\\n >\\n\\n+1. I can write patches for each of this maintenance statement too.\\n\\n\\nIf we're going to go that route, then perhaps it would make more sense to create a command that allows you to apply a second command to every object in a schema. We would have to be careful about PreventTransactionChain commands. Sorry but I don't understand what you meant. Can you explain more about your idea?Regards,Fabrízio Mello-- Fabrízio de Royes MelloConsultoria/Coaching PostgreSQL>> Timbira: http://www.timbira.com.br>> Blog: http://fabriziomello.github.io>> Linkedin: http://br.linkedin.com/in/fabriziomello>> Twitter: http://twitter.com/fabriziomello>> Github: http://github.com/fabriziomello\",\n \"msg_date\": \"Mon, 22 Dec 2014 00:55:14 -0200\",\n \"msg_from\": \"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"On 12/21/2014 10:30 PM, Fabrízio de Royes Mello wrote:\\n> [snip]\\n\\nI do agree that \\\"vacuum schema\\\" might very well be useful (I'll probably \\nuse it myself from time to time, too).\\nANALYZE SCHEMA (specially coupled with some transaction-wide \\\"SET \\nstatistics_target\\\" could be beneficial)\\n\\n>\\n> > And why that, but not\\n> > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\\n> >\\n>\\n> +1. I can write patches for each of this maintenance statement too.\\n\\nHmm... I think Tom might have been a bit rethorical (or even sarcastic \\nwith that), but I can definitely be wrong.\\n\\nDo we really want to have some such operation potentially (and \\ninadvertently) locking for *hours* at a time?\\n\\nCLUSTER SCHEMA somename;\\n\\n ... where schema \\\"somename\\\" contains \\\"myHugeTable\\\"\\n\\n Given that the cluster command exclusively locks and rewrites the \\ntable, it might lock queries and overwhelm the I/O subsystem for quite a \\nlong time.\\n\\n\\nTRUNCATE SCHEMA whatever sounds quite dangerous, too.\\n\\n\\n\\nJust my .02€\\n\\n / J.L.\\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, 22 Dec 2014 16:55:50 +0100\",\n \"msg_from\": \"=?UTF-8?B?Sm9zw6kgTHVpcyBUYWxsw7Nu?= <jltallon@adv-solutions.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"On 2014-12-21 14:18:33 -0500, Tom Lane wrote:\\n> =?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com> writes:\\n> > I work with some customer that have databases with a lot of schemas and\\n> > sometimes we need to run manual VACUUM in one schema, and would be nice to\\n> > have a new option to run vacuum in relations from a specific schema.\\n> \\n> I'm pretty skeptical of this alleged use-case. Manual vacuuming ought\\n> to be mostly a thing of the past, and even if it's not, hitting\\n> *everything* in a schema should seldom be an appropriate thing to do.\\n\\nBased on my experience autovacuum isn't sufficient on bigger high\\nthroughput databases. At the very least manual vacuuming with lower\\nfreeze_table_age settings during low-load times is required lest\\nanti-wraparound vacuums increase load too much during prime business\\nhours.\\nThat said, I don't see how this feature is actually helpful in those\\ncases. In pretty much all of what I've seen you'd want to have more\\ncomplex selection criteria than the schema.\\n\\n> While the feature itself might be fairly innocuous, I'm just wondering\\n> why we need to encourage manual vacuuming. And why that, but not\\n> say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\\n\\nThere's one argument for supporting more for VACUUM than the rest - it\\ncan't be executed directly as the result of a query as the others\\ncan... I wonder if that'd not better be answered by adding a feature to\\nvacuumdb that allows selecting the to-be-vacuumed table by a user\\ndefined query.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 22 Dec 2014 17:05:49 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"Jos� Luis Tall�n wrote:\\n> On 12/21/2014 10:30 PM, Fabr�zio de Royes Mello wrote:\\n> >[snip]\\n> \\n> I do agree that \\\"vacuum schema\\\" might very well be useful (I'll probably use\\n> it myself from time to time, too).\\n> ANALYZE SCHEMA (specially coupled with some transaction-wide \\\"SET\\n> statistics_target\\\" could be beneficial)\\n\\nWe already have transanction-wide SET -- it's spelled SET LOCAL.\\n\\n> >\\n> >> And why that, but not say schema-wide ANALYZE, CLUSTER, TRUNCATE,\\n> >> ...\\n> >\\n> >+1. I can write patches for each of this maintenance statement too.\\n> \\n> Hmm... I think Tom might have been a bit rethorical (or even sarcastic with\\n> that),\\n\\nThat was my impression too.\\n\\n> Do we really want to have some such operation potentially (and\\n> inadvertently) locking for *hours* at a time?\\n> \\n> CLUSTER SCHEMA somename;\\n> \\n> ... where schema \\\"somename\\\" contains \\\"myHugeTable\\\"\\n> \\n> Given that the cluster command exclusively locks and rewrites the table,\\n> it might lock queries and overwhelm the I/O subsystem for quite a long time.\\n\\nMulti-table CLUSTER uses multiple transactions, so this should not be an\\nissue. That said, I don't think there's much point in CLUSTER SCHEMA,\\nmuch less TRUNCATE SCHEMA. Do you normally organize your schemas so\\nthat there are some that contain only tables that need to be truncated\\ntogether? That would be a strange use case.\\n\\nOverall, this whole line of development seems like bloating the parse\\ntables for little gain.\\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\": \"Mon, 22 Dec 2014 13:51:57 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"Re: Alvaro Herrera 2014-12-22 <20141222165157.GD1768@alvh.no-ip.org>\\n> Multi-table CLUSTER uses multiple transactions, so this should not be an\\n> issue. That said, I don't think there's much point in CLUSTER SCHEMA,\\n> much less TRUNCATE SCHEMA. Do you normally organize your schemas so\\n> that there are some that contain only tables that need to be truncated\\n> together? That would be a strange use case.\\n\\nHaving a schema that's only used for importing data in batch jobs\\ndoesn't sound too unreasonable. It could then be cleaned in a simple\\n\\\"TRUNCATE SCHEMA import_area\\\" command.\\n\\n> Overall, this whole line of development seems like bloating the parse\\n> tables for little gain.\\n\\nReading the thread, my impression was that most people opposed the\\nidea because there's ways to script \\\"vacuum schema\\\", or because of\\n\\\"people shouldn't be invoking manual vacuums anyway\\\". I think the\\npatch tries to solve a practical problem, and does have its merits.\\n\\nChristoph\\n-- \\ncb@df7cb.de | http://www.df7cb.de/\\n\\n\\n-- \\nSent via pgsql-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 Dec 2014 17:58:49 +0100\",\n \"msg_from\": \"Christoph Berg <cb@df7cb.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"* Alvaro Herrera (alvherre@2ndquadrant.com) wrote:\\n> Multi-table CLUSTER uses multiple transactions, so this should not be an\\n> issue. That said, I don't think there's much point in CLUSTER SCHEMA,\\n> much less TRUNCATE SCHEMA. Do you normally organize your schemas so\\n> that there are some that contain only tables that need to be truncated\\n> together? That would be a strange use case.\\n\\nI could see it happening in environments which use schemas when doing\\npartitioning. eg: data_2014 contains all of the data_201401-201412\\nmonthly (or perhaps weekly) tables.\\n\\n> Overall, this whole line of development seems like bloating the parse\\n> tables for little gain.\\n\\nStill, I see this point also. I do think it'd be really great if we\\ncould figure out a way to segregate these kinds of DDL / maintenance\\ncommands from the normal select/insert/update/delete SQL parsing, such\\nthat we could add more options, etc, to those longer running and less\\nfrequent commands without impacting parse time for the high-volume\\ncommands.\\n\\nI'm less concerned about the memory impact, except to the extent that it\\nimpacts throughput and performance.\\n\\n\\tThanks,\\n\\n\\t\\tStephen\",\n \"msg_date\": \"Mon, 22 Dec 2014 12:05:42 -0500\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"Stephen Frost wrote:\\n> * Alvaro Herrera (alvherre@2ndquadrant.com) wrote:\\n\\n> > Overall, this whole line of development seems like bloating the parse\\n> > tables for little gain.\\n> \\n> Still, I see this point also. I do think it'd be really great if we\\n> could figure out a way to segregate these kinds of DDL / maintenance\\n> commands from the normal select/insert/update/delete SQL parsing, such\\n> that we could add more options, etc, to those longer running and less\\n> frequent commands without impacting parse time for the high-volume\\n> commands.\\n\\nWe do have a parenthesized options clause in VACUUM. I think adding\\nthis as a clause there would be pretty much free.\\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\": \"Mon, 22 Dec 2014 14:11:07 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"* Andres Freund (andres@2ndquadrant.com) wrote:\\n> On 2014-12-21 14:18:33 -0500, Tom Lane wrote:\\n> > While the feature itself might be fairly innocuous, I'm just wondering\\n> > why we need to encourage manual vacuuming. And why that, but not\\n> > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\\n> \\n> There's one argument for supporting more for VACUUM than the rest - it\\n> can't be executed directly as the result of a query as the others\\n> can... I wonder if that'd not better be answered by adding a feature to\\n> vacuumdb that allows selecting the to-be-vacuumed table by a user\\n> defined query.\\n\\nWow. That's certainly an interesting idea.\\n\\nWe might end up turning the autovacuum process into a generalized\\nscheduler/cron-like entity that way though. I'd rather we just build\\nthat. Users would then be able to run a script periodically which\\nwould add VACUUM commands to be run on whichever tables they want to\\nthe jobs queue, either for immediate execution or at whatever time they\\nwant (or possibly chronically :).\\n\\n\\tThanks!\\n\\n\\t\\tStephen\",\n \"msg_date\": \"Mon, 22 Dec 2014 12:12:12 -0500\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"On 2014-12-22 12:12:12 -0500, Stephen Frost wrote:\\n> * Andres Freund (andres@2ndquadrant.com) wrote:\\n> > On 2014-12-21 14:18:33 -0500, Tom Lane wrote:\\n> > > While the feature itself might be fairly innocuous, I'm just wondering\\n> > > why we need to encourage manual vacuuming. And why that, but not\\n> > > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\\n> > \\n> > There's one argument for supporting more for VACUUM than the rest - it\\n> > can't be executed directly as the result of a query as the others\\n> > can... I wonder if that'd not better be answered by adding a feature to\\n> > vacuumdb that allows selecting the to-be-vacuumed table by a user\\n> > defined query.\\n> \\n> Wow. That's certainly an interesting idea.\\n> \\n> We might end up turning the autovacuum process into a generalized\\n> scheduler/cron-like entity that way though.\\n\\nI'm not talking about autovacuum, just plain vacuumdb.\\n\\n> I'd rather we just build\\n> that. Users would then be able to run a script periodically which\\n> would add VACUUM commands to be run on whichever tables they want to\\n> the jobs queue, either for immediate execution or at whatever time they\\n> want (or possibly chronically :).\\n\\nAnd this discussion just feature creeped beyond anything realistic... :)\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 22 Dec 2014 18:15:51 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"Stephen Frost wrote:\\n> * Andres Freund (andres@2ndquadrant.com) wrote:\\n\\n> > There's one argument for supporting more for VACUUM than the rest - it\\n> > can't be executed directly as the result of a query as the others\\n> > can... I wonder if that'd not better be answered by adding a feature to\\n> > vacuumdb that allows selecting the to-be-vacuumed table by a user\\n> > defined query.\\n> \\n> Wow. That's certainly an interesting idea.\\n\\n+1.\\n\\n> We might end up turning the autovacuum process into a generalized\\n> scheduler/cron-like entity that way though. I'd rather we just build\\n> that. Users would then be able to run a script periodically which\\n> would add VACUUM commands to be run on whichever tables they want to\\n> the jobs queue, either for immediate execution or at whatever time they\\n> want (or possibly chronically :).\\n\\nThis too. I think there's one or two orders of magnitude of difference\\nin implementation effort of these two ideas, however.\\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\": \"Mon, 22 Dec 2014 14:17:15 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"* Andres Freund (andres@2ndquadrant.com) wrote:\\n> On 2014-12-22 12:12:12 -0500, Stephen Frost wrote:\\n> > We might end up turning the autovacuum process into a generalized\\n> > scheduler/cron-like entity that way though.\\n> \\n> I'm not talking about autovacuum, just plain vacuumdb.\\n\\nOh, right, clearly I was thinking of autovacuum. Adding an option like\\nthat to vacuumdb would certainly be a lot more straight-forward.\\n\\n> > I'd rather we just build\\n> > that. Users would then be able to run a script periodically which\\n> > would add VACUUM commands to be run on whichever tables they want to\\n> > the jobs queue, either for immediate execution or at whatever time they\\n> > want (or possibly chronically :).\\n> \\n> And this discussion just feature creeped beyond anything realistic... :)\\n\\nYeah, but I really *want* this... ;)\\n\\n\\tThanks!\\n\\n\\t\\tStephen\",\n \"msg_date\": \"Mon, 22 Dec 2014 12:17:33 -0500\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 22, 2014 at 3:17 PM, Alvaro Herrera <alvherre@2ndquadrant.com>\\nwrote:\\n>\\n> Stephen Frost wrote:\\n> > * Andres Freund (andres@2ndquadrant.com) wrote:\\n>\\n> > > There's one argument for supporting more for VACUUM than the rest - it\\n> > > can't be executed directly as the result of a query as the others\\n> > > can... I wonder if that'd not better be answered by adding a feature\\nto\\n> > > vacuumdb that allows selecting the to-be-vacuumed table by a user\\n> > > defined query.\\n> >\\n> > Wow. That's certainly an interesting idea.\\n>\\n> +1.\\n>\\n\\nThen to simplify can we allow the \\\"--table\\\" option of vacuumdb act similar\\nto the \\\"--table\\\" option of pg_dump??\\n\\nRegards,\\n\\n--\\nFabrízio de Royes Mello\\nConsultoria/Coaching PostgreSQL\\n>> Timbira: http://www.timbira.com.br\\n>> Blog: http://fabriziomello.github.io\\n>> Linkedin: http://br.linkedin.com/in/fabriziomello\\n>> Twitter: http://twitter.com/fabriziomello\\n>> Github: http://github.com/fabriziomello\\n\\nOn Mon, Dec 22, 2014 at 3:17 PM, Alvaro Herrera <alvherre@2ndquadrant.com> wrote:>> Stephen Frost wrote:> > * Andres Freund (andres@2ndquadrant.com) wrote:>> > > There's one argument for supporting more for VACUUM than the rest - it> > > can't be executed directly as the result of a query as the others> > > can... I wonder if that'd not better be answered by adding a feature to> > > vacuumdb that allows selecting the to-be-vacuumed table by a user> > > defined query.> >> > Wow. That's certainly an interesting idea.>> +1.>Then to simplify can we allow the \\\"--table\\\" option of vacuumdb act similar to the \\\"--table\\\" option of pg_dump??Regards,--Fabrízio de Royes MelloConsultoria/Coaching PostgreSQL>> Timbira: http://www.timbira.com.br>> Blog: http://fabriziomello.github.io>> Linkedin: http://br.linkedin.com/in/fabriziomello>> Twitter: http://twitter.com/fabriziomello>> Github: http://github.com/fabriziomello\",\n \"msg_date\": \"Mon, 22 Dec 2014 15:23:07 -0200\",\n \"msg_from\": \"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"\\nOn 12/21/2014 02:18 PM, Tom Lane wrote:\\n> =?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com> writes:\\n>> I work with some customer that have databases with a lot of schemas and\\n>> sometimes we need to run manual VACUUM in one schema, and would be nice to\\n>> have a new option to run vacuum in relations from a specific schema.\\n> I'm pretty skeptical of this alleged use-case. Manual vacuuming ought\\n> to be mostly a thing of the past, and even if it's not, hitting\\n> *everything* in a schema should seldom be an appropriate thing to do.\\n>\\n> While the feature itself might be fairly innocuous, I'm just wondering\\n> why we need to encourage manual vacuuming. And why that, but not\\n> say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\\n>\\n> \\t\\t\\t\\n\\n\\nSadly, manual vacuuming is very far from a thing of the past. Autovacuum \\nsimply doesn't give us enough control in many cases.\\n\\nMaybe this gadget would be useful, but its application seems a bit \\nlimited. Someone mentioned allowing multiple --table options to \\nvacuumdb. That would be mopre flexible.\\n\\nBut really I think we need to work on how we can make autovacuum more \\nuseful. For example, it would be nice not to have to do \\\"ALTER TABLE\\\" to \\nchange the autovac settings. It would be nice to be able to specify \\ntimes of day and days of week when autovacuum should be turned on or off \\nfor a table. I'm sure there are plenty of other ideas.\\n\\ncheers\\n\\nandrew\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 22 Dec 2014 14:35:54 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"On 12/22/14, 10:05 AM, Andres Freund wrote:\\n>> While the feature itself might be fairly innocuous, I'm just wondering\\n>> >why we need to encourage manual vacuuming. And why that, but not\\n>> >say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\\n> There's one argument for supporting more for VACUUM than the rest - it\\n> can't be executed directly as the result of a query as the others\\n> can... I wonder if that'd not better be answered by adding a feature to\\n> vacuumdb that allows selecting the to-be-vacuumed table by a user\\n> defined query.\\n\\nI would MUCH rather that we find a way to special-case executing non-transactional commands dynamically, because VACUUM isn't the only one that suffers from this problem.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 22 Dec 2014 16:00:38 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"On 12/21/14, 8:55 PM, Fabrízio de Royes Mello wrote:\\n> > And why that, but not\\n> > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\\n> >\\n>\\n> +1. I can write patches for each of this maintenance statement too.\\n>\\n>\\n> If we're going to go that route, then perhaps it would make more sense to create a command that allows you to apply a second command to every object in a schema. We would have to be careful about PreventTransactionChain commands.\\n>\\n>\\n> Sorry but I don't understand what you meant. Can you explain more about your idea?\\n\\nThere's a very large number of commands that could be useful to execute on every object in a schema. (RE)INDEX, CLUSTER, ALTER come to mind besides VACUUM.\\n\\nRight now a lot of people just work around this with things like DO blocks, but as mentioned elsewhere in the thread that fails for commands that can't be in a transaction.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 22 Dec 2014 16:02:57 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 22, 2014 at 5:00 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\\n> I would MUCH rather that we find a way to special-case executing\\n> non-transactional commands dynamically, because VACUUM isn't the only one\\n> that suffers from this problem.\\n\\nIs pg_background a solution to this problem?\\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 Dec 2014 08:44:17 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 22, 2014 at 11:51 AM, Alvaro Herrera\\n<alvherre@2ndquadrant.com> wrote:\\n> Multi-table CLUSTER uses multiple transactions, so this should not be an\\n> issue. That said, I don't think there's much point in CLUSTER SCHEMA,\\n> much less TRUNCATE SCHEMA. Do you normally organize your schemas so\\n> that there are some that contain only tables that need to be truncated\\n> together? That would be a strange use case.\\n>\\n> Overall, this whole line of development seems like bloating the parse\\n> tables for little gain.\\n\\nWe added REINDEX SCHEMA less than three weeks ago; if we accept that\\nthat was a good change, but think this is a bad one, it's not clear to\\nme that there is any guiding principle here beyond who happened to\\nweigh in on which threads.\\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 Dec 2014 08:54:22 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 22, 2014 at 8:02 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\\n>\\n> On 12/21/14, 8:55 PM, Fabrízio de Royes Mello wrote:\\n>>\\n>> > And why that, but not\\n>> > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...\\n>> >\\n>>\\n>> +1. I can write patches for each of this maintenance statement\\ntoo.\\n>>\\n>>\\n>> If we're going to go that route, then perhaps it would make more\\nsense to create a command that allows you to apply a second command to\\nevery object in a schema. We would have to be careful about\\nPreventTransactionChain commands.\\n>>\\n>>\\n>> Sorry but I don't understand what you meant. Can you explain more\\nabout your idea?\\n>\\n>\\n> There's a very large number of commands that could be useful to execute\\non every object in a schema. (RE)INDEX, CLUSTER, ALTER come to mind besides\\nVACUUM.\\n>\\n\\nANALYZE too...\\n\\n\\n\\n> Right now a lot of people just work around this with things like DO\\nblocks, but as mentioned elsewhere in the thread that fails for commands\\nthat can't be in a transaction.\\n>\\n\\nI use \\\"dblink\\\" to solve it. :-)\\n\\n\\nRegards,\\n\\n--\\nFabrízio de Royes Mello\\nConsultoria/Coaching PostgreSQL\\n>> Timbira: http://www.timbira.com.br\\n>> Blog: http://fabriziomello.github.io\\n>> Linkedin: http://br.linkedin.com/in/fabriziomello\\n>> Twitter: http://twitter.com/fabriziomello\\n>> Github: http://github.com/fabriziomello\\n\\nOn Mon, Dec 22, 2014 at 8:02 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:>> On 12/21/14, 8:55 PM, Fabrízio de Royes Mello wrote:>>>> > And why that, but not>> > say schema-wide ANALYZE, CLUSTER, TRUNCATE, ...>> >>>>> +1. I can write patches for each of this maintenance statement too.>>>>>> If we're going to go that route, then perhaps it would make more sense to create a command that allows you to apply a second command to every object in a schema. We would have to be careful about PreventTransactionChain commands.>>>>>> Sorry but I don't understand what you meant. Can you explain more about your idea?>>> There's a very large number of commands that could be useful to execute on every object in a schema. (RE)INDEX, CLUSTER, ALTER come to mind besides VACUUM.>ANALYZE too... > Right now a lot of people just work around this with things like DO blocks, but as mentioned elsewhere in the thread that fails for commands that can't be in a transaction.>I use \\\"dblink\\\" to solve it. :-)Regards,--Fabrízio de Royes MelloConsultoria/Coaching PostgreSQL>> Timbira: http://www.timbira.com.br>> Blog: http://fabriziomello.github.io>> Linkedin: http://br.linkedin.com/in/fabriziomello>> Twitter: http://twitter.com/fabriziomello>> Github: http://github.com/fabriziomello\",\n \"msg_date\": \"Tue, 23 Dec 2014 12:54:01 -0200\",\n \"msg_from\": \"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"Robert Haas wrote:\\n> On Mon, Dec 22, 2014 at 11:51 AM, Alvaro Herrera\\n> <alvherre@2ndquadrant.com> wrote:\\n> > Multi-table CLUSTER uses multiple transactions, so this should not be an\\n> > issue. That said, I don't think there's much point in CLUSTER SCHEMA,\\n> > much less TRUNCATE SCHEMA. Do you normally organize your schemas so\\n> > that there are some that contain only tables that need to be truncated\\n> > together? That would be a strange use case.\\n> >\\n> > Overall, this whole line of development seems like bloating the parse\\n> > tables for little gain.\\n> \\n> We added REINDEX SCHEMA less than three weeks ago; if we accept that\\n> that was a good change, but think this is a bad one, it's not clear to\\n> me that there is any guiding principle here beyond who happened to\\n> weigh in on which threads.\\n\\nI didn't think much of REINDEX SCHEMA, TBH.\\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, 23 Dec 2014 12:28:35 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"On 12/23/14, 8:54 AM, Fabrízio de Royes Mello wrote:\\n> > Right now a lot of people just work around this with things like DO blocks, but as mentioned elsewhere in the thread that fails for commands that can't be in a transaction.\\n> >\\n>\\n> I use \\\"dblink\\\" to solve it. :-)\\n\\nSo... how about instead of solving this only for vacuum we create something generic? :) Possibly using Robert's background worker work?\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 23 Dec 2014 19:06:25 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"On 12/23/14, 7:44 AM, Robert Haas wrote:\\n> On Mon, Dec 22, 2014 at 5:00 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\\n>> I would MUCH rather that we find a way to special-case executing\\n>> non-transactional commands dynamically, because VACUUM isn't the only one\\n>> that suffers from this problem.\\n>\\n> Is pg_background a solution to this problem?\\n\\nYes, since it allows you to do \\\"autonomous transactions\\\". It's probably not the most efficient way to solve this, but it should work.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 23 Dec 2014 19:09:27 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"Em terça-feira, 23 de dezembro de 2014, Jim Nasby <Jim.Nasby@bluetreble.com>\\nescreveu:\\n\\n> On 12/23/14, 8:54 AM, Fabrízio de Royes Mello wrote:\\n>\\n>> > Right now a lot of people just work around this with things like DO\\n>> blocks, but as mentioned elsewhere in the thread that fails for commands\\n>> that can't be in a transaction.\\n>> >\\n>>\\n>> I use \\\"dblink\\\" to solve it. :-)\\n>>\\n>\\n> So... how about instead of solving this only for vacuum we create\\n> something generic? :) Possibly using Robert's background worker work?\\n\\n\\n Interesting idea.\\n\\nBut and what about the idea of improve the \\\"--table\\\" option from clients:\\nvaccumdb and clusterdb?\\n\\nRegards,\\n\\nFabrízio Mello\\n\\n\\n\\n-- \\nFabrízio de Royes Mello\\nConsultoria/Coaching PostgreSQL\\n>> Timbira: http://www.timbira.com.br\\n>> Blog: http://fabriziomello.github.io\\n>> Linkedin: http://br.linkedin.com/in/fabriziomello\\n>> Twitter: http://twitter.com/fabriziomello\\n>> Github: http://github.com/fabriziomello\\n\\nEm terça-feira, 23 de dezembro de 2014, Jim Nasby <Jim.Nasby@bluetreble.com> escreveu:On 12/23/14, 8:54 AM, Fabrízio de Royes Mello wrote:\\n\\n > Right now a lot of people just work around this with things like DO blocks, but as mentioned elsewhere in the thread that fails for commands that can't be in a transaction.\\n >\\n\\nI use \\\"dblink\\\" to solve it. :-)\\n\\n\\nSo... how about instead of solving this only for vacuum we create something generic? :) Possibly using Robert's background worker work? Interesting idea. But and what about the idea of improve the \\\"--table\\\" option from clients: vaccumdb and clusterdb?Regards,Fabrízio Mello -- Fabrízio de Royes MelloConsultoria/Coaching PostgreSQL>> Timbira: http://www.timbira.com.br>> Blog: http://fabriziomello.github.io>> Linkedin: http://br.linkedin.com/in/fabriziomello>> Twitter: http://twitter.com/fabriziomello>> Github: http://github.com/fabriziomello\",\n \"msg_date\": \"Wed, 24 Dec 2014 00:49:40 -0200\",\n \"msg_from\": \"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"On 12/23/14, 8:49 PM, Fabrízio de Royes Mello wrote:\\n> Em terça-feira, 23 de dezembro de 2014, Jim Nasby <Jim.Nasby@bluetreble.com <mailto:Jim.Nasby@bluetreble.com>> escreveu:\\n>\\n> On 12/23/14, 8:54 AM, Fabrízio de Royes Mello wrote:\\n>\\n> > Right now a lot of people just work around this with things like DO blocks, but as mentioned elsewhere in the thread that fails for commands that can't be in a transaction.\\n> >\\n>\\n> I use \\\"dblink\\\" to solve it. :-)\\n>\\n>\\n> So... how about instead of solving this only for vacuum we create something generic? :) Possibly using Robert's background worker work?\\n>\\n>\\n> Interesting idea.\\n>\\n> But and what about the idea of improve the \\\"--table\\\" option from clients: vaccumdb and clusterdb?\\n\\nSeems reasonable.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 24 Dec 2014 12:53:06 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n },\n {\n \"msg_contents\": \"21.12.2014, 18:48, Fabrízio de Royes Mello kirjoitti:\\n> I work with some customer that have databases with a lot of schemas and\\n> sometimes we need to run manual VACUUM in one schema, and would be nice\\n> to have a new option to run vacuum in relations from a specific schema.