\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Thanks a lot for hacking on the patch. I'm really happy with the\\ndirection you took for inheritance_planner, as it allows UPDATE/DELETE to\\nuse partition pruning.\\n\\nOn 2019/03/29 7:38, Tom Lane wrote:\\n> I've been hacking on this pretty hard over the last couple days,\\n> because I really didn't like the contortions you'd made to allow\\n> inheritance_planner to call expand_inherited_rtentry in a completely\\n> different context than the regular code path did. I eventually got\\n> rid of that\\n\\nGood riddance.\\n\\n> by having inheritance_planner run one cycle of planning\\n> the query as if it were a SELECT, and extracting the list of unpruned\\n> children from that.\\n\\nThis is somewhat like my earlier patch that we decided to not to pursue,\\nminus all the hackery within query_planner() that was in that patch, which\\nis great.\\n\\n(I can't find the link, but David Rowley had posted a patch for allowing\\nUPDATE/DELETE to use partition pruning in the late stages of PG 11\\ndevelopment, which had taken a similar approach.)\\n\\n> I had to rearrange the generation of the final\\n> rtable a bit to make that work, but I think that inheritance_planner\\n> winds up somewhat cleaner and safer this way; it's making (slightly)\\n> fewer assumptions about how much the results of planning the child\\n> queries can vary.\\n>\\n> Perhaps somebody will object that that's one more planning pass than\\n> we had before, but I'm not very concerned, because\\n> (a) at least for partitioned tables that we can prune successfully,\\n> this should still be better than v11, since we avoid the planning\\n> passes for pruned children.\\n\\nCertainly. Note that previously we'd always scan *all* hash partitions\\nfor UPDATE and DELETE queries, because constraint exclusion can't exclude\\nhash partitions due to the shape of their partition constraint.\\n\\nI ran my usual benchmark with up to 8192 partitions.\\n\\nN: 2..8192\\n\\ncreate table rt (a int, b int, c int) partition by range (a);\\nselect 'create table rt' || x::text || ' partition of rt for values from\\n(' || (x)::text || ') to (' || (x+1)::text || ');' from generate_series(1,\\nN) x;\\n\\\\gexec\\n\\nupdate.sql:\\n\\n\\\\set param random(1, N)\\nupdate rt set a = 0 where a = :param;\\n\\npgbench -n -T 120 -f select.sql\\n\\nnparts v38 HEAD\\n====== ==== ====\\n2\\t 2971 2969\\n8\\t 2980 1949\\n32\\t 2955 733\\n128\\t 2946 145\\n512\\t 2924 11\\n1024\\t 2986 3\\n4096\\t 2702 0\\n8192\\t 2531 OOM\\n\\nObviously, you'll get similar numbers with hash or list partitioning.\\n\\n> (b) inheritance_planner is horridly inefficient anyway, in that it\\n> has to run a near-duplicate planning pass for each child table.\\n> If we're concerned about its cost, we should be working to get rid of\\n> the function altogether, as per [1]. In the meantime, I do not want\\n> to contort other code to make life easier for inheritance_planner.\\n\\nAgreed.\\n\\n> There's still some loose ends:\\n> \\n> 1. I don't like 0003 much, and omitted it from the attached.\\n> I think that what we ought to be doing instead is not having holes\\n> in the rel_parts[] arrays to begin with, ie they should only include\\n> the unpruned partitions. If we are actually encoding any important\\n> information in those array positions, I suspect that is broken anyway\\n> in view of 898e5e329: we can't assume that the association of child\\n> rels with particular PartitionDesc slots will hold still from planning\\n> to execution.\\n\\nIt's useful for part_rels array to be indexed in the same way as\\nPartitionDesc. Firstly, because partition pruning code returns the\\nPartitionDesc-defined indexes of unpruned partitions. Second,\\npartitionwise join code decides two partitioned tables as being compatible\\nfor partitionwise joining, then it must join partitions that have\\nidentical *PartitionDesc* indexes, which is what it does by part_rels\\narrays of both sides in one loop.\\n\\nRegarding the impact of 898e5e329 on this, I think it invented\\nPartitionDirectory exactly to avoid PartitionDesc changing under us\\naffecting the planning or execution of a given query. As for\\nPartitionDesc indexes being different between planning and execution, it\\nonly affects PartitionPruneInfo and the commit did make changes to\\nExecCreatePartitionPruningState to remap the old indexes of unpruned\\npartitions in PartitionPruneInfo (as they were during planning) to the new\\nones.\\n\\n> 2. I seriously dislike what's been done in joinrels.c, too. That\\n> really seems like a kluge (and I haven't had time to study it\\n> closely).\\n\\nThose hunks account for the fact that pruned partitions, for which we no\\nlonger create RangeTblEntry and RelOptInfo, may appear on the nullable\\nside of an outer join. We'll need a RelOptInfo holding a dummy path, so\\nthat outer join paths can be created with one side of join being dummy\\nresult path, which are built in the patch by build_dummy_partition_rel().\\n\\n> 3. It's not entirely clear to me why the patch has to touch\\n> execPartition.c. That implies that the planner-to-executor API\\n> changed, but how so, and why is there no comment update clarifying it?\\n\\nThe change is that make_partitionedrel_pruneinfo() no longer adds the OIDs\\nof pruned partitions to the PartitionPruneInfo.relid_map array, a field\\nwhich 898e5e329 added. 898e5e329 had also added an Assert in\\nExecCreatePartitionPruneState, which you're seeing is being changed in the\\npatch. In the case that partition count hasn't changed between planning\\nand execution (*), it asserts that partition OIDs in\\nPartitionPruneInfo.relid_map, which is essentially an exact copy of the\\npartition OIDs in the PartitionDesc that planner retrieved, are same as\\nthe OIDs in the PartitionDesc that executor retrieves.\\n\\n* this seems a bit shaky, because partition count not having changed\\ndoesn't discount the possibility that partitions themselves haven't changed\\n\\n> Given the short amount of time left in this CF, there may not be\\n> time to address the first two points, and I won't necessarily\\n> insist that those be changed before committing. I'd like at least\\n> a comment about point 3 though.\\n> \\n> Attached is updated patch as a single patch --- I didn't think the\\n> division into multiple patches was terribly helpful, due to the\\n> flapping in expected regression results.\\n\\nThanks again for the new patch. I'm reading it now and will send comments\\nlater today if I find something.\\n\\nThanks,\\nAmit\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 29 Mar 2019 12:44:44 +0900\",\n \"msg_from\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 29, 2019 at 3:45 PM, Amit Langote wrote:\\n> Thanks a lot for hacking on the patch. I'm really happy with the direction\\n> you took for inheritance_planner, as it allows UPDATE/DELETE to use\\n> partition pruning.\\n\\nI was astonished by Tom's awesome works and really thanks him.\\n\\n> Certainly. Note that previously we'd always scan *all* hash partitions\\n> for UPDATE and DELETE queries, because constraint exclusion can't exclude\\n> hash partitions due to the shape of their partition constraint.\\n> \\n> I ran my usual benchmark with up to 8192 partitions.\\n> \\n> N: 2..8192\\n> \\n> create table rt (a int, b int, c int) partition by range (a); select 'create\\n> table rt' || x::text || ' partition of rt for values from (' || (x)::text\\n> || ') to (' || (x+1)::text || ');' from generate_series(1,\\n> N) x;\\n> \\\\gexec\\n> \\n> update.sql:\\n> \\n> \\\\set param random(1, N)\\n> update rt set a = 0 where a = :param;\\n> \\n> pgbench -n -T 120 -f select.sql\\n> \\n> nparts v38 HEAD\\n> ====== ==== ====\\n> 2\\t 2971 2969\\n> 8\\t 2980 1949\\n> 32\\t 2955 733\\n> 128\\t 2946 145\\n> 512\\t 2924 11\\n> 1024\\t 2986 3\\n> 4096\\t 2702 0\\n> 8192\\t 2531 OOM\\n> \\n> Obviously, you'll get similar numbers with hash or list partitioning.\\n\\nI also ran the test for hash partitioning for just make sure.\\n\\n\\nN: 2..8192\\n\\ncreate table ht (a int, b int, c int) partition by hash (a);\\nselect 'create table ht' || x::text ||\\n' partition of ht for values with (MODULUS N, REMAINDER || (x)::text || ');'\\nfrom generate_series(0, N-1) x;\\n\\\\gexec\\n\\nupdate.sql:\\n\\n\\\\set param random(1, N * 100)\\nupdate ht set b = b + 1 where a = :param;\\n\\npgbench -n -T 60 -f update.sql\\n\\n\\n[updating one partition]\\nnparts v38 HEAD\\n====== ==== ====\\n0: 10538 10487\\n2: 6942 7028\\n4: 7043 5645\\n8: 6981 3954\\n16: 6932 2440\\n32: 6897 1243\\n64: 6897 309\\n128: 6753 120\\n256: 6727 46\\n512: 6708 12\\n1024: 6063 3\\n2048: 5894 1\\n4096: 5374 OOM\\n8192: 4572 OOM\\n\\n\\nThe performance for hash is also improved, though drop rate of performance with large partitions seems higher than that of range partitioning.\\n\\nThanks\\n--\\nImai Yoshikazu\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 29 Mar 2019 07:00:48 +0000\",\n \"msg_from\": \"\\\"Imai, Yoshikazu\\\" <imai.yoshikazu@jp.fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Here are some comments on v38.\\n\\nOn 2019/03/29 12:44, Amit Langote wrote:\\n> Thanks again for the new patch. I'm reading it now and will send comments\\n> later today if I find something.\\n\\n- Assert(rte->rtekind == RTE_RELATION ||\\n- rte->rtekind == RTE_SUBQUERY);\\n- add_appendrel_other_rels(root, rel, rti);\\n+ if (rte->rtekind == RTE_RELATION)\\n+ expand_inherited_rtentry(root, rel, rte, rti);\\n+ else\\n+ expand_appendrel_subquery(root, rel, rte, rti);\\n\\nWouldn't it be a good idea to keep the Assert?\\n\\n+ * It's possible that the RTIs we just assigned for the child rels in the\\n+ * final rtable are different from where they were in the SELECT query.\\n\\nIn the 2nd sentence, maybe you meant \\\"...from what they were\\\"\\n\\n+ forboth(lc, old_child_rtis, lc2, new_child_rtis)\\n+ {\\n+ int old_child_rti = lfirst_int(lc);\\n+ int new_child_rti = lfirst_int(lc2);\\n+\\n+ if (old_child_rti == new_child_rti)\\n+ continue; /* nothing to do */\\n+\\n+ Assert(old_child_rti > new_child_rti);\\n+\\n+ ChangeVarNodes((Node *) child_appinfos,\\n+ old_child_rti, new_child_rti, 0);\\n+ }\\n\\nThis seems necessary? RTEs of children of the target table should be in\\nthe same position in the final_rtable as they are in the select_rtable.\\nIt seems that they can be added to parse->rtable simply as:\\n\\n orig_rtable_len = list_length(parse->rtable);\\n parse->rtable = list_concat(parse->rtable,\\n list_copy_tail(select_rtable,\\n orig_rtable_len));\\n\\nThat is, after the block of code that plans the query as SELECT.\\n\\n+ * about the childrens' reltargets, they'll be made later\\n\\nShould it be children's?\\n\\n+ /*\\n+ * If the partitioned table has no partitions, treat this as the\\n+ * non-inheritance case.\\n+ */\\n+ if (partdesc->nparts == 0)\\n+ {\\n+ /* XXX wrong? */\\n+ parentrte->inh = false;\\n+ return;\\n+ }\\n\\nAbout the XXX: I think resetting inh flag is unnecessary, so we should\\njust remove the line. If we do that, we can also get rid of the following\\ncode in set_rel_size():\\n\\n else if (rte->relkind == RELKIND_PARTITIONED_TABLE)\\n {\\n /*\\n * A partitioned table without any partitions is marked as\\n * a dummy rel.\\n */\\n set_dummy_rel_pathlist(rel);\\n }\\n\\n\\nFinally, it's not in the patch, but how about visiting\\nget_relation_constraints() for revising this block of code:\\n\\n/*\\n * Append partition predicates, if any.\\n *\\n * For selects, partition pruning uses the parent table's partition bound\\n * descriptor, instead of constraint exclusion which is driven by the\\n * individual partition's partition constraint.\\n */\\nif (enable_partition_pruning && root->parse->commandType != CMD_SELECT)\\n{\\n List *pcqual = RelationGetPartitionQual(relation);\\n\\n if (pcqual)\\n {\\n /*\\n * Run the partition quals through const-simplification similar to\\n * check constraints. We skip canonicalize_qual, though, because\\n * partition quals should be in canonical form already; also,\\n * since the qual is in implicit-AND format, we'd have to\\n * explicitly convert it to explicit-AND format and back again.\\n */\\n pcqual = (List *) eval_const_expressions(root, (Node *) pcqual);\\n\\n /* Fix Vars to have the desired varno */\\n if (varno != 1)\\n ChangeVarNodes((Node *) pcqual, 1, varno, 0);\\n\\n result = list_concat(result, pcqual);\\n }\\n}\\n\\nWe will no longer need to load the partition constraints for \\\"other rel\\\"\\npartitions, not even for UPDATE and DELETE queries. Now, we won't load\\nthem with the patch applied, because we're cheating by first planning the\\nquery as SELECT, so that's not an issue. But we should change the\\ncondition here to check if the input relation is a \\\"baserel\\\", in which\\ncase, this should still load the partition constraint so that constraint\\nexclusion can use it when running with constraint_exclusion = on. In\\nfact, I recently reported [1] on -hackers that we don't load the partition\\nconstraint even if the partition is being accessed directly as a bug\\nintroduced in PG 11.\\n\\nThanks,\\nAmit\\n\\n[1]\\nhttps://www.postgresql.org/message-id/9813f079-f16b-61c8-9ab7-4363cab28d80%40lab.ntt.co.jp\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 29 Mar 2019 18:08:46 +0900\",\n \"msg_from\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> writes:\\n> Here are some comments on v38.\\n\\nThanks for looking it over! I'll just reply to points worth discussing:\\n\\n> - Assert(rte->rtekind == RTE_RELATION ||\\n> - rte->rtekind == RTE_SUBQUERY);\\n> - add_appendrel_other_rels(root, rel, rti);\\n> + if (rte->rtekind == RTE_RELATION)\\n> + expand_inherited_rtentry(root, rel, rte, rti);\\n> + else\\n> + expand_appendrel_subquery(root, rel, rte, rti);\\n\\n> Wouldn't it be a good idea to keep the Assert?\\n\\nThere's an Assert in expand_appendrel_subquery that what it got is an\\nRTE_SUBQUERY, so I thought the one at the call site was redundant.\\nI suppose another way to do this would be like\\n\\n if (rte->rtekind == RTE_RELATION)\\n expand_inherited_rtentry(root, rel, rte, rti);\\n else if (rte->rtekind == RTE_SUBQUERY)\\n expand_appendrel_subquery(root, rel, rte, rti);\\n else\\n Assert(false);\\n\\nNot sure if that's better or not. Or we could go back to the\\ndesign of just having one function and letting it dispatch the\\ncase it doesn't want to the other function --- though I think\\nI'd make expand_inherited_rtentry be the primary function,\\nrather than the other way around as you had it in v37.\\n\\n> + forboth(lc, old_child_rtis, lc2, new_child_rtis)\\n> + {\\n> + int old_child_rti = lfirst_int(lc);\\n> + int new_child_rti = lfirst_int(lc2);\\n> +\\n> + if (old_child_rti == new_child_rti)\\n> + continue; /* nothing to do */\\n> +\\n> + Assert(old_child_rti > new_child_rti);\\n> +\\n> + ChangeVarNodes((Node *) child_appinfos,\\n> + old_child_rti, new_child_rti, 0);\\n> + }\\n\\n> This seems necessary? RTEs of children of the target table should be in\\n> the same position in the final_rtable as they are in the select_rtable.\\n\\nWell, that's what I'm not very convinced of. I have observed that\\nthe regression tests don't reach this ChangeVarNodes call, but\\nI think that might just be lack of test cases rather than a proof\\nthat it's impossible. The question is whether it'd ever be possible\\nfor the update/delete target to not be the first \\\"inh\\\" table that\\ngets expanded. Since that expansion is done in RTE order, it\\nreduces to \\\"is the target always before any other RTE entries\\nthat could need inheritance expansion?\\\" Certainly that would typically\\nbe true, but I don't feel very comfortable about assuming that it\\nmust be true, when you start thinking about things like updatable\\nviews, rules, WITH queries, and so on.\\n\\nIt might be worth trying to devise a test case that does reach this\\ncode. If we could convince ourselves that it's really impossible,\\nI'd be willing to drop it in favor of putting a test-and-elog check\\nin the earlier loop that the RTI pairs are all equal. But I'm not\\nwilling to do it without more investigation.\\n\\n> + /* XXX wrong? */\\n> + parentrte->inh = false;\\n\\n> About the XXX: I think resetting inh flag is unnecessary, so we should\\n> just remove the line.\\n\\nPossibly. I hadn't had time to follow up the XXX annotation.\\n\\n> If we do that, we can also get rid of the following\\n> code in set_rel_size():\\n\\n> else if (rte->relkind == RELKIND_PARTITIONED_TABLE)\\n> {\\n> /*\\n> * A partitioned table without any partitions is marked as\\n> * a dummy rel.\\n> */\\n> set_dummy_rel_pathlist(rel);\\n> }\\n\\nNot following? Surely we need to mark the childless parent as dummy at\\nsome point, and that seems like as good a place as any.\\n\\n> Finally, it's not in the patch, but how about visiting\\n> get_relation_constraints() for revising this block of code:\\n\\nThat seems like probably an independent patch --- do you want to write it?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 29 Mar 2019 11:29:45 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> writes:\\n>> About the XXX: I think resetting inh flag is unnecessary, so we should\\n>> just remove the line.\\n\\n> Possibly. I hadn't had time to follow up the XXX annotation.\\n\\nNow I have ...\\n\\nYeah, it seems we can just drop that and leave the flag alone. We'll\\nend up running through set_append_rel_size and finding no relevant\\nAppendRelInfos, but that's not going to take long enough to be a problem.\\nIt seems better to have the principle that rte->inh doesn't change after\\nsubquery_planner's initial scan of the rtable, so I'll make it so.\\n\\n>> If we do that, we can also get rid of the following\\n>> code in set_rel_size():\\n\\nNo, we can't --- that's still reachable if somebody says\\n\\\"SELECT FROM ONLY partitioned_table\\\".\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 29 Mar 2019 12:31:09 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> writes:\\n> On 2019/03/29 7:38, Tom Lane wrote:\\n>> 2. I seriously dislike what's been done in joinrels.c, too. That\\n>> really seems like a kluge (and I haven't had time to study it\\n>> closely).\\n\\n> Those hunks account for the fact that pruned partitions, for which we no\\n> longer create RangeTblEntry and RelOptInfo, may appear on the nullable\\n> side of an outer join. We'll need a RelOptInfo holding a dummy path, so\\n> that outer join paths can be created with one side of join being dummy\\n> result path, which are built in the patch by build_dummy_partition_rel().\\n\\nJust for the record, that code is completely broken: it falls over\\nbadly under GEQO. (Try running the regression tests with\\n\\\"alter system set geqo_threshold to 2\\\".) However, the partitionwise\\njoin code was completely broken for GEQO before this patch, too, so\\nI'm just going to log that as an open issue for the moment.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 29 Mar 2019 14:32:57 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> writes:\\n> On 2019/03/29 7:38, Tom Lane wrote:\\n>> 2. I seriously dislike what's been done in joinrels.c, too. That\\n>> really seems like a kluge (and I haven't had time to study it\\n>> closely).\\n\\n> Those hunks account for the fact that pruned partitions, for which we no\\n> longer create RangeTblEntry and RelOptInfo, may appear on the nullable\\n> side of an outer join. We'll need a RelOptInfo holding a dummy path, so\\n> that outer join paths can be created with one side of join being dummy\\n> result path, which are built in the patch by build_dummy_partition_rel().\\n\\nNow that I've had a chance to look closer, there's no way I'm committing\\nthat change in joinrels.c. If it works at all, it's accidental, because\\nit's breaking all sorts of data structure invariants. The business with\\nan AppendRelInfo that maps from the parentrel to itself is particularly\\nugly; and I doubt that you can get away with assuming that\\nroot->append_rel_array[parent->relid] is available for use for that.\\n(What if the parent is an intermediate partitioned table?)\\n\\nThere's also the small problem of the GEQO crash. It's possible that\\nthat could be gotten around by switching into the long-term planner\\ncontext in update_child_rel_info and build_dummy_partition_rel, but\\nthen you're creating a memory leak across GEQO cycles. It'd be much\\nbetter to avoid touching base-relation data structures during join\\nplanning.\\n\\nWhat I propose we do about the GEQO problem is shown in 0001 attached\\n(which would need to be back-patched into v11). This is based on the\\nobservation that, if we know an input relation is empty, we can often\\nprove the join is empty and then skip building it at all. (In the\\nexisting partitionwise-join code, the same cases are detected by\\npopulate_joinrel_with_paths, but we do a fair amount of work before\\ndiscovering that.) The cases where that's not true are where we\\nhave a pruned partition on the inside of a left join, or either side\\nof a full join ... but frankly, what the existing code produces for\\nthose cases is not short of embarrassing:\\n\\n -> Hash Left Join\\n Hash Cond: (pagg_tab1_p1.x = y)\\n Filter: ((pagg_tab1_p1.x > 5) OR (y < 20))\\n -> Seq Scan on pagg_tab1_p1\\n Filter: (x < 20)\\n -> Hash\\n -> Result\\n One-Time Filter: false\\n\\nThat's just dumb. What we *ought* to be doing in such degenerate\\nouter-join cases is just emitting the non-dummy side, ie\\n\\n -> Seq Scan on pagg_tab1_p1\\n Filter: (x < 20) AND ((pagg_tab1_p1.x > 5) OR (y < 20))\\n\\nin this example. I would envision handling this by teaching the\\ncode to generate a path for the joinrel that's basically just a\\nProjectionPath atop a path for the non-dummy input rel, with the\\nprojection set up to emit nulls for the columns of the dummy side.\\n(Note that this would be useful for outer joins against dummy rels\\nin regular planning contexts, not only partitionwise joins.)\\n\\nPending somebody doing the work for that, though, I do not\\nhave a problem with just being unable to generate partitionwise\\njoins in such cases, so 0001 attached just changes the expected\\noutputs for the affected regression test cases.\\n\\n0002 attached is then the rest of the partition-planning patch;\\nit doesn't need to mess with joinrels.c at all. I've addressed\\nthe other points discussed today in that, except for the business\\nabout whether we want your 0003 bitmap-of-live-partitions patch.