\\n> \\n> The new syntax could be:\\n> \\n> VACUUM [ FULL ] [ FREEZE ] [ VERBOSE ] { [ table_name ] | SCHEMA\\n> schema_name }\\n> \\n> Also I'll add a new option to \\\"vacuumdb\\\" client:\\n> \\n> -S, --schema=SCHEMA\\n> \\n> I can work on this feature to 2015/02 CF.\\n> \\n> Thoughts?\\n\\nThis would be useful for ANALYZE to make it easier to run analyze only\\nfor the interesting schemas after a pg_upgrade. I have a database with\\nmost of the actively used data in the \\\"public\\\" schema and a number of\\nrarely accessed large logging and archive tables in other schemas. It'd\\nbe useful to prioritize analyzing the main tables before doing anything\\nabout the rarely used schemas to allow the database to be put back into\\nproduction as soon as possible.\\n\\n/ Oskari\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sun, 28 Dec 2014 17:38:45 +0200\",\n \"msg_from\": \"Oskari Saarenmaa <os@ohmu.fi>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal \\\"VACUUM SCHEMA\\\"\"\n }\n]","isConversational":false}}},{"rowIdx":39993,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nI think ExplainModifyTarget should show the parent of the inheritance\ntree in multi-target-table cases, as described there, but noticed that\nit doesn't always work like that. Here is an example.\n\npostgres=# create table parent (a int check (a < 0) no inherit);\nCREATE TABLE\npostgres=# create table child (a int check (a >= 0));\nCREATE TABLE\npostgres=# alter table child inherit parent;\nALTER TABLE\npostgres=# explain update parent set a = a * 2 where a >= 0;\n QUERY PLAN\n---------------------------------------------------------------\n Update on child (cost=0.00..42.00 rows=800 width=10)\n -> Seq Scan on child (cost=0.00..42.00 rows=800 width=10)\n Filter: (a >= 0)\n(3 rows)\n\nIIUC, I think this is because ExplainModifyTarget doesn't take into\naccount that the parent *can* be excluded by constraint exclusion. So,\nI added a field to ModifyTable to record the parent, apart from\nresultRelations. (More precisely, the parent in its role as a simple\nmember of the inheritance tree is recorded so that appending digits to\nrefname in select_rtable_names_for_explain works as before.) Attached\nis a proposed patch for that.\n\nThanks,\n\nBest regards,\nEtsuro Fujita\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Mon, 22 Dec 2014 15:50:52 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"ExplainModifyTarget doesn't work as expected"},{"msg_contents":"On 12/22/14 12:50 AM, Etsuro Fujita wrote:\n> I think ExplainModifyTarget should show the parent of the inheritance\n> tree in multi-target-table cases, as described there, but noticed that\n> it doesn't always work like that. Here is an example.\n\nAnything ever happen with this?\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 26 Jan 2015 18:15:31 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":false,"msg_subject":"Re: ExplainModifyTarget doesn't work as expected"},{"msg_contents":"On 2015/01/27 9:15, Jim Nasby wrote:\n> On 12/22/14 12:50 AM, Etsuro Fujita wrote:\n>> I think ExplainModifyTarget should show the parent of the inheritance\n>> tree in multi-target-table cases, as described there, but noticed that\n>> it doesn't always work like that. Here is an example.\n>\n> Anything ever happen with this?\n\nNothing. I'll add this to the next CF.\n\nBest regards,\nEtsuro Fujita\n\n\n-- \nSent via 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 Jan 2015 11:33:52 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: ExplainModifyTarget doesn't work as expected"},{"msg_contents":"Hi Fujita-san,\nI agree that it's a problem, and it looks more severe when there are\nmultiple children\npostgres=# create table parent (a int check (a < 0) no inherit);\nCREATE TABLE\npostgres=# create table child1 (a int check (a >= 0));\nCREATE TABLE\npostgres=# create table child2 (a int check (a >= 0));\nCREATE TABLE\npostgres=# create table child3 (a int check (a >= 0));\nCREATE TABLE\npostgres=# alter table child1 inherit parent;\nALTER TABLE\npostgres=# alter table child2 inherit parent;\nALTER TABLE\npostgres=# alter table child3 inherit parent;\nALTER TABLE\npostgres=# explain update parent set a = a * 2 where a >= 0;\n QUERY PLAN\n----------------------------------------------------------------\n Update on child1 (cost=0.00..126.00 rows=2400 width=10)\n -> Seq Scan on child1 (cost=0.00..42.00 rows=800 width=10)\n Filter: (a >= 0)\n -> Seq Scan on child2 (cost=0.00..42.00 rows=800 width=10)\n Filter: (a >= 0)\n -> Seq Scan on child3 (cost=0.00..42.00 rows=800 width=10)\n Filter: (a >= 0)\n(7 rows)\n\nIt's certainly confusing why would an update on child1 cause scan on child*.\n\nBut I also think that showing parent's name with Upate would be misleading\nesp. when user expects it to get filtered because of constraint exclusion.\n\nInstead, can we show all the relations that are being modified e.g Update\non child1, child2, child3. That will disambiguate everything.\n\nOn Mon, Dec 22, 2014 at 12:20 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp\n> wrote:\n\n> Hi,\n>\n> I think ExplainModifyTarget should show the parent of the inheritance\n> tree in multi-target-table cases, as described there, but noticed that\n> it doesn't always work like that. Here is an example.\n>\n> postgres=# create table parent (a int check (a < 0) no inherit);\n> CREATE TABLE\n> postgres=# create table child (a int check (a >= 0));\n> CREATE TABLE\n> postgres=# alter table child inherit parent;\n> ALTER TABLE\n> postgres=# explain update parent set a = a * 2 where a >= 0;\n> QUERY PLAN\n> ---------------------------------------------------------------\n> Update on child (cost=0.00..42.00 rows=800 width=10)\n> -> Seq Scan on child (cost=0.00..42.00 rows=800 width=10)\n> Filter: (a >= 0)\n> (3 rows)\n>\n> IIUC, I think this is because ExplainModifyTarget doesn't take into\n> account that the parent *can* be excluded by constraint exclusion. So,\n> I added a field to ModifyTable to record the parent, apart from\n> resultRelations. (More precisely, the parent in its role as a simple\n> member of the inheritance tree is recorded so that appending digits to\n> refname in select_rtable_names_for_explain works as before.) Attached\n> is a proposed patch for that.\n>\n> Thanks,\n>\n> Best regards,\n> Etsuro Fujita\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\nHi Fujita-san,I agree that it's a problem, and it looks more severe when there are multiple children postgres=# create table parent (a int check (a < 0) no inherit);CREATE TABLEpostgres=# create table child1 (a int check (a >= 0));CREATE TABLEpostgres=# create table child2 (a int check (a >= 0));CREATE TABLEpostgres=# create table child3 (a int check (a >= 0));CREATE TABLEpostgres=# alter table child1 inherit parent;ALTER TABLEpostgres=# alter table child2 inherit parent;ALTER TABLEpostgres=# alter table child3 inherit parent;ALTER TABLEpostgres=# explain update parent set a = a * 2 where a >= 0; QUERY PLAN ---------------------------------------------------------------- Update on child1 (cost=0.00..126.00 rows=2400 width=10) -> Seq Scan on child1 (cost=0.00..42.00 rows=800 width=10) Filter: (a >= 0) -> Seq Scan on child2 (cost=0.00..42.00 rows=800 width=10) Filter: (a >= 0) -> Seq Scan on child3 (cost=0.00..42.00 rows=800 width=10) Filter: (a >= 0)(7 rows)It's certainly confusing why would an update on child1 cause scan on child*.But I also think that showing parent's name with Upate would be misleading esp. when user expects it to get filtered because of constraint exclusion.Instead, can we show all the relations that are being modified e.g Update on child1, child2, child3. That will disambiguate everything. On Mon, Dec 22, 2014 at 12:20 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp> wrote:Hi,\n\nI think ExplainModifyTarget should show the parent of the inheritance\ntree in multi-target-table cases, as described there, but noticed that\nit doesn't always work like that. Here is an example.\n\npostgres=# create table parent (a int check (a < 0) no inherit);\nCREATE TABLE\npostgres=# create table child (a int check (a >= 0));\nCREATE TABLE\npostgres=# alter table child inherit parent;\nALTER TABLE\npostgres=# explain update parent set a = a * 2 where a >= 0;\n QUERY PLAN\n---------------------------------------------------------------\n Update on child (cost=0.00..42.00 rows=800 width=10)\n -> Seq Scan on child (cost=0.00..42.00 rows=800 width=10)\n Filter: (a >= 0)\n(3 rows)\n\nIIUC, I think this is because ExplainModifyTarget doesn't take into\naccount that the parent *can* be excluded by constraint exclusion. So,\nI added a field to ModifyTable to record the parent, apart from\nresultRelations. (More precisely, the parent in its role as a simple\nmember of the inheritance tree is recorded so that appending digits to\nrefname in select_rtable_names_for_explain works as before.) Attached\nis a proposed patch for that.\n\nThanks,\n\nBest regards,\nEtsuro Fujita\n\n--\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":"Tue, 3 Feb 2015 12:02:04 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: ExplainModifyTarget doesn't work as expected"},{"msg_contents":"Hi Ashutosh,\n\nThank you for the review!\n\nOn 2015/02/03 15:32, Ashutosh Bapat wrote:\n> I agree that it's a problem, and it looks more severe when there are\n> multiple children\n> postgres=# create table parent (a int check (a < 0) no inherit);\n> CREATE TABLE\n> postgres=# create table child1 (a int check (a >= 0));\n> CREATE TABLE\n> postgres=# create table child2 (a int check (a >= 0));\n> CREATE TABLE\n> postgres=# create table child3 (a int check (a >= 0));\n> CREATE TABLE\n> postgres=# alter table child1 inherit parent;\n> ALTER TABLE\n> postgres=# alter table child2 inherit parent;\n> ALTER TABLE\n> postgres=# alter table child3 inherit parent;\n> ALTER TABLE\n> postgres=# explain update parent set a = a * 2 where a >= 0;\n> QUERY PLAN\n> ----------------------------------------------------------------\n> Update on child1 (cost=0.00..126.00 rows=2400 width=10)\n> -> Seq Scan on child1 (cost=0.00..42.00 rows=800 width=10)\n> Filter: (a >= 0)\n> -> Seq Scan on child2 (cost=0.00..42.00 rows=800 width=10)\n> Filter: (a >= 0)\n> -> Seq Scan on child3 (cost=0.00..42.00 rows=800 width=10)\n> Filter: (a >= 0)\n> (7 rows)\n>\n> It's certainly confusing why would an update on child1 cause scan on child*.\n\nYeah, I think so too.\n\n> But I also think that showing parent's name with Upate would be\n> misleading esp. when user expects it to get filtered because of\n> constraint exclusion.\n>\n> Instead, can we show all the relations that are being modified e.g\n> Update on child1, child2, child3. That will disambiguate everything.\n\nThat's an idea, but my concern about that is the cases where there are a \nlarge number of child tables as the EXPLAIN would be difficult to read \nin such cases.\n\nBest regards,\nEtsuro Fujita\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 06 Feb 2015 17:05:28 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: ExplainModifyTarget doesn't work as expected"},{"msg_contents":"Well let's see what others think. Also, we might want to separate that\ninformation on result relations heading probably.\n\nOn Fri, Feb 6, 2015 at 1:35 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\nwrote:\n\n> Hi Ashutosh,\n>\n> Thank you for the review!\n>\n>\n> On 2015/02/03 15:32, Ashutosh Bapat wrote:\n>\n>> I agree that it's a problem, and it looks more severe when there are\n>> multiple children\n>> postgres=# create table parent (a int check (a < 0) no inherit);\n>> CREATE TABLE\n>> postgres=# create table child1 (a int check (a >= 0));\n>> CREATE TABLE\n>> postgres=# create table child2 (a int check (a >= 0));\n>> CREATE TABLE\n>> postgres=# create table child3 (a int check (a >= 0));\n>> CREATE TABLE\n>> postgres=# alter table child1 inherit parent;\n>> ALTER TABLE\n>> postgres=# alter table child2 inherit parent;\n>> ALTER TABLE\n>> postgres=# alter table child3 inherit parent;\n>> ALTER TABLE\n>> postgres=# explain update parent set a = a * 2 where a >= 0;\n>> QUERY PLAN\n>> ----------------------------------------------------------------\n>> Update on child1 (cost=0.00..126.00 rows=2400 width=10)\n>> -> Seq Scan on child1 (cost=0.00..42.00 rows=800 width=10)\n>> Filter: (a >= 0)\n>> -> Seq Scan on child2 (cost=0.00..42.00 rows=800 width=10)\n>> Filter: (a >= 0)\n>> -> Seq Scan on child3 (cost=0.00..42.00 rows=800 width=10)\n>> Filter: (a >= 0)\n>> (7 rows)\n>>\n>> It's certainly confusing why would an update on child1 cause scan on\n>> child*.\n>>\n>\n> Yeah, I think so too.\n>\n> But I also think that showing parent's name with Upate would be\n>> misleading esp. when user expects it to get filtered because of\n>> constraint exclusion.\n>>\n>> Instead, can we show all the relations that are being modified e.g\n>> Update on child1, child2, child3. That will disambiguate everything.\n>>\n>\n> That's an idea, but my concern about that is the cases where there are a\n> large number of child tables as the EXPLAIN would be difficult to read in\n> such cases.\n>\n> Best regards,\n> Etsuro Fujita\n>\n\n\n\n-- \nBest Wishes,\nAshutosh Bapat\nEnterpriseDB Corporation\nThe Postgres Database Company\n\nWell let's see what others think. Also, we might want to separate that information on result relations heading probably.On Fri, Feb 6, 2015 at 1:35 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp> wrote:Hi Ashutosh,\n\nThank you for the review!\n\nOn 2015/02/03 15:32, Ashutosh Bapat wrote:\n\nI agree that it's a problem, and it looks more severe when there are\nmultiple children\npostgres=# create table parent (a int check (a < 0) no inherit);\nCREATE TABLE\npostgres=# create table child1 (a int check (a >= 0));\nCREATE TABLE\npostgres=# create table child2 (a int check (a >= 0));\nCREATE TABLE\npostgres=# create table child3 (a int check (a >= 0));\nCREATE TABLE\npostgres=# alter table child1 inherit parent;\nALTER TABLE\npostgres=# alter table child2 inherit parent;\nALTER TABLE\npostgres=# alter table child3 inherit parent;\nALTER TABLE\npostgres=# explain update parent set a = a * 2 where a >= 0;\n QUERY PLAN\n----------------------------------------------------------------\n Update on child1 (cost=0.00..126.00 rows=2400 width=10)\n -> Seq Scan on child1 (cost=0.00..42.00 rows=800 width=10)\n Filter: (a >= 0)\n -> Seq Scan on child2 (cost=0.00..42.00 rows=800 width=10)\n Filter: (a >= 0)\n -> Seq Scan on child3 (cost=0.00..42.00 rows=800 width=10)\n Filter: (a >= 0)\n(7 rows)\n\nIt's certainly confusing why would an update on child1 cause scan on child*.\n\n\nYeah, I think so too.\n\n\nBut I also think that showing parent's name with Upate would be\nmisleading esp. when user expects it to get filtered because of\nconstraint exclusion.\n\nInstead, can we show all the relations that are being modified e.g\nUpdate on child1, child2, child3. That will disambiguate everything.\n\n\nThat's an idea, but my concern about that is the cases where there are a large number of child tables as the EXPLAIN would be difficult to read in such cases.\n\nBest regards,\nEtsuro Fujita\n-- Best Wishes,Ashutosh BapatEnterpriseDB CorporationThe Postgres Database Company","msg_date":"Fri, 6 Feb 2015 14:10:10 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: ExplainModifyTarget doesn't work as expected"},{"msg_contents":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp> writes:\n> On 2015/02/03 15:32, Ashutosh Bapat wrote:\n>> Instead, can we show all the relations that are being modified e.g\n>> Update on child1, child2, child3. That will disambiguate everything.\n\n> That's an idea, but my concern about that is the cases where there are a \n> large number of child tables as the EXPLAIN would be difficult to read \n> in such cases.\n\nI concur, that would *not* be an improvement in readability. Moreover,\nI don't think it really fixes the issue: what we want to show is a table\nname in Modify that matches what the user wrote in the query. Given that\ncontext, the user should be able to understand why some tables are listed\nbelow that and others are not.\n\nIIRC, this code was written at a time when we didn't have NO INHERIT check\nconstraints and so it was impossible for the parent table to get optimized\naway while leaving children. So the comment in ExplainModifyTarget was\ngood at the time. But it no longer is.\n\nI think your basic idea of preserving the original parent table's relid\nis correct; but instead of doing it like this patch does, I'd be inclined\nto make ModifyTable inherit from Scan not Plan, and use the scan.scanrelid\nfield to carry the parent RTI. Then you would probably end up with a net\nsavings of code rather than net addition; certainly ExplainModifyTarget\nwould go away entirely since you'd just treat ModifyTable like any other\nScan in this part of EXPLAIN.\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, 06 Feb 2015 11:09:32 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: ExplainModifyTarget doesn't work as expected"},{"msg_contents":"On 2015/02/07 1:09, Tom Lane wrote:\n> IIRC, this code was written at a time when we didn't have NO INHERIT check\n> constraints and so it was impossible for the parent table to get optimized\n> away while leaving children. So the comment in ExplainModifyTarget was\n> good at the time. But it no longer is.\n> \n> I think your basic idea of preserving the original parent table's relid\n> is correct; but instead of doing it like this patch does, I'd be inclined\n> to make ModifyTable inherit from Scan not Plan, and use the scan.scanrelid\n> field to carry the parent RTI. Then you would probably end up with a net\n> savings of code rather than net addition; certainly ExplainModifyTarget\n> would go away entirely since you'd just treat ModifyTable like any other\n> Scan in this part of EXPLAIN.\n\nWill follow your revision.\n\nThanks!\n\nBest regards,\nEtsuro Fujita\n\n\n-- \nSent via 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, 10 Feb 2015 14:49:50 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: ExplainModifyTarget doesn't work as expected"},{"msg_contents":"On 2015/02/10 14:49, Etsuro Fujita wrote:\n> On 2015/02/07 1:09, Tom Lane wrote:\n>> IIRC, this code was written at a time when we didn't have NO INHERIT check\n>> constraints and so it was impossible for the parent table to get optimized\n>> away while leaving children. So the comment in ExplainModifyTarget was\n>> good at the time. But it no longer is.\n>>\n>> I think your basic idea of preserving the original parent table's relid\n>> is correct; but instead of doing it like this patch does, I'd be inclined\n>> to make ModifyTable inherit from Scan not Plan, and use the scan.scanrelid\n>> field to carry the parent RTI. Then you would probably end up with a net\n>> savings of code rather than net addition; certainly ExplainModifyTarget\n>> would go away entirely since you'd just treat ModifyTable like any other\n>> Scan in this part of EXPLAIN.\n> \n> Will follow your revision.\n\nDone. Attached 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":"Thu, 12 Feb 2015 14:05:31 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: ExplainModifyTarget doesn't work as expected"},{"msg_contents":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp> writes:\n> On 2015/02/10 14:49, Etsuro Fujita wrote:\n>> On 2015/02/07 1:09, Tom Lane wrote:\n>>> I think your basic idea of preserving the original parent table's relid\n>>> is correct; but instead of doing it like this patch does, I'd be inclined\n>>> to make ModifyTable inherit from Scan not Plan, and use the scan.scanrelid\n>>> field to carry the parent RTI. Then you would probably end up with a net\n>>> savings of code rather than net addition; certainly ExplainModifyTarget\n>>> would go away entirely since you'd just treat ModifyTable like any other\n>>> Scan in this part of EXPLAIN.\n\n>> Will follow your revision.\n\n> Done. Attached is an updated version of the patch.\n\nI looked at this and was not really pleased with the results, in\nparticular the way that you'd moved a bare minimum number of the code\nstanzas for struct Scan so that things still compiled. The new placement\nof those stanzas didn't make any sense in context, and it was a\nsignificant violation of our layout rule that files dealing with various\ntypes of Nodes should whenever possible handle those Nodes in the same\norder that they're listed in nodes.h, so that it's clear where new bits of\ncode ought to be placed. (I'm aware that there are historical violations\nof this policy in a few places, but that doesn't make it a bad policy to\nfollow.)\n\nI experimented with relocating ModifyTable down into the group of Scan\nnodes in this global ordering, but soon decided that that would involve\nfar more code churn than the idea was worth.\n\nSo I went back to your v1 patch and have now committed that with some\ncosmetic modifications. Sorry for making you put time into a dead end.\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, 17 Feb 2015 18:13:42 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: ExplainModifyTarget doesn't work as expected"},{"msg_contents":"On 2015/02/18 8:13, Tom Lane wrote:\n> So I went back to your v1 patch and have now committed that with some\n> cosmetic modifications. Sorry for making you put time into a dead end.\n\nI don't mind it. Thanks for committing the patch!\n\nBest regards,\nEtsuro Fujita\n\n\n-- \nSent via 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 Feb 2015 18:28:25 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: ExplainModifyTarget doesn't work as expected"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nI think ExplainModifyTarget should show the parent of the inheritance\\ntree in multi-target-table cases, as described there, but noticed that\\nit doesn't always work like that. Here is an example.\\n\\npostgres=# create table parent (a int check (a < 0) no inherit);\\nCREATE TABLE\\npostgres=# create table child (a int check (a >= 0));\\nCREATE TABLE\\npostgres=# alter table child inherit parent;\\nALTER TABLE\\npostgres=# explain update parent set a = a * 2 where a >= 0;\\n QUERY PLAN\\n---------------------------------------------------------------\\n Update on child (cost=0.00..42.00 rows=800 width=10)\\n -> Seq Scan on child (cost=0.00..42.00 rows=800 width=10)\\n Filter: (a >= 0)\\n(3 rows)\\n\\nIIUC, I think this is because ExplainModifyTarget doesn't take into\\naccount that the parent *can* be excluded by constraint exclusion. So,\\nI added a field to ModifyTable to record the parent, apart from\\nresultRelations. (More precisely, the parent in its role as a simple\\nmember of the inheritance tree is recorded so that appending digits to\\nrefname in select_rtable_names_for_explain works as before.) Attached\\nis a proposed patch for that.\\n\\nThanks,\\n\\nBest regards,\\nEtsuro Fujita\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Mon, 22 Dec 2014 15:50:52 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"ExplainModifyTarget doesn't work as expected\"\n },\n {\n \"msg_contents\": \"On 12/22/14 12:50 AM, Etsuro Fujita wrote:\\n> I think ExplainModifyTarget should show the parent of the inheritance\\n> tree in multi-target-table cases, as described there, but noticed that\\n> it doesn't always work like that. Here is an example.\\n\\nAnything ever happen with this?\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 26 Jan 2015 18:15:31 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ExplainModifyTarget doesn't work as expected\"\n },\n {\n \"msg_contents\": \"On 2015/01/27 9:15, Jim Nasby wrote:\\n> On 12/22/14 12:50 AM, Etsuro Fujita wrote:\\n>> I think ExplainModifyTarget should show the parent of the inheritance\\n>> tree in multi-target-table cases, as described there, but noticed that\\n>> it doesn't always work like that. Here is an example.\\n>\\n> Anything ever happen with this?\\n\\nNothing. I'll add this to the next CF.\\n\\nBest regards,\\nEtsuro Fujita\\n\\n\\n-- \\nSent via pgsql-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 Jan 2015 11:33:52 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ExplainModifyTarget doesn't work as expected\"\n },\n {\n \"msg_contents\": \"Hi Fujita-san,\\nI agree that it's a problem, and it looks more severe when there are\\nmultiple children\\npostgres=# create table parent (a int check (a < 0) no inherit);\\nCREATE TABLE\\npostgres=# create table child1 (a int check (a >= 0));\\nCREATE TABLE\\npostgres=# create table child2 (a int check (a >= 0));\\nCREATE TABLE\\npostgres=# create table child3 (a int check (a >= 0));\\nCREATE TABLE\\npostgres=# alter table child1 inherit parent;\\nALTER TABLE\\npostgres=# alter table child2 inherit parent;\\nALTER TABLE\\npostgres=# alter table child3 inherit parent;\\nALTER TABLE\\npostgres=# explain update parent set a = a * 2 where a >= 0;\\n QUERY PLAN\\n----------------------------------------------------------------\\n Update on child1 (cost=0.00..126.00 rows=2400 width=10)\\n -> Seq Scan on child1 (cost=0.00..42.00 rows=800 width=10)\\n Filter: (a >= 0)\\n -> Seq Scan on child2 (cost=0.00..42.00 rows=800 width=10)\\n Filter: (a >= 0)\\n -> Seq Scan on child3 (cost=0.00..42.00 rows=800 width=10)\\n Filter: (a >= 0)\\n(7 rows)\\n\\nIt's certainly confusing why would an update on child1 cause scan on child*.\\n\\nBut I also think that showing parent's name with Upate would be misleading\\nesp. when user expects it to get filtered because of constraint exclusion.\\n\\nInstead, can we show all the relations that are being modified e.g Update\\non child1, child2, child3. That will disambiguate everything.\\n\\nOn Mon, Dec 22, 2014 at 12:20 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp\\n> wrote:\\n\\n> Hi,\\n>\\n> I think ExplainModifyTarget should show the parent of the inheritance\\n> tree in multi-target-table cases, as described there, but noticed that\\n> it doesn't always work like that. Here is an example.\\n>\\n> postgres=# create table parent (a int check (a < 0) no inherit);\\n> CREATE TABLE\\n> postgres=# create table child (a int check (a >= 0));\\n> CREATE TABLE\\n> postgres=# alter table child inherit parent;\\n> ALTER TABLE\\n> postgres=# explain update parent set a = a * 2 where a >= 0;\\n> QUERY PLAN\\n> ---------------------------------------------------------------\\n> Update on child (cost=0.00..42.00 rows=800 width=10)\\n> -> Seq Scan on child (cost=0.00..42.00 rows=800 width=10)\\n> Filter: (a >= 0)\\n> (3 rows)\\n>\\n> IIUC, I think this is because ExplainModifyTarget doesn't take into\\n> account that the parent *can* be excluded by constraint exclusion. So,\\n> I added a field to ModifyTable to record the parent, apart from\\n> resultRelations. (More precisely, the parent in its role as a simple\\n> member of the inheritance tree is recorded so that appending digits to\\n> refname in select_rtable_names_for_explain works as before.) Attached\\n> is a proposed patch for that.\\n>\\n> Thanks,\\n>\\n> Best regards,\\n> Etsuro Fujita\\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\\nHi Fujita-san,I agree that it's a problem, and it looks more severe when there are multiple children postgres=# create table parent (a int check (a < 0) no inherit);CREATE TABLEpostgres=# create table child1 (a int check (a >= 0));CREATE TABLEpostgres=# create table child2 (a int check (a >= 0));CREATE TABLEpostgres=# create table child3 (a int check (a >= 0));CREATE TABLEpostgres=# alter table child1 inherit parent;ALTER TABLEpostgres=# alter table child2 inherit parent;ALTER TABLEpostgres=# alter table child3 inherit parent;ALTER TABLEpostgres=# explain update parent set a = a * 2 where a >= 0; QUERY PLAN ---------------------------------------------------------------- Update on child1 (cost=0.00..126.00 rows=2400 width=10) -> Seq Scan on child1 (cost=0.00..42.00 rows=800 width=10) Filter: (a >= 0) -> Seq Scan on child2 (cost=0.00..42.00 rows=800 width=10) Filter: (a >= 0) -> Seq Scan on child3 (cost=0.00..42.00 rows=800 width=10) Filter: (a >= 0)(7 rows)It's certainly confusing why would an update on child1 cause scan on child*.But I also think that showing parent's name with Upate would be misleading esp. when user expects it to get filtered because of constraint exclusion.Instead, can we show all the relations that are being modified e.g Update on child1, child2, child3. That will disambiguate everything. On Mon, Dec 22, 2014 at 12:20 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp> wrote:Hi,\\n\\nI think ExplainModifyTarget should show the parent of the inheritance\\ntree in multi-target-table cases, as described there, but noticed that\\nit doesn't always work like that. Here is an example.\\n\\npostgres=# create table parent (a int check (a < 0) no inherit);\\nCREATE TABLE\\npostgres=# create table child (a int check (a >= 0));\\nCREATE TABLE\\npostgres=# alter table child inherit parent;\\nALTER TABLE\\npostgres=# explain update parent set a = a * 2 where a >= 0;\\n QUERY PLAN\\n---------------------------------------------------------------\\n Update on child (cost=0.00..42.00 rows=800 width=10)\\n -> Seq Scan on child (cost=0.00..42.00 rows=800 width=10)\\n Filter: (a >= 0)\\n(3 rows)\\n\\nIIUC, I think this is because ExplainModifyTarget doesn't take into\\naccount that the parent *can* be excluded by constraint exclusion. So,\\nI added a field to ModifyTable to record the parent, apart from\\nresultRelations. (More precisely, the parent in its role as a simple\\nmember of the inheritance tree is recorded so that appending digits to\\nrefname in select_rtable_names_for_explain works as before.) Attached\\nis a proposed patch for that.\\n\\nThanks,\\n\\nBest regards,\\nEtsuro Fujita\\n\\n--\\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\": \"Tue, 3 Feb 2015 12:02:04 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ExplainModifyTarget doesn't work as expected\"\n },\n {\n \"msg_contents\": \"Hi Ashutosh,\\n\\nThank you for the review!\\n\\nOn 2015/02/03 15:32, Ashutosh Bapat wrote:\\n> I agree that it's a problem, and it looks more severe when there are\\n> multiple children\\n> postgres=# create table parent (a int check (a < 0) no inherit);\\n> CREATE TABLE\\n> postgres=# create table child1 (a int check (a >= 0));\\n> CREATE TABLE\\n> postgres=# create table child2 (a int check (a >= 0));\\n> CREATE TABLE\\n> postgres=# create table child3 (a int check (a >= 0));\\n> CREATE TABLE\\n> postgres=# alter table child1 inherit parent;\\n> ALTER TABLE\\n> postgres=# alter table child2 inherit parent;\\n> ALTER TABLE\\n> postgres=# alter table child3 inherit parent;\\n> ALTER TABLE\\n> postgres=# explain update parent set a = a * 2 where a >= 0;\\n> QUERY PLAN\\n> ----------------------------------------------------------------\\n> Update on child1 (cost=0.00..126.00 rows=2400 width=10)\\n> -> Seq Scan on child1 (cost=0.00..42.00 rows=800 width=10)\\n> Filter: (a >= 0)\\n> -> Seq Scan on child2 (cost=0.00..42.00 rows=800 width=10)\\n> Filter: (a >= 0)\\n> -> Seq Scan on child3 (cost=0.00..42.00 rows=800 width=10)\\n> Filter: (a >= 0)\\n> (7 rows)\\n>\\n> It's certainly confusing why would an update on child1 cause scan on child*.\\n\\nYeah, I think so too.\\n\\n> But I also think that showing parent's name with Upate would be\\n> misleading esp. when user expects it to get filtered because of\\n> constraint exclusion.\\n>\\n> Instead, can we show all the relations that are being modified e.g\\n> Update on child1, child2, child3. That will disambiguate everything.\\n\\nThat's an idea, but my concern about that is the cases where there are a \\nlarge number of child tables as the EXPLAIN would be difficult to read \\nin such cases.\\n\\nBest regards,\\nEtsuro Fujita\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 06 Feb 2015 17:05:28 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ExplainModifyTarget doesn't work as expected\"\n },\n {\n \"msg_contents\": \"Well let's see what others think. Also, we might want to separate that\\ninformation on result relations heading probably.\\n\\nOn Fri, Feb 6, 2015 at 1:35 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\\nwrote:\\n\\n> Hi Ashutosh,\\n>\\n> Thank you for the review!\\n>\\n>\\n> On 2015/02/03 15:32, Ashutosh Bapat wrote:\\n>\\n>> I agree that it's a problem, and it looks more severe when there are\\n>> multiple children\\n>> postgres=# create table parent (a int check (a < 0) no inherit);\\n>> CREATE TABLE\\n>> postgres=# create table child1 (a int check (a >= 0));\\n>> CREATE TABLE\\n>> postgres=# create table child2 (a int check (a >= 0));\\n>> CREATE TABLE\\n>> postgres=# create table child3 (a int check (a >= 0));\\n>> CREATE TABLE\\n>> postgres=# alter table child1 inherit parent;\\n>> ALTER TABLE\\n>> postgres=# alter table child2 inherit parent;\\n>> ALTER TABLE\\n>> postgres=# alter table child3 inherit parent;\\n>> ALTER TABLE\\n>> postgres=# explain update parent set a = a * 2 where a >= 0;\\n>> QUERY PLAN\\n>> ----------------------------------------------------------------\\n>> Update on child1 (cost=0.00..126.00 rows=2400 width=10)\\n>> -> Seq Scan on child1 (cost=0.00..42.00 rows=800 width=10)\\n>> Filter: (a >= 0)\\n>> -> Seq Scan on child2 (cost=0.00..42.00 rows=800 width=10)\\n>> Filter: (a >= 0)\\n>> -> Seq Scan on child3 (cost=0.00..42.00 rows=800 width=10)\\n>> Filter: (a >= 0)\\n>> (7 rows)\\n>>\\n>> It's certainly confusing why would an update on child1 cause scan on\\n>> child*.\\n>>\\n>\\n> Yeah, I think so too.\\n>\\n> But I also think that showing parent's name with Upate would be\\n>> misleading esp. when user expects it to get filtered because of\\n>> constraint exclusion.\\n>>\\n>> Instead, can we show all the relations that are being modified e.g\\n>> Update on child1, child2, child3. That will disambiguate everything.\\n>>\\n>\\n> That's an idea, but my concern about that is the cases where there are a\\n> large number of child tables as the EXPLAIN would be difficult to read in\\n> such cases.\\n>\\n> Best regards,\\n> Etsuro Fujita\\n>\\n\\n\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\nEnterpriseDB Corporation\\nThe Postgres Database Company\\n\\nWell let's see what others think. Also, we might want to separate that information on result relations heading probably.On Fri, Feb 6, 2015 at 1:35 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp> wrote:Hi Ashutosh,\\n\\nThank you for the review!\\n\\nOn 2015/02/03 15:32, Ashutosh Bapat wrote:\\n\\nI agree that it's a problem, and it looks more severe when there are\\nmultiple children\\npostgres=# create table parent (a int check (a < 0) no inherit);\\nCREATE TABLE\\npostgres=# create table child1 (a int check (a >= 0));\\nCREATE TABLE\\npostgres=# create table child2 (a int check (a >= 0));\\nCREATE TABLE\\npostgres=# create table child3 (a int check (a >= 0));\\nCREATE TABLE\\npostgres=# alter table child1 inherit parent;\\nALTER TABLE\\npostgres=# alter table child2 inherit parent;\\nALTER TABLE\\npostgres=# alter table child3 inherit parent;\\nALTER TABLE\\npostgres=# explain update parent set a = a * 2 where a >= 0;\\n QUERY PLAN\\n----------------------------------------------------------------\\n Update on child1 (cost=0.00..126.00 rows=2400 width=10)\\n -> Seq Scan on child1 (cost=0.00..42.00 rows=800 width=10)\\n Filter: (a >= 0)\\n -> Seq Scan on child2 (cost=0.00..42.00 rows=800 width=10)\\n Filter: (a >= 0)\\n -> Seq Scan on child3 (cost=0.00..42.00 rows=800 width=10)\\n Filter: (a >= 0)\\n(7 rows)\\n\\nIt's certainly confusing why would an update on child1 cause scan on child*.\\n\\n\\nYeah, I think so too.\\n\\n\\nBut I also think that showing parent's name with Upate would be\\nmisleading esp. when user expects it to get filtered because of\\nconstraint exclusion.\\n\\nInstead, can we show all the relations that are being modified e.g\\nUpdate on child1, child2, child3. That will disambiguate everything.\\n\\n\\nThat's an idea, but my concern about that is the cases where there are a large number of child tables as the EXPLAIN would be difficult to read in such cases.\\n\\nBest regards,\\nEtsuro Fujita\\n-- Best Wishes,Ashutosh BapatEnterpriseDB CorporationThe Postgres Database Company\",\n \"msg_date\": \"Fri, 6 Feb 2015 14:10:10 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ExplainModifyTarget doesn't work as expected\"\n },\n {\n \"msg_contents\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp> writes:\\n> On 2015/02/03 15:32, Ashutosh Bapat wrote:\\n>> Instead, can we show all the relations that are being modified e.g\\n>> Update on child1, child2, child3. That will disambiguate everything.\\n\\n> That's an idea, but my concern about that is the cases where there are a \\n> large number of child tables as the EXPLAIN would be difficult to read \\n> in such cases.\\n\\nI concur, that would *not* be an improvement in readability. Moreover,\\nI don't think it really fixes the issue: what we want to show is a table\\nname in Modify that matches what the user wrote in the query. Given that\\ncontext, the user should be able to understand why some tables are listed\\nbelow that and others are not.\\n\\nIIRC, this code was written at a time when we didn't have NO INHERIT check\\nconstraints and so it was impossible for the parent table to get optimized\\naway while leaving children. So the comment in ExplainModifyTarget was\\ngood at the time. But it no longer is.\\n\\nI think your basic idea of preserving the original parent table's relid\\nis correct; but instead of doing it like this patch does, I'd be inclined\\nto make ModifyTable inherit from Scan not Plan, and use the scan.scanrelid\\nfield to carry the parent RTI. Then you would probably end up with a net\\nsavings of code rather than net addition; certainly ExplainModifyTarget\\nwould go away entirely since you'd just treat ModifyTable like any other\\nScan in this part of EXPLAIN.\\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, 06 Feb 2015 11:09:32 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ExplainModifyTarget doesn't work as expected\"\n },\n {\n \"msg_contents\": \"On 2015/02/07 1:09, Tom Lane wrote:\\n> IIRC, this code was written at a time when we didn't have NO INHERIT check\\n> constraints and so it was impossible for the parent table to get optimized\\n> away while leaving children. So the comment in ExplainModifyTarget was\\n> good at the time. But it no longer is.\\n> \\n> I think your basic idea of preserving the original parent table's relid\\n> is correct; but instead of doing it like this patch does, I'd be inclined\\n> to make ModifyTable inherit from Scan not Plan, and use the scan.scanrelid\\n> field to carry the parent RTI. Then you would probably end up with a net\\n> savings of code rather than net addition; certainly ExplainModifyTarget\\n> would go away entirely since you'd just treat ModifyTable like any other\\n> Scan in this part of EXPLAIN.\\n\\nWill follow your revision.\\n\\nThanks!\\n\\nBest regards,\\nEtsuro Fujita\\n\\n\\n-- \\nSent via pgsql-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, 10 Feb 2015 14:49:50 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ExplainModifyTarget doesn't work as expected\"\n },\n {\n \"msg_contents\": \"On 2015/02/10 14:49, Etsuro Fujita wrote:\\n> On 2015/02/07 1:09, Tom Lane wrote:\\n>> IIRC, this code was written at a time when we didn't have NO INHERIT check\\n>> constraints and so it was impossible for the parent table to get optimized\\n>> away while leaving children. So the comment in ExplainModifyTarget was\\n>> good at the time. But it no longer is.\\n>>\\n>> I think your basic idea of preserving the original parent table's relid\\n>> is correct; but instead of doing it like this patch does, I'd be inclined\\n>> to make ModifyTable inherit from Scan not Plan, and use the scan.scanrelid\\n>> field to carry the parent RTI. Then you would probably end up with a net\\n>> savings of code rather than net addition; certainly ExplainModifyTarget\\n>> would go away entirely since you'd just treat ModifyTable like any other\\n>> Scan in this part of EXPLAIN.\\n> \\n> Will follow your revision.\\n\\nDone. Attached 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\": \"Thu, 12 Feb 2015 14:05:31 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ExplainModifyTarget doesn't work as expected\"\n },\n {\n \"msg_contents\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp> writes:\\n> On 2015/02/10 14:49, Etsuro Fujita wrote:\\n>> On 2015/02/07 1:09, Tom Lane wrote:\\n>>> I think your basic idea of preserving the original parent table's relid\\n>>> is correct; but instead of doing it like this patch does, I'd be inclined\\n>>> to make ModifyTable inherit from Scan not Plan, and use the scan.scanrelid\\n>>> field to carry the parent RTI. Then you would probably end up with a net\\n>>> savings of code rather than net addition; certainly ExplainModifyTarget\\n>>> would go away entirely since you'd just treat ModifyTable like any other\\n>>> Scan in this part of EXPLAIN.\\n\\n>> Will follow your revision.\\n\\n> Done. Attached is an updated version of the patch.\\n\\nI looked at this and was not really pleased with the results, in\\nparticular the way that you'd moved a bare minimum number of the code\\nstanzas for struct Scan so that things still compiled. The new placement\\nof those stanzas didn't make any sense in context, and it was a\\nsignificant violation of our layout rule that files dealing with various\\ntypes of Nodes should whenever possible handle those Nodes in the same\\norder that they're listed in nodes.h, so that it's clear where new bits of\\ncode ought to be placed. (I'm aware that there are historical violations\\nof this policy in a few places, but that doesn't make it a bad policy to\\nfollow.)\\n\\nI experimented with relocating ModifyTable down into the group of Scan\\nnodes in this global ordering, but soon decided that that would involve\\nfar more code churn than the idea was worth.\\n\\nSo I went back to your v1 patch and have now committed that with some\\ncosmetic modifications. Sorry for making you put time into a dead end.\\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, 17 Feb 2015 18:13:42 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ExplainModifyTarget doesn't work as expected\"\n },\n {\n \"msg_contents\": \"On 2015/02/18 8:13, Tom Lane wrote:\\n> So I went back to your v1 patch and have now committed that with some\\n> cosmetic modifications. Sorry for making you put time into a dead end.\\n\\nI don't mind it. Thanks for committing the patch!\\n\\nBest regards,\\nEtsuro Fujita\\n\\n\\n-- \\nSent via pgsql-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, 18 Feb 2015 18:28:25 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ExplainModifyTarget doesn't work as expected\"\n }\n]","isConversational":false}}},{"rowIdx":39994,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi all,\n\nI have been though the patches of the current CF, looking at their\nrelated threads and updating each patch status if needed. After one\nweek in this CF, we have done progress on many patches, more than 2/3\nof them getting some comments, reviews and/or refreshed versions.\n\nLooking at the CF app for more details, we have the following numbers\nshowing up:\n- Needs Review: 43\n- Waiting on Author: 23\n- Ready for Committer: 4\n- Committed: 7\n- Returned with Feedback: 1\nThe number of patches waiting on author is high meaning that many\nreviewers showed up, but there is still much work to do in terms of\nreviews.\n\nNote that there are still 33 patches that do not have a reviewer\nassigned in the CF app. In some cases, no people looked at them, but\nin most of them some people punctually commented on them.\n\nHence, to continue the progress on this CF, be sure to do the following things:\n- As a patch submitter, check your patch status. Resubmit or\ncounter-argue if necessary as you feel. If your patch reviewer is not\nshowing up, be sure to ping him/her. Don't forget to review as well\none patch per patch submitted, of equal difficulty if possible.\nLooking at new areas of the code is highly recommended.\n- As a patch reviewer, if the patch you are reviewing is on \"waiting\non author\" for a long time, be sure to ping the patch submitter.\n- Enjoy Christmas and your vacations with your relatives and friends.\nRegards,\n-- \nMichael\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 22 Dec 2014 16:24:52 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Commit Fest 2014-12, Status after week 1"}],"string":"[\n {\n \"msg_contents\": \"Hi all,\\n\\nI have been though the patches of the current CF, looking at their\\nrelated threads and updating each patch status if needed. After one\\nweek in this CF, we have done progress on many patches, more than 2/3\\nof them getting some comments, reviews and/or refreshed versions.\\n\\nLooking at the CF app for more details, we have the following numbers\\nshowing up:\\n- Needs Review: 43\\n- Waiting on Author: 23\\n- Ready for Committer: 4\\n- Committed: 7\\n- Returned with Feedback: 1\\nThe number of patches waiting on author is high meaning that many\\nreviewers showed up, but there is still much work to do in terms of\\nreviews.\\n\\nNote that there are still 33 patches that do not have a reviewer\\nassigned in the CF app. In some cases, no people looked at them, but\\nin most of them some people punctually commented on them.\\n\\nHence, to continue the progress on this CF, be sure to do the following things:\\n- As a patch submitter, check your patch status. Resubmit or\\ncounter-argue if necessary as you feel. If your patch reviewer is not\\nshowing up, be sure to ping him/her. Don't forget to review as well\\none patch per patch submitted, of equal difficulty if possible.\\nLooking at new areas of the code is highly recommended.\\n- As a patch reviewer, if the patch you are reviewing is on \\\"waiting\\non author\\\" for a long time, be sure to ping the patch submitter.\\n- Enjoy Christmas and your vacations with your relatives and friends.\\nRegards,\\n-- \\nMichael\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 22 Dec 2014 16:24:52 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Commit Fest 2014-12, Status after week 1\"\n }\n]","isConversational":false}}},{"rowIdx":39995,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Peter Geoghegan suggested [1] moving rbtree.c to src/backend/lib, which \nI think makes a lot of sense. Now that we have several other general \npurpose data structures in src/backend/lib (linked lists, a binary heap, \nand a pairing heap), rbtree.c would definitely be better placed in \nsrc/backend/lib, too.\n\nThe usual objection to moving things is that it makes back-patching \nharder. It also might break third-party code that use it (since \npresumably we would also move the .h file). Nevertheless, I feel the \nadvantages outweigh the disadvantages in this case.\n\nAny objections?\n\n- Heikki\n\n\n-- \nSent via 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 Dec 2014 12:16:11 +0200","msg_from":"Heikki Linnakangas <hlinnakangas@vmware.com>","msg_from_op":true,"msg_subject":"Moving src/backend/utils/misc/rbtree.c to src/backend/lib"},{"msg_contents":"Heikki Linnakangas <hlinnakangas@vmware.com> writes:\n> Peter Geoghegan suggested [1] moving rbtree.c to src/backend/lib, which \n> I think makes a lot of sense. Now that we have several other general \n> purpose data structures in src/backend/lib (linked lists, a binary heap, \n> and a pairing heap), rbtree.c would definitely be better placed in \n> src/backend/lib, too.\n\n> The usual objection to moving things is that it makes back-patching \n> harder. It also might break third-party code that use it (since \n> presumably we would also move the .h file). Nevertheless, I feel the \n> advantages outweigh the disadvantages in this case.\n\n> Any objections?\n\nA look at the git history says that rbtree.h/.c have not been touched\n(except by copyright/pgindent commits) since 9.0, so probably the\nbackpatch argument doesn't have much force.\n\nHowever, wasn't there some speculation about removing rbtree entirely?\nIf we're going to end up doing that, moving the files first is just\nunnecessary thrashing.\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 Dec 2014 10:19:48 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Moving src/backend/utils/misc/rbtree.c to src/backend/lib"},{"msg_contents":"On 12/22/2014 05:19 PM, Tom Lane wrote:\n> However, wasn't there some speculation about removing rbtree entirely?\n\nNot that I recall. It's still used for GIN bulk loading. There might be \nbetter ways to do that, but there hasn't been any serious discussion on \nthat.\n\nThere was some discussion on replacing the existing binary heap usage \nwith the pairing heap, in MergeAppend and in tuplesort.c, but that's a \ndifferent story.\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 Dec 2014 17:24:05 +0200","msg_from":"Heikki Linnakangas <hlinnakangas@vmware.com>","msg_from_op":true,"msg_subject":"Re: Moving src/backend/utils/misc/rbtree.c to src/backend/lib"},{"msg_contents":"On Mon, Dec 22, 2014 at 5:16 AM, Heikki Linnakangas\n<hlinnakangas@vmware.com> wrote:\n> Peter Geoghegan suggested [1] moving rbtree.c to src/backend/lib, which I\n> think makes a lot of sense. Now that we have several other general purpose\n> data structures in src/backend/lib (linked lists, a binary heap, and a\n> pairing heap), rbtree.c would definitely be better placed in\n> src/backend/lib, too.\n>\n> The usual objection to moving things is that it makes back-patching harder.\n> It also might break third-party code that use it (since presumably we would\n> also move the .h file). Nevertheless, I feel the advantages outweigh the\n> disadvantages in this case.\n\nI agree.\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 Dec 2014 10:44:50 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Moving src/backend/utils/misc/rbtree.c to src/backend/lib"}],"string":"[\n {\n \"msg_contents\": \"Peter Geoghegan suggested [1] moving rbtree.c to src/backend/lib, which \\nI think makes a lot of sense. Now that we have several other general \\npurpose data structures in src/backend/lib (linked lists, a binary heap, \\nand a pairing heap), rbtree.c would definitely be better placed in \\nsrc/backend/lib, too.\\n\\nThe usual objection to moving things is that it makes back-patching \\nharder. It also might break third-party code that use it (since \\npresumably we would also move the .h file). Nevertheless, I feel the \\nadvantages outweigh the disadvantages in this case.\\n\\nAny objections?\\n\\n- Heikki\\n\\n\\n-- \\nSent via pgsql-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 Dec 2014 12:16:11 +0200\",\n \"msg_from\": \"Heikki Linnakangas <hlinnakangas@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Moving src/backend/utils/misc/rbtree.c to src/backend/lib\"\n },\n {\n \"msg_contents\": \"Heikki Linnakangas <hlinnakangas@vmware.com> writes:\\n> Peter Geoghegan suggested [1] moving rbtree.c to src/backend/lib, which \\n> I think makes a lot of sense. Now that we have several other general \\n> purpose data structures in src/backend/lib (linked lists, a binary heap, \\n> and a pairing heap), rbtree.c would definitely be better placed in \\n> src/backend/lib, too.\\n\\n> The usual objection to moving things is that it makes back-patching \\n> harder. It also might break third-party code that use it (since \\n> presumably we would also move the .h file). Nevertheless, I feel the \\n> advantages outweigh the disadvantages in this case.\\n\\n> Any objections?\\n\\nA look at the git history says that rbtree.h/.c have not been touched\\n(except by copyright/pgindent commits) since 9.0, so probably the\\nbackpatch argument doesn't have much force.\\n\\nHowever, wasn't there some speculation about removing rbtree entirely?\\nIf we're going to end up doing that, moving the files first is just\\nunnecessary thrashing.\\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 Dec 2014 10:19:48 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Moving src/backend/utils/misc/rbtree.c to src/backend/lib\"\n },\n {\n \"msg_contents\": \"On 12/22/2014 05:19 PM, Tom Lane wrote:\\n> However, wasn't there some speculation about removing rbtree entirely?\\n\\nNot that I recall. It's still used for GIN bulk loading. There might be \\nbetter ways to do that, but there hasn't been any serious discussion on \\nthat.\\n\\nThere was some discussion on replacing the existing binary heap usage \\nwith the pairing heap, in MergeAppend and in tuplesort.c, but that's a \\ndifferent story.\\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\": \"Mon, 22 Dec 2014 17:24:05 +0200\",\n \"msg_from\": \"Heikki Linnakangas <hlinnakangas@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Moving src/backend/utils/misc/rbtree.c to src/backend/lib\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 22, 2014 at 5:16 AM, Heikki Linnakangas\\n<hlinnakangas@vmware.com> wrote:\\n> Peter Geoghegan suggested [1] moving rbtree.c to src/backend/lib, which I\\n> think makes a lot of sense. Now that we have several other general purpose\\n> data structures in src/backend/lib (linked lists, a binary heap, and a\\n> pairing heap), rbtree.c would definitely be better placed in\\n> src/backend/lib, too.\\n>\\n> The usual objection to moving things is that it makes back-patching harder.\\n> It also might break third-party code that use it (since presumably we would\\n> also move the .h file). Nevertheless, I feel the advantages outweigh the\\n> disadvantages in this case.\\n\\nI agree.\\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 Dec 2014 10:44:50 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Moving src/backend/utils/misc/rbtree.c to src/backend/lib\"\n }\n]","isConversational":false}}},{"rowIdx":39996,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"The following bug has been logged on the website:\n\nBug reference: 12320\nLogged by: Aaron Botsis\nEmail address: postgres@bt-r.com\nPostgreSQL version: 9.4.0\nOperating system: Mac OSX\nDescription: \n\npostgres=# create table ct (data jsonb);\nCREATE TABLE\npostgres=# copy ct (data) from '/Users/aaron/Desktop/queries.json';\nERROR: invalid input syntax for type json\nDETAIL: Token \"root\" is invalid.\nCONTEXT: JSON data, line 1: ...1418066241619 AND <=1418671041621) AND\nuser:\"root...\nCOPY ct, line 3, column data:\n\"{\"timestamp\":\"2014-12-15T19:17:32.505Z\",\"duration\":7.947,\"query\":{\"query\":{\"filtered\":{\"filter\":{\"qu...\"\n\n\nThis is a sample of the JSON:\n\n{\"timestamp\":\"2014-12-15T19:17:32.505Z\",\"duration\":7.947,\"query\":{\"query\":{\"filtered\":{\"filter\":{\"query\":{\"query_string\":{\"query\":\"organization_id:[REDACTED]\nAND timestamp:(>=1418066241619 AND <=1418671041621) AND\nuser:\\\"root\\\"\"}}}}},\"from\":0,\"size\":20,\"timeout\":10000,\"sort\":{\"timestamp\":{\"order\":\"desc\"}},\"facets\":{}},\"organization_id\":\"[REDACTED]\",\"indices\":[\"logs.[REDACTED].20141209\",\"logs.[REDACTED].20141208\",\"logs.[REDACTED].20141215\",\"logs.[REDACTED].20141210\",\"logs.[REDACTED].20141213\",\"logs.[REDACTED].20141214\",\"logs.[REDACTED].20141211\",\"logs.[REDACTED].20141212\"],\"type\":\"all\",\"level\":\"info\",\"message\":\"\"}\n\n\n\n-- \nSent via pgsql-bugs mailing list (pgsql-bugs@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-bugs\n","msg_date":"Mon, 22 Dec 2014 20:52:48 +0000","msg_from":"postgres@bt-r.com","msg_from_op":true,"msg_subject":"BUG #12320: json parsing with embedded double quotes"},{"msg_contents":"Hi Aaron:\n....\n\n>\n> user:\\\"root\\\"\"}}}}},\"from\":0,\"size\":20,\"timeout\":10000,\"sort\":{\"timestamp\":{\"order\":\"\n\n.....\n\nCopy from does it's own round of backslash scaping, if you look at\n\nhttp://www.postgresql.org/docs/9.4/static/sql-copy.html\n\nThe third paragraph under File formats / text format reads:\n\nBackslash characters (\\) can be used in the COPY data to quote data\ncharacters that might otherwise be taken as row or column delimiters. In\nparticular, the following characters must be preceded by a backslash if\nthey appear as part of a column value: backslash itself, newline, carriage\nreturn, and the current delimiter character.\n\nAs you want the backslash as part of the data you need to double it in the\nfile, otherwise you'll fall into the later case:\n.....Any other backslashed character that is not mentioned in the above\ntable will be taken to represent itself\n\nAnd the backslash will get eaten and the json parser will fail. It's the\nsame problem that you get when you try to put windows paths or perl type\nregular expresions inside doube quoted C/Java/perl/shell/and_many_more\nStrings.\n\nIf your table is always a single column and you cannot edit the file, you\ncan always try to change the scape character.\n\nRegards.\n Francisco Olarte.