\\nI'm still inclined to think that that's not really worth it,\\nespecially in view of your performance results.\\n\\nIf people are OK with this approach to solving the GEQO problem,\\nI think these are committable.\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Fri, 29 Mar 2019 20:17:32 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Thanks for the new patches.\\n\\nOn Sat, Mar 30, 2019 at 9:17 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> writes:\\n> > On 2019/03/29 7:38, Tom Lane wrote:\\n> >> 2. I seriously dislike what's been done in joinrels.c, too. That\\n> >> really seems like a kluge (and I haven't had time to study it\\n> >> closely).\\n>\\n> > Those hunks account for the fact that pruned partitions, for which we no\\n> > longer create RangeTblEntry and RelOptInfo, may appear on the nullable\\n> > side of an outer join. We'll need a RelOptInfo holding a dummy path, so\\n> > that outer join paths can be created with one side of join being dummy\\n> > result path, which are built in the patch by build_dummy_partition_rel().\\n>\\n> Now that I've had a chance to look closer, there's no way I'm committing\\n> that change in joinrels.c. If it works at all, it's accidental, because\\n> it's breaking all sorts of data structure invariants. The business with\\n> an AppendRelInfo that maps from the parentrel to itself is particularly\\n> ugly; and I doubt that you can get away with assuming that\\n> root->append_rel_array[parent->relid] is available for use for that.\\n> (What if the parent is an intermediate partitioned table?)\\n>\\n> There's also the small problem of the GEQO crash. It's possible that\\n> that could be gotten around by switching into the long-term planner\\n> context in update_child_rel_info and build_dummy_partition_rel, but\\n> then you're creating a memory leak across GEQO cycles. It'd be much\\n> better to avoid touching base-relation data structures during join\\n> planning.\\n>\\n> What I propose we do about the GEQO problem is shown in 0001 attached\\n> (which would need to be back-patched into v11). This is based on the\\n> observation that, if we know an input relation is empty, we can often\\n> prove the join is empty and then skip building it at all. (In the\\n> existing partitionwise-join code, the same cases are detected by\\n> populate_joinrel_with_paths, but we do a fair amount of work before\\n> discovering that.) The cases where that's not true are where we\\n> have a pruned partition on the inside of a left join, or either side\\n> of a full join ... but frankly, what the existing code produces for\\n> those cases is not short of embarrassing:\\n>\\n> -> Hash Left Join\\n> Hash Cond: (pagg_tab1_p1.x = y)\\n> Filter: ((pagg_tab1_p1.x > 5) OR (y < 20))\\n> -> Seq Scan on pagg_tab1_p1\\n> Filter: (x < 20)\\n> -> Hash\\n> -> Result\\n> One-Time Filter: false\\n>\\n> That's just dumb. What we *ought* to be doing in such degenerate\\n> outer-join cases is just emitting the non-dummy side, ie\\n>\\n> -> Seq Scan on pagg_tab1_p1\\n> Filter: (x < 20) AND ((pagg_tab1_p1.x > 5) OR (y < 20))\\n>\\n> in this example. I would envision handling this by teaching the\\n> code to generate a path for the joinrel that's basically just a\\n> ProjectionPath atop a path for the non-dummy input rel, with the\\n> projection set up to emit nulls for the columns of the dummy side.\\n> (Note that this would be useful for outer joins against dummy rels\\n> in regular planning contexts, not only partitionwise joins.)\\n>\\n> Pending somebody doing the work for that, though, I do not\\n> have a problem with just being unable to generate partitionwise\\n> joins in such cases, so 0001 attached just changes the expected\\n> outputs for the affected regression test cases.\\n\\nFwiw, I agree that we should fix join planning so that we get the\\nProjectionPath atop scan path of non-nullable relation instead of a\\nfull-fledged join path with dummy path on the nullable side. It seems\\nto me that the \\\"fix\\\" would be mostly be localized to\\ntry_partitionwise_join() at least insofar as detecting whether we\\nshould generate a join or the other plan shape is concerned, right?\\n\\nBy the way, does it make sense to remove the tests whose output\\nchanges altogether and reintroduce them when we fix join planning?\\nEspecially, in partitionwise_aggregate.out, there are comments near\\nchanged plans which are no longer true.\\n\\n> 0002 attached is then the rest of the partition-planning patch;\\n> it doesn't need to mess with joinrels.c at all. I've addressed\\n> the other points discussed today in that, except for the business\\n> about whether we want your 0003 bitmap-of-live-partitions patch.\\n> I'm still inclined to think that that's not really worth it,\\n> especially in view of your performance results.\\n\\nI think the performance results did prove that degradation due to\\nthose loops over part_rels becomes significant for very large\\npartition counts. Is there a better solution than the bitmapset that\\nyou have in mind?\\n\\n> If people are OK with this approach to solving the GEQO problem,\\n> I think these are committable.\\n\\nThanks again. Really appreciate that you are putting so much of your\\ntime into this.\\n\\nRegards,\\nAmit\\n\\n\\n\",\n \"msg_date\": \"Sat, 30 Mar 2019 11:19:58 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Amit Langote <amitlangote09@gmail.com> writes:\\n> On Sat, Mar 30, 2019 at 9:17 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> What I propose we do about the GEQO problem is shown in 0001 attached\\n>> (which would need to be back-patched into v11).\\n>> ...\\n>> That's just dumb. What we *ought* to be doing in such degenerate\\n>> outer-join cases is just emitting the non-dummy side, ie\\n\\n> Fwiw, I agree that we should fix join planning so that we get the\\n> ProjectionPath atop scan path of non-nullable relation instead of a\\n> full-fledged join path with dummy path on the nullable side. It seems\\n> to me that the \\\"fix\\\" would be mostly be localized to\\n> try_partitionwise_join() at least insofar as detecting whether we\\n> should generate a join or the other plan shape is concerned, right?\\n\\nWell, if we're going to do something about that, I would like to see\\nit work for non-partition cases too, ie we're not smart about this\\neither:\\n\\nregression=# explain select * from tenk1 left join (select 1 where false) ss(x) on unique1=x;\\n QUERY PLAN \\n-------------------------------------------------------------------\\n Nested Loop Left Join (cost=0.00..570.00 rows=10000 width=248)\\n Join Filter: (tenk1.unique1 = 1)\\n -> Seq Scan on tenk1 (cost=0.00..445.00 rows=10000 width=244)\\n -> Result (cost=0.00..0.00 rows=0 width=0)\\n One-Time Filter: false\\n(5 rows)\\n\\nA general solution would presumably involve new logic in\\npopulate_joinrel_with_paths for the case where the nullable side is\\ndummy. I'm not sure whether that leaves anything special to do in\\ntry_partitionwise_join or not. Maybe it would, since that would\\nhave a requirement to build the joinrel without any RHS input\\nRelOptInfo, but I don't think that's the place to begin working on\\nthis.\\n\\n> By the way, does it make sense to remove the tests whose output\\n> changes altogether and reintroduce them when we fix join planning?\\n> Especially, in partitionwise_aggregate.out, there are comments near\\n> changed plans which are no longer true.\\n\\nGood point about the comments, but we shouldn't just remove those test\\ncases; they're useful to exercise the give-up-on-partitionwise-join\\ncode paths. I'll tweak the comments.\\n\\n>> 0002 attached is then the rest of the partition-planning patch;\\n>> it doesn't need to mess with joinrels.c at all. I've addressed\\n>> the other points discussed today in that, except for the business\\n>> about whether we want your 0003 bitmap-of-live-partitions patch.\\n>> I'm still inclined to think that that's not really worth it,\\n>> especially in view of your performance results.\\n\\n> I think the performance results did prove that degradation due to\\n> those loops over part_rels becomes significant for very large\\n> partition counts. Is there a better solution than the bitmapset that\\n> you have in mind?\\n\\nHm, I didn't see much degradation in what you posted in\\n<5c83dbca-12b5-1acf-0e85-58299e464a26@lab.ntt.co.jp>.\\nI am curious as to why there seems to be more degradation\\nfor hash cases, as per Yoshikazu-san's results in\\n<0F97FA9ABBDBE54F91744A9B37151A512BAC60@g01jpexmbkw24>,\\nbut whatever's accounting for the difference probably\\nis not that. Anyway I still believe that getting rid of\\nthese sparse arrays would be a better answer.\\n\\nBefore that, though, I remain concerned that the PartitionPruneInfo\\ndata structure the planner is transmitting to the executor is unsafe\\nagainst concurrent ATTACH PARTITION operations. The comment for\\nPartitionedRelPruneInfo says in so many words that it's relying on\\nindexes in the table's PartitionDesc; how is that not broken by\\n898e5e329?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 30 Mar 2019 11:11:27 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On Sat, Mar 30, 2019 at 11:11 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Before that, though, I remain concerned that the PartitionPruneInfo\\n> data structure the planner is transmitting to the executor is unsafe\\n> against concurrent ATTACH PARTITION operations. The comment for\\n> PartitionedRelPruneInfo says in so many words that it's relying on\\n> indexes in the table's PartitionDesc; how is that not broken by\\n> 898e5e329?\\n\\nThe only problem with PartitionPruneInfo structures of which I am\\naware is that they rely on PartitionDesc offsets not changing. But I\\nadded code in that commit in ExecCreatePartitionPruneState to handle\\nthat exact problem. See also paragraph 5 of the commit message, which\\nbegins with \\\"Although in general...\\\"\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Sat, 30 Mar 2019 11:36:56 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Sat, Mar 30, 2019 at 11:11 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> Before that, though, I remain concerned that the PartitionPruneInfo\\n>> data structure the planner is transmitting to the executor is unsafe\\n>> against concurrent ATTACH PARTITION operations. The comment for\\n>> PartitionedRelPruneInfo says in so many words that it's relying on\\n>> indexes in the table's PartitionDesc; how is that not broken by\\n>> 898e5e329?\\n\\n> The only problem with PartitionPruneInfo structures of which I am\\n> aware is that they rely on PartitionDesc offsets not changing. But I\\n> added code in that commit in ExecCreatePartitionPruneState to handle\\n> that exact problem. See also paragraph 5 of the commit message, which\\n> begins with \\\"Although in general...\\\"\\n\\nAh. Grotty, but I guess it will cover the issue.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 30 Mar 2019 11:46:46 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On Sat, Mar 30, 2019 at 11:46 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> > The only problem with PartitionPruneInfo structures of which I am\\n> > aware is that they rely on PartitionDesc offsets not changing. But I\\n> > added code in that commit in ExecCreatePartitionPruneState to handle\\n> > that exact problem. See also paragraph 5 of the commit message, which\\n> > begins with \\\"Although in general...\\\"\\n>\\n> Ah. Grotty, but I guess it will cover the issue.\\n\\nI suppose it is. I am a little suspicious of the decision to make\\nPartitionPruneInfo structures depend on PartitionDesc indexes. First,\\nit's really non-obvious that the dependency exists, and I do not think\\nI would have spotted it had not Alvaro pointed the problem out.\\nSecond, I wonder whether it is really a good idea in general to make a\\nplan depend on array indexes when the array is not stored in the plan.\\nIn one sense, we do that all the time, because attnums are arguably\\njust indexes into what is conceptually an array of attributes.\\nHowever, I feel that's not quite the same, because the attnum is\\nexplicitly stored in the catalogs, and PartitionDesc array indexes are\\nnot stored anywhere, but rather are the result of a fairly complex\\ncalculation. Now I guess it's probably OK because we will probably\\nhave lots of other problems if we don't get the same answer every time\\nwe do that calculation, but it still makes me a little nervous. I\\nwould try to propose something better but I don't have a good idea.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Sat, 30 Mar 2019 11:59:32 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On Sun, Mar 31, 2019 at 12:11 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Amit Langote <amitlangote09@gmail.com> writes:\\n> > I think the performance results did prove that degradation due to\\n> > those loops over part_rels becomes significant for very large\\n> > partition counts. Is there a better solution than the bitmapset that\\n> > you have in mind?\\n>\\n> Hm, I didn't see much degradation in what you posted in\\n> <5c83dbca-12b5-1acf-0e85-58299e464a26@lab.ntt.co.jp>.\\n\\nSorry that I didn't mention the link to begin with, but I meant to\\npoint to numbers that I reported on Monday this week.\\n\\nhttps://www.postgresql.org/message-id/19f54c17-1619-b228-10e5-ca343be6a4e8%40lab.ntt.co.jp\\n\\nYou were complaining of the bitmapset being useless overhead for small\\npartition counts, but the numbers I get tend to suggest that any\\ndegradation in performance is within noise range, whereas the\\nperformance benefit from having them looks pretty significant for very\\nlarge partition counts.\\n\\n> I am curious as to why there seems to be more degradation\\n> for hash cases, as per Yoshikazu-san's results in\\n> <0F97FA9ABBDBE54F91744A9B37151A512BAC60@g01jpexmbkw24>,\\n> but whatever's accounting for the difference probably\\n> is not that.\\n\\nI suspected it may have been the lack of bitmapsets, but maybe only\\nImai-san could've confirmed that by applying the live_parts patch too.\\n\\nThanks,\\nAmit\\n\\n\\n\",\n \"msg_date\": \"Sun, 31 Mar 2019 01:06:26 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On Sun, Mar 31, 2019 at 12:59 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n>\\n> On Sat, Mar 30, 2019 at 11:46 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> > > The only problem with PartitionPruneInfo structures of which I am\\n> > > aware is that they rely on PartitionDesc offsets not changing. But I\\n> > > added code in that commit in ExecCreatePartitionPruneState to handle\\n> > > that exact problem. See also paragraph 5 of the commit message, which\\n> > > begins with \\\"Although in general...\\\"\\n> >\\n> > Ah. Grotty, but I guess it will cover the issue.\\n>\\n> I suppose it is. I am a little suspicious of the decision to make\\n> PartitionPruneInfo structures depend on PartitionDesc indexes.\\n\\nFwiw, I had complained when reviewing the run-time pruning patch that\\ncreating those maps in the planner and putting them in\\nPartitionPruneInfo might not be a good idea, but David insisted that\\nit'd be good for performance (in the context of using cached plans) to\\ncompute this information during planning.\\n\\nThanks,\\nAmit\\n\\n\\n\",\n \"msg_date\": \"Sun, 31 Mar 2019 01:16:12 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On Sat, Mar 30, 2019 at 12:16 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> Fwiw, I had complained when reviewing the run-time pruning patch that\\n> creating those maps in the planner and putting them in\\n> PartitionPruneInfo might not be a good idea, but David insisted that\\n> it'd be good for performance (in the context of using cached plans) to\\n> compute this information during planning.\\n\\nWell, he's not wrong about that, I expect.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Sat, 30 Mar 2019 12:50:38 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On 2019/03/31 1:06, Amit Langote wrote:\\r\\n > On Sun, Mar 31, 2019 at 12:11 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\r\\n >> Amit Langote <amitlangote09@gmail.com> writes:\\r\\n >>> I think the performance results did prove that degradation due to\\r\\n >>> those loops over part_rels becomes significant for very large\\r\\n >>> partition counts. Is there a better solution than the bitmapset that\\r\\n >>> you have in mind?\\r\\n >>\\r\\n >> Hm, I didn't see much degradation in what you posted in\\r\\n >> <5c83dbca-12b5-1acf-0e85-58299e464a26@lab.ntt.co.jp>.\\r\\n >\\r\\n > Sorry that I didn't mention the link to begin with, but I meant to\\r\\n > point to numbers that I reported on Monday this week.\\r\\n >\\r\\n > \\r\\nhttps://www.postgresql.org/message-id/19f54c17-1619-b228-10e5-ca343be6a4e8%40lab.ntt.co.jp\\r\\n >\\r\\n > You were complaining of the bitmapset being useless overhead for small\\r\\n > partition counts, but the numbers I get tend to suggest that any\\r\\n > degradation in performance is within noise range, whereas the\\r\\n > performance benefit from having them looks pretty significant for very\\r\\n > large partition counts.\\r\\n >\\r\\n >> I am curious as to why there seems to be more degradation\\r\\n >> for hash cases, as per Yoshikazu-san's results in\\r\\n >> <0F97FA9ABBDBE54F91744A9B37151A512BAC60@g01jpexmbkw24>,\\r\\n >> but whatever's accounting for the difference probably\\r\\n >> is not that.\\r\\n >\\r\\n > I suspected it may have been the lack of bitmapsets, but maybe only\\r\\n > Imai-san could've confirmed that by applying the live_parts patch too.\\r\\n\\r\\nYeah, I forgot to applying live_parts patch. I did same test again which \\r\\nI did for hash before.\\r\\n(BTW, thanks for committing speeding up patches!)\\r\\n\\r\\n[HEAD(428b260)]\\r\\nnparts TPS\\r\\n====== =====\\r\\n2: 13134 (13240, 13290, 13071, 13172, 12896)\\r\\n1024: 12627 (12489, 12635, 12716, 12732, 12562)\\r\\n8192: 10289 (10216, 10265, 10171, 10278, 10514)\\r\\n\\r\\n[HEAD(428b260) + live_parts.diff]\\r\\nnparts TPS\\r\\n====== =====\\r\\n2: 13277 (13112, 13290, 13241, 13360, 13382)\\r\\n1024: 12821 (12930, 12849, 12909, 12700, 12716)\\r\\n8192: 11102 (11134, 11158, 11114, 10997, 11109)\\r\\n\\r\\n\\r\\nDegradations of performance are below.\\r\\n\\r\\n\\r\\nMy test results from above (with live_parts, HEAD(428b260) + \\r\\nlive_parts.diff)\\r\\nnparts live_parts HEAD\\r\\n====== ========== ====\\r\\n2: 13277 13134\\r\\n1024: 12821 12627\\r\\n8192: 11102 10289\\r\\n\\r\\n11102/13277 = 83.6 %\\r\\n\\r\\n\\r\\nAmit-san's test results (with live_parts)\\r\\n > nparts v38 HEAD\\r\\n > ====== ==== ====\\r\\n > 2 2971 2969\\r\\n > 8 2980 1949\\r\\n > 32 2955 733\\r\\n > 128 2946 145\\r\\n > 512 2924 11\\r\\n > 1024 2986 3\\r\\n > 4096 2702 0\\r\\n > 8192 2531 OOM\\r\\n\\r\\n2531/2971 = 85.2 %\\r\\n\\r\\n\\r\\nMy test results I posted before (without live_parts)\\r\\n > nparts v38 HEAD\\r\\n > ====== ==== ====\\r\\n > 0: 10538 10487\\r\\n > 2: 6942 7028\\r\\n > 4: 7043 5645\\r\\n > 8: 6981 3954\\r\\n > 16: 6932 2440\\r\\n > 32: 6897 1243\\r\\n > 64: 6897 309\\r\\n > 128: 6753 120\\r\\n > 256: 6727 46\\r\\n > 512: 6708 12\\r\\n > 1024: 6063 3\\r\\n > 2048: 5894 1\\r\\n > 4096: 5374 OOM\\r\\n > 8192: 4572 OOM\\r\\n\\r\\n4572/6942 = 65.9 %\\r\\n\\r\\n\\r\\nCertainly, using bitmapset contributes to the performance when scanning \\r\\none partition(few partitions) from large partitions.\\r\\n\\r\\n\\r\\nThanks\\r\\n--\\r\\nImai Yoshikazu\",\n \"msg_date\": \"Sun, 31 Mar 2019 02:45:54 +0000\",\n \"msg_from\": \"Imai Yoshikazu <yoshikazu_i443@live.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On Sun, Mar 31, 2019 at 11:45 AM Imai Yoshikazu <yoshikazu_i443@live.jp> wrote:\\n> On 2019/03/31 1:06, Amit Langote wrote:\\n> > On Sun, Mar 31, 2019 at 12:11 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> >> I am curious as to why there seems to be more degradation\\n> >> for hash cases, as per Yoshikazu-san's results in\\n> >> <0F97FA9ABBDBE54F91744A9B37151A512BAC60@g01jpexmbkw24>,\\n> >> but whatever's accounting for the difference probably\\n> >> is not that.\\n> >\\n> > I suspected it may have been the lack of bitmapsets, but maybe only\\n> > Imai-san could've confirmed that by applying the live_parts patch too.\\n>\\n> Yeah, I forgot to applying live_parts patch. I did same test again which\\n> I did for hash before.\\n> (BTW, thanks for committing speeding up patches!)\\n\\nThanks a lot for committing, Tom. I wish you had listed yourself as\\nan author though.\\n\\nI will send the patch for get_relation_constraints() mentioned\\nupthread tomorrow.\\n\\n> [HEAD(428b260)]\\n> nparts TPS\\n> ====== =====\\n> 2: 13134 (13240, 13290, 13071, 13172, 12896)\\n> 1024: 12627 (12489, 12635, 12716, 12732, 12562)\\n> 8192: 10289 (10216, 10265, 10171, 10278, 10514)\\n>\\n> [HEAD(428b260) + live_parts.diff]\\n> nparts TPS\\n> ====== =====\\n> 2: 13277 (13112, 13290, 13241, 13360, 13382)\\n> 1024: 12821 (12930, 12849, 12909, 12700, 12716)\\n> 8192: 11102 (11134, 11158, 11114, 10997, 11109)\\n>\\n>\\n> Degradations of performance are below.\\n>\\n>\\n> My test results from above (with live_parts, HEAD(428b260) +\\n> live_parts.diff)\\n> nparts live_parts HEAD\\n> ====== ========== ====\\n> 2: 13277 13134\\n> 1024: 12821 12627\\n> 8192: 11102 10289\\n>\\n> 11102/13277 = 83.6 %\\n>\\n>\\n> Amit-san's test results (with live_parts)\\n> > nparts v38 HEAD\\n> > ====== ==== ====\\n> > 2 2971 2969\\n> > 8 2980 1949\\n> > 32 2955 733\\n> > 128 2946 145\\n> > 512 2924 11\\n> > 1024 2986 3\\n> > 4096 2702 0\\n> > 8192 2531 OOM\\n>\\n> 2531/2971 = 85.2 %\\n>\\n>\\n> My test results I posted before (without live_parts)\\n> > nparts v38 HEAD\\n> > ====== ==== ====\\n> > 0: 10538 10487\\n> > 2: 6942 7028\\n> > 4: 7043 5645\\n> > 8: 6981 3954\\n> > 16: 6932 2440\\n> > 32: 6897 1243\\n> > 64: 6897 309\\n> > 128: 6753 120\\n> > 256: 6727 46\\n> > 512: 6708 12\\n> > 1024: 6063 3\\n> > 2048: 5894 1\\n> > 4096: 5374 OOM\\n> > 8192: 4572 OOM\\n>\\n> 4572/6942 = 65.9 %\\n>\\n>\\n> Certainly, using bitmapset contributes to the performance when scanning\\n> one partition(few partitions) from large partitions.\\n\\nThanks Imai-san for testing.\\n\\nRegards,\\nAmit\\n\\n\\n\",\n \"msg_date\": \"Sun, 31 Mar 2019 14:40:46 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Amit Langote <amitlangote09@gmail.com> writes:\\n> On Sun, Mar 31, 2019 at 11:45 AM Imai Yoshikazu <yoshikazu_i443@live.jp> wrote:\\n>> Certainly, using bitmapset contributes to the performance when scanning\\n>> one partition(few partitions) from large partitions.