\n\nHi Aaron:.... \nuser:\\\"root\\\"\"}}}}},\"from\":0,\"size\":20,\"timeout\":10000,\"sort\":{\"timestamp\":{\"order\":\".....Copy from does it's own round of backslash scaping, if you look at http://www.postgresql.org/docs/9.4/static/sql-copy.htmlThe third paragraph under File formats / text format reads:Backslash characters (\\) can be\n used in the COPY data to quote data\n characters that might otherwise be taken as row or column\n delimiters. In particular, the following characters\n must be preceded by a\n backslash if they appear as part of a column value: backslash\n itself, newline, carriage return, and the current delimiter\n character.As you want the backslash as part of the data you need to double it in the file, otherwise you'll fall into the later case:.....Any other backslashed character that is not mentioned in the\n above table will be taken to represent itselfAnd the backslash will get eaten and the json parser will fail. It's the same problem that you get when you try to put windows paths or perl type regular expresions inside doube quoted C/Java/perl/shell/and_many_more Strings.If your table is always a single column and you cannot edit the file, you can always try to change the scape character.Regards. Francisco Olarte.","msg_date":"Tue, 23 Dec 2014 19:17:05 +0100","msg_from":"Francisco Olarte <folarte@peoplecall.com>","msg_from_op":false,"msg_subject":"Re: BUG #12320: json parsing with embedded double quotes"},{"msg_contents":"Hi Francisco, I’m aware, but still consider this to be a bug, or at least a great opportunity for an enhancement. :) \n\nThis had bitten me for the third time while trying to import some json data. It’d be great to bypass the copy escaping (and possibly other meta characters) when the column type is json or jsonb. I’d be happy to try and write it and submit a patch if folks believe this is an acceptable way to go… That said, I should probably read what the process is for this kind of thing :)\n\nAaron\n\n\n> On Dec 23, 2014, at 1:17 PM, Francisco Olarte <folarte@peoplecall.com> wrote:\n> \n> Hi Aaron:\n> .... \n> user:\\\"root\\\"\"}}}}},\"from\":0,\"size\":20,\"timeout\":10000,\"sort\":{\"timestamp\":{\"order\":\"\n> .....\n> \n> Copy from does it's own round of backslash scaping, if you look at \n> \n> http://www.postgresql.org/docs/9.4/static/sql-copy.html <http://www.postgresql.org/docs/9.4/static/sql-copy.html>\n> \n> The third paragraph under File formats / text format reads:\n> Backslash characters (\\) can be used in the COPY data to quote data characters that might otherwise be taken as row or column delimiters. In particular, the following characters must be preceded by a backslash if they appear as part of a column value: backslash itself, newline, carriage return, and the current delimiter character.\n> \n> As you want the backslash as part of the data you need to double it in the file, otherwise you'll fall into the later case:\n> \n> .....Any other backslashed character that is not mentioned in the above table will be taken to represent itself\n> \n> And the backslash will get eaten and the json parser will fail. It's the same problem that you get when you try to put windows paths or perl type regular expresions inside doube quoted C/Java/perl/shell/and_many_more Strings.\n> \n> If your table is always a single column and you cannot edit the file, you can always try to change the scape character.\n> \n> Regards.\n> Francisco Olarte.\n> \n\n\nHi Francisco, I’m aware, but still consider this to be a bug, or at least a great opportunity for an enhancement. :) This had bitten me for the third time while trying to import some json data. It’d be great to bypass the copy escaping (and possibly other meta characters) when the column type is json or jsonb. I’d be happy to try and write it and submit a patch if folks believe this is an acceptable way to go… That said, I should probably read what the process is for this kind of thing :)AaronOn Dec 23, 2014, at 1:17 PM, Francisco Olarte <folarte@peoplecall.com> wrote:Hi Aaron:.... \nuser:\\\"root\\\"\"}}}}},\"from\":0,\"size\":20,\"timeout\":10000,\"sort\":{\"timestamp\":{\"order\":\".....Copy from does it's own round of backslash scaping, if you look at http://www.postgresql.org/docs/9.4/static/sql-copy.htmlThe third paragraph under File formats / text format reads:Backslash characters (\\) can be\n used in the COPY data to quote data\n characters that might otherwise be taken as row or column\n delimiters. In particular, the following characters\n must be preceded by a\n backslash if they appear as part of a column value: backslash\n itself, newline, carriage return, and the current delimiter\n character.As you want the backslash as part of the data you need to double it in the file, otherwise you'll fall into the later case:.....Any other backslashed character that is not mentioned in the\n above table will be taken to represent itselfAnd the backslash will get eaten and the json parser will fail. It's the same problem that you get when you try to put windows paths or perl type regular expresions inside doube quoted C/Java/perl/shell/and_many_more Strings.If your table is always a single column and you cannot edit the file, you can always try to change the scape character.Regards. Francisco Olarte.","msg_date":"Tue, 6 Jan 2015 13:06:38 -0500","msg_from":"Aaron Botsis <aaron@bt-r.com>","msg_from_op":false,"msg_subject":"Re: BUG #12320: json parsing with embedded double quotes"},{"msg_contents":"Hi Aaron:\n\nOn Tue, Jan 6, 2015 at 7:06 PM, Aaron Botsis <aaron@bt-r.com> wrote:\n\n> Hi Francisco, I’m aware, but still consider this to be a bug, or at least\n> a great opportunity for an enhancement. :)\n>\n\nMaybe, but you are going to have a problem.\n\n\n> This had bitten me for the third time while trying to import some json\n> data. It’d be great to bypass the copy escaping (and possibly other meta\n> characters) when the column type is json or jsonb. I’d be happy to try and\n> write it and submit a patch if folks believe this is an acceptable way to\n> go… That said, I should probably read what the process is for this kind of\n> thing :)\n>\n\nReading this, you are talking about 'the column being json'. COPY needs to\ndo the escaping at the same time it's constructing the columns. The present\nway is easy to do, read char by char, if it's a escape, process next char\nacumulating into current field, otherwise see whether it is a field or\nrecord separator and act accordingly. It's also layered, when you construct\nthe records for copy you get all the field data, turn them into escaped\nstrings, join them by the field separator and spit them out followed by a\nrecord separator ( practical implementations may do this virtually ).\nIntermixing this with the 'I'm in a json column' would need to pass\ninformation from the upper layer, and make it more difficult and, specially\nerror prone. What do you do if ( using the standard delimiters ) your json\nvalue has embeded newlines and tabs ( which, IIRC, are legal in several\nplaces inside the json ). And all this to make some incorrectly formatted\nfiles read ( which can be correctly formatted with a perl one liner or\nsomething similar ). I'm not the one to decide, but I will vote against\nincluding that ( but do not trust me too much, I would also vote against\nincluding 'csv' which I consider the root of many evils ).\n\nFrancisco Olarte.\n\nHi Aaron:On Tue, Jan 6, 2015 at 7:06 PM, Aaron Botsis <aaron@bt-r.com> wrote:Hi Francisco, I’m aware, but still consider this to be a bug, or at least a great opportunity for an enhancement. :) Maybe, but you are going to have a problem. This had bitten me for the third time while trying to import some json data. It’d be great to bypass the copy escaping (and possibly other meta characters) when the column type is json or jsonb. I’d be happy to try and write it and submit a patch if folks believe this is an acceptable way to go… That said, I should probably read what the process is for this kind of thing :)Reading this, you are talking about 'the column being json'. COPY needs to do the escaping at the same time it's constructing the columns. The present way is easy to do, read char by char, if it's a escape, process next char acumulating into current field, otherwise see whether it is a field or record separator and act accordingly. It's also layered, when you construct the records for copy you get all the field data, turn them into escaped strings, join them by the field separator and spit them out followed by a record separator ( practical implementations may do this virtually ). Intermixing this with the 'I'm in a json column' would need to pass information from the upper layer, and make it more difficult and, specially error prone. What do you do if ( using the standard delimiters ) your json value has embeded newlines and tabs ( which, IIRC, are legal in several places inside the json ). And all this to make some incorrectly formatted files read ( which can be correctly formatted with a perl one liner or something similar ). I'm not the one to decide, but I will vote against including that ( but do not trust me too much, I would also vote against including 'csv' which I consider the root of many evils ).Francisco Olarte.","msg_date":"Tue, 6 Jan 2015 19:52:28 +0100","msg_from":"Francisco Olarte <folarte@peoplecall.com>","msg_from_op":false,"msg_subject":"Re: BUG #12320: json parsing with embedded double quotes"},{"msg_contents":"Hi folks, I was having a problem importing json data with COPY. Lots of things export data nicely as one json blob per line. This is excellent for directly importing into a JSON/JSONB column for analysis. \n\n...Except when there’s an embedded doublequote. Or anything that’s escaped. COPY handles this, but by the time the escaped char hit the JSON parser, it's not escaped anymore. This breaks the JSON parsing. This means I need to manipulate the input data to double-escape it. See bug #12320 for an example. Yuck.\n\nI propose this small patch that simply allows specifying COPY … ESCAPE without requiring the CSV parser. It will make it much easier to directly use json formatted export data for folks going forward. This seemed like the simplest route.\n\nUsage is simply:\n\npostgres=# copy t1 from '/Users/nok/Desktop/queries.json';\nERROR: invalid input syntax for type json\nDETAIL: Token \"root\" is invalid.\nCONTEXT: JSON data, line 1: ...1418066241619 AND <=1418671041621) AND user:\"root...\nCOPY t1, line 3, column bleh: \"{\"timestamp\":\"2014-12-15T19:17:32.505Z\",\"duration\":7.947,\"query\":{\"query\":{\"filtered\":{\"filter\":{\"qu...\"\npostgres=# copy t1 from '/Users/nok/Desktop/queries.json' escape '';\nCOPY 1966\n\nI’ve included regression tests, and all existing tests pass. This is my first contribution, so be kind to me. :)\n\n\n\n> Begin forwarded message:\n> \n> From: Francisco Olarte <folarte@peoplecall.com>\n> Date: January 6, 2015 at 1:52:28 PM EST\n> Subject: Re: [BUGS] BUG #12320: json parsing with embedded double quotes\n> To: Aaron Botsis <aaron@bt-r.com>\n> Cc: postgres@bt-r.com, \"pgsql-bugs@postgresql.org\" <pgsql-bugs@postgresql.org>\n> \n> Hi Aaron:\n> \n> On Tue, Jan 6, 2015 at 7:06 PM, Aaron Botsis <aaron@bt-r.com <mailto:aaron@bt-r.com>> wrote:\n> Hi Francisco, I’m aware, but still consider this to be a bug, or at least a great opportunity for an enhancement. :) \n> \n> Maybe, but you are going to have a problem.\n> \n> This had bitten me for the third time while trying to import some json data. It’d be great to bypass the copy escaping (and possibly other meta characters) when the column type is json or jsonb. I’d be happy to try and write it and submit a patch if folks believe this is an acceptable way to go… That said, I should probably read what the process is for this kind of thing :)\n> \n> Reading this, you are talking about 'the column being json'. COPY needs to do the escaping at the same time it's constructing the columns. The present way is easy to do, read char by char, if it's a escape, process next char acumulating into current field, otherwise see whether it is a field or record separator and act accordingly. It's also layered, when you construct the records for copy you get all the field data, turn them into escaped strings, join them by the field separator and spit them out followed by a record separator ( practical implementations may do this virtually ). Intermixing this with the 'I'm in a json column' would need to pass information from the upper layer, and make it more difficult and, specially error prone. What do you do if ( using the standard delimiters ) your json value has embeded newlines and tabs ( which, IIRC, are legal in several places inside the json ). And all this to make some incorrectly formatted files read ( which can be correctly formatted with a perl one liner or something similar ). I'm not the one to decide, but I will vote against including that ( but do not trust me too much, I would also vote against including 'csv' which I consider the root of many evils ).\n> \n> Francisco Olarte.\n> \n>","msg_date":"Wed, 7 Jan 2015 08:25:34 -0500","msg_from":"Aaron Botsis <aaron@bt-r.com>","msg_from_op":false,"msg_subject":"Patch: [BUGS] BUG #12320: json parsing with embedded double quotes"},{"msg_contents":"Ack. Original had a bug. Let’s try again.\n\n\n\n\n> On Jan 7, 2015, at 8:25 AM, Aaron Botsis <aaron@bt-r.com> wrote:\n> \n> Hi folks, I was having a problem importing json data with COPY. Lots of things export data nicely as one json blob per line. This is excellent for directly importing into a JSON/JSONB column for analysis. \n> \n> ...Except when there’s an embedded doublequote. Or anything that’s escaped. COPY handles this, but by the time the escaped char hit the JSON parser, it's not escaped anymore. This breaks the JSON parsing. This means I need to manipulate the input data to double-escape it. See bug #12320 for an example. Yuck.\n> \n> I propose this small patch that simply allows specifying COPY … ESCAPE without requiring the CSV parser. It will make it much easier to directly use json formatted export data for folks going forward. This seemed like the simplest route.\n> \n> Usage is simply:\n> \n> postgres=# copy t1 from '/Users/nok/Desktop/queries.json';\n> ERROR: invalid input syntax for type json\n> DETAIL: Token \"root\" is invalid.\n> CONTEXT: JSON data, line 1: ...1418066241619 AND <=1418671041621) AND user:\"root...\n> COPY t1, line 3, column bleh: \"{\"timestamp\":\"2014-12-15T19:17:32.505Z\",\"duration\":7.947,\"query\":{\"query\":{\"filtered\":{\"filter\":{\"qu...\"\n> postgres=# copy t1 from '/Users/nok/Desktop/queries.json' escape '';\n> COPY 1966\n> \n> I’ve included regression tests, and all existing tests pass. This is my first contribution, so be kind to me. :)\n> \n> <escape-without-csv.patch>\n> \n> \n>> Begin forwarded message:\n>> \n>> From: Francisco Olarte <folarte@peoplecall.com <mailto:folarte@peoplecall.com>>\n>> Date: January 6, 2015 at 1:52:28 PM EST\n>> Subject: Re: [BUGS] BUG #12320: json parsing with embedded double quotes\n>> To: Aaron Botsis <aaron@bt-r.com <mailto:aaron@bt-r.com>>\n>> Cc: postgres@bt-r.com <mailto:postgres@bt-r.com>, \"pgsql-bugs@postgresql.org <mailto:pgsql-bugs@postgresql.org>\" <pgsql-bugs@postgresql.org <mailto:pgsql-bugs@postgresql.org>>\n>> \n>> Hi Aaron:\n>> \n>> On Tue, Jan 6, 2015 at 7:06 PM, Aaron Botsis <aaron@bt-r.com <mailto:aaron@bt-r.com>> wrote:\n>> Hi Francisco, I’m aware, but still consider this to be a bug, or at least a great opportunity for an enhancement. :) \n>> \n>> Maybe, but you are going to have a problem.\n>> \n>> This had bitten me for the third time while trying to import some json data. It’d be great to bypass the copy escaping (and possibly other meta characters) when the column type is json or jsonb. I’d be happy to try and write it and submit a patch if folks believe this is an acceptable way to go… That said, I should probably read what the process is for this kind of thing :)\n>> \n>> Reading this, you are talking about 'the column being json'. COPY needs to do the escaping at the same time it's constructing the columns. The present way is easy to do, read char by char, if it's a escape, process next char acumulating into current field, otherwise see whether it is a field or record separator and act accordingly. It's also layered, when you construct the records for copy you get all the field data, turn them into escaped strings, join them by the field separator and spit them out followed by a record separator ( practical implementations may do this virtually ). Intermixing this with the 'I'm in a json column' would need to pass information from the upper layer, and make it more difficult and, specially error prone. What do you do if ( using the standard delimiters ) your json value has embeded newlines and tabs ( which, IIRC, are legal in several places inside the json ). And all this to make some incorrectly formatted files read ( which can be correctly formatted with a perl one liner or something similar ). I'm not the one to decide, but I will vote against including that ( but do not trust me too much, I would also vote against including 'csv' which I consider the root of many evils ).\n>> \n>> Francisco Olarte.\n>> \n>> \n>","msg_date":"Wed, 7 Jan 2015 10:37:58 -0500","msg_from":"Aaron Botsis <aaron@bt-r.com>","msg_from_op":false,"msg_subject":"Re: Patch: [BUGS] BUG #12320: json parsing with embedded double\n quotes"},{"msg_contents":"\n\nOn 01/07/2015 08:25 AM, Aaron Botsis wrote:\n> Hi folks, I was having a problem importing json data with COPY. Lots \n> of things export data nicely as one json blob per line. This is \n> excellent for directly importing into a JSON/JSONB column for analysis.\n>\n> ...Except when there’s an embedded doublequote. Or anything that’s \n> escaped. COPY handles this, but by the time the escaped char hit the \n> JSON parser, it's not escaped anymore. This breaks the JSON parsing. \n> This means I need to manipulate the input data to double-escape it. \n> See bug #12320 for an example. Yuck.\n>\n> I propose this small patch that simply allows specifying COPY … ESCAPE \n> without requiring the CSV parser. It will make it much easier to \n> directly use json formatted export data for folks going forward. This \n> seemed like the simplest route.\n>\n> Usage is simply:\n>\n> postgres=# copy t1 from '/Users/nok/Desktop/queries.json';\n> ERROR: invalid input syntax for type json\n> DETAIL: Token \"root\" is invalid.\n> CONTEXT: JSON data, line 1: ...1418066241619 AND <=1418671041621) AND \n> user:\"root...\n> COPY t1, line 3, column bleh: \n> \"{\"timestamp\":\"2014-12-15T19:17:32.505Z\",\"duration\":7.947,\"query\":{\"query\":{\"filtered\":{\"filter\":{\"qu...\"\n> postgres=# copy t1 from '/Users/nok/Desktop/queries.json' escape '';\n> COPY 1966\n>\n>\n\n\nThis isn't a bug. Neither CSV format nor TEXT format are partucularly \nsuitable for json. I'm quite certain I could compose legal json that \nwill break your proposal (for example, with an embedded newline in the \nwhite space.)\n\nIt's also unnecessary. CSV format, while not designed for this, is \nnevertheless sufficiently flexible to allow successful import of json \ndata meeting certain criteria (essentially no newlines), like this:\n\n copy the_table(jsonfield)\n from '/path/to/jsondata'\n csv quote e'\\x01' delimiter e'\\x02';\n\n\nYou aren't the first person to encounter this problem. See \n<http://adpgtech.blogspot.com/2014/09/importing-json-data.html>\n\nMaybe we need to add something like this to the docs, or to the wiki.\n\nNote too my comment in that blog post:\n\n Now this solution is a bit of a hack. I wonder if there's a case for\n a COPY mode that simply treats each line as a single datum. I also\n wonder if we need some more specialized tools for importing JSON,\n possibly one or more Foreign Data Wrappers. Such things could\n handle, say, embedded newline punctuation.\n\n\ncheers\n\nandrew\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, 07 Jan 2015 14:45:51 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Patch: [BUGS] BUG #12320: json parsing with embedded double\n quotes"},{"msg_contents":"> On Jan 7, 2015, at 2:45 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\n> \n> \n> On 01/07/2015 08:25 AM, Aaron Botsis wrote:\n>> Hi folks, I was having a problem importing json data with COPY. Lots of things export data nicely as one json blob per line. This is excellent for directly importing into a JSON/JSONB column for analysis.\n>> \n>> ...Except when there’s an embedded doublequote. Or anything that’s escaped. COPY handles this, but by the time the escaped char hit the JSON parser, it's not escaped anymore. This breaks the JSON parsing. This means I need to manipulate the input data to double-escape it. See bug #12320 for an example. Yuck.\n>> \n>> I propose this small patch that simply allows specifying COPY … ESCAPE without requiring the CSV parser. It will make it much easier to directly use json formatted export data for folks going forward. This seemed like the simplest route.\n> \n> This isn't a bug. Neither CSV format nor TEXT format are partucularly suitable for json. I'm quite certain I could compose legal json that will break your proposal (for example, with an embedded newline in the white space.)\n\nSorry - though I originally reported it as a bug, I didn’t mean to imply it ultimately was one. :) The patch is a feature enhancement.\n\n> It's also unnecessary. CSV format, while not designed for this, is nevertheless sufficiently flexible to allow successful import of json data meeting certain criteria (essentially no newlines), like this:\n> \n> copy the_table(jsonfield)\n> from '/path/to/jsondata'\n> csv quote e'\\x01' delimiter e'\\x02’;\n\nWhile perhaps unnecessary, given the size and simplicity of the patch, IMO it’s a no brainer to merge (it actually makes the code smaller by 3 lines). It also enables non-json use cases anytime one might want to preserve embedded escapes, or use different ones entirely. Do you see other reasons not to commit it?\n\n> You aren't the first person to encounter this problem. See <http://adpgtech.blogspot.com/2014/09/importing-json-data.html <http://adpgtech.blogspot.com/2014/09/importing-json-data.html>>\n> \n> Maybe we need to add something like this to the docs, or to the wiki.\n\nIn your post you acknowledged a text mode copy with null escape character would have solved your problem, and to me, that was the intuitive/obvious choice as well. *shrug*\n\n> Note too my comment in that blog post:\n> \n> Now this solution is a bit of a hack. I wonder if there's a case for\n> a COPY mode that simply treats each line as a single datum. I also\n> wonder if we need some more specialized tools for importing JSON,\n> possibly one or more Foreign Data Wrappers. Such things could\n> handle, say, embedded newline punctuation.\n\nAgree. Though https://github.com/citusdata/json_fdw <https://github.com/citusdata/json_fdw> already exists (I haven’t used it). How do folks feel about a COPY mode that reads a single datum until it finds a single character record terminator alone on a line? something like:\n\n=# COPY test(data) from stdin terminator ‘}’;\nEnd with a backslash and a period on a line by itself.\n>>{\n>> \"a\": 123, \"c\": \"string\", \"b\": [1, 2, 3\n>> ]\n>>}\n>>{\n>>\t[1,2,3]\n>>}\n>>\\.\nCOPY 2\n\nYou could also get fancy and support multi-character record terminators which would allow the same thing in a slightly different way:\nCOPY test(data) from stdin terminator ‘\\n}’;\n\nCOPY test(data) from stdin terminator ‘\\n}’;\n\nI don’t know how crazy you could get without a lot of rework. It might have to be used in conjunction with a more constrained mode like \"COPY…RAW”.\n\nAaron\nOn Jan 7, 2015, at 2:45 PM, Andrew Dunstan <andrew@dunslane.net> wrote:On 01/07/2015 08:25 AM, Aaron Botsis wrote:Hi folks, I was having a problem importing json data with COPY. Lots of things export data nicely as one json blob per line. This is excellent for directly importing into a JSON/JSONB column for analysis....Except when there’s an embedded doublequote. Or anything that’s escaped. COPY handles this, but by the time the escaped char hit the JSON parser, it's not escaped anymore. This breaks the JSON parsing. This means I need to manipulate the input data to double-escape it. See bug #12320 for an example. Yuck.I propose this small patch that simply allows specifying COPY … ESCAPE without requiring the CSV parser. It will make it much easier to directly use json formatted export data for folks going forward. This seemed like the simplest route.This isn't a bug. Neither CSV format nor TEXT format are partucularly suitable for json. I'm quite certain I could compose legal json that will break your proposal (for example, with an embedded newline in the white space.)Sorry - though I originally reported it as a bug, I didn’t mean to imply it ultimately was one. :) The patch is a feature enhancement.It's also unnecessary. CSV format, while not designed for this, is nevertheless sufficiently flexible to allow successful import of json data meeting certain criteria (essentially no newlines), like this: copy the_table(jsonfield) from '/path/to/jsondata' csv quote e'\\x01' delimiter e'\\x02’;While perhaps unnecessary, given the size and simplicity of the patch, IMO it’s a no brainer to merge (it actually makes the code smaller by 3 lines). It also enables non-json use cases anytime one might want to preserve embedded escapes, or use different ones entirely. Do you see other reasons not to commit it?You aren't the first person to encounter this problem. See <http://adpgtech.blogspot.com/2014/09/importing-json-data.html>Maybe we need to add something like this to the docs, or to the wiki.In your post you acknowledged a text mode copy with null escape character would have solved your problem, and to me, that was the intuitive/obvious choice as well. *shrug*Note too my comment in that blog post: Now this solution is a bit of a hack. I wonder if there's a case for a COPY mode that simply treats each line as a single datum. I also wonder if we need some more specialized tools for importing JSON, possibly one or more Foreign Data Wrappers. Such things could handle, say, embedded newline punctuation.Agree. Though https://github.com/citusdata/json_fdw already exists (I haven’t used it). How do folks feel about a COPY mode that reads a single datum until it finds a single character record terminator alone on a line? something like:=# COPY test(data) from stdin terminator ‘}’;End with a backslash and a period on a line by itself.>>{>> \"a\": 123, \"c\": \"string\", \"b\": [1, 2, 3>> ]>>}>>{>> [1,2,3]>>}>>\\.COPY 2You could also get fancy and support multi-character record terminators which would allow the same thing in a slightly different way:COPY test(data) from stdin terminator ‘\\n}’;COPY test(data) from stdin terminator ‘\\n}’;I don’t know how crazy you could get without a lot of rework. It might have to be used in conjunction with a more constrained mode like \"COPY…RAW”.Aaron","msg_date":"Thu, 8 Jan 2015 15:05:10 -0500","msg_from":"Aaron Botsis <aaron@bt-r.com>","msg_from_op":false,"msg_subject":"Re: Patch: [BUGS] BUG #12320: json parsing with embedded double\n quotes"},{"msg_contents":"\nOn 01/08/2015 03:05 PM, Aaron Botsis wrote:\n>\n>\n>> It's also unnecessary. CSV format, while not designed for this, is \n>> nevertheless sufficiently flexible to allow successful import of json \n>> data meeting certain criteria (essentially no newlines), like this:\n>>\n>> copy the_table(jsonfield)\n>> from '/path/to/jsondata'\n>> csv quote e'\\x01' delimiter e'\\x02’;\n>\n> While perhaps unnecessary, given the size and simplicity of the patch, \n> IMO it’s a no brainer to merge (it actually makes the code smaller by \n> 3 lines). It also enables non-json use cases anytime one might want to \n> preserve embedded escapes, or use different ones entirely. Do you see \n> other reasons not to commit it?\n\n\nWell, for one thing it's seriously incomplete. You need to be able to \nchange the delimiter as well. Otherwise, any embedded tab in the json \nwill cause you major grief.\n\nCurrently the delimiter and the escape MUST be a single byte non-nul \ncharacter, and there is a check for this in csv mode. Your patch would \nallow any arbitrary string (including one of zero length) for the escape \nin text mode, and would then silently ignore all but the first byte. \nThat's not the way we like to do things.\n\nAnd, frankly, I would need to spend quite a lot more time thinking about \nother implications than I have given it so far. This is an area where I \ntend to be VERY cautious about making changes. This is a fairly fragile \necosystem.\n\ncheers\n\nandrew\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 08 Jan 2015 15:44:44 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Patch: [BUGS] BUG #12320: json parsing with embedded double\n quotes"},{"msg_contents":"> On Jan 8, 2015, at 3:44 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\n> \n> \n> On 01/08/2015 03:05 PM, Aaron Botsis wrote:\n>> \n>> \n>>> It's also unnecessary. CSV format, while not designed for this, is nevertheless sufficiently flexible to allow successful import of json data meeting certain criteria (essentially no newlines), like this:\n>>> \n>>> copy the_table(jsonfield)\n>>> from '/path/to/jsondata'\n>>> csv quote e'\\x01' delimiter e'\\x02’;\n>> \n>> While perhaps unnecessary, given the size and simplicity of the patch, IMO it’s a no brainer to merge (it actually makes the code smaller by 3 lines). It also enables non-json use cases anytime one might want to preserve embedded escapes, or use different ones entirely. Do you see other reasons not to commit it?\n> \n> \n> Well, for one thing it's seriously incomplete. You need to be able to change the delimiter as well. Otherwise, any embedded tab in the json will cause you major grief.