\\n\\n> Thanks Imai-san for testing.\\n\\nI tried to replicate these numbers with the code as-committed, and\\ncould not. What I get, using the same table-creation code as you\\nposted and a pgbench script file like\\n\\n\\\\set param random(1, :N)\\nselect * from rt where a = :param;\\n\\nis scaling like this:\\n\\nN\\ttps, range\\ttps, hash\\n\\n2\\t10520.519932\\t10415.230400\\n8\\t10443.361457\\t10480.987665\\n32\\t10341.196768\\t10462.551167\\n128\\t10370.953849\\t10383.885128\\n512\\t10207.578413\\t10214.049394\\n1024\\t10042.794340\\t10121.683993\\n4096\\t8937.561825\\t9214.993778\\n8192\\t8247.614040\\t8486.728918\\n\\nIf I use \\\"-M prepared\\\" the numbers go up a bit for lower N, but\\ndrop at high N:\\n\\nN\\ttps, range\\ttps, hash\\n\\n2\\t11449.920527\\t11462.253871\\n8\\t11530.513146\\t11470.812476\\n32\\t11372.412999\\t11450.213753\\n128\\t11289.351596\\t11322.698856\\n512\\t11095.428451\\t11200.683771\\n1024\\t10757.646108\\t10805.052480\\n4096\\t8689.165875\\t8930.690887\\n8192\\t7301.609147\\t7502.806455\\n\\nDigging into that, it seems like the degradation with -M prepared is\\nmostly in LockReleaseAll's hash_seq_search over the locallock hash table.\\nWhat I think must be happening is that with -M prepared, at some point the\\nplancache decides to try a generic plan, which causes opening/locking all\\nthe partitions, resulting in permanent bloat in the locallock hash table.\\nWe immediately go back to using custom plans, but hash_seq_search has\\nmore buckets to look through for the remainder of the process' lifetime.\\n\\nI do see some cycles getting spent in apply_scanjoin_target_to_paths\\nthat look to be due to scanning over the long part_rels array,\\nwhich your proposal would ameliorate. But (a) that's pretty small\\ncompared to other effects, and (b) IMO, apply_scanjoin_target_to_paths\\nis a remarkable display of brute force inefficiency to begin with.\\nI think we should see if we can't nuke that function altogether in\\nfavor of generating the paths with the right target the first time.\\n\\nBTW, the real elephant in the room is the O(N^2) cost of creating\\nthese tables in the first place. The runtime for the table-creation\\nscripts looks like\\n\\nN\\trange\\t\\thash\\n\\n2\\t0m0.011s\\t0m0.011s\\n8\\t0m0.015s\\t0m0.014s\\n32\\t0m0.032s\\t0m0.030s\\n128\\t0m0.132s\\t0m0.099s\\n512\\t0m0.969s\\t0m0.524s\\n1024\\t0m3.306s\\t0m1.442s\\n4096\\t0m46.058s\\t0m15.522s\\n8192\\t3m11.995s\\t0m58.720s\\n\\nThis seems to be down to the expense of doing RelationBuildPartitionDesc\\nto rebuild the parent's relcache entry for each child CREATE TABLE.\\nNot sure we can avoid that, but maybe we should consider adopting a\\ncheaper-to-read representation of partition descriptors. The fact that\\nrange-style entries seem to be 3X more expensive to load than hash-style\\nentries is strange.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 31 Mar 2019 13:04:27 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"One thing that I intentionally left out of the committed patch was changes\\nto stop short of scanning the whole simple_rel_array when looking only for\\nbaserels. I thought that had been done in a rather piecemeal fashion\\nand it'd be better to address it holistically, which I've now done in the\\nattached proposed patch.\\n\\nThis probably makes little if any difference in the test cases we've\\nmostly focused on in this thread, since there wouldn't be very many\\notherrels anyway now that we don't create them for pruned partitions.\\nHowever, in a case where there's a lot of partitions that we can't prune,\\nthis could be useful.\\n\\nI have not done any performance testing to see if this is actually\\nworth the trouble, though. Anybody want to do that?\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Sun, 31 Mar 2019 14:46:51 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On Sun, 31 Mar 2019 at 05:50, Robert Haas <robertmhaas@gmail.com> wrote:\\n>\\n> On Sat, Mar 30, 2019 at 12:16 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > Fwiw, I had complained when reviewing the run-time pruning patch that\\n> > creating those maps in the planner and putting them in\\n> > PartitionPruneInfo might not be a good idea, but David insisted that\\n> > it'd be good for performance (in the context of using cached plans) to\\n> > compute this information during planning.\\n>\\n> Well, he's not wrong about that, I expect.\\n\\nI'm aware that there have been combinations of objects to either\\nhaving these arrays and/or editing them during execution.\\n\\nI don't recall Amit's complaint, but I do recall Tom's. He suggested\\nwe not resequence the arrays in the executor and just maintain NULL\\nelements in the Append/MergeAppend subplans. I did consider this when\\nwriting run-time pruning but found that performance suffers. I\\ndemonstrated this on a thread somewhere.\\n\\nIIRC, I wrote this code because there was no way to translate the\\nresult of the pruning code into Append/MergeAppend subplan indexes.\\nRobert has since added a map of Oids to allow the executor to have\\nthose details, so it perhaps would be possible to take the result of\\nthe pruning code then lookup the Oids of the partitions that survived\\npruning, then map those to the subplans using the array Robert added.\\nUsing the array for that wouldn't be very efficient due to a lookup\\nbeing O(n) per surviving partition. Maybe it could be thrown into a\\nhashtable to make that faster. This solution would need to take into\\naccount mixed hierarchy Appends. e.g SELECT * FROM partitioned_table\\nWHERE partkey = $1 UNION ALL SELECT * FROM something_else; so it would\\nlikely need to be a hashtable per partitioned table. If the pruning\\ncode returned a partition whose Oid we didn't know about, then it must\\nbe from a partition that was added concurrently since the plan was\\nbuilt... However, that shouldn't happen today since Robert went to\\ngreat lengths for it not to.\\n\\nFurther discussions are likely best put in their own thread. As far as\\nI know, nothing is broken with the code today.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Mon, 1 Apr 2019 13:24:56 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"(I've closed the CF entry: https://commitfest.postgresql.org/22/1778/)\\n\\nOn 2019/04/01 2:04, Tom Lane wrote:\\n> Amit Langote <amitlangote09@gmail.com> writes:\\n>> On Sun, Mar 31, 2019 at 11:45 AM Imai Yoshikazu <yoshikazu_i443@live.jp> wrote:\\n>>> Certainly, using bitmapset contributes to the performance when scanning\\n>>> one partition(few partitions) from large partitions.\\n> \\n>> Thanks Imai-san for testing.\\n> \\n> I tried to replicate these numbers with the code as-committed, and\\n> could not.\\n\\nThanks for that.\\n\\n> What I get, using the same table-creation code as you\\n> posted and a pgbench script file like\\n> \\n> \\\\set param random(1, :N)\\n> select * from rt where a = :param;\\n> \\n> is scaling like this:\\n> \\n> N\\ttps, range\\ttps, hash\\n> \\n> 2\\t10520.519932\\t10415.230400\\n> 8\\t10443.361457\\t10480.987665\\n> 32\\t10341.196768\\t10462.551167\\n> 128\\t10370.953849\\t10383.885128\\n> 512\\t10207.578413\\t10214.049394\\n> 1024\\t10042.794340\\t10121.683993\\n> 4096\\t8937.561825\\t9214.993778\\n> 8192\\t8247.614040\\t8486.728918\\n> \\n> If I use \\\"-M prepared\\\" the numbers go up a bit for lower N, but\\n> drop at high N:\\n> \\n> N\\ttps, range\\ttps, hash\\n> \\n> 2\\t11449.920527\\t11462.253871\\n> 8\\t11530.513146\\t11470.812476\\n> 32\\t11372.412999\\t11450.213753\\n> 128\\t11289.351596\\t11322.698856\\n> 512\\t11095.428451\\t11200.683771\\n> 1024\\t10757.646108\\t10805.052480\\n> 4096\\t8689.165875\\t8930.690887\\n> 8192\\t7301.609147\\t7502.806455\\n> \\n> Digging into that, it seems like the degradation with -M prepared is\\n> mostly in LockReleaseAll's hash_seq_search over the locallock hash table.\\n> What I think must be happening is that with -M prepared, at some point the\\n> plancache decides to try a generic plan, which causes opening/locking all\\n> the partitions, resulting in permanent bloat in the locallock hash table.\\n> We immediately go back to using custom plans, but hash_seq_search has\\n> more buckets to look through for the remainder of the process' lifetime.\\n\\nAh, we did find this to be a problem upthread [1] and Tsunakawa-san then\\neven posted a patch which is being discussed at:\\n\\nhttps://commitfest.postgresql.org/22/1993/\\n\\n> I do see some cycles getting spent in apply_scanjoin_target_to_paths\\n> that look to be due to scanning over the long part_rels array,\\n> which your proposal would ameliorate. But (a) that's pretty small\\n> compared to other effects, and (b) IMO, apply_scanjoin_target_to_paths\\n> is a remarkable display of brute force inefficiency to begin with.\\n> I think we should see if we can't nuke that function altogether in\\n> favor of generating the paths with the right target the first time.\\n\\nThat's an option if we can make it work.\\n\\nShouldn't we look at *all* of the places that have code that now look like\\nthis:\\n\\n for (i = 0; i < rel->nparts; i++)\\n {\\n RelOptInfo *partrel = rel->part_rels[i];\\n\\n if (partrel == NULL)\\n continue;\\n ...\\n }\\n\\nBeside apply_scanjoin_target_to_paths(), there are:\\n\\ncreate_partitionwise_grouping_paths()\\nmake_partitionedrel_pruneinfo()\\n\\n\\n> BTW, the real elephant in the room is the O(N^2) cost of creating\\n> these tables in the first place. The runtime for the table-creation\\n> scripts looks like\\n> \\n> N\\trange\\t\\thash\\n> \\n> 2\\t0m0.011s\\t0m0.011s\\n> 8\\t0m0.015s\\t0m0.014s\\n> 32\\t0m0.032s\\t0m0.030s\\n> 128\\t0m0.132s\\t0m0.099s\\n> 512\\t0m0.969s\\t0m0.524s\\n> 1024\\t0m3.306s\\t0m1.442s\\n> 4096\\t0m46.058s\\t0m15.522s\\n> 8192\\t3m11.995s\\t0m58.720s\\n> \\n> This seems to be down to the expense of doing RelationBuildPartitionDesc\\n> to rebuild the parent's relcache entry for each child CREATE TABLE.\\n> Not sure we can avoid that, but maybe we should consider adopting a\\n> cheaper-to-read representation of partition descriptors. The fact that\\n> range-style entries seem to be 3X more expensive to load than hash-style\\n> entries is strange.\\n\\nI've noticed this many times too, but never prioritized doing something\\nabout it. I'll try sometime.\\n\\nThanks,\\nAmit\\n\\n[1]\\nhttps://www.postgresql.org/message-id/CAKJS1f-dn1hDZqObwdMrYdV7-cELJwWCPRWet6EQX_WaV8JLgw%40mail.gmail.com\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 1 Apr 2019 13:55:50 +0900\",\n \"msg_from\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On 2019/04/01 3:46, Tom Lane wrote:\\n> One thing that I intentionally left out of the committed patch was changes\\n> to stop short of scanning the whole simple_rel_array when looking only for\\n> baserels. I thought that had been done in a rather piecemeal fashion\\n> and it'd be better to address it holistically, which I've now done in the\\n> attached proposed patch.\\n> \\n> This probably makes little if any difference in the test cases we've\\n> mostly focused on in this thread, since there wouldn't be very many\\n> otherrels anyway now that we don't create them for pruned partitions.\\n> However, in a case where there's a lot of partitions that we can't prune,\\n> this could be useful.\\n> \\n> I have not done any performance testing to see if this is actually\\n> worth the trouble, though. Anybody want to do that?\\n\\nThanks for creating the patch.\\n\\nI spent some time looking for cases where this patch would provide\\nrecognizable benefit, but couldn't find one. As one would suspect, it's\\nhard to notice it if only looking at the overall latency of the query,\\nbecause time spent doing other things with such plans tends to be pretty\\nhuge anyway (both in the planner itself and other parts of the backend).\\nI even devised a query on a partitioned table such that planner has to\\nprocess all partitions, but ExecInitAppend can prune all but one, thus\\nreducing the time spent in the executor, but still wasn't able to see an\\nimprovement in the overall latency of the query due to planner not doing\\nlooping over the long simple_rel_array.\\n\\nThanks,\\nAmit\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 1 Apr 2019 17:37:23 +0900\",\n \"msg_from\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On 2019/03/30 0:29, Tom Lane wrote:\\n> Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> writes:\\n>> Finally, it's not in the patch, but how about visiting\\n>> get_relation_constraints() for revising this block of code:\\n> \\n> That seems like probably an independent patch --- do you want to write it?\\n\\nHere is that patch.\\n\\nIt revises get_relation_constraints() such that the partition constraint\\nis loaded in only the intended cases. To summarize:\\n\\n* PG 11 currently misses one such intended case (select * from partition)\\ncausing a *bug* that constraint exclusion fails to exclude the partition\\nwith constraint_exclusion = on\\n\\n* HEAD loads the partition constraint even in some cases where 428b260f87\\nrendered doing that unnecessary\\n\\nThanks,\\nAmit\",\n \"msg_date\": \"Mon, 1 Apr 2019 19:02:44 +0900\",\n \"msg_from\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> writes:\\n> On 2019/04/01 3:46, Tom Lane wrote:\\n>> One thing that I intentionally left out of the committed patch was changes\\n>> to stop short of scanning the whole simple_rel_array when looking only for\\n>> baserels. I thought that had been done in a rather piecemeal fashion\\n>> and it'd be better to address it holistically, which I've now done in the\\n>> attached proposed patch.\\n>> I have not done any performance testing to see if this is actually\\n>> worth the trouble, though. Anybody want to do that?\\n\\n> Thanks for creating the patch.\\n\\n> I spent some time looking for cases where this patch would provide\\n> recognizable benefit, but couldn't find one.\\n\\nYeah, I was afraid of that. In cases where we do have a ton of otherrels,\\nthe processing that's actually needed on them would probably swamp any\\nsavings from this patch.\\n\\nThe only place where that might possibly not be true is\\ncreate_lateral_join_info, since that has nested loops that could\\npotentially impose an O(N^2) cost. However, since your patch went in,\\nthat runs before inheritance expansion anyway.\\n\\nSo this probably isn't worth even the minuscule cost it imposes.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 01 Apr 2019 10:29:14 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> writes:\\n> On 2019/03/30 0:29, Tom Lane wrote:\\n>> That seems like probably an independent patch --- do you want to write it?\\n\\n> Here is that patch.\\n> It revises get_relation_constraints() such that the partition constraint\\n> is loaded in only the intended cases.\\n\\nSo I see the problem you're trying to solve here, but I don't like this\\npatch a bit, because it depends on root->inhTargetKind which IMO is a\\nbroken bit of junk that we need to get rid of. Here is an example of\\nwhy, with this patch applied:\\n\\nregression=# create table p (a int) partition by list (a);\\nCREATE TABLE\\nregression=# create table p1 partition of p for values in (1);\\nCREATE TABLE\\nregression=# set constraint_exclusion to on;\\nSET\\nregression=# explain select * from p1 where a = 2;\\n QUERY PLAN \\n------------------------------------------\\n Result (cost=0.00..0.00 rows=0 width=0)\\n One-Time Filter: false\\n(2 rows)\\n\\nSo far so good, but watch what happens when we include the same case\\nin an UPDATE on some other partitioned table:\\n\\nregression=# create table prtab (a int, b int) partition by list (a);\\nCREATE TABLE\\nregression=# create table prtab2 partition of prtab for values in (2);\\nCREATE TABLE\\nregression=# explain update prtab set b=b+1 from p1 where prtab.a=p1.a and p1.a=2;\\n QUERY PLAN \\n---------------------------------------------------------------------------\\n Update on prtab (cost=0.00..82.30 rows=143 width=20)\\n Update on prtab2\\n -> Nested Loop (cost=0.00..82.30 rows=143 width=20)\\n -> Seq Scan on p1 (cost=0.00..41.88 rows=13 width=10)\\n Filter: (a = 2)\\n -> Materialize (cost=0.00..38.30 rows=11 width=14)\\n -> Seq Scan on prtab2 (cost=0.00..38.25 rows=11 width=14)\\n Filter: (a = 2)\\n(8 rows)\\n\\nNo constraint exclusion, while in v10 you get\\n\\n Update on prtab (cost=0.00..0.00 rows=0 width=0)\\n -> Result (cost=0.00..0.00 rows=0 width=0)\\n One-Time Filter: false\\n\\nThe reason is that this logic supposes that root->inhTargetKind describes\\n*all* partitioned tables in the query, which is obviously wrong.\\n\\nNow maybe we could make it work by doing something like\\n\\n\\tif (rel->reloptkind == RELOPT_BASEREL &&\\n (root->inhTargetKind == INHKIND_NONE ||\\n rel->relid != root->parse->resultRelation))\\n\\nbut I find that pretty messy, plus it's violating the concept that we\\nshouldn't be allowing messiness from inheritance_planner to leak into\\nother places. What I'd rather do is have this test just read\\n\\n\\tif (rel->reloptkind == RELOPT_BASEREL)\\n\\nMaking it be that way causes some changes in the partition_prune results,\\nas attached, which suggest that removing the enable_partition_pruning\\ncheck as you did wasn't such a great idea either. However, if I add\\nthat back in, then it breaks the proposed new regression test case.\\n\\nI'm not at all clear on what we think the interaction between\\nenable_partition_pruning and constraint_exclusion ought to be,\\nso I'm not sure what the appropriate resolution is here. Thoughts?\\n\\nBTW, just about all the other uses of root->inhTargetKind seem equally\\nbroken from here; none of them are accounting for whether the rel in\\nquestion is the query target.\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Mon, 01 Apr 2019 13:34:45 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Thanks for taking a look.\\n\\nOn 2019/04/02 2:34, Tom Lane wrote:\\n> Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> writes:\\n>> On 2019/03/30 0:29, Tom Lane wrote:\\n>>> That seems like probably an independent patch --- do you want to write it?\\n> \\n>> Here is that patch.\\n>> It revises get_relation_constraints() such that the partition constraint\\n>> is loaded in only the intended cases.\\n> \\n> So I see the problem you're trying to solve here, but I don't like this\\n> patch a bit, because it depends on root->inhTargetKind which IMO is a\\n> broken bit of junk that we need to get rid of. Here is an example of\\n> why, with this patch applied:\\n> \\n> regression=# create table p (a int) partition by list (a);\\n> CREATE TABLE\\n> regression=# create table p1 partition of p for values in (1);\\n> CREATE TABLE\\n> regression=# set constraint_exclusion to on;\\n> SET\\n> regression=# explain select * from p1 where a = 2;\\n> QUERY PLAN \\n> ------------------------------------------\\n> Result (cost=0.00..0.00 rows=0 width=0)\\n> One-Time Filter: false\\n> (2 rows)\\n> \\n> So far so good, but watch what happens when we include the same case\\n> in an UPDATE on some other partitioned table:\\n> \\n> regression=# create table prtab (a int, b int) partition by list (a);\\n> CREATE TABLE\\n> regression=# create table prtab2 partition of prtab for values in (2);\\n> CREATE TABLE\\n> regression=# explain update prtab set b=b+1 from p1 where prtab.a=p1.a and p1.a=2;\\n> QUERY PLAN \\n> ---------------------------------------------------------------------------\\n> Update on prtab (cost=0.00..82.30 rows=143 width=20)\\n> Update on prtab2\\n> -> Nested Loop (cost=0.00..82.30 rows=143 width=20)\\n> -> Seq Scan on p1 (cost=0.00..41.88 rows=13 width=10)\\n> Filter: (a = 2)\\n> -> Materialize (cost=0.00..38.30 rows=11 width=14)\\n> -> Seq Scan on prtab2 (cost=0.00..38.25 rows=11 width=14)\\n> Filter: (a = 2)\\n> (8 rows)\\n> \\n> No constraint exclusion, while in v10 you get\\n> \\n> Update on prtab (cost=0.00..0.00 rows=0 width=0)\\n> -> Result (cost=0.00..0.00 rows=0 width=0)\\n> One-Time Filter: false\\n> \\n> The reason is that this logic supposes that root->inhTargetKind describes\\n> *all* partitioned tables in the query, which is obviously wrong.\\n> \\n> Now maybe we could make it work by doing something like\\n> \\n> \\tif (rel->reloptkind == RELOPT_BASEREL &&\\n> (root->inhTargetKind == INHKIND_NONE ||\\n> rel->relid != root->parse->resultRelation))\\n\\nAh, you're right. inhTargetKind has to be checked in conjunction with\\nchecking whether the relation is the target relation.\\n\\n> but I find that pretty messy, plus it's violating the concept that we\\n> shouldn't be allowing messiness from inheritance_planner to leak into\\n> other places.\\n\\nI'm afraid that we'll have to live with this particular hack as long as we\\nhave inheritance_planner(), but we maybe could somewhat reduce the extent\\nto which the hack is spread into other planner files.\\n\\nHow about we move the part of get_relation_constraints() that loads the\\npartition constraint to its only caller\\nrelation_excluded_by_constraints()? If we do that, all uses of\\nroot->inhTargetKind will be confined to one place. Attached updated patch\\ndoes that.\\n\\n> What I'd rather do is have this test just read\\n> \\n> \\tif (rel->reloptkind == RELOPT_BASEREL)\\n> \\n> Making it be that way causes some changes in the partition_prune results,\\n> as attached, which suggest that removing the enable_partition_pruning\\n> check as you did wasn't such a great idea either. However, if I add\\n> that back in, then it breaks the proposed new regression test case.\\n> \\n> I'm not at all clear on what we think the interaction between\\n> enable_partition_pruning and constraint_exclusion ought to be,\\n> so I'm not sure what the appropriate resolution is here. Thoughts?\\n\\nPrior to 428b260f87 (that is, in PG 11), partition pruning for UPDATE and\\nDELETE queries is realized by applying constraint exclusion to the\\npartition constraint of the target partition. The conclusion of the\\ndiscussion when adding the enable_partition_pruning GUC [1] was that\\nwhether or not constraint exclusion is carried out (to facilitate\\npartition pruning) should be governed by the new GUC, not\\nconstraint_exclusion, if only to avoid confusing users.\\n\\n428b260f87 has obviated the need to check enable_partition_pruning in\\nrelation_excluded_by_constraints(), because inheritance_planner() now runs\\nthe query as if it were SELECT, which does partition pruning using\\npartprune.c, governed by the setting of enable_partition_pruning. So,\\nthere's no need to check it again in relation_excluded_by_constraints(),\\nbecause we won't be consulting the partition constraint again; well, at\\nleast after applying the proposed patch.\\n\\n> BTW, just about all the other uses of root->inhTargetKind seem equally\\n> broken from here; none of them are accounting for whether the rel in\\n> question is the query target.\\n\\nThere's only one other use of its value, AFAICS:\\n\\n switch (constraint_exclusion)\\n {\\n case CONSTRAINT_EXCLUSION_OFF:\\n\\n /*\\n * Don't prune if feature turned off -- except if the relation is\\n * a partition. While partprune.c-style partition pruning is not\\n * yet in use for all cases (update/delete is not handled), it\\n * would be a UI horror to use different user-visible controls\\n * depending on such a volatile implementation detail. Therefore,\\n * for partitioned tables we use enable_partition_pruning to\\n * control this behavior.\\n */\\n if (root->inhTargetKind == INHKIND_PARTITIONED)\\n break;\\n\\nUpdated patch removes it though. Which other uses are there?\\n\\nAttached patch is only for HEAD this time. I'll post one for PG 11 (if\\nyou'd like) once we reach consensus on the best thing to do here is.\\n\\nThanks,\\nAmit\\n\\n[1]\\nhttps://www.postgresql.org/message-id/flat/CAFjFpRcwq7G16J_w%2Byy_xiE7daD0Bm6iYTnhz81f79yrSOn4DA%40mail.gmail.com\",\n \"msg_date\": \"Tue, 2 Apr 2019 14:50:59 +0900\",\n \"msg_from\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Hi all,\\n\\nFirst of all I would like to thank everyone involved in this patch for their hard work on this. This is a big step forward. I've done some performance and functionality testing with the patch that was committed to master and it looks very good.\\nI had a question about the performance of pruning of functions like now() and current_date. I know these are handled differently, as they cannot be excluded during the first phases of planning. However, curerntly, this new patch makes the performance difference between the static timestamp variant and now() very obvious (even more than before). Consider\\nselect * from partitioned_table where ts >= now()\\nor\\nselect * from partitioned_table where ts >= '2019-04-04'\\n\\nThe second plans in less than a millisecond, whereas the first takes +- 180ms for a table with 1000 partitions. Both end up with the same plan.\\n\\nI'm not too familiar with the code that handles this, but is there a possibility for improvement in this area? Or is the stage at which exclusion for now()/current_date occurs already too far in the process to make any good improvements to this? My apologies if this is considered off-topic for this patch, but I ran into this issue specifically when I was testing this patch, so I thought I'd ask here about it. I do think a large number of use-cases for tables with a large number of partitions involve a timestamp for partition key, and naturally people will start writing queries for this that use functions such as now() and current_date.\\n\\nThanks again for your work on this patch!\\n\\n-Floris\\n\\n________________________________________\\nFrom: Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>\\nSent: Tuesday, April 2, 2019 7:50 AM\\nTo: Tom Lane\\nCc: David Rowley; Imai Yoshikazu; jesper.pedersen@redhat.com; Imai, Yoshikazu; Amit Langote; Alvaro Herrera; Robert Haas; Justin Pryzby; Pg Hackers\\nSubject: Re: speeding up planning with partitions [External]\\n\\nThanks for taking a look.\\n\\nOn 2019/04/02 2:34, Tom Lane wrote:\\n> Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> writes:\\n>> On 2019/03/30 0:29, Tom Lane wrote:\\n>>> That seems like probably an independent patch --- do you want to write it?\\n>\\n>> Here is that patch.\\n>> It revises get_relation_constraints() such that the partition constraint\\n>> is loaded in only the intended cases.\\n>\\n> So I see the problem you're trying to solve here, but I don't like this\\n> patch a bit, because it depends on root->inhTargetKind which IMO is a\\n> broken bit of junk that we need to get rid of. Here is an example of\\n> why, with this patch applied:\\n>\\n> regression=# create table p (a int) partition by list (a);\\n> CREATE TABLE\\n> regression=# create table p1 partition of p for values in (1);\\n> CREATE TABLE\\n> regression=# set constraint_exclusion to on;\\n> SET\\n> regression=# explain select * from p1 where a = 2;\\n> QUERY PLAN\\n> ------------------------------------------\\n> Result (cost=0.00..0.00 rows=0 width=0)\\n> One-Time Filter: false\\n> (2 rows)\\n>\\n> So far so good, but watch what happens when we include the same case\\n> in an UPDATE on some other partitioned table:\\n>\\n> regression=# create table prtab (a int, b int) partition by list (a);\\n> CREATE TABLE\\n> regression=# create table prtab2 partition of prtab for values in (2);\\n> CREATE TABLE\\n> regression=# explain update prtab set b=b+1 from p1 where prtab.a=p1.a and p1.a=2;\\n> QUERY PLAN\\n> ---------------------------------------------------------------------------\\n> Update on prtab (cost=0.00..82.30 rows=143 width=20)\\n> Update on prtab2\\n> -> Nested Loop (cost=0.00..82.30 rows=143 width=20)\\n> -> Seq Scan on p1 (cost=0.00..41.88 rows=13 width=10)\\n> Filter: (a = 2)\\n> -> Materialize (cost=0.00..38.30 rows=11 width=14)\\n> -> Seq Scan on prtab2 (cost=0.00..38.25 rows=11 width=14)\\n> Filter: (a = 2)\\n> (8 rows)\\n>\\n> No constraint exclusion, while in v10 you get\\n>\\n> Update on prtab (cost=0.00..0.00 rows=0 width=0)\\n> -> Result (cost=0.00..0.00 rows=0 width=0)\\n> One-Time Filter: false\\n>\\n> The reason is that this logic supposes that root->inhTargetKind describes\\n> *all* partitioned tables in the query, which is obviously wrong.\\n>\\n> Now maybe we could make it work by doing something like\\n>\\n> if (rel->reloptkind == RELOPT_BASEREL &&\\n> (root->inhTargetKind == INHKIND_NONE ||\\n> rel->relid != root->parse->resultRelation))\\n\\nAh, you're right. inhTargetKind has to be checked in conjunction with\\nchecking whether the relation is the target relation.\\n\\n> but I find that pretty messy, plus it's violating the concept that we\\n> shouldn't be allowing messiness from inheritance_planner to leak into\\n> other places.\\n\\nI'm afraid that we'll have to live with this particular hack as long as we\\nhave inheritance_planner(), but we maybe could somewhat reduce the extent\\nto which the hack is spread into other planner files.\\n\\nHow about we move the part of get_relation_constraints() that loads the\\npartition constraint to its only caller\\nrelation_excluded_by_constraints()? If we do that, all uses of\\nroot->inhTargetKind will be confined to one place. Attached updated patch\\ndoes that.\\n\\n> What I'd rather do is have this test just read\\n>\\n> if (rel->reloptkind == RELOPT_BASEREL)\\n>\\n> Making it be that way causes some changes in the partition_prune results,\\n> as attached, which suggest that removing the enable_partition_pruning\\n> check as you did wasn't such a great idea either. However, if I add\\n> that back in, then it breaks the proposed new regression test case.\\n>\\n> I'm not at all clear on what we think the interaction between\\n> enable_partition_pruning and constraint_exclusion ought to be,\\n> so I'm not sure what the appropriate resolution is here. Thoughts?\\n\\nPrior to 428b260f87 (that is, in PG 11), partition pruning for UPDATE and\\nDELETE queries is realized by applying constraint exclusion to the\\npartition constraint of the target partition. The conclusion of the\\ndiscussion when adding the enable_partition_pruning GUC [1] was that\\nwhether or not constraint exclusion is carried out (to facilitate\\npartition pruning) should be governed by the new GUC, not\\nconstraint_exclusion, if only to avoid confusing users.\\n\\n428b260f87 has obviated the need to check enable_partition_pruning in\\nrelation_excluded_by_constraints(), because inheritance_planner() now runs\\nthe query as if it were SELECT, which does partition pruning using\\npartprune.c, governed by the setting of enable_partition_pruning. So,\\nthere's no need to check it again in relation_excluded_by_constraints(),\\nbecause we won't be consulting the partition constraint again; well, at\\nleast after applying the proposed patch.\\n\\n> BTW, just about all the other uses of root->inhTargetKind seem equally\\n> broken from here; none of them are accounting for whether the rel in\\n> question is the query target.\\n\\nThere's only one other use of its value, AFAICS:\\n\\n switch (constraint_exclusion)\\n {\\n case CONSTRAINT_EXCLUSION_OFF:\\n\\n /*\\n * Don't prune if feature turned off -- except if the relation is\\n * a partition. While partprune.c-style partition pruning is not\\n * yet in use for all cases (update/delete is not handled), it\\n * would be a UI horror to use different user-visible controls\\n * depending on such a volatile implementation detail. Therefore,\\n * for partitioned tables we use enable_partition_pruning to\\n * control this behavior.\\n */\\n if (root->inhTargetKind == INHKIND_PARTITIONED)\\n break;\\n\\nUpdated patch removes it though. Which other uses are there?\\n\\nAttached patch is only for HEAD this time. I'll post one for PG 11 (if\\nyou'd like) once we reach consensus on the best thing to do here is.\\n\\nThanks,\\nAmit\\n\\n[1]\\nhttps://www.postgresql.org/message-id/flat/CAFjFpRcwq7G16J_w%2Byy_xiE7daD0Bm6iYTnhz81f79yrSOn4DA%40mail.gmail.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 4 Apr 2019 18:33:55 +0000\",\n \"msg_from\": \"Floris Van Nee <florisvannee@Optiver.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On Fri, 5 Apr 2019 at 07:33, Floris Van Nee <florisvannee@optiver.com> wrote:\\n> I had a question about the performance of pruning of functions like now() and current_date. I know these are handled differently, as they cannot be excluded during the first phases of planning. However, curerntly, this new patch makes the performance difference between the static timestamp variant and now() very obvious (even more than before). Consider\\n> select * from partitioned_table where ts >= now()\\n> or\\n> select * from partitioned_table where ts >= '2019-04-04'\\n>\\n> The second plans in less than a millisecond, whereas the first takes +- 180ms for a table with 1000 partitions. Both end up with the same plan.\\n\\nThe patch here only aims to improve the performance of queries to\\npartitioned tables when partitions can be pruned during planning. The\\nnow() version of the query is unable to do that since we don't know\\nwhat that value will be during the execution of the query. In that\\nversion, you're most likely seeing \\\"Subplans Removed: <n>\\\". This means\\nrun-time pruning did some pruning and the planner generated subplans\\nfor what you see plus <n> others. Since planning for all partitions is\\nstill slow, you're getting a larger performance difference than\\nbefore, but only due to the fact that the other plan is now faster to\\ngenerate.\\n\\nIf you're never using prepared statements, i.e, always planning right\\nbefore execution, then you might want to consider using \\\"where ts >=\\n'today'::timestamp\\\". This will evaluate to the current date during\\nparse and make the value available to the planner. You'll need to be\\npretty careful with that though, as if you do prepare queries or\\nchange to do that in the future then the bugs in your application\\ncould be very subtle and only do the wrong thing just after midnight\\non some day when the current time progresses over your partition\\nboundary.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 Apr 2019 10:59:21 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On 2019/04/05 6:59, David Rowley wrote:\\n> On Fri, 5 Apr 2019 at 07:33, Floris Van Nee <florisvannee@optiver.com> wrote:\\n>> I had a question about the performance of pruning of functions like now() and current_date. I know these are handled differently, as they cannot be excluded during the first phases of planning. However, curerntly, this new patch makes the performance difference between the static timestamp variant and now() very obvious (even more than before). Consider\\n>> select * from partitioned_table where ts >= now()\\n>> or\\n>> select * from partitioned_table where ts >= '2019-04-04'\\n>>\\n>> The second plans in less than a millisecond, whereas the first takes +- 180ms for a table with 1000 partitions. Both end up with the same plan.\\n> \\n> The patch here only aims to improve the performance of queries to\\n> partitioned tables when partitions can be pruned during planning. The\\n> now() version of the query is unable to do that since we don't know\\n> what that value will be during the execution of the query. In that\\n> version, you're most likely seeing \\\"Subplans Removed: <n>\\\". This means\\n> run-time pruning did some pruning and the planner generated subplans\\n> for what you see plus <n> others. Since planning for all partitions is\\n> still slow, you're getting a larger performance difference than\\n> before, but only due to the fact that the other plan is now faster to\\n> generate.\\n\\nYeah, the time for generating plan for a query that *can* use pruning but\\nnot during planning is still very much dependent on the number of\\npartitions, because access plans must be created for all partitions, even\\nif only one of those plans will actually be used and the rest pruned away\\nduring execution.\\n\\n> If you're never using prepared statements,\\n\\nOr if using prepared statements is an option, the huge planning cost\\nmentioned above need not be paid repeatedly. Although, we still have ways\\nto go in terms of scaling generic plan execution to larger partition\\ncounts, solution(s) for which have been proposed by David but haven't made\\nit into master yet.\\n\\nThanks,\\nAmit\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 Apr 2019 12:09:16 +0900\",\n \"msg_from\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On Fri, 5 Apr 2019 at 16:09, Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> wrote:\\n> Although, we still have ways\\n> to go in terms of scaling generic plan execution to larger partition\\n> counts, solution(s) for which have been proposed by David but haven't made\\n> it into master yet.\\n\\nIs that a reference to the last paragraph in [1]? That idea has not\\ngone beyond me writing that text yet! :-( It was more of a passing\\ncomment on the only way I could think of to solve the problem.\\n\\n[1] https://www.postgresql.org/message-id/CAKJS1f-y1HQK+VjG7=C==vGcLnzxjN8ysD5NmaN8Wh4=VsYipw@mail.gmail.com\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 Apr 2019 16:18:41 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On 2019/04/05 12:18, David Rowley wrote:\\n> On Fri, 5 Apr 2019 at 16:09, Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> wrote:\\n>> Although, we still have ways\\n>> to go in terms of scaling generic plan execution to larger partition\\n>> counts, solution(s) for which have been proposed by David but haven't made\\n>> it into master yet.\\n> \\n> Is that a reference to the last paragraph in [1]? That idea has not\\n> gone beyond me writing that text yet! :-( It was more of a passing\\n> comment on the only way I could think of to solve the problem.\\n> \\n> [1] https://www.postgresql.org/message-id/CAKJS1f-y1HQK+VjG7=C==vGcLnzxjN8ysD5NmaN8Wh4=VsYipw@mail.gmail.com\\n\\nActually, I meant to refer to the following:\\n\\nhttps://commitfest.postgresql.org/22/1897/\\n\\nOf course, we should pursue all available options. :)\\n\\nThanks,\\nAmit\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 Apr 2019 12:25:15 +0900\",\n \"msg_from\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On 2019/04/02 14:50, Amit Langote wrote:\\n> Attached patch is only for HEAD this time. I'll post one for PG 11 (if\\n> you'd like) once we reach consensus on the best thing to do here is.\\n\\nWhile we're on the topic of the relation between constraint exclusion and\\npartition pruning, I'd like to (re-) propose this documentation update\\npatch. The partitioning chapter in ddl.sgml says update/delete of\\npartitioned tables uses constraint exclusion internally to emulate\\npartition pruning, which is no longer true as of 428b260f8.\\n\\nThe v2-0001 patch hasn't changed.\\n\\nThanks,\\nAmit\",\n \"msg_date\": \"Fri, 5 Apr 2019 15:50:25 +0900\",\n \"msg_from\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Thanks for the details! Indeed the versions with now()/current_date use the runtime pruning rather than planning time. I wasn't aware of the use of 'today' though - that could be useful in case we're sure statements won't be prepared.\\n\\nPreviously (v10/ partly v11) it was necessary to make sure that statements on partioned tables were never prepared, because run-time pruning wasn't available - using a generic plan was almost always a bad option. Now in v12 it seems to be a tradeoff between whether or not run-time pruning can occur. If pruning is possible at planning time it's probably still better not to prepare statements, whereas if run-time pruning has to occur, it's better to prepare them.\\n\\nOne unrelated thing I noticed (but I'm not sure if it's worth a separate email thread) is that the changed default of jit=on in v12 doesn't work very well with a large number of partitions at run-time, for which a large number get excluded at run-time. A query that has an estimated cost above jit_optimize_above_cost takes about 30 seconds to run (for a table with 1000 partitions), because JIT is optimizing the full plan. Without JIT it's barely 20ms (+400ms planning). I can give more details in a separate thread if it's deemed interesting.\\n\\nPlanning Time: 411.321 ms\\nJIT:\\n Functions: 5005\\n Options: Inlining false, Optimization true, Expressions true, Deforming true\\n Timing: Generation 721.472 ms, Inlining 0.000 ms, Optimization 16312.195 ms, Emission 12533.611 ms, Total 29567.278 ms\\n\\n-Floris\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 Apr 2019 09:13:29 +0000\",\n \"msg_from\": \"Floris Van Nee <florisvannee@Optiver.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2019/04/05 18:13, Floris Van Nee wrote:\\n> One unrelated thing I noticed (but I'm not sure if it's worth a separate email thread) is that the changed default of jit=on in v12 doesn't work very well with a large number of partitions at run-time, for which a large number get excluded at run-time. A query that has an estimated cost above jit_optimize_above_cost takes about 30 seconds to run (for a table with 1000 partitions), because JIT is optimizing the full plan. Without JIT it's barely 20ms (+400ms planning). I can give more details in a separate thread if it's deemed interesting.\\n> \\n> Planning Time: 411.321 ms\\n> JIT:\\n> Functions: 5005\\n> Options: Inlining false, Optimization true, Expressions true, Deforming true\\n> Timing: Generation 721.472 ms, Inlining 0.000 ms, Optimization 16312.195 ms, Emission 12533.611 ms, Total 29567.278 ms\\n\\nI've noticed a similar problem but in the context of interaction with\\nparallel query mechanism. The planner, seeing that all partitions will be\\nscanned (after failing to prune with clauses containing CURRENT_TIMESTAMP\\netc.), prepares a parallel plan (containing Parallel Append in this case).\\n As you can imagine, parallel query initialization (Gather+workers) will\\ntake large amount of time relative to the time it will take to scan the\\npartitions that remain after pruning (often just one).\\n\\nThe problem in this case is that the planner is oblivious to the\\npossibility of partition pruning occurring during execution, which may be\\ncommon to both parallel query and JIT. If it wasn't oblivious, it\\nwould've set the cost of pruning-capable Append such that parallel query\\nand/or JIT won't be invoked. We are going to have to fix that sooner or\\nlater.\\n\\nThanks,\\nAmit\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 Apr 2019 19:07:17 +0900\",\n \"msg_from\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On Fri, 5 Apr 2019 at 23:07, Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> wrote:\\n>\\n> Hi,\\n>\\n> On 2019/04/05 18:13, Floris Van Nee wrote:\\n> > One unrelated thing I noticed (but I'm not sure if it's worth a separate email thread) is that the changed default of jit=on in v12 doesn't work very well with a large number of partitions at run-time, for which a large number get excluded at run-time. A query that has an estimated cost above jit_optimize_above_cost takes about 30 seconds to run (for a table with 1000 partitions), because JIT is optimizing the full plan. Without JIT it's barely 20ms (+400ms planning). I can give more details in a separate thread if it's deemed interesting.\\n> >\\n> > Planning Time: 411.321 ms\\n> > JIT:\\n> > Functions: 5005\\n> > Options: Inlining false, Optimization true, Expressions true, Deforming true\\n> > Timing: Generation 721.472 ms, Inlining 0.000 ms, Optimization 16312.195 ms, Emission 12533.611 ms, Total 29567.278 ms\\n>\\n> I've noticed a similar problem but in the context of interaction with\\n> parallel query mechanism. The planner, seeing that all partitions will be\\n> scanned (after failing to prune with clauses containing CURRENT_TIMESTAMP\\n> etc.), prepares a parallel plan (containing Parallel Append in this case).\\n> As you can imagine, parallel query initialization (Gather+workers) will\\n> take large amount of time relative to the time it will take to scan the\\n> partitions that remain after pruning (often just one).\\n>\\n> The problem in this case is that the planner is oblivious to the\\n> possibility of partition pruning occurring during execution, which may be\\n> common to both parallel query and JIT. If it wasn't oblivious, it\\n> would've set the cost of pruning-capable Append such that parallel query\\n> and/or JIT won't be invoked. We are going to have to fix that sooner or\\n> later.