\n> \n> Currently the delimiter and the escape MUST be a single byte non-nul character, and there is a check for this in csv mode. Your patch would allow any arbitrary string (including one of zero length) for the escape in text mode, and would then silently ignore all but the first byte. That's not the way we like to do things.\n> \n> And, frankly, I would need to spend quite a lot more time thinking about other implications than I have given it so far. This is an area where I tend to be VERY cautious about making changes. This is a fairly fragile ecosystem.\n\nGood point.\n\nThis version:\n\n* doesn't allow ENCODING in BINARY mode (existing bug)\n* doesn’t allow ESCAPE in BINARY mode\n* makes COPY TO work with escape\n* ensures escape char length is < 2 for text mode, 1 for CSV\n\nCouple more things to realize: setting both the escape and delimiter characters to null won’t be any different than how you fiddled with them in CSV mode. The code paths will be the same because we should never encounter a null in the middle of the string. And even if we did (and the encoding didn’t catch it), we’d treat it just like any other field delimiter or escape character.\n\nSo I’m going to do a bit more testing with another patch tomorrow with delimiters removed. If you can think of any specific cases you think will break it let me know and I’ll make sure to add regression tests for them as well. \n\nCheers!\n\nAaron\n\n\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":"Thu, 8 Jan 2015 20:42:00 -0500","msg_from":"Aaron Botsis <aaron@bt-r.com>","msg_from_op":false,"msg_subject":"Re: Patch: [BUGS] BUG #12320: json parsing with embedded double\n quotes"},{"msg_contents":"\nOn 01/08/2015 08:42 PM, Aaron Botsis wrote:\n>\n>> On Jan 8, 2015, at 3:44 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\n>>\n>>\n>> On 01/08/2015 03:05 PM, Aaron Botsis wrote:\n>>>\n>>>> It's also unnecessary. CSV format, while not designed for this, is nevertheless sufficiently flexible to allow successful import of json data meeting certain criteria (essentially no newlines), like this:\n>>>>\n>>>> copy the_table(jsonfield)\n>>>> from '/path/to/jsondata'\n>>>> csv quote e'\\x01' delimiter e'\\x02’;\n>>> While perhaps unnecessary, given the size and simplicity of the patch, IMO it’s a no brainer to merge (it actually makes the code smaller by 3 lines). It also enables non-json use cases anytime one might want to preserve embedded escapes, or use different ones entirely. Do you see other reasons not to commit it?\n>>\n>> Well, for one thing it's seriously incomplete. You need to be able to change the delimiter as well. Otherwise, any embedded tab in the json will cause you major grief.\n>>\n>> Currently the delimiter and the escape MUST be a single byte non-nul character, and there is a check for this in csv mode. Your patch would allow any arbitrary string (including one of zero length) for the escape in text mode, and would then silently ignore all but the first byte. That's not the way we like to do things.\n>>\n>> And, frankly, I would need to spend quite a lot more time thinking about other implications than I have given it so far. This is an area where I tend to be VERY cautious about making changes. This is a fairly fragile ecosystem.\n> Good point.\n>\n> This version:\n>\n> * doesn't allow ENCODING in BINARY mode (existing bug)\n> * doesn’t allow ESCAPE in BINARY mode\n> * makes COPY TO work with escape\n> * ensures escape char length is < 2 for text mode, 1 for CSV\n>\n> Couple more things to realize: setting both the escape and delimiter characters to null won’t be any different than how you fiddled with them in CSV mode. The code paths will be the same because we should never encounter a null in the middle of the string. And even if we did (and the encoding didn’t catch it), we’d treat it just like any other field delimiter or escape character.\n>\n> So I’m going to do a bit more testing with another patch tomorrow with delimiters removed. If you can think of any specific cases you think will break it let me know and I’ll make sure to add regression tests for them as well.\n>\n>\n\nWell, I still need convincing that this is the best solution to the \nproblem. As I said, I need to spend more time thinking about it.\n\ncheers\n\nandrew\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 09 Jan 2015 11:37:19 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Patch: [BUGS] BUG #12320: json parsing with embedded double\n quotes"},{"msg_contents":"> On Jan 9, 2015, at 11:37 AM, Andrew Dunstan <andrew@dunslane.net> wrote:\n> \n> \n> On 01/08/2015 08:42 PM, Aaron Botsis wrote:\n>> \n>>> On Jan 8, 2015, at 3:44 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\n>>> \n>>> \n>>> On 01/08/2015 03:05 PM, Aaron Botsis wrote:\n>>>> \n>>>>> It's also unnecessary. CSV format, while not designed for this, is nevertheless sufficiently flexible to allow successful import of json data meeting certain criteria (essentially no newlines), like this:\n>>>>> \n>>>>> copy the_table(jsonfield)\n>>>>> from '/path/to/jsondata'\n>>>>> csv quote e'\\x01' delimiter e'\\x02’;\n>>>> While perhaps unnecessary, given the size and simplicity of the patch, IMO it’s a no brainer to merge (it actually makes the code smaller by 3 lines). It also enables non-json use cases anytime one might want to preserve embedded escapes, or use different ones entirely. Do you see other reasons not to commit it?\n>>> \n>>> Well, for one thing it's seriously incomplete. You need to be able to change the delimiter as well. Otherwise, any embedded tab in the json will cause you major grief.\n>>> \n>>> Currently the delimiter and the escape MUST be a single byte non-nul character, and there is a check for this in csv mode. Your patch would allow any arbitrary string (including one of zero length) for the escape in text mode, and would then silently ignore all but the first byte. That's not the way we like to do things.\n>>> \n>>> And, frankly, I would need to spend quite a lot more time thinking about other implications than I have given it so far. This is an area where I tend to be VERY cautious about making changes. This is a fairly fragile ecosystem.\n>> \n>> So I’m going to do a bit more testing with another patch tomorrow with delimiters removed. If you can think of any specific cases you think will break it let me know and I’ll make sure to add regression tests for them as well.\n>> \n> Well, I still need convincing that this is the best solution to the problem. As I said, I need to spend more time thinking about it.\n\n\nI offered an alternative (RAW mode w/record terminator) that you ignored. So in lieu of a productive discussion about “the best solution to the problem”, I went ahead and continued to complete the patch since I believe it’s a useful FE that it could be helpful for this and other use cases. There have been multiple times (not even w/json) I wished COPY would stop being so smart. \n\nFWIW, (if anyone’s interested in it) I also hacked up some python that’ll read a json file, and outputs a binary file suitable for use with COPY BINARY that gets around all this stuff. Obviously this only works for json (not jsonb) columns (though you could SELECT INTO a jsonb column). Happy to pass it along.\n\nAaron\n\n\nOn Jan 9, 2015, at 11:37 AM, Andrew Dunstan <andrew@dunslane.net> wrote:On 01/08/2015 08:42 PM, Aaron Botsis wrote:On Jan 8, 2015, at 3:44 PM, Andrew Dunstan <andrew@dunslane.net> wrote:On 01/08/2015 03:05 PM, Aaron Botsis wrote:It's also unnecessary. CSV format, while not designed for this, is nevertheless sufficiently flexible to allow successful import of json data meeting certain criteria (essentially no newlines), like this: copy the_table(jsonfield) from '/path/to/jsondata' csv quote e'\\x01' delimiter e'\\x02’;While perhaps unnecessary, given the size and simplicity of the patch, IMO it’s a no brainer to merge (it actually makes the code smaller by 3 lines). It also enables non-json use cases anytime one might want to preserve embedded escapes, or use different ones entirely. Do you see other reasons not to commit it?Well, for one thing it's seriously incomplete. You need to be able to change the delimiter as well. Otherwise, any embedded tab in the json will cause you major grief.Currently the delimiter and the escape MUST be a single byte non-nul character, and there is a check for this in csv mode. Your patch would allow any arbitrary string (including one of zero length) for the escape in text mode, and would then silently ignore all but the first byte. That's not the way we like to do things.And, frankly, I would need to spend quite a lot more time thinking about other implications than I have given it so far. This is an area where I tend to be VERY cautious about making changes. This is a fairly fragile ecosystem.So I’m going to do a bit more testing with another patch tomorrow with delimiters removed. If you can think of any specific cases you think will break it let me know and I’ll make sure to add regression tests for them as well.Well, I still need convincing that this is the best solution to the problem. As I said, I need to spend more time thinking about it.I offered an alternative (RAW mode w/record terminator) that you ignored. So in lieu of a productive discussion about “the best solution to the problem”, I went ahead and continued to complete the patch since I believe it’s a useful FE that it could be helpful for this and other use cases. There have been multiple times (not even w/json) I wished COPY would stop being so smart. FWIW, (if anyone’s interested in it) I also hacked up some python that’ll read a json file, and outputs a binary file suitable for use with COPY BINARY that gets around all this stuff. Obviously this only works for json (not jsonb) columns (though you could SELECT INTO a jsonb column). Happy to pass it along.Aaron","msg_date":"Fri, 9 Jan 2015 12:57:31 -0500","msg_from":"Aaron Botsis <aaron@bt-r.com>","msg_from_op":false,"msg_subject":"Re: Patch: [BUGS] BUG #12320: json parsing with embedded double\n quotes"},{"msg_contents":"On Thu, Jan 8, 2015 at 8:42 PM, Aaron Botsis <aaron@bt-r.com> wrote:\n> This version:\n>\n> * doesn't allow ENCODING in BINARY mode (existing bug)\n> * doesn’t allow ESCAPE in BINARY mode\n\nHmm, it does look like ENCODING and ESCAPE don't do anything in BINARY\nmode, so adding error checks for that makes sense to me. I wondered\nwhether the ENCODING option would affect the interpretation of text\ndata that might be stored in binary-format COPY data, but I can't find\nany evidence that it does.\n\n> * makes COPY TO work with escape\n\nI don't see anything horribly wrong with this, but it seems pretty\nuseless. I mean, does anyone really want a newline encoded in the\noutput as %n or !n or qn rather than \\n? What problem are you trying\nto solve here?\n\n> * ensures escape char length is < 2 for text mode, 1 for CSV\n\nSo, is the idea here that we're going to have no escape character at\nall? So, the characters that would require the use of an escape\ncharacter to represent simply can't be represented at all, but we're\nOK with that, because the input data doesn't contain any such\ncharacters?\n\nOne general problem with this patch is that it doesn't update the\ndocumentation, which makes it a bit harder than it might be to see\nwhat you're trying to accomplish.\n\nAlso, if you want to decrease the chances that this patch will get\nforgotten about, you can register it here:\n\nhttps://commitfest.postgresql.org/action/commitfest_view/open\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 14 Jan 2015 09:05:52 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Patch: [BUGS] BUG #12320: json parsing with embedded double\n quotes"}],"string":"[\n {\n \"msg_contents\": \"The following bug has been logged on the website:\\n\\nBug reference: 12320\\nLogged by: Aaron Botsis\\nEmail address: postgres@bt-r.com\\nPostgreSQL version: 9.4.0\\nOperating system: Mac OSX\\nDescription: \\n\\npostgres=# create table ct (data jsonb);\\nCREATE TABLE\\npostgres=# copy ct (data) from '/Users/aaron/Desktop/queries.json';\\nERROR: invalid input syntax for type json\\nDETAIL: Token \\\"root\\\" is invalid.\\nCONTEXT: JSON data, line 1: ...1418066241619 AND <=1418671041621) AND\\nuser:\\\"root...\\nCOPY ct, line 3, column data:\\n\\\"{\\\"timestamp\\\":\\\"2014-12-15T19:17:32.505Z\\\",\\\"duration\\\":7.947,\\\"query\\\":{\\\"query\\\":{\\\"filtered\\\":{\\\"filter\\\":{\\\"qu...\\\"\\n\\n\\nThis is a sample of the JSON:\\n\\n{\\\"timestamp\\\":\\\"2014-12-15T19:17:32.505Z\\\",\\\"duration\\\":7.947,\\\"query\\\":{\\\"query\\\":{\\\"filtered\\\":{\\\"filter\\\":{\\\"query\\\":{\\\"query_string\\\":{\\\"query\\\":\\\"organization_id:[REDACTED]\\nAND timestamp:(>=1418066241619 AND <=1418671041621) AND\\nuser:\\\\\\\"root\\\\\\\"\\\"}}}}},\\\"from\\\":0,\\\"size\\\":20,\\\"timeout\\\":10000,\\\"sort\\\":{\\\"timestamp\\\":{\\\"order\\\":\\\"desc\\\"}},\\\"facets\\\":{}},\\\"organization_id\\\":\\\"[REDACTED]\\\",\\\"indices\\\":[\\\"logs.[REDACTED].20141209\\\",\\\"logs.[REDACTED].20141208\\\",\\\"logs.[REDACTED].20141215\\\",\\\"logs.[REDACTED].20141210\\\",\\\"logs.[REDACTED].20141213\\\",\\\"logs.[REDACTED].20141214\\\",\\\"logs.[REDACTED].20141211\\\",\\\"logs.[REDACTED].20141212\\\"],\\\"type\\\":\\\"all\\\",\\\"level\\\":\\\"info\\\",\\\"message\\\":\\\"\\\"}\\n\\n\\n\\n-- \\nSent via pgsql-bugs mailing list (pgsql-bugs@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-bugs\\n\",\n \"msg_date\": \"Mon, 22 Dec 2014 20:52:48 +0000\",\n \"msg_from\": \"postgres@bt-r.com\",\n \"msg_from_op\": true,\n \"msg_subject\": \"BUG #12320: json parsing with embedded double quotes\"\n },\n {\n \"msg_contents\": \"Hi Aaron:\\n....\\n\\n>\\n> user:\\\\\\\"root\\\\\\\"\\\"}}}}},\\\"from\\\":0,\\\"size\\\":20,\\\"timeout\\\":10000,\\\"sort\\\":{\\\"timestamp\\\":{\\\"order\\\":\\\"\\n\\n.....\\n\\nCopy from does it's own round of backslash scaping, if you look at\\n\\nhttp://www.postgresql.org/docs/9.4/static/sql-copy.html\\n\\nThe third paragraph under File formats / text format reads:\\n\\nBackslash characters (\\\\) can be used in the COPY data to quote data\\ncharacters that might otherwise be taken as row or column delimiters. In\\nparticular, the following characters must be preceded by a backslash if\\nthey appear as part of a column value: backslash itself, newline, carriage\\nreturn, and the current delimiter character.\\n\\nAs you want the backslash as part of the data you need to double it in the\\nfile, otherwise you'll fall into the later case:\\n.....Any other backslashed character that is not mentioned in the above\\ntable will be taken to represent itself\\n\\nAnd the backslash will get eaten and the json parser will fail. It's the\\nsame problem that you get when you try to put windows paths or perl type\\nregular expresions inside doube quoted C/Java/perl/shell/and_many_more\\nStrings.\\n\\nIf your table is always a single column and you cannot edit the file, you\\ncan always try to change the scape character.\\n\\nRegards.\\n Francisco Olarte.\\n\\nHi Aaron:.... \\nuser:\\\\\\\"root\\\\\\\"\\\"}}}}},\\\"from\\\":0,\\\"size\\\":20,\\\"timeout\\\":10000,\\\"sort\\\":{\\\"timestamp\\\":{\\\"order\\\":\\\".....Copy from does it's own round of backslash scaping, if you look at http://www.postgresql.org/docs/9.4/static/sql-copy.htmlThe third paragraph under File formats / text format reads:Backslash characters (\\\\) can be\\n used in the COPY data to quote data\\n characters that might otherwise be taken as row or column\\n delimiters. In particular, the following characters\\n must be preceded by a\\n backslash if they appear as part of a column value: backslash\\n itself, newline, carriage return, and the current delimiter\\n character.As you want the backslash as part of the data you need to double it in the file, otherwise you'll fall into the later case:.....Any other backslashed character that is not mentioned in the\\n above table will be taken to represent itselfAnd the backslash will get eaten and the json parser will fail. It's the same problem that you get when you try to put windows paths or perl type regular expresions inside doube quoted C/Java/perl/shell/and_many_more Strings.If your table is always a single column and you cannot edit the file, you can always try to change the scape character.Regards. Francisco Olarte.\",\n \"msg_date\": \"Tue, 23 Dec 2014 19:17:05 +0100\",\n \"msg_from\": \"Francisco Olarte <folarte@peoplecall.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: BUG #12320: json parsing with embedded double quotes\"\n },\n {\n \"msg_contents\": \"Hi Francisco, I’m aware, but still consider this to be a bug, or at least a great opportunity for an enhancement. :) \\n\\nThis had bitten me for the third time while trying to import some json data. It’d be great to bypass the copy escaping (and possibly other meta characters) when the column type is json or jsonb. I’d be happy to try and write it and submit a patch if folks believe this is an acceptable way to go… That said, I should probably read what the process is for this kind of thing :)\\n\\nAaron\\n\\n\\n> On Dec 23, 2014, at 1:17 PM, Francisco Olarte <folarte@peoplecall.com> wrote:\\n> \\n> Hi Aaron:\\n> .... \\n> user:\\\\\\\"root\\\\\\\"\\\"}}}}},\\\"from\\\":0,\\\"size\\\":20,\\\"timeout\\\":10000,\\\"sort\\\":{\\\"timestamp\\\":{\\\"order\\\":\\\"\\n> .....\\n> \\n> Copy from does it's own round of backslash scaping, if you look at \\n> \\n> http://www.postgresql.org/docs/9.4/static/sql-copy.html <http://www.postgresql.org/docs/9.4/static/sql-copy.html>\\n> \\n> The third paragraph under File formats / text format reads:\\n> Backslash characters (\\\\) can be used in the COPY data to quote data characters that might otherwise be taken as row or column delimiters. In particular, the following characters must be preceded by a backslash if they appear as part of a column value: backslash itself, newline, carriage return, and the current delimiter character.\\n> \\n> As you want the backslash as part of the data you need to double it in the file, otherwise you'll fall into the later case:\\n> \\n> .....Any other backslashed character that is not mentioned in the above table will be taken to represent itself\\n> \\n> And the backslash will get eaten and the json parser will fail. It's the same problem that you get when you try to put windows paths or perl type regular expresions inside doube quoted C/Java/perl/shell/and_many_more Strings.\\n> \\n> If your table is always a single column and you cannot edit the file, you can always try to change the scape character.\\n> \\n> Regards.\\n> Francisco Olarte.\\n> \\n\\n\\nHi Francisco, I’m aware, but still consider this to be a bug, or at least a great opportunity for an enhancement. :) This had bitten me for the third time while trying to import some json data. It’d be great to bypass the copy escaping (and possibly other meta characters) when the column type is json or jsonb. I’d be happy to try and write it and submit a patch if folks believe this is an acceptable way to go… That said, I should probably read what the process is for this kind of thing :)AaronOn Dec 23, 2014, at 1:17 PM, Francisco Olarte <folarte@peoplecall.com> wrote:Hi Aaron:.... \\nuser:\\\\\\\"root\\\\\\\"\\\"}}}}},\\\"from\\\":0,\\\"size\\\":20,\\\"timeout\\\":10000,\\\"sort\\\":{\\\"timestamp\\\":{\\\"order\\\":\\\".....Copy from does it's own round of backslash scaping, if you look at http://www.postgresql.org/docs/9.4/static/sql-copy.htmlThe third paragraph under File formats / text format reads:Backslash characters (\\\\) can be\\n used in the COPY data to quote data\\n characters that might otherwise be taken as row or column\\n delimiters. In particular, the following characters\\n must be preceded by a\\n backslash if they appear as part of a column value: backslash\\n itself, newline, carriage return, and the current delimiter\\n character.As you want the backslash as part of the data you need to double it in the file, otherwise you'll fall into the later case:.....Any other backslashed character that is not mentioned in the\\n above table will be taken to represent itselfAnd the backslash will get eaten and the json parser will fail. It's the same problem that you get when you try to put windows paths or perl type regular expresions inside doube quoted C/Java/perl/shell/and_many_more Strings.If your table is always a single column and you cannot edit the file, you can always try to change the scape character.Regards. Francisco Olarte.\",\n \"msg_date\": \"Tue, 6 Jan 2015 13:06:38 -0500\",\n \"msg_from\": \"Aaron Botsis <aaron@bt-r.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: BUG #12320: json parsing with embedded double quotes\"\n },\n {\n \"msg_contents\": \"Hi Aaron:\\n\\nOn Tue, Jan 6, 2015 at 7:06 PM, Aaron Botsis <aaron@bt-r.com> wrote:\\n\\n> Hi Francisco, I’m aware, but still consider this to be a bug, or at least\\n> a great opportunity for an enhancement. :)\\n>\\n\\nMaybe, but you are going to have a problem.\\n\\n\\n> This had bitten me for the third time while trying to import some json\\n> data. It’d be great to bypass the copy escaping (and possibly other meta\\n> characters) when the column type is json or jsonb. I’d be happy to try and\\n> write it and submit a patch if folks believe this is an acceptable way to\\n> go… That said, I should probably read what the process is for this kind of\\n> thing :)\\n>\\n\\nReading this, you are talking about 'the column being json'. COPY needs to\\ndo the escaping at the same time it's constructing the columns. The present\\nway is easy to do, read char by char, if it's a escape, process next char\\nacumulating into current field, otherwise see whether it is a field or\\nrecord separator and act accordingly. It's also layered, when you construct\\nthe records for copy you get all the field data, turn them into escaped\\nstrings, join them by the field separator and spit them out followed by a\\nrecord separator ( practical implementations may do this virtually ).\\nIntermixing this with the 'I'm in a json column' would need to pass\\ninformation from the upper layer, and make it more difficult and, specially\\nerror prone. What do you do if ( using the standard delimiters ) your json\\nvalue has embeded newlines and tabs ( which, IIRC, are legal in several\\nplaces inside the json ). And all this to make some incorrectly formatted\\nfiles read ( which can be correctly formatted with a perl one liner or\\nsomething similar ). I'm not the one to decide, but I will vote against\\nincluding that ( but do not trust me too much, I would also vote against\\nincluding 'csv' which I consider the root of many evils ).\\n\\nFrancisco Olarte.\\n\\nHi Aaron:On Tue, Jan 6, 2015 at 7:06 PM, Aaron Botsis <aaron@bt-r.com> wrote:Hi Francisco, I’m aware, but still consider this to be a bug, or at least a great opportunity for an enhancement. :) Maybe, but you are going to have a problem. This had bitten me for the third time while trying to import some json data. It’d be great to bypass the copy escaping (and possibly other meta characters) when the column type is json or jsonb. I’d be happy to try and write it and submit a patch if folks believe this is an acceptable way to go… That said, I should probably read what the process is for this kind of thing :)Reading this, you are talking about 'the column being json'. COPY needs to do the escaping at the same time it's constructing the columns. The present way is easy to do, read char by char, if it's a escape, process next char acumulating into current field, otherwise see whether it is a field or record separator and act accordingly. It's also layered, when you construct the records for copy you get all the field data, turn them into escaped strings, join them by the field separator and spit them out followed by a record separator ( practical implementations may do this virtually ). Intermixing this with the 'I'm in a json column' would need to pass information from the upper layer, and make it more difficult and, specially error prone. What do you do if ( using the standard delimiters ) your json value has embeded newlines and tabs ( which, IIRC, are legal in several places inside the json ). And all this to make some incorrectly formatted files read ( which can be correctly formatted with a perl one liner or something similar ). I'm not the one to decide, but I will vote against including that ( but do not trust me too much, I would also vote against including 'csv' which I consider the root of many evils ).Francisco Olarte.\",\n \"msg_date\": \"Tue, 6 Jan 2015 19:52:28 +0100\",\n \"msg_from\": \"Francisco Olarte <folarte@peoplecall.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: BUG #12320: json parsing with embedded double quotes\"\n },\n {\n \"msg_contents\": \"Hi folks, I was having a problem importing json data with COPY. Lots of things export data nicely as one json blob per line. This is excellent for directly importing into a JSON/JSONB column for analysis. \\n\\n...Except when there’s an embedded doublequote. Or anything that’s escaped. COPY handles this, but by the time the escaped char hit the JSON parser, it's not escaped anymore. This breaks the JSON parsing. This means I need to manipulate the input data to double-escape it. See bug #12320 for an example. Yuck.\\n\\nI propose this small patch that simply allows specifying COPY … ESCAPE without requiring the CSV parser. It will make it much easier to directly use json formatted export data for folks going forward. This seemed like the simplest route.\\n\\nUsage is simply:\\n\\npostgres=# copy t1 from '/Users/nok/Desktop/queries.json';\\nERROR: invalid input syntax for type json\\nDETAIL: Token \\\"root\\\" is invalid.\\nCONTEXT: JSON data, line 1: ...1418066241619 AND <=1418671041621) AND user:\\\"root...\\nCOPY t1, line 3, column bleh: \\\"{\\\"timestamp\\\":\\\"2014-12-15T19:17:32.505Z\\\",\\\"duration\\\":7.947,\\\"query\\\":{\\\"query\\\":{\\\"filtered\\\":{\\\"filter\\\":{\\\"qu...\\\"\\npostgres=# copy t1 from '/Users/nok/Desktop/queries.json' escape '';\\nCOPY 1966\\n\\nI’ve included regression tests, and all existing tests pass. This is my first contribution, so be kind to me. :)\\n\\n\\n\\n> Begin forwarded message:\\n> \\n> From: Francisco Olarte <folarte@peoplecall.com>\\n> Date: January 6, 2015 at 1:52:28 PM EST\\n> Subject: Re: [BUGS] BUG #12320: json parsing with embedded double quotes\\n> To: Aaron Botsis <aaron@bt-r.com>\\n> Cc: postgres@bt-r.com, \\\"pgsql-bugs@postgresql.org\\\" <pgsql-bugs@postgresql.org>\\n> \\n> Hi Aaron:\\n> \\n> On Tue, Jan 6, 2015 at 7:06 PM, Aaron Botsis <aaron@bt-r.com <mailto:aaron@bt-r.com>> wrote:\\n> Hi Francisco, I’m aware, but still consider this to be a bug, or at least a great opportunity for an enhancement. :) \\n> \\n> Maybe, but you are going to have a problem.\\n> \\n> This had bitten me for the third time while trying to import some json data. It’d be great to bypass the copy escaping (and possibly other meta characters) when the column type is json or jsonb. I’d be happy to try and write it and submit a patch if folks believe this is an acceptable way to go… That said, I should probably read what the process is for this kind of thing :)\\n> \\n> Reading this, you are talking about 'the column being json'. COPY needs to do the escaping at the same time it's constructing the columns. The present way is easy to do, read char by char, if it's a escape, process next char acumulating into current field, otherwise see whether it is a field or record separator and act accordingly. It's also layered, when you construct the records for copy you get all the field data, turn them into escaped strings, join them by the field separator and spit them out followed by a record separator ( practical implementations may do this virtually ). Intermixing this with the 'I'm in a json column' would need to pass information from the upper layer, and make it more difficult and, specially error prone. What do you do if ( using the standard delimiters ) your json value has embeded newlines and tabs ( which, IIRC, are legal in several places inside the json ). And all this to make some incorrectly formatted files read ( which can be correctly formatted with a perl one liner or something similar ). I'm not the one to decide, but I will vote against including that ( but do not trust me too much, I would also vote against including 'csv' which I consider the root of many evils ).\\n> \\n> Francisco Olarte.\\n> \\n>\",\n \"msg_date\": \"Wed, 7 Jan 2015 08:25:34 -0500\",\n \"msg_from\": \"Aaron Botsis <aaron@bt-r.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Patch: [BUGS] BUG #12320: json parsing with embedded double quotes\"\n },\n {\n \"msg_contents\": \"Ack. Original had a bug. Let’s try again.\\n\\n\\n\\n\\n> On Jan 7, 2015, at 8:25 AM, Aaron Botsis <aaron@bt-r.com> wrote:\\n> \\n> Hi folks, I was having a problem importing json data with COPY. Lots of things export data nicely as one json blob per line. This is excellent for directly importing into a JSON/JSONB column for analysis. \\n> \\n> ...Except when there’s an embedded doublequote. Or anything that’s escaped. COPY handles this, but by the time the escaped char hit the JSON parser, it's not escaped anymore. This breaks the JSON parsing. This means I need to manipulate the input data to double-escape it. See bug #12320 for an example. Yuck.\\n> \\n> I propose this small patch that simply allows specifying COPY … ESCAPE without requiring the CSV parser. It will make it much easier to directly use json formatted export data for folks going forward. This seemed like the simplest route.\\n> \\n> Usage is simply:\\n> \\n> postgres=# copy t1 from '/Users/nok/Desktop/queries.json';\\n> ERROR: invalid input syntax for type json\\n> DETAIL: Token \\\"root\\\" is invalid.\\n> CONTEXT: JSON data, line 1: ...1418066241619 AND <=1418671041621) AND user:\\\"root...\\n> COPY t1, line 3, column bleh: \\\"{\\\"timestamp\\\":\\\"2014-12-15T19:17:32.505Z\\\",\\\"duration\\\":7.947,\\\"query\\\":{\\\"query\\\":{\\\"filtered\\\":{\\\"filter\\\":{\\\"qu...\\\"\\n> postgres=# copy t1 from '/Users/nok/Desktop/queries.json' escape '';\\n> COPY 1966\\n> \\n> I’ve included regression tests, and all existing tests pass. This is my first contribution, so be kind to me. :)\\n> \\n> <escape-without-csv.patch>\\n> \\n> \\n>> Begin forwarded message:\\n>> \\n>> From: Francisco Olarte <folarte@peoplecall.com <mailto:folarte@peoplecall.com>>\\n>> Date: January 6, 2015 at 1:52:28 PM EST\\n>> Subject: Re: [BUGS] BUG #12320: json parsing with embedded double quotes\\n>> To: Aaron Botsis <aaron@bt-r.com <mailto:aaron@bt-r.com>>\\n>> Cc: postgres@bt-r.com <mailto:postgres@bt-r.com>, \\\"pgsql-bugs@postgresql.org <mailto:pgsql-bugs@postgresql.org>\\\" <pgsql-bugs@postgresql.org <mailto:pgsql-bugs@postgresql.org>>\\n>> \\n>> Hi Aaron:\\n>> \\n>> On Tue, Jan 6, 2015 at 7:06 PM, Aaron Botsis <aaron@bt-r.com <mailto:aaron@bt-r.com>> wrote:\\n>> Hi Francisco, I’m aware, but still consider this to be a bug, or at least a great opportunity for an enhancement. :) \\n>> \\n>> Maybe, but you are going to have a problem.\\n>> \\n>> This had bitten me for the third time while trying to import some json data. It’d be great to bypass the copy escaping (and possibly other meta characters) when the column type is json or jsonb. I’d be happy to try and write it and submit a patch if folks believe this is an acceptable way to go… That said, I should probably read what the process is for this kind of thing :)\\n>> \\n>> Reading this, you are talking about 'the column being json'. COPY needs to do the escaping at the same time it's constructing the columns. The present way is easy to do, read char by char, if it's a escape, process next char acumulating into current field, otherwise see whether it is a field or record separator and act accordingly. It's also layered, when you construct the records for copy you get all the field data, turn them into escaped strings, join them by the field separator and spit them out followed by a record separator ( practical implementations may do this virtually ). Intermixing this with the 'I'm in a json column' would need to pass information from the upper layer, and make it more difficult and, specially error prone. What do you do if ( using the standard delimiters ) your json value has embeded newlines and tabs ( which, IIRC, are legal in several places inside the json ). And all this to make some incorrectly formatted files read ( which can be correctly formatted with a perl one liner or something similar ). I'm not the one to decide, but I will vote against including that ( but do not trust me too much, I would also vote against including 'csv' which I consider the root of many evils ).\\n>> \\n>> Francisco Olarte.\\n>> \\n>> \\n>\",\n \"msg_date\": \"Wed, 7 Jan 2015 10:37:58 -0500\",\n \"msg_from\": \"Aaron Botsis <aaron@bt-r.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch: [BUGS] BUG #12320: json parsing with embedded double\\n quotes\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 01/07/2015 08:25 AM, Aaron Botsis wrote:\\n> Hi folks, I was having a problem importing json data with COPY. Lots \\n> of things export data nicely as one json blob per line. This is \\n> excellent for directly importing into a JSON/JSONB column for analysis.\\n>\\n> ...Except when there’s an embedded doublequote. Or anything that’s \\n> escaped. COPY handles this, but by the time the escaped char hit the \\n> JSON parser, it's not escaped anymore. This breaks the JSON parsing. \\n> This means I need to manipulate the input data to double-escape it. \\n> See bug #12320 for an example. Yuck.\\n>\\n> I propose this small patch that simply allows specifying COPY … ESCAPE \\n> without requiring the CSV parser. It will make it much easier to \\n> directly use json formatted export data for folks going forward. This \\n> seemed like the simplest route.\\n>\\n> Usage is simply:\\n>\\n> postgres=# copy t1 from '/Users/nok/Desktop/queries.json';\\n> ERROR: invalid input syntax for type json\\n> DETAIL: Token \\\"root\\\" is invalid.\\n> CONTEXT: JSON data, line 1: ...1418066241619 AND <=1418671041621) AND \\n> user:\\\"root...\\n> COPY t1, line 3, column bleh: \\n> \\\"{\\\"timestamp\\\":\\\"2014-12-15T19:17:32.505Z\\\",\\\"duration\\\":7.947,\\\"query\\\":{\\\"query\\\":{\\\"filtered\\\":{\\\"filter\\\":{\\\"qu...\\\"\\n> postgres=# copy t1 from '/Users/nok/Desktop/queries.json' escape '';\\n> COPY 1966\\n>\\n>\\n\\n\\nThis isn't a bug. Neither CSV format nor TEXT format are partucularly \\nsuitable for json. I'm quite certain I could compose legal json that \\nwill break your proposal (for example, with an embedded newline in the \\nwhite space.)\\n\\nIt's also unnecessary. CSV format, while not designed for this, is \\nnevertheless sufficiently flexible to allow successful import of json \\ndata meeting certain criteria (essentially no newlines), like this:\\n\\n copy the_table(jsonfield)\\n from '/path/to/jsondata'\\n csv quote e'\\\\x01' delimiter e'\\\\x02';\\n\\n\\nYou aren't the first person to encounter this problem. See \\n<http://adpgtech.blogspot.com/2014/09/importing-json-data.html>\\n\\nMaybe we need to add something like this to the docs, or to the wiki.\\n\\nNote too my comment in that blog post:\\n\\n Now this solution is a bit of a hack. I wonder if there's a case for\\n a COPY mode that simply treats each line as a single datum. I also\\n wonder if we need some more specialized tools for importing JSON,\\n possibly one or more Foreign Data Wrappers. Such things could\\n handle, say, embedded newline punctuation.\\n\\n\\ncheers\\n\\nandrew\\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\": \"Wed, 07 Jan 2015 14:45:51 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch: [BUGS] BUG #12320: json parsing with embedded double\\n quotes\"\n },\n {\n \"msg_contents\": \"> On Jan 7, 2015, at 2:45 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n> \\n> \\n> On 01/07/2015 08:25 AM, Aaron Botsis wrote:\\n>> Hi folks, I was having a problem importing json data with COPY. Lots of things export data nicely as one json blob per line. This is excellent for directly importing into a JSON/JSONB column for analysis.\\n>> \\n>> ...Except when there’s an embedded doublequote. Or anything that’s escaped. COPY handles this, but by the time the escaped char hit the JSON parser, it's not escaped anymore. This breaks the JSON parsing. This means I need to manipulate the input data to double-escape it. See bug #12320 for an example. Yuck.\\n>> \\n>> I propose this small patch that simply allows specifying COPY … ESCAPE without requiring the CSV parser. It will make it much easier to directly use json formatted export data for folks going forward. This seemed like the simplest route.\\n> \\n> This isn't a bug. Neither CSV format nor TEXT format are partucularly suitable for json. I'm quite certain I could compose legal json that will break your proposal (for example, with an embedded newline in the white space.)\\n\\nSorry - though I originally reported it as a bug, I didn’t mean to imply it ultimately was one. :) The patch is a feature enhancement.\\n\\n> It's also unnecessary. CSV format, while not designed for this, is nevertheless sufficiently flexible to allow successful import of json data meeting certain criteria (essentially no newlines), like this:\\n> \\n> copy the_table(jsonfield)\\n> from '/path/to/jsondata'\\n> csv quote e'\\\\x01' delimiter e'\\\\x02’;\\n\\nWhile perhaps unnecessary, given the size and simplicity of the patch, IMO it’s a no brainer to merge (it actually makes the code smaller by 3 lines). It also enables non-json use cases anytime one might want to preserve embedded escapes, or use different ones entirely. Do you see other reasons not to commit it?\\n\\n> You aren't the first person to encounter this problem. See <http://adpgtech.blogspot.com/2014/09/importing-json-data.html <http://adpgtech.blogspot.com/2014/09/importing-json-data.html>>\\n> \\n> Maybe we need to add something like this to the docs, or to the wiki.\\n\\nIn your post you acknowledged a text mode copy with null escape character would have solved your problem, and to me, that was the intuitive/obvious choice as well. *shrug*\\n\\n> Note too my comment in that blog post:\\n> \\n> Now this solution is a bit of a hack. I wonder if there's a case for\\n> a COPY mode that simply treats each line as a single datum. I also\\n> wonder if we need some more specialized tools for importing JSON,\\n> possibly one or more Foreign Data Wrappers. Such things could\\n> handle, say, embedded newline punctuation.\\n\\nAgree. Though https://github.com/citusdata/json_fdw <https://github.com/citusdata/json_fdw> already exists (I haven’t used it). How do folks feel about a COPY mode that reads a single datum until it finds a single character record terminator alone on a line? something like:\\n\\n=# COPY test(data) from stdin terminator ‘}’;\\nEnd with a backslash and a period on a line by itself.\\n>>{\\n>> \\\"a\\\": 123, \\\"c\\\": \\\"string\\\", \\\"b\\\": [1, 2, 3\\n>> ]\\n>>}\\n>>{\\n>>\\t[1,2,3]\\n>>}\\n>>\\\\.\\nCOPY 2\\n\\nYou could also get fancy and support multi-character record terminators which would allow the same thing in a slightly different way:\\nCOPY test(data) from stdin terminator ‘\\\\n}’;\\n\\nCOPY test(data) from stdin terminator ‘\\\\n}’;\\n\\nI don’t know how crazy you could get without a lot of rework. It might have to be used in conjunction with a more constrained mode like \\\"COPY…RAW”.\\n\\nAaron\\nOn Jan 7, 2015, at 2:45 PM, Andrew Dunstan <andrew@dunslane.net> wrote:On 01/07/2015 08:25 AM, Aaron Botsis wrote:Hi folks, I was having a problem importing json data with COPY. Lots of things export data nicely as one json blob per line. This is excellent for directly importing into a JSON/JSONB column for analysis....Except when there’s an embedded doublequote. Or anything that’s escaped. COPY handles this, but by the time the escaped char hit the JSON parser, it's not escaped anymore. This breaks the JSON parsing. This means I need to manipulate the input data to double-escape it. See bug #12320 for an example. Yuck.I propose this small patch that simply allows specifying COPY … ESCAPE without requiring the CSV parser. It will make it much easier to directly use json formatted export data for folks going forward. This seemed like the simplest route.This isn't a bug. Neither CSV format nor TEXT format are partucularly suitable for json. I'm quite certain I could compose legal json that will break your proposal (for example, with an embedded newline in the white space.)Sorry - though I originally reported it as a bug, I didn’t mean to imply it ultimately was one. :) The patch is a feature enhancement.It's also unnecessary. CSV format, while not designed for this, is nevertheless sufficiently flexible to allow successful import of json data meeting certain criteria (essentially no newlines), like this: copy the_table(jsonfield) from '/path/to/jsondata' csv quote e'\\\\x01' delimiter e'\\\\x02’;While perhaps unnecessary, given the size and simplicity of the patch, IMO it’s a no brainer to merge (it actually makes the code smaller by 3 lines). It also enables non-json use cases anytime one might want to preserve embedded escapes, or use different ones entirely. Do you see other reasons not to commit it?You aren't the first person to encounter this problem. See <http://adpgtech.blogspot.com/2014/09/importing-json-data.html>Maybe we need to add something like this to the docs, or to the wiki.In your post you acknowledged a text mode copy with null escape character would have solved your problem, and to me, that was the intuitive/obvious choice as well. *shrug*Note too my comment in that blog post: Now this solution is a bit of a hack. I wonder if there's a case for a COPY mode that simply treats each line as a single datum. I also wonder if we need some more specialized tools for importing JSON, possibly one or more Foreign Data Wrappers. Such things could handle, say, embedded newline punctuation.Agree. Though https://github.com/citusdata/json_fdw already exists (I haven’t used it). How do folks feel about a COPY mode that reads a single datum until it finds a single character record terminator alone on a line? something like:=# COPY test(data) from stdin terminator ‘}’;End with a backslash and a period on a line by itself.>>{>> \\\"a\\\": 123, \\\"c\\\": \\\"string\\\", \\\"b\\\": [1, 2, 3>> ]>>}>>{>> [1,2,3]>>}>>\\\\.COPY 2You could also get fancy and support multi-character record terminators which would allow the same thing in a slightly different way:COPY test(data) from stdin terminator ‘\\\\n}’;COPY test(data) from stdin terminator ‘\\\\n}’;I don’t know how crazy you could get without a lot of rework. It might have to be used in conjunction with a more constrained mode like \\\"COPY…RAW”.Aaron\",\n \"msg_date\": \"Thu, 8 Jan 2015 15:05:10 -0500\",\n \"msg_from\": \"Aaron Botsis <aaron@bt-r.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch: [BUGS] BUG #12320: json parsing with embedded double\\n quotes\"\n },\n {\n \"msg_contents\": \"\\nOn 01/08/2015 03:05 PM, Aaron Botsis wrote:\\n>\\n>\\n>> It's also unnecessary. CSV format, while not designed for this, is \\n>> nevertheless sufficiently flexible to allow successful import of json \\n>> data meeting certain criteria (essentially no newlines), like this:\\n>>\\n>> copy the_table(jsonfield)\\n>> from '/path/to/jsondata'\\n>> csv quote e'\\\\x01' delimiter e'\\\\x02’;\\n>\\n> While perhaps unnecessary, given the size and simplicity of the patch, \\n> IMO it’s a no brainer to merge (it actually makes the code smaller by \\n> 3 lines). It also enables non-json use cases anytime one might want to \\n> preserve embedded escapes, or use different ones entirely. Do you see \\n> other reasons not to commit it?\\n\\n\\nWell, for one thing it's seriously incomplete. You need to be able to \\nchange the delimiter as well. Otherwise, any embedded tab in the json \\nwill cause you major grief.\\n\\nCurrently the delimiter and the escape MUST be a single byte non-nul \\ncharacter, and there is a check for this in csv mode. Your patch would \\nallow any arbitrary string (including one of zero length) for the escape \\nin text mode, and would then silently ignore all but the first byte. \\nThat's not the way we like to do things.\\n\\nAnd, frankly, I would need to spend quite a lot more time thinking about \\nother implications than I have given it so far. This is an area where I \\ntend to be VERY cautious about making changes. This is a fairly fragile \\necosystem.\\n\\ncheers\\n\\nandrew\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 08 Jan 2015 15:44:44 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch: [BUGS] BUG #12320: json parsing with embedded double\\n quotes\"\n },\n {\n \"msg_contents\": \"> On Jan 8, 2015, at 3:44 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n> \\n> \\n> On 01/08/2015 03:05 PM, Aaron Botsis wrote:\\n>> \\n>> \\n>>> It's also unnecessary. CSV format, while not designed for this, is nevertheless sufficiently flexible to allow successful import of json data meeting certain criteria (essentially no newlines), like this:\\n>>> \\n>>> copy the_table(jsonfield)\\n>>> from '/path/to/jsondata'\\n>>> csv quote e'\\\\x01' delimiter e'\\\\x02’;\\n>> \\n>> While perhaps unnecessary, given the size and simplicity of the patch, IMO it’s a no brainer to merge (it actually makes the code smaller by 3 lines). It also enables non-json use cases anytime one might want to preserve embedded escapes, or use different ones entirely. Do you see other reasons not to commit it?\\n> \\n> \\n> Well, for one thing it's seriously incomplete. You need to be able to change the delimiter as well. Otherwise, any embedded tab in the json will cause you major grief.\\n> \\n> Currently the delimiter and the escape MUST be a single byte non-nul character, and there is a check for this in csv mode. Your patch would allow any arbitrary string (including one of zero length) for the escape in text mode, and would then silently ignore all but the first byte. That's not the way we like to do things.\\n> \\n> And, frankly, I would need to spend quite a lot more time thinking about other implications than I have given it so far. This is an area where I tend to be VERY cautious about making changes. This is a fairly fragile ecosystem.\\n\\nGood point.\\n\\nThis version:\\n\\n* doesn't allow ENCODING in BINARY mode (existing bug)\\n* doesn’t allow ESCAPE in BINARY mode\\n* makes COPY TO work with escape\\n* ensures escape char length is < 2 for text mode, 1 for CSV\\n\\nCouple more things to realize: setting both the escape and delimiter characters to null won’t be any different than how you fiddled with them in CSV mode. The code paths will be the same because we should never encounter a null in the middle of the string. And even if we did (and the encoding didn’t catch it), we’d treat it just like any other field delimiter or escape character.\\n\\nSo I’m going to do a bit more testing with another patch tomorrow with delimiters removed. If you can think of any specific cases you think will break it let me know and I’ll make sure to add regression tests for them as well. \\n\\nCheers!\\n\\nAaron\\n\\n\\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, 8 Jan 2015 20:42:00 -0500\",\n \"msg_from\": \"Aaron Botsis <aaron@bt-r.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch: [BUGS] BUG #12320: json parsing with embedded double\\n quotes\"\n },\n {\n \"msg_contents\": \"\\nOn 01/08/2015 08:42 PM, Aaron Botsis wrote:\\n>\\n>> On Jan 8, 2015, at 3:44 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>>\\n>>\\n>> On 01/08/2015 03:05 PM, Aaron Botsis wrote:\\n>>>\\n>>>> It's also unnecessary. CSV format, while not designed for this, is nevertheless sufficiently flexible to allow successful import of json data meeting certain criteria (essentially no newlines), like this:\\n>>>>\\n>>>> copy the_table(jsonfield)\\n>>>> from '/path/to/jsondata'\\n>>>> csv quote e'\\\\x01' delimiter e'\\\\x02’;\\n>>> While perhaps unnecessary, given the size and simplicity of the patch, IMO it’s a no brainer to merge (it actually makes the code smaller by 3 lines). It also enables non-json use cases anytime one might want to preserve embedded escapes, or use different ones entirely. Do you see other reasons not to commit it?\\n>>\\n>> Well, for one thing it's seriously incomplete. You need to be able to change the delimiter as well. Otherwise, any embedded tab in the json will cause you major grief.\\n>>\\n>> Currently the delimiter and the escape MUST be a single byte non-nul character, and there is a check for this in csv mode. Your patch would allow any arbitrary string (including one of zero length) for the escape in text mode, and would then silently ignore all but the first byte. That's not the way we like to do things.\\n>>\\n>> And, frankly, I would need to spend quite a lot more time thinking about other implications than I have given it so far. This is an area where I tend to be VERY cautious about making changes. This is a fairly fragile ecosystem.\\n> Good point.\\n>\\n> This version:\\n>\\n> * doesn't allow ENCODING in BINARY mode (existing bug)\\n> * doesn’t allow ESCAPE in BINARY mode\\n> * makes COPY TO work with escape\\n> * ensures escape char length is < 2 for text mode, 1 for CSV\\n>\\n> Couple more things to realize: setting both the escape and delimiter characters to null won’t be any different than how you fiddled with them in CSV mode. The code paths will be the same because we should never encounter a null in the middle of the string. And even if we did (and the encoding didn’t catch it), we’d treat it just like any other field delimiter or escape character.\\n>\\n> So I’m going to do a bit more testing with another patch tomorrow with delimiters removed. If you can think of any specific cases you think will break it let me know and I’ll make sure to add regression tests for them as well.\\n>\\n>\\n\\nWell, I still need convincing that this is the best solution to the \\nproblem. As I said, I need to spend more time thinking about it.\\n\\ncheers\\n\\nandrew\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 09 Jan 2015 11:37:19 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch: [BUGS] BUG #12320: json parsing with embedded double\\n quotes\"\n },\n {\n \"msg_contents\": \"> On Jan 9, 2015, at 11:37 AM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n> \\n> \\n> On 01/08/2015 08:42 PM, Aaron Botsis wrote:\\n>> \\n>>> On Jan 8, 2015, at 3:44 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>>> \\n>>> \\n>>> On 01/08/2015 03:05 PM, Aaron Botsis wrote:\\n>>>> \\n>>>>> It's also unnecessary. CSV format, while not designed for this, is nevertheless sufficiently flexible to allow successful import of json data meeting certain criteria (essentially no newlines), like this:\\n>>>>> \\n>>>>> copy the_table(jsonfield)\\n>>>>> from '/path/to/jsondata'\\n>>>>> csv quote e'\\\\x01' delimiter e'\\\\x02’;\\n>>>> While perhaps unnecessary, given the size and simplicity of the patch, IMO it’s a no brainer to merge (it actually makes the code smaller by 3 lines). It also enables non-json use cases anytime one might want to preserve embedded escapes, or use different ones entirely. Do you see other reasons not to commit it?\\n>>> \\n>>> Well, for one thing it's seriously incomplete. You need to be able to change the delimiter as well. Otherwise, any embedded tab in the json will cause you major grief.\\n>>> \\n>>> Currently the delimiter and the escape MUST be a single byte non-nul character, and there is a check for this in csv mode. Your patch would allow any arbitrary string (including one of zero length) for the escape in text mode, and would then silently ignore all but the first byte. That's not the way we like to do things.\\n>>> \\n>>> And, frankly, I would need to spend quite a lot more time thinking about other implications than I have given it so far. This is an area where I tend to be VERY cautious about making changes. This is a fairly fragile ecosystem.\\n>> \\n>> So I’m going to do a bit more testing with another patch tomorrow with delimiters removed. If you can think of any specific cases you think will break it let me know and I’ll make sure to add regression tests for them as well.\\n>> \\n> Well, I still need convincing that this is the best solution to the problem. As I said, I need to spend more time thinking about it.\\n\\n\\nI offered an alternative (RAW mode w/record terminator) that you ignored. So in lieu of a productive discussion about “the best solution to the problem”, I went ahead and continued to complete the patch since I believe it’s a useful FE that it could be helpful for this and other use cases. There have been multiple times (not even w/json) I wished COPY would stop being so smart. \\n\\nFWIW, (if anyone’s interested in it) I also hacked up some python that’ll read a json file, and outputs a binary file suitable for use with COPY BINARY that gets around all this stuff. Obviously this only works for json (not jsonb) columns (though you could SELECT INTO a jsonb column). Happy to pass it along.\\n\\nAaron\\n\\n\\nOn Jan 9, 2015, at 11:37 AM, Andrew Dunstan <andrew@dunslane.net> wrote:On 01/08/2015 08:42 PM, Aaron Botsis wrote:On Jan 8, 2015, at 3:44 PM, Andrew Dunstan <andrew@dunslane.net> wrote:On 01/08/2015 03:05 PM, Aaron Botsis wrote:It's also unnecessary. CSV format, while not designed for this, is nevertheless sufficiently flexible to allow successful import of json data meeting certain criteria (essentially no newlines), like this: copy the_table(jsonfield) from '/path/to/jsondata' csv quote e'\\\\x01' delimiter e'\\\\x02’;While perhaps unnecessary, given the size and simplicity of the patch, IMO it’s a no brainer to merge (it actually makes the code smaller by 3 lines). It also enables non-json use cases anytime one might want to preserve embedded escapes, or use different ones entirely. Do you see other reasons not to commit it?Well, for one thing it's seriously incomplete. You need to be able to change the delimiter as well. Otherwise, any embedded tab in the json will cause you major grief.Currently the delimiter and the escape MUST be a single byte non-nul character, and there is a check for this in csv mode. Your patch would allow any arbitrary string (including one of zero length) for the escape in text mode, and would then silently ignore all but the first byte. That's not the way we like to do things.And, frankly, I would need to spend quite a lot more time thinking about other implications than I have given it so far. This is an area where I tend to be VERY cautious about making changes. This is a fairly fragile ecosystem.So I’m going to do a bit more testing with another patch tomorrow with delimiters removed. If you can think of any specific cases you think will break it let me know and I’ll make sure to add regression tests for them as well.Well, I still need convincing that this is the best solution to the problem. As I said, I need to spend more time thinking about it.I offered an alternative (RAW mode w/record terminator) that you ignored. So in lieu of a productive discussion about “the best solution to the problem”, I went ahead and continued to complete the patch since I believe it’s a useful FE that it could be helpful for this and other use cases. There have been multiple times (not even w/json) I wished COPY would stop being so smart. FWIW, (if anyone’s interested in it) I also hacked up some python that’ll read a json file, and outputs a binary file suitable for use with COPY BINARY that gets around all this stuff. Obviously this only works for json (not jsonb) columns (though you could SELECT INTO a jsonb column). Happy to pass it along.Aaron\",\n \"msg_date\": \"Fri, 9 Jan 2015 12:57:31 -0500\",\n \"msg_from\": \"Aaron Botsis <aaron@bt-r.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch: [BUGS] BUG #12320: json parsing with embedded double\\n quotes\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 8, 2015 at 8:42 PM, Aaron Botsis <aaron@bt-r.com> wrote:\\n> This version:\\n>\\n> * doesn't allow ENCODING in BINARY mode (existing bug)\\n> * doesn’t allow ESCAPE in BINARY mode\\n\\nHmm, it does look like ENCODING and ESCAPE don't do anything in BINARY\\nmode, so adding error checks for that makes sense to me. I wondered\\nwhether the ENCODING option would affect the interpretation of text\\ndata that might be stored in binary-format COPY data, but I can't find\\nany evidence that it does.\\n\\n> * makes COPY TO work with escape\\n\\nI don't see anything horribly wrong with this, but it seems pretty\\nuseless. I mean, does anyone really want a newline encoded in the\\noutput as %n or !n or qn rather than \\\\n? What problem are you trying\\nto solve here?\\n\\n> * ensures escape char length is < 2 for text mode, 1 for CSV\\n\\nSo, is the idea here that we're going to have no escape character at\\nall? So, the characters that would require the use of an escape\\ncharacter to represent simply can't be represented at all, but we're\\nOK with that, because the input data doesn't contain any such\\ncharacters?\\n\\nOne general problem with this patch is that it doesn't update the\\ndocumentation, which makes it a bit harder than it might be to see\\nwhat you're trying to accomplish.