\\n\\nRobert and I had a go at discussing this in [1]. Some ideas were\\nthrown around in the nature of contorting the Append/MergeAppend's\\ntotal_cost in a similar way to how clauselist_selectivity does its\\nestimates for unknown values. Perhaps it is possible to actually\\nmultiplying the total_cost by the clauselist_selectivity for the\\nrun-time pruning quals. That's pretty crude and highly unusual, but\\nit's probably going to give something more sane than what's there\\ntoday. The run-time prune quals would likely need to be determined\\nearlier than createplan.c for that to work though. IIRC the reason it\\nwas done there is, because at the time, there wasn't a need to do it\\nper path.\\n\\nI don't really have any better ideas right now, so if someone does\\nthen maybe we should take it up on a new thread. It would be good to\\nleave this thread alone for unrelated things. It's long enough\\nalready.\\n\\n[1] https://www.postgresql.org/message-id/CA%2BTgmobhXJGMuHxKjbaKcEJXacxVZHG4%3DhEGFfPF_FrGt37T_Q%40mail.gmail.com\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 Apr 2019 23:47:47 +1300\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On Fri, 5 Apr 2019 at 19:50, Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> wrote:\\n> While we're on the topic of the relation between constraint exclusion and\\n> partition pruning, I'd like to (re-) propose this documentation update\\n> patch. The partitioning chapter in ddl.sgml says update/delete of\\n> partitioned tables uses constraint exclusion internally to emulate\\n> partition pruning, which is no longer true as of 428b260f8.\\n\\nUpdate-docs-that-update-delete-no-longer-use-cons.patch looks good to\\nme. It should be changed as what the docs say is no longer true.\\n\\n-- \\n David Rowley http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Apr 2019 17:03:07 +1200\",\n \"msg_from\": \"David Rowley <david.rowley@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On 2019/04/11 14:03, David Rowley wrote:\\n> On Fri, 5 Apr 2019 at 19:50, Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> wrote:\\n>> While we're on the topic of the relation between constraint exclusion and\\n>> partition pruning, I'd like to (re-) propose this documentation update\\n>> patch. The partitioning chapter in ddl.sgml says update/delete of\\n>> partitioned tables uses constraint exclusion internally to emulate\\n>> partition pruning, which is no longer true as of 428b260f8.\\n> \\n> Update-docs-that-update-delete-no-longer-use-cons.patch looks good to\\n> me. It should be changed as what the docs say is no longer true.\\n\\nThanks for the quick review. :)\\n\\nRegards,\\nAmit\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Apr 2019 14:09:14 +0900\",\n \"msg_from\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> writes:\\n> On 2019/04/02 2:34, Tom Lane wrote:\\n>> I'm not at all clear on what we think the interaction between\\n>> enable_partition_pruning and constraint_exclusion ought to be,\\n>> so I'm not sure what the appropriate resolution is here. Thoughts?\\n\\n> Prior to 428b260f87 (that is, in PG 11), partition pruning for UPDATE and\\n> DELETE queries is realized by applying constraint exclusion to the\\n> partition constraint of the target partition. The conclusion of the\\n> discussion when adding the enable_partition_pruning GUC [1] was that\\n> whether or not constraint exclusion is carried out (to facilitate\\n> partition pruning) should be governed by the new GUC, not\\n> constraint_exclusion, if only to avoid confusing users.\\n\\nI got back to thinking about how this ought to work. It appears to me\\nthat we've got half a dozen different behaviors that depend on one or both\\nof these settings:\\n\\n1. Use of ordinary table constraints (CHECK, NOT NULL) in baserel pruning,\\n that is relation_excluded_by_constraints for baserels.\\n This is enabled by constraint_exclusion = on.\\n\\n2. Use of partition constraints in baserel pruning (applicable only\\n when a partition is accessed directly).\\n This is currently partly broken, and it's what your patch wants to\\n change.\\n\\n3. Use of ordinary table constraints in appendrel pruning,\\n that is relation_excluded_by_constraints for appendrel members.\\n This is enabled by constraint_exclusion >= partition.\\n\\n4. Use of partition constraints in appendrel pruning.\\n This is enabled by the combination of enable_partition_pruning AND\\n constraint_exclusion >= partition. However, it looks to me like this\\n is now nearly if not completely useless because of #5.\\n\\n5. Use of partition constraints in expand_partitioned_rtentry.\\n Enabled by enable_partition_pruning (see prune_append_rel_partitions).\\n\\n6. Use of partition constraints in run-time partition pruning.\\n This is also enabled by enable_partition_pruning, cf\\n create_append_plan, create_merge_append_plan.\\n\\nNow in addition to what I mention above, there are assorted random\\ndifferences in behavior depending on whether we are in an inherited\\nUPDATE/DELETE or not. I consider these differences to be so bogus\\nthat I'm not even going to include them in this taxonomy; they should\\nnot exist. The UPDATE/DELETE target ought to act the same as a baserel.\\n\\nI think this is ridiculously overcomplicated even without said random\\ndifferences. I propose that we do the following:\\n\\n* Get rid of point 4 by not considering partition constraints for\\nappendrel members in relation_excluded_by_constraints. It's just\\nuseless cycles in view of point 5, or nearly so. (Possibly there\\nare corner cases where we could prove contradictions between a\\nrelation's partition constraints and regular constraints ... but is\\nit really worth spending planner cycles to look for that?)\\n\\n* Make point 2 like point 1 by treating partition constraints for\\nbaserels like ordinary table constraints, ie, they are considered\\nonly when constraint_exclusion = on (independently of whether\\nenable_partition_pruning is on).\\n\\n* Treat an inherited UPDATE/DELETE target table as if it were an\\nappendrel member for the purposes of relation_excluded_by_constraints,\\nthus removing the behavioral differences between SELECT and UPDATE/DELETE.\\n\\nWith this, constraint_exclusion would act pretty much as it traditionally\\nhas, and in most cases would not have any special impact on partitions\\ncompared to old-style inheritance. The behaviors that\\nenable_partition_pruning would control are expand_partitioned_rtentry\\npruning and run-time pruning, neither of which have any applicability to\\nold-style inheritance.\\n\\nThoughts?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 22 Apr 2019 18:08:14 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On 2019/04/23 7:08, Tom Lane wrote:\\n> Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> writes:\\n>> On 2019/04/02 2:34, Tom Lane wrote:\\n>>> I'm not at all clear on what we think the interaction between\\n>>> enable_partition_pruning and constraint_exclusion ought to be,\\n>>> so I'm not sure what the appropriate resolution is here. Thoughts?\\n> \\n>> Prior to 428b260f87 (that is, in PG 11), partition pruning for UPDATE and\\n>> DELETE queries is realized by applying constraint exclusion to the\\n>> partition constraint of the target partition. The conclusion of the\\n>> discussion when adding the enable_partition_pruning GUC [1] was that\\n>> whether or not constraint exclusion is carried out (to facilitate\\n>> partition pruning) should be governed by the new GUC, not\\n>> constraint_exclusion, if only to avoid confusing users.\\n> \\n> I got back to thinking about how this ought to work.\\n\\nThanks a lot for taking time to look at this.\\n\\n> It appears to me\\n> that we've got half a dozen different behaviors that depend on one or both\\n> of these settings:\\n> \\n> 1. Use of ordinary table constraints (CHECK, NOT NULL) in baserel pruning,\\n> that is relation_excluded_by_constraints for baserels.\\n> This is enabled by constraint_exclusion = on.\\n> \\n> 2. Use of partition constraints in baserel pruning (applicable only\\n> when a partition is accessed directly).\\n> This is currently partly broken, and it's what your patch wants to\\n> change.\\n\\nYes. Any fix we come up with for this will need to be back-patched to 11,\\nbecause it's a regression introduced in 11 when the then new partition\\npruning feature was committed (9fdb675fc).\\n\\n> 3. Use of ordinary table constraints in appendrel pruning,\\n> that is relation_excluded_by_constraints for appendrel members.\\n> This is enabled by constraint_exclusion >= partition.\\n> \\n> 4. Use of partition constraints in appendrel pruning.\\n> This is enabled by the combination of enable_partition_pruning AND\\n> constraint_exclusion >= partition. However, it looks to me like this\\n> is now nearly if not completely useless because of #5.\\n> \\n> 5. Use of partition constraints in expand_partitioned_rtentry.\\n> Enabled by enable_partition_pruning (see prune_append_rel_partitions).\\n\\nRight, #5 obviates #4.\\n\\n> 6. Use of partition constraints in run-time partition pruning.\\n> This is also enabled by enable_partition_pruning, cf\\n> create_append_plan, create_merge_append_plan.\\n> \\n> Now in addition to what I mention above, there are assorted random\\n> differences in behavior depending on whether we are in an inherited\\n> UPDATE/DELETE or not. I consider these differences to be so bogus\\n> that I'm not even going to include them in this taxonomy; they should\\n> not exist. The UPDATE/DELETE target ought to act the same as a baserel.\\n\\nThe *partition* constraint of UPDATE/DELETE targets would never be refuted\\nby the query, because we process only those partition targets that remain\\nafter applying partition pruning during the initial planning of the query\\nas if it were SELECT. I'm saying we should distinguish such targets as\\nsuch when addressing #2.\\n\\nNot sure if you'll like it but maybe we could ignore even regular\\ninheritance child targets that are proven to be empty (is_dummy_rel()) for\\na given query during the initial SELECT planning. That way, we can avoid\\nre-running relation_excluded_by_constraints() a second time for *all*\\nchild target relations.\\n\\n> I think this is ridiculously overcomplicated even without said random\\n> differences. I propose that we do the following:\\n> \\n> * Get rid of point 4 by not considering partition constraints for\\n> appendrel members in relation_excluded_by_constraints. It's just\\n> useless cycles in view of point 5, or nearly so. (Possibly there\\n> are corner cases where we could prove contradictions between a\\n> relation's partition constraints and regular constraints ... but is\\n> it really worth spending planner cycles to look for that?)\\n\\nI guess not. If partition constraint contradicts regular constraints,\\nthere wouldn't be any data in such partitions to begin with, no?\\n\\n> * Make point 2 like point 1 by treating partition constraints for\\n> baserels like ordinary table constraints, ie, they are considered\\n> only when constraint_exclusion = on (independently of whether\\n> enable_partition_pruning is on).\\n\\nRight, enable_partition_pruning should only apply to appendrel pruning.\\nIf a partition is accessed directly and hence a baserel to the planner, we\\nonly consider constraint_exclusion and perform it only if the setting is on.\\n\\nAnother opinion on this is that we treat partition constraints differently\\nfrom regular constraints and don't mind the setting of\\nconstraint_exclusion, that is, always perform constraint exclusion using\\npartition constraints.\\n\\n> * Treat an inherited UPDATE/DELETE target table as if it were an\\n> appendrel member for the purposes of relation_excluded_by_constraints,\\n> thus removing the behavioral differences between SELECT and UPDATE/DELETE.\\n\\nAs I mentioned above, planner encounters any given UPDATE/DELETE *child*\\ntarget *twice*. Once during the initial SELECT planning and then again\\nduring when planning the query with a given child target relation as its\\nresultRelation. For partition targets, since the initial run only selects\\nthose that survive pruning, their partition constraint need not be\\nconsidered in the 2nd encounter as the query's baserel.\\n\\nAlso, it's during the 2nd encounter that we check inhTargetKind setting to\\ndistinguish partition target baserels from SELECT baserels. Its value is\\nINHKIND_INHERITED or INHKIND_PARTITIONED for the former, whereas it's\\nINHKIND_NONE for the latter.\\n\\n> With this, constraint_exclusion would act pretty much as it traditionally\\n> has, and in most cases would not have any special impact on partitions\\n> compared to old-style inheritance. The behaviors that\\n> enable_partition_pruning would control are expand_partitioned_rtentry\\n> pruning and run-time pruning, neither of which have any applicability to\\n> old-style inheritance.\\n\\nThat's right.\\n\\nDo you want me to update my patch considering the above summary?\\n\\nThanks,\\nAmit\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 23 Apr 2019 13:26:45 +0900\",\n \"msg_from\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> writes:\\n> On 2019/04/23 7:08, Tom Lane wrote:\\n>> [ a bunch of stuff ]\\n\\n> Not sure if you'll like it but maybe we could ignore even regular\\n> inheritance child targets that are proven to be empty (is_dummy_rel()) for\\n> a given query during the initial SELECT planning. That way, we can avoid\\n> re-running relation_excluded_by_constraints() a second time for *all*\\n> child target relations.\\n\\nMy thought was to keep traditional inheritance working more or less\\nas it has. To do what you're suggesting, we'd have to move generic\\nconstraint-exclusion logic up into the RTE expansion phase, and I don't\\nthink that's a particularly great idea. I think what we should be\\ndoing is applying partition pruning (which is a very specialized form\\nof constraint exclusion) during RTE expansion, then applying generic\\nconstraint exclusion in relation_excluded_by_constraints, and not\\nexamining partition constraints again there if we already used them.\\n\\n> Do you want me to update my patch considering the above summary?\\n\\nYes please. However, I wonder whether you're thinking differently in\\nlight of what you wrote in [1]:\\n\\n>>> Pruning in 10.2 works using internally generated partition constraints\\n>>> (which for this purpose are same as CHECK constraints). With the new\\n>>> pruning logic introduced in 11, planner no longer considers partition\\n>>> constraint because it's redundant to check them in most cases, because\\n>>> pruning would've selected the right set of partitions. Given that the new\\n>>> pruning logic is still unable to handle the cases like above, maybe we\\n>>> could change the planner to consider them, at least until we fix the\\n>>> pruning logic to handle such cases.\\n\\nIf we take that seriously then it would suggest not ignoring partition\\nconstraints in relation_excluded_by_constraints. However, I'm of the\\nopinion that we shouldn't let temporary deficiencies in the\\npartition-pruning logic drive what we do here. I don't think the set\\nof cases where we could get a win by reconsidering the partition\\nconstraints is large enough to justify the cycles expended in doing so;\\nand it'll get even smaller as pruning gets smarter.\\n\\n\\t\\t\\tregards, tom lane\\n\\n[1] https://www.postgresql.org/message-id/358cd54d-c018-60f8-7d76-55780eef6678@lab.ntt.co.jp\\n\\n\\n\",\n \"msg_date\": \"Sat, 27 Apr 2019 19:10:10 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On Sun, Apr 28, 2019 at 8:10 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> writes:\\n> > Not sure if you'll like it but maybe we could ignore even regular\\n> > inheritance child targets that are proven to be empty (is_dummy_rel()) for\\n> > a given query during the initial SELECT planning. That way, we can avoid\\n> > re-running relation_excluded_by_constraints() a second time for *all*\\n> > child target relations.\\n>\\n> My thought was to keep traditional inheritance working more or less\\n> as it has. To do what you're suggesting, we'd have to move generic\\n> constraint-exclusion logic up into the RTE expansion phase, and I don't\\n> think that's a particularly great idea. I think what we should be\\n> doing is applying partition pruning (which is a very specialized form\\n> of constraint exclusion) during RTE expansion, then applying generic\\n> constraint exclusion in relation_excluded_by_constraints, and not\\n> examining partition constraints again there if we already used them.\\n\\nJust to clarify, I wasn't suggesting that we change query_planner(),\\nbut the blocks in inheritance_planner() that perform initial planning\\nas if the query was SELECT and gather child target relations from that\\nplanner run; the following consecutive blocks:\\n\\n /*\\n * Before generating the real per-child-relation plans, do a cycle of\\n * planning as though the query were a SELECT.\\n...\\n */\\n {\\n PlannerInfo *subroot;\\nand:\\n\\n /*----------\\n * Since only one rangetable can exist in the final plan, we need to make\\n * sure that it contains all the RTEs needed for any child plan.\\n...\\n child_appinfos = NIL;\\n old_child_rtis = NIL;\\n new_child_rtis = NIL;\\n parent_relids = bms_make_singleton(top_parentRTindex);\\n foreach(lc, select_appinfos)\\n {\\n AppendRelInfo *appinfo = lfirst_node(AppendRelInfo, lc);\\n RangeTblEntry *child_rte;\\n\\n /* append_rel_list contains all append rels; ignore others */\\n if (!bms_is_member(appinfo->parent_relid, parent_relids))\\n continue;\\n\\n /* remember relevant AppendRelInfos for use below */\\n child_appinfos = lappend(child_appinfos, appinfo);\\n\\nI'm suggesting that we don't add the child relations that are dummy\\ndue to constraint exclusion to child_appinfos. Maybe, we'll need to\\ncombine the two blocks so that the latter can use the PlannerInfo\\ndefined in the former to look up the child relation to check if dummy.\\n\\n> > Do you want me to update my patch considering the above summary?\\n>\\n> Yes please.\\n\\nI will try to get that done hopefully by tomorrow.\\n\\n(On extended holidays that those of us who are in Japan have around\\nthis time of year.)\\n\\n> However, I wonder whether you're thinking differently in\\n> light of what you wrote in [1]:\\n\\nThanks for checking that thread.\\n\\n> >>> Pruning in 10.2 works using internally generated partition constraints\\n> >>> (which for this purpose are same as CHECK constraints). With the new\\n> >>> pruning logic introduced in 11, planner no longer considers partition\\n> >>> constraint because it's redundant to check them in most cases, because\\n> >>> pruning would've selected the right set of partitions. Given that the new\\n> >>> pruning logic is still unable to handle the cases like above, maybe we\\n> >>> could change the planner to consider them, at least until we fix the\\n> >>> pruning logic to handle such cases.\\n>\\n> If we take that seriously then it would suggest not ignoring partition\\n> constraints in relation_excluded_by_constraints. However, I'm of the\\n> opinion that we shouldn't let temporary deficiencies in the\\n> partition-pruning logic drive what we do here. I don't think the set\\n> of cases where we could get a win by reconsidering the partition\\n> constraints is large enough to justify the cycles expended in doing so;\\n> and it'll get even smaller as pruning gets smarter.\\n\\nYeah, maybe we could away with that by telling users to define\\nequivalent CHECK constraints for corner cases like that although\\nthat's not really great.\\n\\nThanks,\\nAmit\\n\\n\\n\",\n \"msg_date\": \"Mon, 29 Apr 2019 00:26:08 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Amit Langote <amitlangote09@gmail.com> writes:\\n> Here is the patch. I've also included the patch to update the text in\\n> ddl.sgml regarding constraint exclusion and partition pruning.\\n\\nI thought this was a bit messy. In particular, IMV the reason to\\nhave a split between get_relation_constraints and its only caller\\nrelation_excluded_by_constraints is to create a policy vs mechanism\\nseparation: relation_excluded_by_constraints figures out what kinds\\nof constraints we need to look at, while get_relation_constraints does\\nthe gruntwork of digging them out of the catalog data. Somebody had\\nalready ignored this principle to the extent of putting this\\nvery-much-policy test into get_relation_constraints:\\n\\n if (enable_partition_pruning && root->parse->commandType != CMD_SELECT)\\n\\nbut the way to improve that is to add another flag parameter to convey\\nthe policy choice, not to move the whole chunk of mechanism out to the\\ncaller.\\n\\nIt also struck me while looking at the code that we were being\\nunnecessarily stupid about non-inheritable constraints: rather than\\njust throwing up our hands for traditional inheritance situations,\\nwe can still apply constraint exclusion, as long as we consider only\\nconstraints that aren't marked ccnoinherit. (attnotnull constraints\\nhave to be considered as if they were ccnoinherit, for ordinary\\ntables but not partitioned ones.)\\n\\nSo, I propose the attached revised patch.\\n\\nI'm not sure how much of this, if anything, we should back-patch to\\nv11. It definitely doesn't seem like we should back-patch the\\nimprovement just explained. I tried diking out that change, as\\nin the v11 variant attached, and found that this still causes quite a\\nfew other changes in v11's expected results, most of them not for the\\nbetter. So I'm thinking that we'd better conclude that v11's ship\\nhas sailed. Its behavior is in some ways weird, but I am not sure\\nthat anyone will appreciate our changing it on the fourth minor\\nrelease.\\n\\nIt's somewhat interesting that we get these other changes in v11\\nbut not HEAD. I think the reason is that we reimplemented so much\\nof inheritance_planner in 428b260f8; that is, it seems the weird\\ndecisions we find in relation_excluded_by_constraints are mostly\\nthere to band-aid over the old weird behavior of inheritance_planner.\\n\\nAnyway, my current thought is to apply this to HEAD and do nothing\\nin v11. I include the v11 patch just for amusement. (I did not\\ncheck v11's behavior outside the core regression tests; it might\\npossibly have additional test diffs in contrib.)\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Mon, 29 Apr 2019 17:12:29 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"Amit Langote <amitlangote09@gmail.com> writes:\\n> On Tue, Apr 30, 2019 at 1:26 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n>> It would be nice if at least we fix the bug that directly accessed\\n>> partitions are not excluded with constraint_exclusion = on, without\\n>> removing PG 11's contortions in relation_excluded_by_constraints to\\n>> work around the odd requirements imposed by inheritance_planner, which\\n>> is causing the additional diffs in the regression expected output.