\\n\\nAlso, if you want to decrease the chances that this patch will get\\nforgotten about, you can register it here:\\n\\nhttps://commitfest.postgresql.org/action/commitfest_view/open\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 14 Jan 2015 09:05:52 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch: [BUGS] BUG #12320: json parsing with embedded double\\n quotes\"\n }\n]","isConversational":false}}},{"rowIdx":39997,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nI've played with trying to find out which minimal set of files I need\nfrom the old version to make pg_upgrade work. Interestingly, this\nincludes the good old postmaster binary:\n\n$ sudo -u postgres pgsql/bin/pg_upgrade -b /var/tmp/pgsql/bin/ -B /usr/lib/postgresql/9.5/bin/ -d /etc/postgresql/9.5/main -D /tmp/9.5/data\nFinding the real data directory for the old cluster sh: 1: /var/tmp/pgsql/bin/postmaster: not found\n\nCould not get data directory using \"/var/tmp/pgsql/bin/postmaster\" -D \"/etc/postgresql/9.5/main\" -C data_directory: No such file or directory\nFailure, exiting\n\nI think it should just use \"postgres\" there, patch attached. (If we\nneed to be compatible with postmaster-only PG versions from the last\ncentury, it should try both names.)\n\nChristoph\n-- \ncb@df7cb.de | http://www.df7cb.de/\n\n\n\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 Dec 2014 23:48:52 +0100","msg_from":"Christoph Berg <cb@df7cb.de>","msg_from_op":true,"msg_subject":"pg_upgrade needs postmaster [sic]"},{"msg_contents":"On Mon, Dec 22, 2014 at 11:48:52PM +0100, Christoph Berg wrote:\n> Hi,\n> \n> I've played with trying to find out which minimal set of files I need\n> from the old version to make pg_upgrade work. Interestingly, this\n> includes the good old postmaster binary:\n> \n> $ sudo -u postgres pgsql/bin/pg_upgrade -b /var/tmp/pgsql/bin/ -B /usr/lib/postgresql/9.5/bin/ -d /etc/postgresql/9.5/main -D /tmp/9.5/data\n> Finding the real data directory for the old cluster sh: 1: /var/tmp/pgsql/bin/postmaster: not found\n> \n> Could not get data directory using \"/var/tmp/pgsql/bin/postmaster\" -D \"/etc/postgresql/9.5/main\" -C data_directory: No such file or directory\n> Failure, exiting\n> \n> I think it should just use \"postgres\" there, patch attached. (If we\n> need to be compatible with postmaster-only PG versions from the last\n> century, it should try both names.)\n\nYes, I think you are right. I see pg_ctl using the 'postgres' binary,\nso pg_upgrade should do the same. I will apply this patch soon, and see\nif there are any other 'postmaster' mentions that need updating.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n + Everyone has their own god. +\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 22 Dec 2014 18:18:35 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: pg_upgrade needs postmaster [sic]"},{"msg_contents":"On Mon, Dec 22, 2014 at 06:18:35PM -0500, Bruce Momjian wrote:\n> On Mon, Dec 22, 2014 at 11:48:52PM +0100, Christoph Berg wrote:\n> > Hi,\n> > \n> > I've played with trying to find out which minimal set of files I need\n> > from the old version to make pg_upgrade work. Interestingly, this\n> > includes the good old postmaster binary:\n> > \n> > $ sudo -u postgres pgsql/bin/pg_upgrade -b /var/tmp/pgsql/bin/ -B /usr/lib/postgresql/9.5/bin/ -d /etc/postgresql/9.5/main -D /tmp/9.5/data\n> > Finding the real data directory for the old cluster sh: 1: /var/tmp/pgsql/bin/postmaster: not found\n> > \n> > Could not get data directory using \"/var/tmp/pgsql/bin/postmaster\" -D \"/etc/postgresql/9.5/main\" -C data_directory: No such file or directory\n> > Failure, exiting\n> > \n> > I think it should just use \"postgres\" there, patch attached. (If we\n> > need to be compatible with postmaster-only PG versions from the last\n> > century, it should try both names.)\n> \n> Yes, I think you are right. I see pg_ctl using the 'postgres' binary,\n> so pg_upgrade should do the same. I will apply this patch soon, and see\n> if there are any other 'postmaster' mentions that need updating.\n\nPatch applied. Thanks.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n + Everyone has their own god. +\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 1 Apr 2015 18:26:00 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: pg_upgrade needs postmaster [sic]"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nI've played with trying to find out which minimal set of files I need\\nfrom the old version to make pg_upgrade work. Interestingly, this\\nincludes the good old postmaster binary:\\n\\n$ sudo -u postgres pgsql/bin/pg_upgrade -b /var/tmp/pgsql/bin/ -B /usr/lib/postgresql/9.5/bin/ -d /etc/postgresql/9.5/main -D /tmp/9.5/data\\nFinding the real data directory for the old cluster sh: 1: /var/tmp/pgsql/bin/postmaster: not found\\n\\nCould not get data directory using \\\"/var/tmp/pgsql/bin/postmaster\\\" -D \\\"/etc/postgresql/9.5/main\\\" -C data_directory: No such file or directory\\nFailure, exiting\\n\\nI think it should just use \\\"postgres\\\" there, patch attached. (If we\\nneed to be compatible with postmaster-only PG versions from the last\\ncentury, it should try both names.)\\n\\nChristoph\\n-- \\ncb@df7cb.de | http://www.df7cb.de/\\n\\n\\n\\n-- \\nSent via 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 Dec 2014 23:48:52 +0100\",\n \"msg_from\": \"Christoph Berg <cb@df7cb.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"pg_upgrade needs postmaster [sic]\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 22, 2014 at 11:48:52PM +0100, Christoph Berg wrote:\\n> Hi,\\n> \\n> I've played with trying to find out which minimal set of files I need\\n> from the old version to make pg_upgrade work. Interestingly, this\\n> includes the good old postmaster binary:\\n> \\n> $ sudo -u postgres pgsql/bin/pg_upgrade -b /var/tmp/pgsql/bin/ -B /usr/lib/postgresql/9.5/bin/ -d /etc/postgresql/9.5/main -D /tmp/9.5/data\\n> Finding the real data directory for the old cluster sh: 1: /var/tmp/pgsql/bin/postmaster: not found\\n> \\n> Could not get data directory using \\\"/var/tmp/pgsql/bin/postmaster\\\" -D \\\"/etc/postgresql/9.5/main\\\" -C data_directory: No such file or directory\\n> Failure, exiting\\n> \\n> I think it should just use \\\"postgres\\\" there, patch attached. (If we\\n> need to be compatible with postmaster-only PG versions from the last\\n> century, it should try both names.)\\n\\nYes, I think you are right. I see pg_ctl using the 'postgres' binary,\\nso pg_upgrade should do the same. I will apply this patch soon, and see\\nif there are any other 'postmaster' mentions that need updating.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n + Everyone has their own god. +\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 22 Dec 2014 18:18:35 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_upgrade needs postmaster [sic]\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 22, 2014 at 06:18:35PM -0500, Bruce Momjian wrote:\\n> On Mon, Dec 22, 2014 at 11:48:52PM +0100, Christoph Berg wrote:\\n> > Hi,\\n> > \\n> > I've played with trying to find out which minimal set of files I need\\n> > from the old version to make pg_upgrade work. Interestingly, this\\n> > includes the good old postmaster binary:\\n> > \\n> > $ sudo -u postgres pgsql/bin/pg_upgrade -b /var/tmp/pgsql/bin/ -B /usr/lib/postgresql/9.5/bin/ -d /etc/postgresql/9.5/main -D /tmp/9.5/data\\n> > Finding the real data directory for the old cluster sh: 1: /var/tmp/pgsql/bin/postmaster: not found\\n> > \\n> > Could not get data directory using \\\"/var/tmp/pgsql/bin/postmaster\\\" -D \\\"/etc/postgresql/9.5/main\\\" -C data_directory: No such file or directory\\n> > Failure, exiting\\n> > \\n> > I think it should just use \\\"postgres\\\" there, patch attached. (If we\\n> > need to be compatible with postmaster-only PG versions from the last\\n> > century, it should try both names.)\\n> \\n> Yes, I think you are right. I see pg_ctl using the 'postgres' binary,\\n> so pg_upgrade should do the same. I will apply this patch soon, and see\\n> if there are any other 'postmaster' mentions that need updating.\\n\\nPatch applied. Thanks.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n + Everyone has their own god. +\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 1 Apr 2015 18:26:00 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_upgrade needs postmaster [sic]\"\n }\n]","isConversational":false}}},{"rowIdx":39998,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Currently the buildfarm animal crake (my development instance) is \nrunning the bin check, but not any other animal. These tests still take \nfor too long, not least because each set of tests requires a separate \ninstall.\n\nI have complained about this before, but we don't seem to have made any \nprogress. I am very reluctant to put this check into a buildfarm client \nrelease until that is addressed. Is there really nothing we can do about \nit? This is getting more important since we now have another feature \nthat we want to get into a buildfarm client release.\n\ncheers\n\nandrew\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 22 Dec 2014 19:56:07 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":true,"msg_subject":"bin checks taking too long."},{"msg_contents":"On 12/22/14 7:56 PM, Andrew Dunstan wrote:\n> Currently the buildfarm animal crake (my development instance) is\n> running the bin check, but not any other animal. These tests still take\n> for too long, not least because each set of tests requires a separate\n> install.\n\nYou can avoid that by using make installcheck.\n\n> Is there really nothing we can do about it?\n\nThere is, but it's not straightforward, as it turns out. Ideas are welcome.\n\nAlso, as you can imagine, any build system changes are bottlenecked on\nWindows support.\n\n\n\n-- \nSent via 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 Dec 2014 22:41:26 -0500","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":false,"msg_subject":"Re: bin checks taking too long."},{"msg_contents":"\nOn 12/22/2014 10:41 PM, Peter Eisentraut wrote:\n> On 12/22/14 7:56 PM, Andrew Dunstan wrote:\n>> Currently the buildfarm animal crake (my development instance) is\n>> running the bin check, but not any other animal. These tests still take\n>> for too long, not least because each set of tests requires a separate\n>> install.\n> You can avoid that by using make installcheck.\n\n\nNot working too well:\n\n make -C pg_ctl installcheck\n make[1]: Entering directory\n `/home/bf/bfr/root/HEAD/pgsql.build/src/bin/pg_ctl'\n cd /home/bf/bfr/root/HEAD/pgsql.build/../pgsql/src/bin/pg_ctl &&\n TESTDIR='/home/bf/bfr/root/HEAD/pgsql.build/src/bin/pg_ctl'\n PATH=\"/home/bf/bfr/root/HEAD/inst/bin:$PATH\" PGPORT='65648' prove -I\n /home/bf/bfr/root/HEAD/pgsql.build/../pgsql/src/test/perl/ --verbose\n t/*.pl\n Use of uninitialized value $ENV{\"top_builddir\"} in concatenation (.)\n or string at t/001_start_stop.pl line 17.\n file not found: /src/test/regress/pg_regress at\n /home/bf/bfr/root/HEAD/pgsql.build/../pgsql/src/test/perl/TestLib.pm\n line 142.\n\nThis was invoked like this:\n\n cd $pgsql/src/bin && make NO_LOCALE=1 installcheck\n\nNote that it's a vpath build.\n\n\ncheers\n\nandrew\n\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 23 Dec 2014 11:11:42 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":true,"msg_subject":"Re: bin checks taking too long."},{"msg_contents":"On 12/23/2014 11:11 AM, Andrew Dunstan wrote:\n>\n> On 12/22/2014 10:41 PM, Peter Eisentraut wrote:\n>> On 12/22/14 7:56 PM, Andrew Dunstan wrote:\n>>> Currently the buildfarm animal crake (my development instance) is\n>>> running the bin check, but not any other animal. These tests still take\n>>> for too long, not least because each set of tests requires a separate\n>>> install.\n>> You can avoid that by using make installcheck.\n>\n>\n> Not working too well:\n>\n> make -C pg_ctl installcheck\n> make[1]: Entering directory\n> `/home/bf/bfr/root/HEAD/pgsql.build/src/bin/pg_ctl'\n> cd /home/bf/bfr/root/HEAD/pgsql.build/../pgsql/src/bin/pg_ctl &&\n> TESTDIR='/home/bf/bfr/root/HEAD/pgsql.build/src/bin/pg_ctl'\n> PATH=\"/home/bf/bfr/root/HEAD/inst/bin:$PATH\" PGPORT='65648' prove -I\n> /home/bf/bfr/root/HEAD/pgsql.build/../pgsql/src/test/perl/ --verbose\n> t/*.pl\n> Use of uninitialized value $ENV{\"top_builddir\"} in concatenation (.)\n> or string at t/001_start_stop.pl line 17.\n> file not found: /src/test/regress/pg_regress at\n> /home/bf/bfr/root/HEAD/pgsql.build/../pgsql/src/test/perl/TestLib.pm\n> line 142.\n>\n> This was invoked like this:\n>\n> cd $pgsql/src/bin && make NO_LOCALE=1 installcheck\n>\n> Note that it's a vpath build.\n>\n>\n\nThe attached patch seems to fix this. However, running installcheck also \nmakes very little difference to the time. These checks still take around \n100 seconds on my machine, including around 60 for the script tests. \nThat seems ridiculously long, when the whole core regression suite takes \n23 seconds.\n\ncheers\n\nandrew\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":"Tue, 23 Dec 2014 13:55:46 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":true,"msg_subject":"Re: bin checks taking too long."},{"msg_contents":"On Tue, Dec 23, 2014 at 01:55:46PM -0500, Andrew Dunstan wrote:\n> On 12/23/2014 11:11 AM, Andrew Dunstan wrote:\n> > Use of uninitialized value $ENV{\"top_builddir\"} in concatenation (.)\n> > or string at t/001_start_stop.pl line 17.\n> > file not found: /src/test/regress/pg_regress at\n> >/home/bf/bfr/root/HEAD/pgsql.build/../pgsql/src/test/perl/TestLib.pm\n> > line 142.\n> >\n> >This was invoked like this:\n> >\n> > cd $pgsql/src/bin && make NO_LOCALE=1 installcheck\n> >\n> >Note that it's a vpath build.\n> >\n> >\n> \n> The attached patch seems to fix this.\n\nOops. Your fix looks good.\n\n> However, running installcheck also\n> makes very little difference to the time. These checks still take around 100\n> seconds on my machine, including around 60 for the script tests. That seems\n> ridiculously long, when the whole core regression suite takes 23 seconds.\n\nI am satisfied with the performance of the TAP suites.\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 24 Dec 2014 02:56:31 -0500","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":false,"msg_subject":"Re: bin checks taking too long."}],"string":"[\n {\n \"msg_contents\": \"Currently the buildfarm animal crake (my development instance) is \\nrunning the bin check, but not any other animal. These tests still take \\nfor too long, not least because each set of tests requires a separate \\ninstall.\\n\\nI have complained about this before, but we don't seem to have made any \\nprogress. I am very reluctant to put this check into a buildfarm client \\nrelease until that is addressed. Is there really nothing we can do about \\nit? This is getting more important since we now have another feature \\nthat we want to get into a buildfarm client release.\\n\\ncheers\\n\\nandrew\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 22 Dec 2014 19:56:07 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"bin checks taking too long.\"\n },\n {\n \"msg_contents\": \"On 12/22/14 7:56 PM, Andrew Dunstan wrote:\\n> Currently the buildfarm animal crake (my development instance) is\\n> running the bin check, but not any other animal. These tests still take\\n> for too long, not least because each set of tests requires a separate\\n> install.\\n\\nYou can avoid that by using make installcheck.\\n\\n> Is there really nothing we can do about it?\\n\\nThere is, but it's not straightforward, as it turns out. Ideas are welcome.\\n\\nAlso, as you can imagine, any build system changes are bottlenecked on\\nWindows support.\\n\\n\\n\\n-- \\nSent via pgsql-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 Dec 2014 22:41:26 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: bin checks taking too long.\"\n },\n {\n \"msg_contents\": \"\\nOn 12/22/2014 10:41 PM, Peter Eisentraut wrote:\\n> On 12/22/14 7:56 PM, Andrew Dunstan wrote:\\n>> Currently the buildfarm animal crake (my development instance) is\\n>> running the bin check, but not any other animal. These tests still take\\n>> for too long, not least because each set of tests requires a separate\\n>> install.\\n> You can avoid that by using make installcheck.\\n\\n\\nNot working too well:\\n\\n make -C pg_ctl installcheck\\n make[1]: Entering directory\\n `/home/bf/bfr/root/HEAD/pgsql.build/src/bin/pg_ctl'\\n cd /home/bf/bfr/root/HEAD/pgsql.build/../pgsql/src/bin/pg_ctl &&\\n TESTDIR='/home/bf/bfr/root/HEAD/pgsql.build/src/bin/pg_ctl'\\n PATH=\\\"/home/bf/bfr/root/HEAD/inst/bin:$PATH\\\" PGPORT='65648' prove -I\\n /home/bf/bfr/root/HEAD/pgsql.build/../pgsql/src/test/perl/ --verbose\\n t/*.pl\\n Use of uninitialized value $ENV{\\\"top_builddir\\\"} in concatenation (.)\\n or string at t/001_start_stop.pl line 17.\\n file not found: /src/test/regress/pg_regress at\\n /home/bf/bfr/root/HEAD/pgsql.build/../pgsql/src/test/perl/TestLib.pm\\n line 142.\\n\\nThis was invoked like this:\\n\\n cd $pgsql/src/bin && make NO_LOCALE=1 installcheck\\n\\nNote that it's a vpath build.\\n\\n\\ncheers\\n\\nandrew\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 23 Dec 2014 11:11:42 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: bin checks taking too long.\"\n },\n {\n \"msg_contents\": \"On 12/23/2014 11:11 AM, Andrew Dunstan wrote:\\n>\\n> On 12/22/2014 10:41 PM, Peter Eisentraut wrote:\\n>> On 12/22/14 7:56 PM, Andrew Dunstan wrote:\\n>>> Currently the buildfarm animal crake (my development instance) is\\n>>> running the bin check, but not any other animal. These tests still take\\n>>> for too long, not least because each set of tests requires a separate\\n>>> install.\\n>> You can avoid that by using make installcheck.\\n>\\n>\\n> Not working too well:\\n>\\n> make -C pg_ctl installcheck\\n> make[1]: Entering directory\\n> `/home/bf/bfr/root/HEAD/pgsql.build/src/bin/pg_ctl'\\n> cd /home/bf/bfr/root/HEAD/pgsql.build/../pgsql/src/bin/pg_ctl &&\\n> TESTDIR='/home/bf/bfr/root/HEAD/pgsql.build/src/bin/pg_ctl'\\n> PATH=\\\"/home/bf/bfr/root/HEAD/inst/bin:$PATH\\\" PGPORT='65648' prove -I\\n> /home/bf/bfr/root/HEAD/pgsql.build/../pgsql/src/test/perl/ --verbose\\n> t/*.pl\\n> Use of uninitialized value $ENV{\\\"top_builddir\\\"} in concatenation (.)\\n> or string at t/001_start_stop.pl line 17.\\n> file not found: /src/test/regress/pg_regress at\\n> /home/bf/bfr/root/HEAD/pgsql.build/../pgsql/src/test/perl/TestLib.pm\\n> line 142.\\n>\\n> This was invoked like this:\\n>\\n> cd $pgsql/src/bin && make NO_LOCALE=1 installcheck\\n>\\n> Note that it's a vpath build.\\n>\\n>\\n\\nThe attached patch seems to fix this. However, running installcheck also \\nmakes very little difference to the time. These checks still take around \\n100 seconds on my machine, including around 60 for the script tests. \\nThat seems ridiculously long, when the whole core regression suite takes \\n23 seconds.\\n\\ncheers\\n\\nandrew\\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\": \"Tue, 23 Dec 2014 13:55:46 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: bin checks taking too long.\"\n },\n {\n \"msg_contents\": \"On Tue, Dec 23, 2014 at 01:55:46PM -0500, Andrew Dunstan wrote:\\n> On 12/23/2014 11:11 AM, Andrew Dunstan wrote:\\n> > Use of uninitialized value $ENV{\\\"top_builddir\\\"} in concatenation (.)\\n> > or string at t/001_start_stop.pl line 17.\\n> > file not found: /src/test/regress/pg_regress at\\n> >/home/bf/bfr/root/HEAD/pgsql.build/../pgsql/src/test/perl/TestLib.pm\\n> > line 142.\\n> >\\n> >This was invoked like this:\\n> >\\n> > cd $pgsql/src/bin && make NO_LOCALE=1 installcheck\\n> >\\n> >Note that it's a vpath build.\\n> >\\n> >\\n> \\n> The attached patch seems to fix this.\\n\\nOops. Your fix looks good.\\n\\n> However, running installcheck also\\n> makes very little difference to the time. These checks still take around 100\\n> seconds on my machine, including around 60 for the script tests. That seems\\n> ridiculously long, when the whole core regression suite takes 23 seconds.\\n\\nI am satisfied with the performance of the TAP suites.\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 24 Dec 2014 02:56:31 -0500\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: bin checks taking too long.\"\n }\n]","isConversational":false}}},{"rowIdx":39999,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"There is a new column added in pg_authid (rolbypassrls)\nand the updation for same is missed in below places:\n\na. System catalog page for pg_authid\nhttp://www.postgresql.org/docs/devel/static/catalog-pg-authid.html\nb. Do we want to add this new column in pg_shadow view?\n\nWith Regards,\nAmit Kapila.\nEnterpriseDB: http://www.enterprisedb.com\n\nThere is a new column added in pg_authid (rolbypassrls)and the updation for same is missed in below places:a. System catalog page for pg_authidhttp://www.postgresql.org/docs/devel/static/catalog-pg-authid.htmlb. Do we want to add this new column in pg_shadow view?With Regards,Amit Kapila.EnterpriseDB: http://www.enterprisedb.com","msg_date":"Tue, 23 Dec 2014 11:48:07 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":true,"msg_subject":"Missing updates at few places for row level security"},{"msg_contents":"Amit,\n\n* Amit Kapila (amit.kapila16@gmail.com) wrote:\n> There is a new column added in pg_authid (rolbypassrls)\n> and the updation for same is missed in below places:\n> \n> a. System catalog page for pg_authid\n> http://www.postgresql.org/docs/devel/static/catalog-pg-authid.html\n\nYup, thanks, will fix.\n\n> b. Do we want to add this new column in pg_shadow view?\n\nThis was intentionally not done as I had really viewed pg_user, pg_group\nand pg_shadow as deprecated and only for backwards-compatibility.\nThat's certainly why those views were added originally. On the other\nhand, I do see that 'userepl' was added, so we've been keeping it\nupdated.\n\nI am inclined to suggest that we actually mark pg_user, pg_group, and\npg_shadow as deprecated and planned for removal. We can't simply remove\nthem as we haven't actually said that but I don't think we should\ncontinue to carry these pre-8.1 views around forever. Still, until we\ndo that, we should keep them updated.\n\nI'll make these changes in the next couple of days (after the role\nattribute bitmask patch, as that touches a lot of the same places).\n\n\tThanks,\n\n\t\tStephen","msg_date":"Tue, 23 Dec 2014 07:11:32 -0500","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: Missing updates at few places for row level security"},{"msg_contents":"Amit,\n\n* Amit Kapila (amit.kapila16@gmail.com) wrote:\n> There is a new column added in pg_authid (rolbypassrls)\n> and the updation for same is missed in below places:\n> \n> a. System catalog page for pg_authid\n> http://www.postgresql.org/docs/devel/static/catalog-pg-authid.html\n> b. Do we want to add this new column in pg_shadow view?\n\nThanks for this. Pushed with a few additional changes (included it in\nboth pg_user and pg_shadow, few other doc updates, etc).\n\n\tThanks again!\n\n\t\tStephen","msg_date":"Wed, 28 Jan 2015 21:48:06 -0500","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: Missing updates at few places for row level security"}],"string":"[\n {\n \"msg_contents\": \"There is a new column added in pg_authid (rolbypassrls)\\nand the updation for same is missed in below places:\\n\\na. System catalog page for pg_authid\\nhttp://www.postgresql.org/docs/devel/static/catalog-pg-authid.html\\nb. Do we want to add this new column in pg_shadow view?\\n\\nWith Regards,\\nAmit Kapila.\\nEnterpriseDB: http://www.enterprisedb.com\\n\\nThere is a new column added in pg_authid (rolbypassrls)and the updation for same is missed in below places:a. System catalog page for pg_authidhttp://www.postgresql.org/docs/devel/static/catalog-pg-authid.htmlb. Do we want to add this new column in pg_shadow view?With Regards,Amit Kapila.EnterpriseDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Tue, 23 Dec 2014 11:48:07 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Missing updates at few places for row level security\"\n },\n {\n \"msg_contents\": \"Amit,\\n\\n* Amit Kapila (amit.kapila16@gmail.com) wrote:\\n> There is a new column added in pg_authid (rolbypassrls)\\n> and the updation for same is missed in below places:\\n> \\n> a. System catalog page for pg_authid\\n> http://www.postgresql.org/docs/devel/static/catalog-pg-authid.html\\n\\nYup, thanks, will fix.\\n\\n> b. Do we want to add this new column in pg_shadow view?\\n\\nThis was intentionally not done as I had really viewed pg_user, pg_group\\nand pg_shadow as deprecated and only for backwards-compatibility.\\nThat's certainly why those views were added originally. On the other\\nhand, I do see that 'userepl' was added, so we've been keeping it\\nupdated.\\n\\nI am inclined to suggest that we actually mark pg_user, pg_group, and\\npg_shadow as deprecated and planned for removal. We can't simply remove\\nthem as we haven't actually said that but I don't think we should\\ncontinue to carry these pre-8.1 views around forever. Still, until we\\ndo that, we should keep them updated.\\n\\nI'll make these changes in the next couple of days (after the role\\nattribute bitmask patch, as that touches a lot of the same places).\\n\\n\\tThanks,\\n\\n\\t\\tStephen\",\n \"msg_date\": \"Tue, 23 Dec 2014 07:11:32 -0500\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Missing updates at few places for row level security\"\n },\n {\n \"msg_contents\": \"Amit,\\n\\n* Amit Kapila (amit.kapila16@gmail.com) wrote:\\n> There is a new column added in pg_authid (rolbypassrls)\\n> and the updation for same is missed in below places:\\n> \\n> a. System catalog page for pg_authid\\n> http://www.postgresql.org/docs/devel/static/catalog-pg-authid.html\\n> b. Do we want to add this new column in pg_shadow view?\\n\\nThanks for this. Pushed with a few additional changes (included it in\\nboth pg_user and pg_shadow, few other doc updates, etc).\\n\\n\\tThanks again!\\n\\n\\t\\tStephen\",\n \"msg_date\": \"Wed, 28 Jan 2015 21:48:06 -0500\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Missing updates at few places for row level security\"\n }\n]","isConversational":false}}}],"truncated":false,"partial":false},"paginationData":{"pageIndex":399,"numItemsPerPage":100,"numTotalItems":58371,"offset":39900,"length":100}},"jwt":"eyJhbGciOiJFZERTQSIsImtpZCI6IjVHZDBvd0g5MTM2eDZjc1FvbE1zcktNWUZoRVFoUm5PVVVybEpjOUhaUEEifQ.eyJyZWFkIjp0cnVlLCJwZXJtaXNzaW9ucyI6eyJyZXBvLmNvbnRlbnQucmVhZCI6dHJ1ZX0sImlhdCI6MTc4NTYwNzc5NywianRpIjoiMThjZmVhNTctOTc3YS00MmYxLWJlNTUtZTRlNzJiMmMxZjFjIiwic3ViIjoiL2RhdGFzZXRzL3dhbnNoZW5sL3Bnc3FsLWhhY2tlcnMiLCJleHAiOjE3ODU2MTEzOTcsImlzcyI6Imh0dHBzOi8vaHVnZ2luZ2ZhY2UuY28ifQ.CkjMlys0DAR2CCzPhau108so_YPrKyPDjlKLZDWnfY1uzW1ayyOAKirSbVe81o4rekFXWrwsAheCbqlAAK_EDA","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": "There doesn't seem to be documentation on *= (or search isn't finding it). Is this intentional?\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo ma... |
[
{
"msg_contents": "On Wed, Nov 26, 2014 at 4:13 PM, Jeff Janes <jeff.janes@gmail.com> wrote:\n\n> On Tue, Nov 25, 2014 at 9:30 PM, Jeff Janes <jeff.janes@gmail.com> wrote:\n>\n>> Using both 9.2.9 and 9_2_STABLE 9b468bcec15f1ea7433d4, I get a fairly\n>> reproducible startup failure.\n>>\n>> What I was doing is r... |
[
{
"msg_contents": "Hi all,\n\nwhile working on the patch decreasing amount of memory consumed by\narray_agg [1], I've ran into some strange OOM issues. Reproducing them\nusing the attached SQL script is rather simple.\n\n [1] https://commitfest.postgresql.org/action/patch_view?id=1652\n\nAt first I thought the... |
[
{
"msg_contents": "Test configuration:\nHardware:\n\t4P intel server, 60 core 120 hard thread.\n\tMemory:512G\n\tSSD:2.4T\n\nPG:\nmax_connections = 160\t\t\t# (change requires restart)\nshared_buffers = 32GB\t\t\t\nwork_mem = 128MB\t\t\t\t\nmaintenance_work_mem = 32MB\t\t\t\nbgwriter_delay = 100ms\t\t\t# 10-100... |
[
{
"msg_contents": "Hello,\n\n \n\nI'm using PostgreSQL .9.2.8 on Windows from a .NET application using\nNpgsql.\n\nI'm working in the Radiology Information System field.\n\n \n\nWe have thousands of users against a big accounting database.\n\nWe're using the SERIALIZABLE isolation level to ensure data consisten... |
[
{
"msg_contents": "Hello,О©╫How can I get all depend objects for oid of object? For example, I have oid of db db_id and I want to get all oids of namespaces, which contains this db. Thank you!О©╫-- О©╫ О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫, О©╫О©╫О©╫О©╫О©╫О©╫О©╫ О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫\n",
"msg_date": "Tue, 02 Dec 2... |
[
{
"msg_contents": "Hi,\n\npg_stat_statement's query normalization fails when temporary tables are\nused in a query. That's because JumbleRangeTable() uses the relid in the\nRTE to build the query fingerprint. I think in this case pgss from 9.1\nactually would do better than 9.2+ as the hash lookup previously di... |
[
{
"msg_contents": "Hi,\n\nAs of now postgres is reporting only \"really\" important variables, but\namong them\nthere are also not so important, like 'application_name'. The only reason\nto report\nit, was: \"help pgbouncer, and perhaps other connection poolers\".\nChange was introduced by commit: 59ed94ad0c9f7... |