\\n\\n> FWIW, attached is a delta patch that applies on top of your patch for\\n> v11 branch that shows what may be one way to go about this.\\n\\nOK, I tweaked that a bit and pushed both versions.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 30 Apr 2019 15:08:11 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n },\n {\n \"msg_contents\": \"On Wed, May 1, 2019 at 4:08 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> OK, I tweaked that a bit and pushed both versions.\\n\\nThank you.\\n\\nRegards,\\nAmit\\n\\n\\n\",\n \"msg_date\": \"Thu, 2 May 2019 18:33:12 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: speeding up planning with partitions\"\n }\n]","isConversational":false}}},{"rowIdx":47189,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"I've seen this message now a handful of times recently. It seems to happen\novernight, during a maintenance job which reindex things, including system\ncatalog indices.\n\nIt's easy to reproduce error under 10.5, but not under 10.3 nor 10.4.\n\nwhile :; do for a in pg_class_oid_index pg_class_relname_nsp_index pg_class_tblspc_relfilenode_index; do psql ts -qc \"REINDEX INDEX $a\"; done; done&\n\n[pryzbyj@database ~]$ while :; do psql ts -qc ''; done\npsql: FATAL: could not read block 1 in file \"base/16400/313430687\": read only 0 of 8192 bytes\npsql: FATAL: could not read block 0 in file \"base/16400/313430708\": read only 0 of 8192 bytes\npsql: FATAL: could not read block 0 in file \"base/16400/313430711\": read only 0 of 8192 bytes\n^C\n\npostgres=# SELECT * FROM postgres_log WHERE error_severity ='FATAL';\nlog_time | 2018-08-28 22:19:58.822-05\nuser_name | telsasoft\ndatabase | ts\npid | 22445\nconnection_from | 192.168.122.12:58318\nsession_id | 5b8610de.57ad\nsession_line | 1\ncommand_tag | startup\nsession_start_time | 2018-08-28 22:19:58-05\nvirtual_transaction_id | 26/280967\ntransaction_id | 0\nerror_severity | FATAL\nsql_state_code | XX001\nmessage | could not read block 0 in file \"base/16400/313316300\": read only 0 of 8192 bytes\n\nJustin\n\n","msg_date":"Wed, 29 Aug 2018 09:01:49 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":true,"msg_subject":"10.5 but not 10.4: backend startup during reindex system: could not\n read block 0 in file \"base/16400/..\": read only 0 of 8192 bytes"},{"msg_contents":"Justin Pryzby <pryzby@telsasoft.com> writes:\n> I've seen this message now a handful of times recently. It seems to happen\n> overnight, during a maintenance job which reindex things, including system\n> catalog indices.\n> It's easy to reproduce error under 10.5, but not under 10.3 nor 10.4.\n> while :; do for a in pg_class_oid_index pg_class_relname_nsp_index pg_class_tblspc_relfilenode_index; do psql ts -qc \"REINDEX INDEX $a\"; done; done&\n\nThis looks suspiciously like the issue under discussion in\n\nhttps://www.postgresql.org/message-id/12259.1532117714%40sss.pgh.pa.us\n\nAs far as we can tell, that bug is a dozen years old, so it's not clear\nwhy you find that you can reproduce it only in 10.5. But there might be\nsome subtle timing change accounting for that.\n\n\t\t\tregards, tom lane\n\n","msg_date":"Wed, 29 Aug 2018 11:35:51 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: 10.5 but not 10.4: backend startup during reindex system: could\n not read block 0 in file \"base/16400/..\": read only 0 of 8192 bytes"},{"msg_contents":"On Wed, Aug 29, 2018 at 11:35:51AM -0400, Tom Lane wrote:\n> Justin Pryzby <pryzby@telsasoft.com> writes:\n> > I've seen this message now a handful of times recently. It seems to happen\n> > overnight, during a maintenance job which reindex things, including system\n> > catalog indices.\n> > It's easy to reproduce error under 10.5, but not under 10.3 nor 10.4.\n> > while :; do for a in pg_class_oid_index pg_class_relname_nsp_index pg_class_tblspc_relfilenode_index; do psql ts -qc \"REINDEX INDEX $a\"; done; done&\n> > time while psql ts -qc ''; do :; done\n> \n> This looks suspiciously like the issue under discussion in\n> \n> https://www.postgresql.org/message-id/12259.1532117714%40sss.pgh.pa.us\n> \n> As far as we can tell, that bug is a dozen years old, so it's not clear\n> why you find that you can reproduce it only in 10.5. But there might be\n> some subtle timing change accounting for that.\n\nI confirmed the stack appears to be same as here, modulo differences between\n9.6 to 10.\nhttps://www.postgresql.org/message-id/fcab55ec-660a-bc43-7fe1-09af3665e89a%402ndquadrant.com\n\nIt seems to me there's one root problem occurring in (at least) two slightly\ndifferent ways. The issue/symptom that I've been seeing occurs in 10.5 but not\n10.4, and specifically at commit 2ce64ca, but not before. \n\n|commit 2ce64caaf793615d0f7a7ce4380e7255077d130c\n|Author: Andres Freund <andres@anarazel.de>\n|Date: Tue Jun 12 11:13:21 2018 -0700\n|\n|Fix bugs in vacuum of shared rels, by keeping their relcache entries current.\n\nHowever, the same message CAN happen under 10.4, though less easily, due to\nautovacuum but (afaict) not user queries. Maybe that's a coincidence, but my\nexperience mirrors the build logs: the problem since 10.5 is easily tickled in\na few seconds rather than an hour.\n\nThe issue I see frequently under 10.5 has this stack:\n\nCore was generated by `postgres: pryzbyj postgres [local] startup '.\n#2 0x0000000000709425 in mdread ()\n#3 0x00000000006e2bc8 in ReadBuffer_common ()\n#4 0x00000000006e3511 in ReadBufferExtended ()\n#5 0x00000000004c2da9 in _bt_getbuf ()\n#6 0x00000000004c32b2 in _bt_getroot ()\n#7 0x00000000004c799b in _bt_search ()\n#8 0x00000000004c8568 in _bt_first ()\n#9 0x00000000004c52c7 in btgettuple ()\n#10 0x00000000004bf936 in index_getnext_tid ()\n#11 0x00000000004bfb43 in index_getnext ()\n#12 0x00000000004beeae in systable_getnext ()\n#13 0x00000000007fa074 in ScanPgRelation ()\n#14 0x00000000007fe371 in RelationClearRelation ()\n#15 0x00000000007fe84b in RelationCacheInvalidateEntry ()\n#16 0x00000000007f7e15 in LocalExecuteInvalidationMessage ()\n#17 0x00000000006f3b38 in ReceiveSharedInvalidMessages ()\n#18 0x00000000006f6eb5 in LockRelationOid ()\n#19 0x00000000004ae495 in relation_open ()\n#20 0x00000000004bf336 in index_open ()\n#21 0x00000000004bee3c in systable_beginscan ()\n#22 0x00000000007fa069 in ScanPgRelation ()\n#23 0x00000000007fce84 in RelationBuildDesc ()\n#24 0x00000000007fe6f8 in RelationIdGetRelation ()\n#25 0x00000000004ae463 in relation_open ()\n#26 0x00000000004ae6a6 in heap_open ()\n#27 0x0000000000819237 in InitPostgres ()\n#28 0x000000000070e2e9 in PostgresMain ()\n#29 0x0000000000477f36 in ServerLoop ()\n#30 0x00000000006a5e3e in PostmasterMain ()\n#31 0x0000000000478a08 in main ()\n\n From autovacum running REL_10_4 I've now seen (since trying to tickle the\nproblem yesterday):\n\n#1 0x00007fbfb5b9d01a in __GI_abort () at abort.c:89\n#2 0x0000000000708489 in mdread ()\n#3 0x00000000006e1c48 in ReadBuffer_common ()\n#4 0x00000000006e2591 in ReadBufferExtended ()\n#5 0x00000000004c2369 in _bt_getbuf ()\n#6 0x00000000004c2810 in _bt_getroot ()\n#7 0x00000000004c6f5b in _bt_search ()\n#8 0x00000000004c7b28 in _bt_first ()\n#9 0x00000000004c4887 in btgettuple ()\n#10 0x00000000004bf00e in index_getnext_tid ()\n#11 0x00000000004bf1f3 in index_getnext ()\n#12 0x00000000004be679 in systable_getnext ()\n#13 0x00000000007f9084 in ScanPgRelation ()\n#14 0x00000000007fbe54 in RelationBuildDesc ()\n#15 0x00000000007fcc86 in RelationClearRelation ()\n#16 0x00000000007fd7bb in RelationCacheInvalidateEntry ()\n#17 0x00000000007f6e25 in LocalExecuteInvalidationMessage ()\n#18 0x00000000006f2bb8 in ReceiveSharedInvalidMessages ()\n#19 0x00000000006f5f35 in LockRelationOid ()\n#20 0x00000000004ade05 in relation_open ()\n#21 0x00000000004bea96 in index_open ()\n#22 0x00000000004be606 in systable_beginscan ()\n#23 0x00000000007f9079 in ScanPgRelation ()\n#24 0x00000000007fbe54 in RelationBuildDesc ()\n#25 0x00000000007fd668 in RelationIdGetRelation ()\n#26 0x00000000004addd3 in relation_open ()\n#27 0x00000000004bea96 in index_open ()\n#28 0x00000000004be606 in systable_beginscan ()\n#29 0x00000000007fef60 in RelationGetIndexList ()\n#30 0x00000000005dba6c in vac_open_indexes ()\n#31 0x00000000005dc652 in lazy_vacuum_rel ()\n#32 0x00000000005da7b4 in vacuum_rel ()\n#33 0x00000000005db661 in vacuum ()\n#34 0x00000000004770db in do_autovacuum ()\n#35 0x0000000000696242 in AutoVacWorkerMain.isra.6 ()\n#36 0x0000000000696a49 in StartAutoVacWorker ()\n#37 0x00000000006a3f0a in sigusr1_handler ()\n#38 <signal handler called>\n#39 0x00007fbfb5c62cf3 in __select_nocancel ()\n at ../sysdeps/unix/syscall-template.S:84\n#40 0x000000000047760e in ServerLoop ()\n#41 0x00000000006a4ebe in PostmasterMain ()\n#42 0x0000000000478a78 in main ()\n\nJustin\n\n","msg_date":"Thu, 30 Aug 2018 15:00:28 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":true,"msg_subject":"Re: 10.5 but not 10.4: backend startup during reindex system: could\n not read block 0 in file \"base/16400/..\": read only 0 of 8192 bytes"},{"msg_contents":"Hi,\n\nOn 2018-08-30 15:00:28 -0500, Justin Pryzby wrote:\n> On Wed, Aug 29, 2018 at 11:35:51AM -0400, Tom Lane wrote:\n> > This looks suspiciously like the issue under discussion in\n> > \n> > https://www.postgresql.org/message-id/12259.1532117714%40sss.pgh.pa.us\n> > \n> > As far as we can tell, that bug is a dozen years old, so it's not clear\n> > why you find that you can reproduce it only in 10.5. But there might be\n> > some subtle timing change accounting for that.\n> \n> I confirmed the stack appears to be same as here, modulo differences between\n> 9.6 to 10.\n\nCould you check if both of the proposed attempts at fixing the issue\nalso solves your problem?\n\nThanks,\n\nAndres Freund\n\n","msg_date":"Thu, 30 Aug 2018 13:18:59 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: 10.5 but not 10.4: backend startup during reindex system: could\n not read block 0 in file \"base/16400/..\": read only 0 of 8192 bytes"},{"msg_contents":"Justin Pryzby <pryzby@telsasoft.com> writes:\n> On Wed, Aug 29, 2018 at 11:35:51AM -0400, Tom Lane wrote:\n>> As far as we can tell, that bug is a dozen years old, so it's not clear\n>> why you find that you can reproduce it only in 10.5. But there might be\n>> some subtle timing change accounting for that.\n\n> It seems to me there's one root problem occurring in (at least) two slightly\n> different ways. The issue/symptom that I've been seeing occurs in 10.5 but not\n> 10.4, and specifically at commit 2ce64ca, but not before. \n\nYeah, as you probably saw in the other thread, we later realized that\n2ce64ca created an additional pathway for ScanPgRelation to recurse;\na pathway that's evidently easier to hit than the pre-existing ones.\nI note that both of your stack traces display ScanPgRelation recursion,\nso I'm feeling pretty confident that what you're seeing is the same\nthing.\n\nBut, as Andres says, it'd be great if you could confirm whether the\ndraft patches fix it for you.\n\n\t\t\tregards, tom lane\n\n","msg_date":"Thu, 30 Aug 2018 17:30:30 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: 10.5 but not 10.4: backend startup during reindex system: could\n not read block 0 in file \"base/16400/..\": read only 0 of 8192 bytes"},{"msg_contents":"On Thu, Aug 30, 2018 at 05:30:30PM -0400, Tom Lane wrote:\n> Justin Pryzby <pryzby@telsasoft.com> writes:\n> > On Wed, Aug 29, 2018 at 11:35:51AM -0400, Tom Lane wrote:\n> >> As far as we can tell, that bug is a dozen years old, so it's not clear\n> >> why you find that you can reproduce it only in 10.5. But there might be\n> >> some subtle timing change accounting for that.\n> \n> > It seems to me there's one root problem occurring in (at least) two slightly\n> > different ways. The issue/symptom that I've been seeing occurs in 10.5 but not\n> > 10.4, and specifically at commit 2ce64ca, but not before. \n> \n> Yeah, as you probably saw in the other thread, we later realized that\n> 2ce64ca created an additional pathway for ScanPgRelation to recurse;\n> a pathway that's evidently easier to hit than the pre-existing ones.\n> I note that both of your stack traces display ScanPgRelation recursion,\n> so I'm feeling pretty confident that what you're seeing is the same\n> thing.\n> \n> But, as Andres says, it'd be great if you could confirm whether the\n> draft patches fix it for you.\n\nI tested with relcache-rebuild.diff which hasn't broken in 15min, so I'm\nconfident that doesn't hit the additional recusive pathway, but have to wait\nawhile and see if autovacuum survives, too.\n\nI tried to apply fix-missed-inval-msg-accepts-1.patch on top of PG10.5 but\npatch didn't apply, so I can test HEAD after the first patch soaks awhile.\n\nJust curious, is there really any difficulty in reproducing this? Once I\nrealized this was a continuing issue and started to suspect pg10.5, it takes\njust about nothing to reproduce anywhere I've tried. I just tested 5 servers,\nand only one took more than a handful of seconds to fail. I gave up waiting\nfor a 6th server, because I found it was waiting on a pre-existing lock.\n\n[pryzbyj@database ~]$ while :; do for a in pg_class_oid_index pg_class_relname_nsp_index pg_class_tblspc_relfilenode_index; do psql ts -qc \"REINDEX INDEX $a\"; done; done&\n[pryzbyj@database ~]$ a=0; time while psql ts -qc ''; do a=$((1+a)); done ; echo \"$a\"\npsql: FATAL: could not read block 0 in file \"base/16400/313581263\": read only 0 of 8192 bytes\n\nreal 0m1.772s\nuser 0m0.076s\nsys 0m0.116s\n47\n\nJustin\n\n","msg_date":"Thu, 30 Aug 2018 16:57:12 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":true,"msg_subject":"Re: 10.5 but not 10.4: backend startup during reindex system: could\n not read block 0 in file \"base/16400/..\": read only 0 of 8192 bytes"},{"msg_contents":"Justin Pryzby <pryzby@telsasoft.com> writes:\n> Just curious, is there really any difficulty in reproducing this?\n\nOnce you have the right test case, it's not hard. But it took us two\nmonths to find one ...\n\n\t\t\tregards, tom lane\n\n","msg_date":"Thu, 30 Aug 2018 18:03:27 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: 10.5 but not 10.4: backend startup during reindex system: could\n not read block 0 in file \"base/16400/..\": read only 0 of 8192 bytes"},{"msg_contents":"On Thu, Aug 30, 2018 at 01:25:00PM -0700, Andres Freund wrote:\n> On August 30, 2018 1:24:12 PM PDT, Justin Pryzby <pryzby@telsasoft.com> wrote:\n> >On Thu, Aug 30, 2018 at 01:18:59PM -0700, Andres Freund wrote:\n> >> Could you check if both of the proposed attempts at fixing the issue\n> >> also solves your problem?\n> >\n> >Just checking that you're referring to:\n> >\n> >1) relcache-rebuild.diff\n> >https://www.postgresql.org/message-id/20180829083730.n645apqhb2gyih3g%40alap3.anarazel.de\n> >2) fix-missed-inval-msg-accepts-1.patch\n> >https://www.postgresql.org/message-id/25024.1535579899%40sss.pgh.pa.us\n> \n> Right. Either should fix the issue.\n\nConfirmed that relcache-rebuild.diff seems to fix/avoid the issue under 10.5,\nand fix-missed-inval-msg-accepts-1.patch does so for HEAD (but doesn't apply\ncleanly to 10.5 so not tested). That's for both the the new easy-to-hit\nsymptom and the pre-existing hard-to-hit issue affecting pg10.4 and earlier.\n\nThanks,\nJustin\n\n","msg_date":"Fri, 31 Aug 2018 03:08:37 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":true,"msg_subject":"Re: 10.5 but not 10.4: backend startup during reindex system: could\n not read block 0 in file \"base/16400/..\": read only 0 of 8192 bytes"}],"string":"[\n {\n \"msg_contents\": \"I've seen this message now a handful of times recently. It seems to happen\\novernight, during a maintenance job which reindex things, including system\\ncatalog indices.\\n\\nIt's easy to reproduce error under 10.5, but not under 10.3 nor 10.4.\\n\\nwhile :; do for a in pg_class_oid_index pg_class_relname_nsp_index pg_class_tblspc_relfilenode_index; do psql ts -qc \\\"REINDEX INDEX $a\\\"; done; done&\\n\\n[pryzbyj@database ~]$ while :; do psql ts -qc ''; done\\npsql: FATAL: could not read block 1 in file \\\"base/16400/313430687\\\": read only 0 of 8192 bytes\\npsql: FATAL: could not read block 0 in file \\\"base/16400/313430708\\\": read only 0 of 8192 bytes\\npsql: FATAL: could not read block 0 in file \\\"base/16400/313430711\\\": read only 0 of 8192 bytes\\n^C\\n\\npostgres=# SELECT * FROM postgres_log WHERE error_severity ='FATAL';\\nlog_time | 2018-08-28 22:19:58.822-05\\nuser_name | telsasoft\\ndatabase | ts\\npid | 22445\\nconnection_from | 192.168.122.12:58318\\nsession_id | 5b8610de.57ad\\nsession_line | 1\\ncommand_tag | startup\\nsession_start_time | 2018-08-28 22:19:58-05\\nvirtual_transaction_id | 26/280967\\ntransaction_id | 0\\nerror_severity | FATAL\\nsql_state_code | XX001\\nmessage | could not read block 0 in file \\\"base/16400/313316300\\\": read only 0 of 8192 bytes\\n\\nJustin\\n\\n\",\n \"msg_date\": \"Wed, 29 Aug 2018 09:01:49 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"10.5 but not 10.4: backend startup during reindex system: could not\\n read block 0 in file \\\"base/16400/..\\\": read only 0 of 8192 bytes\"\n },\n {\n \"msg_contents\": \"Justin Pryzby <pryzby@telsasoft.com> writes:\\n> I've seen this message now a handful of times recently. It seems to happen\\n> overnight, during a maintenance job which reindex things, including system\\n> catalog indices.\\n> It's easy to reproduce error under 10.5, but not under 10.3 nor 10.4.\\n> while :; do for a in pg_class_oid_index pg_class_relname_nsp_index pg_class_tblspc_relfilenode_index; do psql ts -qc \\\"REINDEX INDEX $a\\\"; done; done&\\n\\nThis looks suspiciously like the issue under discussion in\\n\\nhttps://www.postgresql.org/message-id/12259.1532117714%40sss.pgh.pa.us\\n\\nAs far as we can tell, that bug is a dozen years old, so it's not clear\\nwhy you find that you can reproduce it only in 10.5. But there might be\\nsome subtle timing change accounting for that.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Wed, 29 Aug 2018 11:35:51 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 10.5 but not 10.4: backend startup during reindex system: could\\n not read block 0 in file \\\"base/16400/..\\\": read only 0 of 8192 bytes\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 29, 2018 at 11:35:51AM -0400, Tom Lane wrote:\\n> Justin Pryzby <pryzby@telsasoft.com> writes:\\n> > I've seen this message now a handful of times recently. It seems to happen\\n> > overnight, during a maintenance job which reindex things, including system\\n> > catalog indices.\\n> > It's easy to reproduce error under 10.5, but not under 10.3 nor 10.4.\\n> > while :; do for a in pg_class_oid_index pg_class_relname_nsp_index pg_class_tblspc_relfilenode_index; do psql ts -qc \\\"REINDEX INDEX $a\\\"; done; done&\\n> > time while psql ts -qc ''; do :; done\\n> \\n> This looks suspiciously like the issue under discussion in\\n> \\n> https://www.postgresql.org/message-id/12259.1532117714%40sss.pgh.pa.us\\n> \\n> As far as we can tell, that bug is a dozen years old, so it's not clear\\n> why you find that you can reproduce it only in 10.5. But there might be\\n> some subtle timing change accounting for that.\\n\\nI confirmed the stack appears to be same as here, modulo differences between\\n9.6 to 10.\\nhttps://www.postgresql.org/message-id/fcab55ec-660a-bc43-7fe1-09af3665e89a%402ndquadrant.com\\n\\nIt seems to me there's one root problem occurring in (at least) two slightly\\ndifferent ways. The issue/symptom that I've been seeing occurs in 10.5 but not\\n10.4, and specifically at commit 2ce64ca, but not before. \\n\\n|commit 2ce64caaf793615d0f7a7ce4380e7255077d130c\\n|Author: Andres Freund <andres@anarazel.de>\\n|Date: Tue Jun 12 11:13:21 2018 -0700\\n|\\n|Fix bugs in vacuum of shared rels, by keeping their relcache entries current.\\n\\nHowever, the same message CAN happen under 10.4, though less easily, due to\\nautovacuum but (afaict) not user queries. Maybe that's a coincidence, but my\\nexperience mirrors the build logs: the problem since 10.5 is easily tickled in\\na few seconds rather than an hour.\\n\\nThe issue I see frequently under 10.5 has this stack:\\n\\nCore was generated by `postgres: pryzbyj postgres [local] startup '.\\n#2 0x0000000000709425 in mdread ()\\n#3 0x00000000006e2bc8 in ReadBuffer_common ()\\n#4 0x00000000006e3511 in ReadBufferExtended ()\\n#5 0x00000000004c2da9 in _bt_getbuf ()\\n#6 0x00000000004c32b2 in _bt_getroot ()\\n#7 0x00000000004c799b in _bt_search ()\\n#8 0x00000000004c8568 in _bt_first ()\\n#9 0x00000000004c52c7 in btgettuple ()\\n#10 0x00000000004bf936 in index_getnext_tid ()\\n#11 0x00000000004bfb43 in index_getnext ()\\n#12 0x00000000004beeae in systable_getnext ()\\n#13 0x00000000007fa074 in ScanPgRelation ()\\n#14 0x00000000007fe371 in RelationClearRelation ()\\n#15 0x00000000007fe84b in RelationCacheInvalidateEntry ()\\n#16 0x00000000007f7e15 in LocalExecuteInvalidationMessage ()\\n#17 0x00000000006f3b38 in ReceiveSharedInvalidMessages ()\\n#18 0x00000000006f6eb5 in LockRelationOid ()\\n#19 0x00000000004ae495 in relation_open ()\\n#20 0x00000000004bf336 in index_open ()\\n#21 0x00000000004bee3c in systable_beginscan ()\\n#22 0x00000000007fa069 in ScanPgRelation ()\\n#23 0x00000000007fce84 in RelationBuildDesc ()\\n#24 0x00000000007fe6f8 in RelationIdGetRelation ()\\n#25 0x00000000004ae463 in relation_open ()\\n#26 0x00000000004ae6a6 in heap_open ()\\n#27 0x0000000000819237 in InitPostgres ()\\n#28 0x000000000070e2e9 in PostgresMain ()\\n#29 0x0000000000477f36 in ServerLoop ()\\n#30 0x00000000006a5e3e in PostmasterMain ()\\n#31 0x0000000000478a08 in main ()\\n\\n From autovacum running REL_10_4 I've now seen (since trying to tickle the\\nproblem yesterday):\\n\\n#1 0x00007fbfb5b9d01a in __GI_abort () at abort.c:89\\n#2 0x0000000000708489 in mdread ()\\n#3 0x00000000006e1c48 in ReadBuffer_common ()\\n#4 0x00000000006e2591 in ReadBufferExtended ()\\n#5 0x00000000004c2369 in _bt_getbuf ()\\n#6 0x00000000004c2810 in _bt_getroot ()\\n#7 0x00000000004c6f5b in _bt_search ()\\n#8 0x00000000004c7b28 in _bt_first ()\\n#9 0x00000000004c4887 in btgettuple ()\\n#10 0x00000000004bf00e in index_getnext_tid ()\\n#11 0x00000000004bf1f3 in index_getnext ()\\n#12 0x00000000004be679 in systable_getnext ()\\n#13 0x00000000007f9084 in ScanPgRelation ()\\n#14 0x00000000007fbe54 in RelationBuildDesc ()\\n#15 0x00000000007fcc86 in RelationClearRelation ()\\n#16 0x00000000007fd7bb in RelationCacheInvalidateEntry ()\\n#17 0x00000000007f6e25 in LocalExecuteInvalidationMessage ()\\n#18 0x00000000006f2bb8 in ReceiveSharedInvalidMessages ()\\n#19 0x00000000006f5f35 in LockRelationOid ()\\n#20 0x00000000004ade05 in relation_open ()\\n#21 0x00000000004bea96 in index_open ()\\n#22 0x00000000004be606 in systable_beginscan ()\\n#23 0x00000000007f9079 in ScanPgRelation ()\\n#24 0x00000000007fbe54 in RelationBuildDesc ()\\n#25 0x00000000007fd668 in RelationIdGetRelation ()\\n#26 0x00000000004addd3 in relation_open ()\\n#27 0x00000000004bea96 in index_open ()\\n#28 0x00000000004be606 in systable_beginscan ()\\n#29 0x00000000007fef60 in RelationGetIndexList ()\\n#30 0x00000000005dba6c in vac_open_indexes ()\\n#31 0x00000000005dc652 in lazy_vacuum_rel ()\\n#32 0x00000000005da7b4 in vacuum_rel ()\\n#33 0x00000000005db661 in vacuum ()\\n#34 0x00000000004770db in do_autovacuum ()\\n#35 0x0000000000696242 in AutoVacWorkerMain.isra.6 ()\\n#36 0x0000000000696a49 in StartAutoVacWorker ()\\n#37 0x00000000006a3f0a in sigusr1_handler ()\\n#38 <signal handler called>\\n#39 0x00007fbfb5c62cf3 in __select_nocancel ()\\n at ../sysdeps/unix/syscall-template.S:84\\n#40 0x000000000047760e in ServerLoop ()\\n#41 0x00000000006a4ebe in PostmasterMain ()\\n#42 0x0000000000478a78 in main ()\\n\\nJustin\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 15:00:28 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: 10.5 but not 10.4: backend startup during reindex system: could\\n not read block 0 in file \\\"base/16400/..\\\": read only 0 of 8192 bytes\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2018-08-30 15:00:28 -0500, Justin Pryzby wrote:\\n> On Wed, Aug 29, 2018 at 11:35:51AM -0400, Tom Lane wrote:\\n> > This looks suspiciously like the issue under discussion in\\n> > \\n> > https://www.postgresql.org/message-id/12259.1532117714%40sss.pgh.pa.us\\n> > \\n> > As far as we can tell, that bug is a dozen years old, so it's not clear\\n> > why you find that you can reproduce it only in 10.5. But there might be\\n> > some subtle timing change accounting for that.\\n> \\n> I confirmed the stack appears to be same as here, modulo differences between\\n> 9.6 to 10.\\n\\nCould you check if both of the proposed attempts at fixing the issue\\nalso solves your problem?\\n\\nThanks,\\n\\nAndres Freund\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 13:18:59 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 10.5 but not 10.4: backend startup during reindex system: could\\n not read block 0 in file \\\"base/16400/..\\\": read only 0 of 8192 bytes\"\n },\n {\n \"msg_contents\": \"Justin Pryzby <pryzby@telsasoft.com> writes:\\n> On Wed, Aug 29, 2018 at 11:35:51AM -0400, Tom Lane wrote:\\n>> As far as we can tell, that bug is a dozen years old, so it's not clear\\n>> why you find that you can reproduce it only in 10.5. But there might be\\n>> some subtle timing change accounting for that.\\n\\n> It seems to me there's one root problem occurring in (at least) two slightly\\n> different ways. The issue/symptom that I've been seeing occurs in 10.5 but not\\n> 10.4, and specifically at commit 2ce64ca, but not before. \\n\\nYeah, as you probably saw in the other thread, we later realized that\\n2ce64ca created an additional pathway for ScanPgRelation to recurse;\\na pathway that's evidently easier to hit than the pre-existing ones.\\nI note that both of your stack traces display ScanPgRelation recursion,\\nso I'm feeling pretty confident that what you're seeing is the same\\nthing.\\n\\nBut, as Andres says, it'd be great if you could confirm whether the\\ndraft patches fix it for you.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 17:30:30 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 10.5 but not 10.4: backend startup during reindex system: could\\n not read block 0 in file \\\"base/16400/..\\\": read only 0 of 8192 bytes\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 30, 2018 at 05:30:30PM -0400, Tom Lane wrote:\\n> Justin Pryzby <pryzby@telsasoft.com> writes:\\n> > On Wed, Aug 29, 2018 at 11:35:51AM -0400, Tom Lane wrote:\\n> >> As far as we can tell, that bug is a dozen years old, so it's not clear\\n> >> why you find that you can reproduce it only in 10.5. But there might be\\n> >> some subtle timing change accounting for that.\\n> \\n> > It seems to me there's one root problem occurring in (at least) two slightly\\n> > different ways. The issue/symptom that I've been seeing occurs in 10.5 but not\\n> > 10.4, and specifically at commit 2ce64ca, but not before. \\n> \\n> Yeah, as you probably saw in the other thread, we later realized that\\n> 2ce64ca created an additional pathway for ScanPgRelation to recurse;\\n> a pathway that's evidently easier to hit than the pre-existing ones.\\n> I note that both of your stack traces display ScanPgRelation recursion,\\n> so I'm feeling pretty confident that what you're seeing is the same\\n> thing.\\n> \\n> But, as Andres says, it'd be great if you could confirm whether the\\n> draft patches fix it for you.\\n\\nI tested with relcache-rebuild.diff which hasn't broken in 15min, so I'm\\nconfident that doesn't hit the additional recusive pathway, but have to wait\\nawhile and see if autovacuum survives, too.\\n\\nI tried to apply fix-missed-inval-msg-accepts-1.patch on top of PG10.5 but\\npatch didn't apply, so I can test HEAD after the first patch soaks awhile.\\n\\nJust curious, is there really any difficulty in reproducing this? Once I\\nrealized this was a continuing issue and started to suspect pg10.5, it takes\\njust about nothing to reproduce anywhere I've tried. I just tested 5 servers,\\nand only one took more than a handful of seconds to fail. I gave up waiting\\nfor a 6th server, because I found it was waiting on a pre-existing lock.\\n\\n[pryzbyj@database ~]$ while :; do for a in pg_class_oid_index pg_class_relname_nsp_index pg_class_tblspc_relfilenode_index; do psql ts -qc \\\"REINDEX INDEX $a\\\"; done; done&\\n[pryzbyj@database ~]$ a=0; time while psql ts -qc ''; do a=$((1+a)); done ; echo \\\"$a\\\"\\npsql: FATAL: could not read block 0 in file \\\"base/16400/313581263\\\": read only 0 of 8192 bytes\\n\\nreal 0m1.772s\\nuser 0m0.076s\\nsys 0m0.116s\\n47\\n\\nJustin\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 16:57:12 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: 10.5 but not 10.4: backend startup during reindex system: could\\n not read block 0 in file \\\"base/16400/..\\\": read only 0 of 8192 bytes\"\n },\n {\n \"msg_contents\": \"Justin Pryzby <pryzby@telsasoft.com> writes:\\n> Just curious, is there really any difficulty in reproducing this?\\n\\nOnce you have the right test case, it's not hard. But it took us two\\nmonths to find one ...\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 18:03:27 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 10.5 but not 10.4: backend startup during reindex system: could\\n not read block 0 in file \\\"base/16400/..\\\": read only 0 of 8192 bytes\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 30, 2018 at 01:25:00PM -0700, Andres Freund wrote:\\n> On August 30, 2018 1:24:12 PM PDT, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> >On Thu, Aug 30, 2018 at 01:18:59PM -0700, Andres Freund wrote:\\n> >> Could you check if both of the proposed attempts at fixing the issue\\n> >> also solves your problem?\\n> >\\n> >Just checking that you're referring to:\\n> >\\n> >1) relcache-rebuild.diff\\n> >https://www.postgresql.org/message-id/20180829083730.n645apqhb2gyih3g%40alap3.anarazel.de\\n> >2) fix-missed-inval-msg-accepts-1.patch\\n> >https://www.postgresql.org/message-id/25024.1535579899%40sss.pgh.pa.us\\n> \\n> Right. Either should fix the issue.\\n\\nConfirmed that relcache-rebuild.diff seems to fix/avoid the issue under 10.5,\\nand fix-missed-inval-msg-accepts-1.patch does so for HEAD (but doesn't apply\\ncleanly to 10.5 so not tested). That's for both the the new easy-to-hit\\nsymptom and the pre-existing hard-to-hit issue affecting pg10.4 and earlier.\\n\\nThanks,\\nJustin\\n\\n\",\n \"msg_date\": \"Fri, 31 Aug 2018 03:08:37 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: 10.5 but not 10.4: backend startup during reindex system: could\\n not read block 0 in file \\\"base/16400/..\\\": read only 0 of 8192 bytes\"\n }\n]","isConversational":false}}},{"rowIdx":47190,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hello all.\n\nbackground worker can use SPI to read a database, but it can call\nBackgroundWorkerInitializeConnection(dbname) only once.\n\nI wonder if there is a way to let a child process of postmaster to access\nall the databases one by one?\n\n\n-- \nThanks\n\nHubert Zhang\n\nHello all.background worker can use SPI to read a database, but it can call BackgroundWorkerInitializeConnection(dbname) only once.I wonder if there is a way to let a child process of postmaster to access all the databases one by one?-- ThanksHubert Zhang","msg_date":"Wed, 29 Aug 2018 22:04:26 +0800","msg_from":"Hubert Zhang <hzhang@pivotal.io>","msg_from_op":true,"msg_subject":"Is child process of postmaster able to access all the databases?"},{"msg_contents":"Hubert Zhang <hzhang@pivotal.io> writes:\n> I wonder if there is a way to let a child process of postmaster to access\n> all the databases one by one?\n\nNo. For starters, you'd need some way to flush all database-specific\ninformation from relcache, catcache, and a boatload of other places;\nbut that logic does not exist.\n\n\t\t\tregards, tom lane\n\n","msg_date":"Wed, 29 Aug 2018 11:33:07 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Is child process of postmaster able to access all the databases?"}],"string":"[\n {\n \"msg_contents\": \"Hello all.\\n\\nbackground worker can use SPI to read a database, but it can call\\nBackgroundWorkerInitializeConnection(dbname) only once.\\n\\nI wonder if there is a way to let a child process of postmaster to access\\nall the databases one by one?\\n\\n\\n-- \\nThanks\\n\\nHubert Zhang\\n\\nHello all.background worker can use SPI to read a database, but it can call BackgroundWorkerInitializeConnection(dbname) only once.I wonder if there is a way to let a child process of postmaster to access all the databases one by one?-- ThanksHubert Zhang\",\n \"msg_date\": \"Wed, 29 Aug 2018 22:04:26 +0800\",\n \"msg_from\": \"Hubert Zhang <hzhang@pivotal.io>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Is child process of postmaster able to access all the databases?\"\n },\n {\n \"msg_contents\": \"Hubert Zhang <hzhang@pivotal.io> writes:\\n> I wonder if there is a way to let a child process of postmaster to access\\n> all the databases one by one?\\n\\nNo. For starters, you'd need some way to flush all database-specific\\ninformation from relcache, catcache, and a boatload of other places;\\nbut that logic does not exist.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Wed, 29 Aug 2018 11:33:07 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Is child process of postmaster able to access all the databases?\"\n }\n]","isConversational":false}}},{"rowIdx":47191,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hello,\n\nTo test postgres 11, I still regularly run series of short sessions of \npgbench-over-logical-replication (basically the same thing that I used \nlast year [1] - now in a perl incarnation). Most of the time the \nreplication is stable and finishes correctly but sometimes (rarely) I \nget:\n\nTRAP: FailedAssertion(\"!(TransactionIdPrecedesOrEquals(safeXid, \nsnap->xmin))\", File: \"snapbuild.c\", Line: 580)\n\nThis will probably be difficult to reproduce and to act upon but I \nwanted to report it anyway as in the course of the last few months I \nhave seen it several times, on several machines. Always rarely, always \npostgres 11 (I did not try other versions).\n\n\nErik Rijkers\n\n\n[1] \nhttps://www.postgresql.org/message-id/3897361c7010c4ac03f358173adbcd60%40xs4all.nl\n\n\nsource/version: bf2d0462cd73 / REL_11_STABLE (compiled 20180828_1629)\n\n1 primary, 4 replicas (on 1 host).\nBasically: pgbench --port=6515 --quiet --initialize --scale=5 postgres\n then: pgbench -M prepared -c 93 -j 8 -T 10 -P 2 -p 6515 postgres\n then: wait for log-repl sync.\n (I also added extra data type columns to the basic pgbench tables to \ntest all the different data types)\n\nBelow is a grep for the assert message, with 15 surrounding lines from \nthe latest occurrence.\n\n$ p=REL_11_STABLE; cd /var/data1/pg_stuff/tmp/cascade/$p ; grep -A 15 -B \n15 -Ei 'trap|assert' 65*/logfile.65* | less\n6516JTq_E8/logfile.6516-2018-08-29 14:28:13.747 CEST [139409] LOG: \nlogical decoding found consistent point at 0/89E038E0\n6516JTq_E8/logfile.6516-2018-08-29 14:28:13.747 CEST [139409] DETAIL: \nThere are no running transactions.\n6516JTq_E8/logfile.6516-2018-08-29 14:28:14.522 CEST [139539] LOG: \nreceived replication command: CREATE_REPLICATION_SLOT \n\"sub2_6517_6517_18834_sync_18804\" TEMPORARY LOGICAL pgoutput \nUSE_SNAPSHOT\n6516JTq_E8/logfile.6516-2018-08-29 14:28:15.681 CEST [139539] LOG: \nlogical decoding found initial starting point at 0/8AF5D590\n6516JTq_E8/logfile.6516-2018-08-29 14:28:15.681 CEST [139539] DETAIL: \nWaiting for transactions (approximately 1) older than 886297 to end.\n6516JTq_E8/logfile.6516-2018-08-29 14:28:15.720 CEST [139539] LOG: \nlogical decoding found consistent point at 0/8B018310\n6516JTq_E8/logfile.6516-2018-08-29 14:28:15.720 CEST [139539] DETAIL: \nThere are no running transactions.\n6516JTq_E8/logfile.6516-2018-08-29 14:28:15.860 CEST [139407] LOG: \nreceived replication command: START_REPLICATION SLOT \n\"sub2_6517_6517_18834_sync_18793\" LOGICAL 0/89E03918 (proto_version '1', \npublication_names '\"pub1_6516\"')\n6516JTq_E8/logfile.6516-2018-08-29 14:28:15.861 CEST [139407] LOG: \nstarting logical decoding for slot \"sub2_6517_6517_18834_sync_18793\"\n6516JTq_E8/logfile.6516-2018-08-29 14:28:15.861 CEST [139407] DETAIL: \nStreaming transactions committing after 0/89E03918, reading WAL from \n0/841EE960.\n6516JTq_E8/logfile.6516-2018-08-29 14:28:16.064 CEST [139407] LOG: \nlogical decoding found consistent point at 0/841EF3E0\n6516JTq_E8/logfile.6516-2018-08-29 14:28:16.064 CEST [139407] DETAIL: \nLogical decoding will begin using saved snapshot.\n6516JTq_E8/logfile.6516-2018-08-29 14:28:18.953 CEST [139541] LOG: \nreceived replication command: CREATE_REPLICATION_SLOT \n\"sub2_6517_6517_18834_sync_18814\" TEMPORARY LOGICAL pgoutput \nUSE_SNAPSHOT\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.002 CEST [139541] LOG: \nlogical decoding found consistent point at 0/8C6BA408\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.002 CEST [139541] DETAIL: \nThere are no running transactions.\n6516JTq_E8/logfile.6516:TRAP: \nFailedAssertion(\"!(TransactionIdPrecedesOrEquals(safeXid, snap->xmin))\", \nFile: \"snapbuild.c\", Line: 580)\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130398] LOG: \nserver process (PID 139541) was terminated by signal 6: Aborted\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130398] DETAIL: \nFailed process was running: BEGIN READ ONLY ISOLATION LEVEL REPEATABLE \nREAD\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130398] LOG: \nterminating any other active server processes\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [139405] WARNING: \nterminating connection because of crash of another server process\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [139405] DETAIL: \nThe postmaster has commanded this server process to roll back the \ncurrent transaction and exit, because another server process exited \nabnormally and possibly corrupted shared memory.\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [139405] HINT: In \na moment you should be able to reconnect to the database and repeat your \ncommand.\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [139405] CONTEXT: \nslot \"sub2_6517_6517\", output plugin \"pgoutput\", in the change callback, \nassociated LSN 0/8B11CF00\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130411] WARNING: \nterminating connection because of crash of another server process\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130411] DETAIL: \nThe postmaster has commanded this server process to roll back the \ncurrent transaction and exit, because another server process exited \nabnormally and possibly corrupted shared memory.\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130411] HINT: In \na moment you should be able to reconnect to the database and repeat your \ncommand.\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130405] WARNING: \nterminating connection because of crash of another server process\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130405] DETAIL: \nThe postmaster has commanded this server process to roll back the \ncurrent transaction and exit, because another server process exited \nabnormally and possibly corrupted shared memory.\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130405] HINT: In \na moment you should be able to reconnect to the database and repeat your \ncommand.\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.116 CEST [130398] LOG: all \nserver processes terminated; reinitializing\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.515 CEST [139548] LOG: \ndatabase system was interrupted; last known up at 2018-08-29 14:26:14 \nCEST\n\n\n\n\n","msg_date":"Wed, 29 Aug 2018 17:43:17 +0200","msg_from":"Erik Rijkers <er@xs4all.nl>","msg_from_op":true,"msg_subject":"rare crash - FailedAssertion snapbuild.c Line: 580"},{"msg_contents":"Hi,\n\nOn 2018-08-29 17:43:17 +0200, Erik Rijkers wrote:\n> To test postgres 11, I still regularly run series of short sessions of\n> pgbench-over-logical-replication (basically the same thing that I used last\n> year [1] - now in a perl incarnation). Most of the time the replication is\n> stable and finishes correctly but sometimes (rarely) I get:\n> \n> TRAP: FailedAssertion(\"!(TransactionIdPrecedesOrEquals(safeXid,\n> snap->xmin))\", File: \"snapbuild.c\", Line: 580)\n> \n> This will probably be difficult to reproduce and to act upon but I wanted to\n> report it anyway as in the course of the last few months I have seen it\n> several times, on several machines. Always rarely, always postgres 11 (I did\n> not try other versions).\n\nThanks for testing! Could you possibly run the tests with core files\nenabled, so we at get a backtrace in case of trouble? Knowing what the\nvalues here are would be tremendously helpful...\n\nGreetings,\n\nAndres Freund\n\n","msg_date":"Wed, 29 Aug 2018 12:15:26 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: rare crash - FailedAssertion snapbuild.c Line: 580"},{"msg_contents":"On 2018-08-29 21:15, Andres Freund wrote:\n> Hi,\n> \n> On 2018-08-29 17:43:17 +0200, Erik Rijkers wrote:\n>> To test postgres 11, I still regularly run series of short sessions of\n>> pgbench-over-logical-replication (basically the same thing that I used \n>> last\n>> year [1] - now in a perl incarnation). Most of the time the \n>> replication is\n>> stable and finishes correctly but sometimes (rarely) I get:\n>> \n>> TRAP: FailedAssertion(\"!(TransactionIdPrecedesOrEquals(safeXid,\n>> snap->xmin))\", File: \"snapbuild.c\", Line: 580)\n>> \n>> This will probably be difficult to reproduce and to act upon but I \n>> wanted to\n>> report it anyway as in the course of the last few months I have seen \n>> it\n>> several times, on several machines. Always rarely, always postgres 11 \n>> (I did\n>> not try other versions).\n> \n> Thanks for testing! Could you possibly run the tests with core files\n> enabled, so we at get a backtrace in case of trouble? Knowing what the\n> values here are would be tremendously helpful...\n\nok, is this any use?\n\n$ gdb --quiet \n/var/data1/pg_stuff/pg_installations/pgsql.REL_11_STABLE/bin/postgres \n/var/data1/pg_stuff/tmp/cascade/REL_11_STABLE/6516_gW1Cl/data/core\nReading symbols from \n/var/data1/pg_stuff/pg_installations/pgsql.REL_11_STABLE/bin/postgres...done.\n[New LWP 147484]\n[Thread debugging using libthread_db enabled]\nUsing host libthread_db library \n\"/lib/x86_64-linux-gnu/libthread_db.so.1\".\nCore was generated by `postgres: walsender rijkers [local] idle in \ntransaction '.\nProgram terminated with signal SIGABRT, Aborted.\n#0 0x00007f0fd20e7067 in __GI_raise (sig=sig@entry=6) at \n../nptl/sysdeps/unix/sysv/linux/raise.c:56\n56 ../nptl/sysdeps/unix/sysv/linux/raise.c: No such file or \ndirectory.\n(gdb) bt\n#0 0x00007f0fd20e7067 in __GI_raise (sig=sig@entry=6) at \n../nptl/sysdeps/unix/sysv/linux/raise.c:56\n#1 0x00007f0fd20e8448 in __GI_abort () at abort.c:89\n#2 0x00000000008880bf in ExceptionalCondition \n(conditionName=conditionName@entry=0xa417f8 \n\"!(TransactionIdPrecedesOrEquals(safeXid, snap->xmin))\", \nerrorType=errorType@entry=0x8d365d \"FailedAssertion\", \nfileName=fileName@entry=0xa41223 \"snapbuild.c\", \nlineNumber=lineNumber@entry=580) at assert.c:54\n#3 0x000000000072676e in SnapBuildInitialSnapshot () at snapbuild.c:580\n#4 0x000000000072ed54 in CreateReplicationSlot (cmd=0x223bae0) at \nwalsender.c:951\n#5 exec_replication_command (cmd_string=cmd_string@entry=0x21a1cf8 \n\"CREATE_REPLICATION_SLOT \\\"sub2_6517_6517_18748_sync_18728\\\" TEMPORARY \nLOGICAL pgoutput USE_SNAPSHOT\") at walsender.c:1527\n#6 0x000000000077e8ee in PostgresMain (argc=<optimized out>, \nargv=argv@entry=0x21cbbc8, dbname=<optimized out>, username=<optimized \nout>) at postgres.c:4155\n#7 0x0000000000704fde in BackendRun (port=0x21c4520) at \npostmaster.c:4361\n#8 BackendStartup (port=0x21c4520) at postmaster.c:4033\n#9 ServerLoop () at postmaster.c:1706\n#10 0x0000000000705e0f in PostmasterMain (argc=argc@entry=12, \nargv=argv@entry=0x219c470) at postmaster.c:1379\n#11 0x0000000000478d80 in main (argc=12, argv=0x219c470) at main.c:228\n(gdb)\n\n\n\n\n","msg_date":"Thu, 30 Aug 2018 14:51:45 +0200","msg_from":"Erik Rijkers <er@xs4all.nl>","msg_from_op":true,"msg_subject":"Re: rare crash - FailedAssertion snapbuild.c Line: 580"},{"msg_contents":"On 2018-Aug-30, Erik Rijkers wrote:\n\n> ok, is this any use?