[
{
"msg_contents": "Hi all,\n\nSince commit 5d93ce2d, the output of xmllint is checked by passing\n--valid to it. Isn't that a regression with what we were doing for\npre-9.4 versions? For example, with 9.4 and older versions it is\npossible to compile man pages even if the xml spec is not entirely\nvalid when u... |
[
{
"msg_contents": "Fix whitespace\n\nBranch\n------\nmaster\n\nDetails\n-------\nhttp://git.postgresql.org/pg/commitdiff/bc2f43eaa4cef83798692e57e9d0364f0bcce021\n\nModified Files\n--------------\nsrc/test/modules/dummy_seclabel/dummy_seclabel--1.0.sql | 1 -\n1 file changed, 1 deletion(-)\n\n\n-- \nSent via ... |
[
{
"msg_contents": "> On Tue, Nov 25, 2014 at 3:44 AM, Kouhei Kaigai <kaigai@ak.jp.nec.com> wrote:\r\n> > Today, I had a talk with Hanada-san to clarify which can be a common\r\n> > portion of them and how to implement it. Then, we concluded both of\r\n> > features can be shared most of the infrastructure.\r\n> ... |
[
{
"msg_contents": "Keep track of transaction commit timestamps\n\nTransactions can now set their commit timestamp directly as they commit,\nor an external transaction commit timestamp can be fed from an outside\nsystem using the new function TransactionTreeSetCommitTsData(). This\ndata is crash-safe, and trunc... |
[
{
"msg_contents": "-----BEGIN PGP SIGNED MESSAGE-----\nHash: SHA1\n\nI have a need to provide a way to change a table's primary key\ncolumns, in the possible presence of foreign keys, and with concurrent\nuse (as much as possible).\n\nThe best way I have come up with is roughly:\na. create the to-be-new-pk inde... |
[
{
"msg_contents": "Folks,\n\nI've been trying out 9.5-to-be's PostgreSQL FDW, and I noticed that it\ndoesn't seem to handle enum types. Would this be a trivial fix?\n\nRepro:\n Create an enum type in a 9.3 instance.\n Use the type in a table.\n Create a foreign server pointing to this instance.\n E... |
[
{
"msg_contents": "Hi,\n\nThe recent discussion about pipelining in the jodbc driver prompted me to look \nat what it would take for libpq.\n\nI have a proof of concept patch working. The results are even more promising \nthan I expected.\n\nWhile it's true that many applications and frameworks won't easily be... |
[
{
"msg_contents": "As per discussion on another thread related to using\ncustom scan nodes for prototype of parallel sequence scan,\nI have developed the same, but directly by adding\nnew nodes for parallel sequence scan. There might be\nsome advantages for developing this as a contrib\nmodule by using custom ... |
[
{
"msg_contents": "After due consideration, the core committee has agreed it's time to\npush this puppy out the door. We will wrap 9.4.0 on Monday Dec 15\nfor public announcement Thursday 18th.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make chan... |
[
{
"msg_contents": "While perusing checkInsertTargets I noticed that it allows duplicated\nassignments to the same member of a composite:\n\nalvherre=# create type f as (a int, b int);\nCREATE TYPE\nalvherre=# create table t (col f);\nCREATE TABLE\nalvherre=# insert into t (col.a, col.b, col.a) values (42, 43, 4... |
[
{
"msg_contents": "Attached patch makes minor modification to the pg_stat_all_tables\ndocumentation. This clarifies that pg_stat_*_tables.n_tup_upd includes\nHOT updates.\n\n-- \nPeter Geoghegan\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:... |
[
{
"msg_contents": "Hi,\n\nWe've recently observed a case where, after a promotion, a postgres\nserver suddenly started to archive a large amount of old WAL.\n\nAfter some digging the problem is this:\n\npg_basebackup -X creates files in pg_xlog/ without creating the\ncorresponding .done file. Note that walrecei... |
[
{
"msg_contents": "Hi all,\n\nWhile reading the code in this area this morning, I noticed that\nwal_log_hints and track_commit_timestamp are not mentioned in the desc\nroutine of XLOG_CHANGE_PARAMETER. Also, it is not mentioned in\npostgresql.conf.sample that a value update of wal_log_hints requires a\nsystem r... |
[
{
"msg_contents": "All,\n\nI've noticed that 'check-world' fails for dummy_seclabel after a 'clean'.\nI believe that in commit da34731, the EXTRA_CLEAN statement should have\nbeen removed from 'src/test/modules/dummy_seclabel/Makefile' as well.\n\nAttached is a proposed patch that addresses this issue.\n\nThank... |
[
{
"msg_contents": "Hackers,\n\nThis is not a complete enough report for a diagnosis. I'm posting it\nhere just in case someone else sees something like it, and having an\nadditional report will help figure out the underlying issue.\n\n* 700GB database with around 5,000 writes per second\n* 8 replicas handling ... |
[
{
"msg_contents": "Hi,\n\nback when we were discussing the hashjoin patches (now committed),\nRobert proposed that maybe it'd be a good idea to sometimes increase the\nnumber of tuples per bucket instead of batching.\n\nThat is, while initially sizing the hash table - if the hash table with\nenough buckets to s... |
[
{
"msg_contents": "Hi,\n\nI've been reading the FSM README file lately\n(src/backend/storage/freespace/README), and I'm puzzled by one of the graph\n(the binary tree structure of an FSM file). Here it is:\n\n 4\n 4 2\n3 4 0 2 <- This level represents heap pages\n\nShouldn't the last line be:\n4 3 ... |
[
{
"msg_contents": "This patch implements the first wiki/Todo Configuration Files item\n\"Consider normalizing fractions in postgresql.conf, perhaps using '%'\".\n\nThe \"Fractions in GUC variables\" discussion is here.\n\nhttp://www.postgresql.org/message-id/467132CF.9020501@enterprisedb.com\n\nThis patch imple... |
[
{
"msg_contents": "Hi all,\n\nI have been through a certain number of patches and marked the ones\nthat needed some love from their authors as returned with feedback (at\nleast the ones marked as such in the CF app), but you may have noticed\nthat :)\n\nThis email is a call to move on and close soon the CF curr... |
[
{
"msg_contents": "> On Thu, Dec 4, 2014 at 8:37 PM, Michael Paquier wrote\n> I pondered something that Andres mentioned upthread: we may not do the\n>compression in WAL record only for blocks, but also at record level. Hence\n>joining the two ideas together I think that we should definitely have\n>a different\... |
[
{
"msg_contents": "Hello all,\n\n\nWe experienced some casting issues with domains. We experienced the\nproblem while querying the information_schema btw, but here is a simpler\ntest case :\n\npostgres=# create table test1 (a text);\nCREATE TABLE\npostgres=# insert into test1 select generate_series(1,100000);\n... |
[
{
"msg_contents": "Hi,\n\nThe pg_recvlogical docs was rewritten but someone forgot to tweak the\nhelp description. It is a bit late in the 9.4 cycle but let be consistent.\n\n\nRegards,\n\n\n-- \n Euler Taveira Timbira - http://www.timbira.com.br/\n PostgreSQL: Consultoria, Desenvolvimento... |
[
{
"msg_contents": "Hello.\n\nAs of now, the only way to restore database options and ACLs is to use \npg_dumpall without the globals options. The often recommended pg_dumpall -g + \nindividual dumps of the target databases doesn't restore those.\n\nSince pg_dump/pg_restore offer the ability to create the databa... |
[
{
"msg_contents": "In the past, building under MinGW produced so many warnings that I never\nbothered to read them.\n\nNow most of them have been removed, so the ones that are left might be\nworth reporting.\n\nUsing gcc.exe (GCC) 4.6.2 on REL9_4_STABLE\neadd80c08ddfc485db84b9af7cca54a0d50ebe6d I get:\n\nmingwc... |
[
{
"msg_contents": "The call:\n\n\treserveFromBuffer(&buffer, sizeof(VARHDRSZ))\n\nis assuming that the size of varlena header is the same\nsize as the type used to return that size, which happens\nto be so, but someone could easily change that macro\nto:\n\n\t#define VARHDRSZ ((int64) sizeof(int32))\n\nAnd you'... |
[
{
"msg_contents": "Let's take another crack at moving stuff out of contrib. Nobody likes\ncontrib. The task is only finding something that most people like better.\n\nLast time this was attempted, the discussion got lost in exactly which\nextensions are worthy enough to be considered official or something lik... |
[
{
"msg_contents": "Hi all,\n\nAs mentioned in $subject, we are missing an entry in pg_xlogdump's .gitignore.\nPatch attached.\nRegards,\n-- \nMichael\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-ha... |
[
{
"msg_contents": "So I've been updating my very old patch to allow logical and physical\ncolumn reordering. Here's a WIP first cut for examination. There are\nplenty of rough edges here; most importantly there is no UI at all for\ncolumn reordering other than direct UPDATEs of pg_attribute, which most\nlikel... |
[
{
"msg_contents": "Is there any particular reason we don't allow comparing char and varchar arrays? If not I'll submit a patch.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresq... |
[
{
"msg_contents": "Hi all,\n\nA little typo in vacuumdb tests.\n\nRegards,\n\n-- \nFabrízio de Royes Mello\nConsultoria/Coaching PostgreSQL\n>> Timbira: http://www.timbira.com.br\n>> Blog: http://fabriziomello.github.io\n>> Linkedin: http://br.linkedin.com/in/fabriziomello\n>> Twitter: http://twitter.com/fabriz... |
[
{
"msg_contents": "At line 1787 of outfuncs.c, the line:\n\n\tWRITE_UINT_FIELD(reltablespace)\n\nshould probably say\n\n\tWRITE_OID_FIELD(reltablespace)\n\nsince that variable is of type Oid, not uint32.\nGranted, these two macros are interchangeable,\nbut they won't be if we ever move to 64-bit Oids.\n\nmark\n... |
[
{
"msg_contents": "Hello,\n\nAttached is a basic implementation of TABLESAMPLE clause. It's SQL \nstandard clause and couple of people tried to submit it before so I \nthink I don't need to explain in length what it does - basically returns \n\"random\" sample of a table using a specified sampling method.\n\nI ... |
[
{
"msg_contents": "initdb currently does\n\n PG_CMD_PUTS(\"ANALYZE;\\nVACUUM FULL;\\nVACUUM FREEZE;\\n\");\n\nFREEZE is now part of FULL, so this seems redundant. Also, ANALYZE can\nbe run as part of VACUUM. So this could be\n\n PG_CMD_PUTS(\"VACUUM FULL ANALYZE;\\n\");\n\nThere has been some concerns a... |
[
{
"msg_contents": "Hi,\n\nA customer recently reported getting \"backup_label contains data\ninconsistent with control file\" after taking a basebackup from a standby\nand starting it with a typo in primary_conninfo.\n\nWhen starting postgres from a basebackup StartupXLOG() has the follow\ncode to deal with bac... |
[
{
"msg_contents": "I have heard repeated concerns about the commitfest process in the past\nfew months. The fact we have been in a continual commitfest since\nAugust also is concerning.\n\nI think we are reaching the limits on how much we can do with our\nexisting commitfest structure, and it might be time to ... |
[
{
"msg_contents": "Hello, we found (and maybe fixed) two wrong behavior related to\ninheritance and FOR UPDATE. This report is about one of them.\n\nThis behavior could be observed on master and back to 8.4 but 8.4\nbehaves a bit more funny. I haven't checked on 8.3.\n\nThis issue is that some query returns dup... |
[
{
"msg_contents": "Hello, this is about the second issue.\n\nSELECT FROM <inheritance parent> WHERE <cond> FOR UPDATE may\nreturn results which does not match the <cond>. The following\nsteps will reproduce the problematic behavior (A and B are\nindividual sessions) on master and back to 9.0 but 8.4 gives\ncorr... |
[
{
"msg_contents": "I found a few places in the code where a variable of\ntype Oid is printed using \"%d\" rather than \"%u\" and\nchanged them in the attached patch.\n\nIn src/backend/replication/logical/reorderbuffer.c,\ncirca line 2494, chunk_seq is of type Oid, but\nent->last_chunk_seq is of type int32, lead... |
[
{
"msg_contents": "When using a vpath build pg_regress writes the processed input/*.source\nfiles into the *source* tree, which isn't supposed to happen.\n\nThis appears to be a thinko introduced in this patch:\ne3fc4a97bc8ee82a78605b5ffe79bd4cf3c6213b\n\nThe attached patch fixes it.\n\n\n\n-- \nSent via pgsql-... |
[
{
"msg_contents": "There have been a few threads on the behavior of WAL archiving, after a \nstandby server is promoted [1] [2]. In short, it doesn't work as you \nmight expect. The standby will start archiving after it's promoted, but \nit will not archive files that were replicated from the old master via \ns... |
[
{
"msg_contents": "Hi,\n\nI'd like to include pg_rewind in contrib. I originally wrote it as an \nexternal project so that I could quickly get it working with the \nexisting versions, and because I didn't feel it was quite ready for \nproduction use yet. Now, with the WAL format changes in master, it is a \nlot... |
[
{
"msg_contents": "Attached is a patch that adds two new concepts: parallel mode, and\nparallel contexts. The idea of this code is to provide a framework\nfor specific parallel things that you want to do, such as parallel\nsequential scan or parallel sort. When you're in parallel mode,\ncertain operations - l... |
[
{
"msg_contents": "Just got a report on IRC of a bug in the array version of\npercentile_cont; if two of the requested percentiles were between the\nsame pair of input rows, the result could be wrong or an error would\nbe generated.\n\ne.g.\n\nselect percentile_cont(array[0.4,0.6]) within group (order by gs)\n ... |
[
{
"msg_contents": "Hi all\n\nWhile working on BDR, I've run into a situation that I think highlights\na limitation of the dynamic bgworker API that should be fixed.\nSpecifically, when the postmaster crashes and restarts shared memory\ngets cleared but dynamic bgworkers don't get unregistered, and that's a\nmes... |
[
{
"msg_contents": "Currently pgbench -f (run custom script) executes vacuum against\npgbench_* tables before stating bench marking if -n (or --no-vacuum)\nis not specified. If those tables do not exist, pgbench fails. To\nprevent this, -n must be specified. For me this behavior seems insane\nbecause \"-f\" does... |
[
{
"msg_contents": "Hi\n\na Orafce package on pgFoundry is obsolete - we migrated to github few years\nago. Please, can somebody modify a description about this migration? Or\ndrop this project there.\n\nRegards\n\nPavel\n\nHia Orafce package on pgFoundry is obsolete - we migrated to github few years ago. Please... |
[
{
"msg_contents": "In commit 2c03216d831160bedd72d45f712601b6f7d03f1c, the\nfollowing define occurs twice in src/include/access/xlogrecord.h:\n\n#define SizeOfXLogRecordDataHeaderLong (sizeof(uint8) + sizeof(uint32))\n\nIt is no big deal, as the definitions don't contradict each other. We\ncould probably live ... |
[
{
"msg_contents": "Hi all,\n\nThis week, we heard about a user willing to use a custom timestamp\nformat across a set of services to improve the debugability of the\nwhole set, Postgres being one of them. Unfortunately datestyle does\nnot take into account the logs. Would it be worth adding a new GUC\nable to c... |
[
{
"msg_contents": "Hi all,\n\nLooking at the CF app as of today, there is the following status for\npending patches:\n- Needs Review: 18\n- Waiting on Author: 1\n- Ready for Committer: 8\n\nAnd the next coming fest that should begin on Monday has this status:\n- Needs Review: 34\n- Waiting on Author: 3\n- Ready... |
[
{
"msg_contents": "The following comment above #define PAGES_PER_CHUNK in tibbitmap.c appears\nto be incorrect:\n\n\"But we\n * also want PAGES_PER_CHUNK to be a power of 2 to avoid expensive integer\n * remainder operations. So, define it like this:\"\n\nI don't quite follow this as it does nothing of the kin... |
[
{
"msg_contents": "Hi,\n\nThis week I had a problem where I wanted to drop only the privileges a \ncertain role had in the system, while keeping all the objects. I \ncouldn't figure out a reasonable way to do that, so I've attached a \npatch for this to this email. Please consider it for inclusion into \n9.5.... |
[
{
"msg_contents": "Hi all,\n\nLet's begin the commit fest of December:\nhttps://commitfest.postgresql.org/action/commitfest_view?id=25\n\nLooking at the stats at the top of the above page, there are currently\n56 patches to be dealt with:\n- Needs Review: 48\n- Waiting on Author: 5\n- Ready for Committer: 3\nIn... |
[
{
"msg_contents": "Hello,\n\nwe've made few helper functions for making logical replication easier, I \nbundled it into contrib module as this is mainly for discussion at this \ntime (I don't expect this to get committed any time soon, but it is good \nway to iron out protocol, etc).\n\nI created sample logical... |
[
{
"msg_contents": "Hi,\n\nFind attached that does:\n\n- * mt_info is laid out in the following fashion:\n+ * bt_info is laid out in the following fashion:\n<snip-comment>\n uint8 bt_info;\n} BrinTuple;\n\nThanks,\nAmit\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgre... |
[
{
"msg_contents": "This attached patch adds a code example of custom-scan interface.\r\n\r\nThis custom-scan provider (\"ctidscan\") performs almost same as\r\nbuilt-in SeqScan plan, but can produce same results with less\r\npage scan in case when qualifier of relation scan has inequality\r\noperators (>, >=, <... |
[
{
"msg_contents": "Hello, this is the 2nd session of 'intoroducing parallelism using\npostgres_fdw'.\n\nThe two patch attached are as following,\n\n- 0001-Async-exec-of-postgres_fdw.patch\n Main patch, which includes all functions.\n\n- 0002-rename-PGConn-variable.patch\n Renaming the variable conn for readab... |
[
{
"msg_contents": "Hi,\n\na fellow Debian Developer found a minor glitch in\nsrc/timezone/localtime.c, where binary search is used. Now I don't\nthink there is an actual problem (unless there's > 2^30 timezones),\nbut it would at least make sense to mark the code as okayish so that\npeople running code scanners... |
[
{
"msg_contents": "The CLOBBER_CACHE_ALWAYS buildfarm members occasionally fail in\ncontrib/test_decoding with\n\nTRAP: FailedAssertion(\"!(((bool)((relation)->rd_refcnt == 0)))\", File: \"relcache.c\", Line: 1981)\n\nfor example here:\nhttp://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=markhor&dt=2014-12-1... |
[
{
"msg_contents": "Currently, WALReceiver writes and fsyncs data it receives. Clearly,\nwhile we are waiting for an fsync we aren't doing any other useful\nwork.\n\nFollowing patch starts WALWriter during recovery and makes it\nresponsible for fsyncing data, allowing WALReceiver to progress other\nuseful action... |
[
{
"msg_contents": "Currently, it doesn't.\n\nThis patch is a WIP version of doing that, but only currently attempts\nto do this in the WALSender.\n\nObjective is to allow cascaded logical replication.\n\nVery WIP, but here for comments.\n\n-- \n Simon Riggs http://www.2ndQuadrant.com/\n Postgr... |
[
{
"msg_contents": "https://commitfest.postgresql.org/action/patch_view?id=1661 (apologies for not replying to the thread; I can't find it in my inbox)\n\nPatch applies and passes make check. Code formatting looks good.\n\nThe regression test partially tests this. It does not cover 2PC, nor does it test rolling ... |
[
{
"msg_contents": "config.cache is created when you pass -C to configure, which speeds it up considerably (3.5s vs 16.5 on my laptop). It would be nice to just ignore the cache file it generates.\n\nOriginally this patch also ignored the regression output files, until I realized why that was a bad idea. Add a c... |
[
{
"msg_contents": "Hello Friends,\n\nCan you please tell me the how can I track the which bugs are fixed in which release and when they will be fixed,\nIf I want to track the analysis and status of the bug raised on Postgres. Can I get this information.\n\n From last few days we are struggling with following is... |
[
{
"msg_contents": "Hi all,\n\nWhile looking at the Windows docs and the MSVC scripts, I noticed that\nthe following run modes of vcregress.pl are either not mentioned in\nthe WIndows installation docs or in the help output of vcregress.pl:\n- ecpgcheck\n- isolationcheck\n- upgradecheck\nAttached is a patch fixi... |
[
{
"msg_contents": "\nOn 16 Dec 2014 7:43 am, Andres Freund <andres@2ndquadrant.com> wrote:\n>\n> On 2014-12-15 21:18:40 -0800, Josh Berkus wrote: \n> > On 12/15/2014 07:34 PM, Andres Freund wrote: \n> > > On 2014-12-15 16:14:30 -0800, Josh Berkus wrote: \n> > >> Read the thread on this list where I suggested cr... |
[
{
"msg_contents": "Hi all,\n\nAs mentioned in $subject, I noticed that those automatically-generated\nfiles are not ignored in the tree when running vcregress on Windows,\nwe do ignore their .sh version though. I think that it would be good\nto back-patch the patch attached to prevent the inclusion of those\nfi... |
[
{
"msg_contents": "Folks,\n\nWhile noodling with some weighted statistics\n<https://github.com/davidfetter/weighted_stats>, I noticed I was\nhaving to jump through a lot of hoops because of all the private\nmethods in numeric.c, especially NumericVar. Would there be some\nmajor objection to exposing NumericVar... |
[
{
"msg_contents": "Hello,\n\nThe comment before declaration of XLogRecordBlockHeader says\n\n>* 'data_length' is the length of the payload data associated with this,\n> * and includes the possible full-page image, and rmgr-specific data. It\n\nIIUC, data_length does not include associated full page image length... |
[
{
"msg_contents": "Hi Andrew,\n\nDid you have a chance to review this?\n\nAlvaro Herrera wrote:\n> Andrew Dunstan wrote:\n> > \n> > On 11/29/2014 10:09 PM, Alvaro Herrera wrote:\n\n> > >Anyway I just pushed this src/test/modules/ patch, which has\n> > >implications for buildfarm: these new test modules are not ... |
[
{
"msg_contents": "> For some datatypes, the compress method might be useful even if the leaf\n > type is the same as the column type. For example, you could allow\n > indexing text datums larger than the page size, with a compress function\n > that just truncates the input.\n\nAgree, and patch allows to use co... |
[
{
"msg_contents": "Folks,\n\nI've noticed that psql's \\c function handles service= requests in a\nway that I can only characterize as broken.\n\nThis came up in the context of connecting to a cloud hosting service\nnamed after warriors or a river or something, whose default hostnames\nare long, confusing, and... |
[
{
"msg_contents": "I just happened to look into bufmgr.c for the first time in awhile, and\nnoticed the privaterefcount-is-no-longer-a-simple-array stuff. It doesn't\nlook too well thought out to me. In particular, PinBuffer_Locked calls\nGetPrivateRefCountEntry while holding a buffer-header spinlock. That\n... |
[
{
"msg_contents": "KaiGai, David Rowley and myself have all made mention of various ways\nwe could optimize aggregates.\n\nFollowing WIP patch adds an extra function called a \"combining\nfunction\", that is intended to allow the user to specify a\nsemantically correct way of breaking down an aggregate into mul... |
[
{
"msg_contents": "Hi there,\n\nI was doing some translation on postgres.po and found a string which\nlooks mistaken.\n\n#: libpq/auth.c:1593\n#, fuzzy, c-format\nmsgid \"could not to look up local user ID %ld: %s\"\n\nIt looks like there is an extra *to* there , so the string should be:\n\n\"could not look up ... |
[
{
"msg_contents": "At the end of archive recovery, we copy the last segment from the old \ntimeline, to initialize the first segment on the new timeline. For \nexample, if the timeline switch happens in the middle of WAL segment \n000000010000000000000005, the whole 000000010000000000000005 segment is \ncopied ... |
[
{
"msg_contents": "Hi,\n\nThe attached patch just removes one bad-looking blank line in the\ncomments at the top of a function in explain.c.\n\nThanks,\n\nBest regards,\nEtsuro Fujita\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://ww... |
[
{
"msg_contents": "Hi all,\n\nToday I spent a bit of time looking at the activity of autovacuum for\none table particularly bloated. log_autovacuum_min_duration was\nenabled and set to a high value but even with that I had to deal with\nsome spam from the jobs of other tables. It would be cool to have the\nposs... |
[
{
"msg_contents": "Change how first WAL segment on new timeline after promotion is created.\n\nTwo changes:\n\n1. When copying a WAL segment from old timeline to create the first segment\non the new timeline, only copy up to the point where the timeline switch\nhappens, and zero-fill the rest. This avoids corne... |
[
{
"msg_contents": "One of our servers is currently running on postgres 9.2 using the 9.2.9-1.pgdg70+1 packages from pgdg.\n\nAfter an apt update this morning which brought in the libpq5 package version 9.4.0-1.pgdg70+1, connections to the database started failing with SSL errors logged on the server:\n\n [unk... |
[
{
"msg_contents": "When streaming replication was introduced in 9.0, we started to recycle \nold WAL segments in archive recovery, like we do during normal \noperation. The WAL segments are recycled on the current timeline. There \nis no guarantee that they are useful, if the current timeline changes, \nbecause... |
[
{
"msg_contents": "Hi,\n\nWhen debugging lwlock issues I found PRINT_LWDEBUG/LOG_LWDEBUG rather\npainful to use because of the amount of elog contexts/statements\nemitted. Given the number of lwlock acquirations that's just not doable.\n\nTo solve that during development I've solved that by basically\nreplacing... |
[
{
"msg_contents": "Right now, both table and domain constraints share the same\nclassification in the ObjType enum, OBJECT_CONSTRAINT. For my patch to\nimplement improved object drop reporting in event triggers, it is\nnecessary to separate them, as per in the attach patch.\n\nOne emerging side-effect is that ... |
[
{
"msg_contents": "Buildfarm members magpie, treepie and fulmar went absent on 2014-10-29. Since\nreturning on 2014-11-16, they have consistently failed with 'initdb: invalid\nlocale name \"cs_CZ.WIN-1250\"'. No commits in that period readily explain a\nregression in this area. Did the underlying system conf... |
[
{
"msg_contents": "PostgreSQL 9.2 and later support Python 3.3 and Python 3.4; PostgreSQL 9.1 and\n9.0 do not. Buildfarm member raccoon upgraded from Python 2.7 to Python 3.3\nfor its 2013-08-18 run, and it has failed for REL9_1_STABLE and REL9_0_STABLE\never since. Please omit --with-python for those branche... |
[
{
"msg_contents": "Hi all,\n\nI work with some customer that have databases with a lot of schemas and\nsometimes we need to run manual VACUUM in one schema, and would be nice to\nhave a new option to run vacuum in relations from a specific schema.\n\nThe new syntax could be:\n\nVACUUM [ FULL ] [ FREEZE ] [ VERB... |
[
{
"msg_contents": "Hi,\n\nI think ExplainModifyTarget should show the parent of the inheritance\ntree in multi-target-table cases, as described there, but noticed that\nit doesn't always work like that. Here is an example.\n\npostgres=# create table parent (a int check (a < 0) no inherit);\nCREATE TABLE\npostg... |
[
{
"msg_contents": "Hi all,\n\nI have been though the patches of the current CF, looking at their\nrelated threads and updating each patch status if needed. After one\nweek in this CF, we have done progress on many patches, more than 2/3\nof them getting some comments, reviews and/or refreshed versions.\n\nLooki... |
[
{
"msg_contents": "Peter Geoghegan suggested [1] moving rbtree.c to src/backend/lib, which \nI think makes a lot of sense. Now that we have several other general \npurpose data structures in src/backend/lib (linked lists, a binary heap, \nand a pairing heap), rbtree.c would definitely be better placed in \nsrc/... |
[
{
"msg_contents": "The following bug has been logged on the website:\n\nBug reference: 12320\nLogged by: Aaron Botsis\nEmail address: postgres@bt-r.com\nPostgreSQL version: 9.4.0\nOperating system: Mac OSX\nDescription: \n\npostgres=# create table ct (data jsonb);\nCREATE TABLE\npost... |
[
{
"msg_contents": "Hi,\n\nI've played with trying to find out which minimal set of files I need\nfrom the old version to make pg_upgrade work. Interestingly, this\nincludes the good old postmaster binary:\n\n$ sudo -u postgres pgsql/bin/pg_upgrade -b /var/tmp/pgsql/bin/ -B /usr/lib/postgresql/9.5/bin/ -d /etc/p... |
[
{
"msg_contents": "Currently the buildfarm animal crake (my development instance) is \nrunning the bin check, but not any other animal. These tests still take \nfor too long, not least because each set of tests requires a separate \ninstall.\n\nI have complained about this before, but we don't seem to have made... |
[
{
"msg_contents": "There is a new column added in pg_authid (rolbypassrls)\nand the updation for same is missed in below places:\n\na. System catalog page for pg_authid\nhttp://www.postgresql.org/docs/devel/static/catalog-pg-authid.html\nb. Do we want to add this new column in pg_shadow view?\n\nWith Regards,\n... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.