\n\nSeems mostly good, but the Xids are not printed. Could you please do\n\"bt full\"? Also:\n\nframe 3\nprint *snap\n\nThanks,\n\n-- \n�lvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n","msg_date":"Thu, 30 Aug 2018 11:44:34 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: rare crash - FailedAssertion snapbuild.c Line: 580"},{"msg_contents":"On 2018-08-30 16:44, Alvaro Herrera wrote:\n> On 2018-Aug-30, Erik Rijkers wrote:\n> \n>> ok, is this any use?\n> \n> Seems mostly good, but the Xids are not printed. Could you please do\n> \"bt full\"? Also:\n> \n> frame 3\n> print *snap\n\nSee the attached.","msg_date":"Thu, 30 Aug 2018 16:57:26 +0200","msg_from":"Erik Rijkers <er@xs4all.nl>","msg_from_op":true,"msg_subject":"Re: rare crash - FailedAssertion snapbuild.c Line: 580"}],"string":"[\n {\n \"msg_contents\": \"Hello,\\n\\nTo test postgres 11, I still regularly run series of short sessions of \\npgbench-over-logical-replication (basically the same thing that I used \\nlast year [1] - now in a perl incarnation). Most of the time the \\nreplication is stable and finishes correctly but sometimes (rarely) I \\nget:\\n\\nTRAP: FailedAssertion(\\\"!(TransactionIdPrecedesOrEquals(safeXid, \\nsnap->xmin))\\\", File: \\\"snapbuild.c\\\", Line: 580)\\n\\nThis will probably be difficult to reproduce and to act upon but I \\nwanted to report it anyway as in the course of the last few months I \\nhave seen it several times, on several machines. Always rarely, always \\npostgres 11 (I did not try other versions).\\n\\n\\nErik Rijkers\\n\\n\\n[1] \\nhttps://www.postgresql.org/message-id/3897361c7010c4ac03f358173adbcd60%40xs4all.nl\\n\\n\\nsource/version: bf2d0462cd73 / REL_11_STABLE (compiled 20180828_1629)\\n\\n1 primary, 4 replicas (on 1 host).\\nBasically: pgbench --port=6515 --quiet --initialize --scale=5 postgres\\n then: pgbench -M prepared -c 93 -j 8 -T 10 -P 2 -p 6515 postgres\\n then: wait for log-repl sync.\\n (I also added extra data type columns to the basic pgbench tables to \\ntest all the different data types)\\n\\nBelow is a grep for the assert message, with 15 surrounding lines from \\nthe latest occurrence.\\n\\n$ p=REL_11_STABLE; cd /var/data1/pg_stuff/tmp/cascade/$p ; grep -A 15 -B \\n15 -Ei 'trap|assert' 65*/logfile.65* | less\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:13.747 CEST [139409] LOG: \\nlogical decoding found consistent point at 0/89E038E0\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:13.747 CEST [139409] DETAIL: \\nThere are no running transactions.\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:14.522 CEST [139539] LOG: \\nreceived replication command: CREATE_REPLICATION_SLOT \\n\\\"sub2_6517_6517_18834_sync_18804\\\" TEMPORARY LOGICAL pgoutput \\nUSE_SNAPSHOT\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:15.681 CEST [139539] LOG: \\nlogical decoding found initial starting point at 0/8AF5D590\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:15.681 CEST [139539] DETAIL: \\nWaiting for transactions (approximately 1) older than 886297 to end.\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:15.720 CEST [139539] LOG: \\nlogical decoding found consistent point at 0/8B018310\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:15.720 CEST [139539] DETAIL: \\nThere are no running transactions.\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:15.860 CEST [139407] LOG: \\nreceived replication command: START_REPLICATION SLOT \\n\\\"sub2_6517_6517_18834_sync_18793\\\" LOGICAL 0/89E03918 (proto_version '1', \\npublication_names '\\\"pub1_6516\\\"')\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:15.861 CEST [139407] LOG: \\nstarting logical decoding for slot \\\"sub2_6517_6517_18834_sync_18793\\\"\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:15.861 CEST [139407] DETAIL: \\nStreaming transactions committing after 0/89E03918, reading WAL from \\n0/841EE960.\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:16.064 CEST [139407] LOG: \\nlogical decoding found consistent point at 0/841EF3E0\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:16.064 CEST [139407] DETAIL: \\nLogical decoding will begin using saved snapshot.\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:18.953 CEST [139541] LOG: \\nreceived replication command: CREATE_REPLICATION_SLOT \\n\\\"sub2_6517_6517_18834_sync_18814\\\" TEMPORARY LOGICAL pgoutput \\nUSE_SNAPSHOT\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.002 CEST [139541] LOG: \\nlogical decoding found consistent point at 0/8C6BA408\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.002 CEST [139541] DETAIL: \\nThere are no running transactions.\\n6516JTq_E8/logfile.6516:TRAP: \\nFailedAssertion(\\\"!(TransactionIdPrecedesOrEquals(safeXid, snap->xmin))\\\", \\nFile: \\\"snapbuild.c\\\", Line: 580)\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130398] LOG: \\nserver process (PID 139541) was terminated by signal 6: Aborted\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130398] DETAIL: \\nFailed process was running: BEGIN READ ONLY ISOLATION LEVEL REPEATABLE \\nREAD\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130398] LOG: \\nterminating any other active server processes\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [139405] WARNING: \\nterminating connection because of crash of another server process\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [139405] DETAIL: \\nThe postmaster has commanded this server process to roll back the \\ncurrent transaction and exit, because another server process exited \\nabnormally and possibly corrupted shared memory.\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [139405] HINT: In \\na moment you should be able to reconnect to the database and repeat your \\ncommand.\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [139405] CONTEXT: \\nslot \\\"sub2_6517_6517\\\", output plugin \\\"pgoutput\\\", in the change callback, \\nassociated LSN 0/8B11CF00\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130411] WARNING: \\nterminating connection because of crash of another server process\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130411] DETAIL: \\nThe postmaster has commanded this server process to roll back the \\ncurrent transaction and exit, because another server process exited \\nabnormally and possibly corrupted shared memory.\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130411] HINT: In \\na moment you should be able to reconnect to the database and repeat your \\ncommand.\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130405] WARNING: \\nterminating connection because of crash of another server process\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130405] DETAIL: \\nThe postmaster has commanded this server process to roll back the \\ncurrent transaction and exit, because another server process exited \\nabnormally and possibly corrupted shared memory.\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.112 CEST [130405] HINT: In \\na moment you should be able to reconnect to the database and repeat your \\ncommand.\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.116 CEST [130398] LOG: all \\nserver processes terminated; reinitializing\\n6516JTq_E8/logfile.6516-2018-08-29 14:28:20.515 CEST [139548] LOG: \\ndatabase system was interrupted; last known up at 2018-08-29 14:26:14 \\nCEST\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 29 Aug 2018 17:43:17 +0200\",\n \"msg_from\": \"Erik Rijkers <er@xs4all.nl>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"rare crash - FailedAssertion snapbuild.c Line: 580\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2018-08-29 17:43:17 +0200, Erik Rijkers wrote:\\n> To test postgres 11, I still regularly run series of short sessions of\\n> pgbench-over-logical-replication (basically the same thing that I used last\\n> year [1] - now in a perl incarnation). Most of the time the replication is\\n> stable and finishes correctly but sometimes (rarely) I get:\\n> \\n> TRAP: FailedAssertion(\\\"!(TransactionIdPrecedesOrEquals(safeXid,\\n> snap->xmin))\\\", File: \\\"snapbuild.c\\\", Line: 580)\\n> \\n> This will probably be difficult to reproduce and to act upon but I wanted to\\n> report it anyway as in the course of the last few months I have seen it\\n> several times, on several machines. Always rarely, always postgres 11 (I did\\n> not try other versions).\\n\\nThanks for testing! Could you possibly run the tests with core files\\nenabled, so we at get a backtrace in case of trouble? Knowing what the\\nvalues here are would be tremendously helpful...\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\",\n \"msg_date\": \"Wed, 29 Aug 2018 12:15:26 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: rare crash - FailedAssertion snapbuild.c Line: 580\"\n },\n {\n \"msg_contents\": \"On 2018-08-29 21:15, Andres Freund wrote:\\n> Hi,\\n> \\n> On 2018-08-29 17:43:17 +0200, Erik Rijkers wrote:\\n>> To test postgres 11, I still regularly run series of short sessions of\\n>> pgbench-over-logical-replication (basically the same thing that I used \\n>> last\\n>> year [1] - now in a perl incarnation). Most of the time the \\n>> replication is\\n>> stable and finishes correctly but sometimes (rarely) I get:\\n>> \\n>> TRAP: FailedAssertion(\\\"!(TransactionIdPrecedesOrEquals(safeXid,\\n>> snap->xmin))\\\", File: \\\"snapbuild.c\\\", Line: 580)\\n>> \\n>> This will probably be difficult to reproduce and to act upon but I \\n>> wanted to\\n>> report it anyway as in the course of the last few months I have seen \\n>> it\\n>> several times, on several machines. Always rarely, always postgres 11 \\n>> (I did\\n>> not try other versions).\\n> \\n> Thanks for testing! Could you possibly run the tests with core files\\n> enabled, so we at get a backtrace in case of trouble? Knowing what the\\n> values here are would be tremendously helpful...\\n\\nok, is this any use?\\n\\n$ gdb --quiet \\n/var/data1/pg_stuff/pg_installations/pgsql.REL_11_STABLE/bin/postgres \\n/var/data1/pg_stuff/tmp/cascade/REL_11_STABLE/6516_gW1Cl/data/core\\nReading symbols from \\n/var/data1/pg_stuff/pg_installations/pgsql.REL_11_STABLE/bin/postgres...done.\\n[New LWP 147484]\\n[Thread debugging using libthread_db enabled]\\nUsing host libthread_db library \\n\\\"/lib/x86_64-linux-gnu/libthread_db.so.1\\\".\\nCore was generated by `postgres: walsender rijkers [local] idle in \\ntransaction '.\\nProgram terminated with signal SIGABRT, Aborted.\\n#0 0x00007f0fd20e7067 in __GI_raise (sig=sig@entry=6) at \\n../nptl/sysdeps/unix/sysv/linux/raise.c:56\\n56 ../nptl/sysdeps/unix/sysv/linux/raise.c: No such file or \\ndirectory.\\n(gdb) bt\\n#0 0x00007f0fd20e7067 in __GI_raise (sig=sig@entry=6) at \\n../nptl/sysdeps/unix/sysv/linux/raise.c:56\\n#1 0x00007f0fd20e8448 in __GI_abort () at abort.c:89\\n#2 0x00000000008880bf in ExceptionalCondition \\n(conditionName=conditionName@entry=0xa417f8 \\n\\\"!(TransactionIdPrecedesOrEquals(safeXid, snap->xmin))\\\", \\nerrorType=errorType@entry=0x8d365d \\\"FailedAssertion\\\", \\nfileName=fileName@entry=0xa41223 \\\"snapbuild.c\\\", \\nlineNumber=lineNumber@entry=580) at assert.c:54\\n#3 0x000000000072676e in SnapBuildInitialSnapshot () at snapbuild.c:580\\n#4 0x000000000072ed54 in CreateReplicationSlot (cmd=0x223bae0) at \\nwalsender.c:951\\n#5 exec_replication_command (cmd_string=cmd_string@entry=0x21a1cf8 \\n\\\"CREATE_REPLICATION_SLOT \\\\\\\"sub2_6517_6517_18748_sync_18728\\\\\\\" TEMPORARY \\nLOGICAL pgoutput USE_SNAPSHOT\\\") at walsender.c:1527\\n#6 0x000000000077e8ee in PostgresMain (argc=<optimized out>, \\nargv=argv@entry=0x21cbbc8, dbname=<optimized out>, username=<optimized \\nout>) at postgres.c:4155\\n#7 0x0000000000704fde in BackendRun (port=0x21c4520) at \\npostmaster.c:4361\\n#8 BackendStartup (port=0x21c4520) at postmaster.c:4033\\n#9 ServerLoop () at postmaster.c:1706\\n#10 0x0000000000705e0f in PostmasterMain (argc=argc@entry=12, \\nargv=argv@entry=0x219c470) at postmaster.c:1379\\n#11 0x0000000000478d80 in main (argc=12, argv=0x219c470) at main.c:228\\n(gdb)\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 14:51:45 +0200\",\n \"msg_from\": \"Erik Rijkers <er@xs4all.nl>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: rare crash - FailedAssertion snapbuild.c Line: 580\"\n },\n {\n \"msg_contents\": \"On 2018-Aug-30, Erik Rijkers wrote:\\n\\n> ok, is this any use?\\n\\nSeems mostly good, but the Xids are not printed. Could you please do\\n\\\"bt full\\\"? Also:\\n\\nframe 3\\nprint *snap\\n\\nThanks,\\n\\n-- \\n�lvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 11:44:34 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: rare crash - FailedAssertion snapbuild.c Line: 580\"\n },\n {\n \"msg_contents\": \"On 2018-08-30 16:44, Alvaro Herrera wrote:\\n> On 2018-Aug-30, Erik Rijkers wrote:\\n> \\n>> ok, is this any use?\\n> \\n> Seems mostly good, but the Xids are not printed. Could you please do\\n> \\\"bt full\\\"? Also:\\n> \\n> frame 3\\n> print *snap\\n\\nSee the attached.\",\n \"msg_date\": \"Thu, 30 Aug 2018 16:57:26 +0200\",\n \"msg_from\": \"Erik Rijkers <er@xs4all.nl>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: rare crash - FailedAssertion snapbuild.c Line: 580\"\n }\n]","isConversational":false}}},{"rowIdx":47192,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nI'm trying to upgrade an old PG 9.4 cluster to PG 10, and it's failing on\ncreating an index that uses the unaccent(text) function.\n\nThat function is part of the unaccent extension, which is installed in the\nold DB cluster. I expect pg_upgrade to create that extension as part of the\nupgrade. It does create other extensions that are installed in the old DB\ncluster. I don't get why this one isn't included.\n\nHere are the commands I run, and their output:\n\n----- snip -----\n$ sudo rm -rf /opt/local/var/db/postgresql10/defaultdb\n\n$ sudo mkdir -p /opt/local/var/db/postgresql10/defaultdb\n\n$ sudo chown postgres:postgres /opt/local/var/db/postgresql10/defaultdb\n\n$ sudo su postgres -c '/opt/local/lib/postgresql10/bin/initdb -D\n/opt/local/var/db/postgresql10/defaultdb --locale en_US.UTF-8'\nThe files belonging to this database system will be owned by user\n\"postgres\".\nThis user must also own the server process.\n\nThe database cluster will be initialized with locale \"en_US.UTF-8\".\nThe default database encoding has accordingly been set to \"UTF8\".\nThe default text search configuration will be set to \"english\".\n\nData page checksums are disabled.\n\nfixing permissions on existing directory\n/opt/local/var/db/postgresql10/defaultdb ... ok\ncreating subdirectories ... ok\nselecting default max_connections ... 100\nselecting default shared_buffers ... 128MB\nselecting dynamic shared memory implementation ... posix\ncreating configuration files ... ok\nrunning bootstrap script ... ok\nperforming post-bootstrap initialization ... ok\nsyncing data to disk ... ok\n\nWARNING: enabling \"trust\" authentication for local connections\nYou can change this by editing pg_hba.conf or using the option -A, or\n--auth-local and --auth-host, the next time you run initdb.\n\nSuccess. You can now start the database server using:\n\n /opt/local/lib/postgresql10/bin/pg_ctl -D\n/opt/local/var/db/postgresql10/defaultdb -l logfile start\n\n$ sudo su postgres -c '/opt/local/lib/postgresql10/bin/pg_upgrade\n--old-bindir /opt/local/lib/postgresql94/bin --old-datadir\n/opt/local/var/db/postgresql94/defaultdb --new-bindir\n/opt/local/lib/postgresql10/bin --new-datadir\n/opt/local/var/db/postgresql10/defaultdb'Performing Consistency Checks\n-----------------------------\nChecking cluster versions ok\nChecking database user is the install user ok\nChecking database connection settings ok\nChecking for prepared transactions ok\nChecking for reg* data types in user tables ok\nChecking for contrib/isn with bigint-passing mismatch ok\nChecking for invalid \"unknown\" user columns ok\nChecking for roles starting with \"pg_\" ok\nCreating dump of global objects ok\nCreating dump of database schemas\n ok\nChecking for presence of required libraries ok\nChecking database user is the install user ok\nChecking for prepared transactions ok\n\nIf pg_upgrade fails after this point, you must re-initdb the\nnew cluster before continuing.\n\nPerforming Upgrade\n------------------\nAnalyzing all rows in the new cluster ok\nFreezing all rows in the new cluster ok\nDeleting files from new pg_xact ok\nCopying old pg_clog to new server ok\nSetting next transaction ID and epoch for new cluster ok\nDeleting files from new pg_multixact/offsets ok\nCopying old pg_multixact/offsets to new server ok\nDeleting files from new pg_multixact/members ok\nCopying old pg_multixact/members to new server ok\nSetting next multixact ID and offset for new cluster ok\nResetting WAL archives ok\nSetting frozenxid and minmxid counters in new cluster ok\nRestoring global objects in the new cluster ok\nRestoring database schemas in the new cluster\n dm_test\n*failure*\n\nConsult the last few lines of \"pg_upgrade_dump_409041.log\" for\nthe probable cause of the failure.\nFailure, exiting\n----- snip -----\n\nAnd here is that pg_upgrade_dump_409041.log file:\n\n----- snip -----\ncommand: \"/opt/local/lib/postgresql10/bin/pg_dump\" --host /private/tmp\n--port 50432 --username postgres --schema-only --quote-all-identifiers\n--binary-upgrade --format=custom --file=\"pg_upgrade_dump_409041.custom\"\n'dbname=dm_test' >> \"pg_upgrade_dump_409041.log\" 2>&1\n\n\ncommand: \"/opt/local/lib/postgresql10/bin/pg_restore\" --host /private/tmp\n--port 50432 --username postgres --exit-on-error --verbose --dbname\n'dbname=dm_test' \"pg_upgrade_dump_409041.custom\" >>\n\"pg_upgrade_dump_409041.log\" 2>&1\npg_restore: connecting to database for restore\npg_restore: creating pg_largeobject \"pg_largeobject\"\npg_restore: creating pg_largeobject_metadata \"pg_largeobject_metadata\"\npg_restore: creating SCHEMA \"acl_admin\"\npg_restore: creating COMMENT \"SCHEMA \"acl_admin\"\"\npg_restore: creating SCHEMA \"public\"\npg_restore: creating COMMENT \"SCHEMA \"public\"\"\npg_restore: creating EXTENSION \"fuzzystrmatch\"\npg_restore: creating COMMENT \"EXTENSION \"fuzzystrmatch\"\"\npg_restore: creating EXTENSION \"hstore\"\npg_restore: creating COMMENT \"EXTENSION \"hstore\"\"\npg_restore: creating EXTENSION \"sslinfo\"\npg_restore: creating COMMENT \"EXTENSION \"sslinfo\"\"\npg_restore: creating TYPE \"public.dataimport_job_state\"\npg_restore: creating TYPE \"public.dblink_pkey_results\"...\n[...]\npg_restore: creating INDEX \"public.ix_semantic_mapping_lower_title\"\npg_restore: creating INDEX \"public.ix_semantic_mapping_normalize_title\"\npg_restore: [archiver (db)] Error while PROCESSING TOC:\npg_restore: [archiver (db)] Error from TOC entry 5240; 1259 427215 INDEX\nix_semantic_mapping_normalize_title dm_admin\npg_restore: [archiver (db)] could not execute query: ERROR: function\nunaccent(text) does not exist\nLINE 2: SELECT lower(unaccent(btrim(regexp_replace($1, '\\s+', ' ',...\n ^\nHINT: No function matches the given name and argument types. You might\nneed to add explicit type casts.\nQUERY:\n SELECT lower(unaccent(btrim(regexp_replace($1, '\\s+', ' ', 'g'), ' \"')))\n\nCONTEXT: SQL function \"semantic_normalize\" during inlining\n Command was:\n-- For binary upgrade, must preserve pg_class oids\nSELECT\npg_catalog.binary_upgrade_set_next_index_pg_class_oid('427215'::pg_catalog.oid);\n\nCREATE INDEX \"ix_semantic_mapping_normalize_title\" ON\n\"public\".\"semantic_mapping\" USING \"btree\"\n(\"public\".\"semantic_normalize\"(\"title\"));\n----- snip -----\n\nThese occurrences of unaccent at the end are the only ones in the file:\n\n$ sudo grep -n unaccent pg_upgrade_dump_409041.log g\n1713:pg_restore: [archiver (db)] could not execute query: ERROR: function\nunaccent(text) does not exist\n1714:LINE 2: SELECT lower(unaccent(btrim(regexp_replace($1, '\\s+', ' ',...\n1718: SELECT lower(unaccent(btrim(regexp_replace($1, '\\s+', ' ', 'g'), '\n\"')))\n\nHere is the definition of that index in the old DB:\n\n\\d ix_semantic_mapping_normalize_title\n Index \"public.ix_semantic_mapping_normalize_title\"\n Column | Type | Definition\n--------------------+------+---------------------------\n semantic_normalize | text | semantic_normalize(title)\nbtree, for table \"public.semantic_mapping\"\n\nand the semantic_normalize function it uses:\n\n\\x \\df+ semantic_normalize\nExpanded display is on.\nList of functions\n-[ RECORD 1\n]-------+---------------------------------------------------------------------------\nSchema | public\nName | semantic_normalize\nResult data type | text\nArgument data types | title text\nType | normal\nVolatility | immutable\nOwner | dm_admin\nSecurity | invoker\nAccess privileges |\nLanguage | sql\nSource code\n|\n+\n | SELECT lower(unaccent(btrim(regexp_replace($1,\n'\\s+', ' ', 'g'), ' \"')))+\n |\nDescription |\n\nAny more information I can provide, to help troubleshoot this?\n\nCheers,\nGulli\n\nHi,I'm trying to upgrade an old PG 9.4 cluster to PG 10, and it's failing on creating an index that uses the unaccent(text) function.That function is part of the unaccent extension, which is installed in the old DB cluster. I expect pg_upgrade to create that extension as part of the upgrade. It does create other extensions that are installed in the old DB cluster. I don't get why this one isn't included.Here are the commands I run, and their output:----- snip -----$ sudo rm -rf /opt/local/var/db/postgresql10/defaultdb$ sudo mkdir -p /opt/local/var/db/postgresql10/defaultdb$ sudo chown postgres:postgres /opt/local/var/db/postgresql10/defaultdb$ sudo su postgres -c '/opt/local/lib/postgresql10/bin/initdb -D /opt/local/var/db/postgresql10/defaultdb --locale en_US.UTF-8'The files belonging to this database system will be owned by user \"postgres\".This user must also own the server process.The database cluster will be initialized with locale \"en_US.UTF-8\".The default database encoding has accordingly been set to \"UTF8\".The default text search configuration will be set to \"english\".Data page checksums are disabled.fixing permissions on existing directory /opt/local/var/db/postgresql10/defaultdb ... okcreating subdirectories ... okselecting default max_connections ... 100selecting default shared_buffers ... 128MBselecting dynamic shared memory implementation ... posixcreating configuration files ... okrunning bootstrap script ... okperforming post-bootstrap initialization ... oksyncing data to disk ... okWARNING: enabling \"trust\" authentication for local connectionsYou can change this by editing pg_hba.conf or using the option -A, or--auth-local and --auth-host, the next time you run initdb.Success. You can now start the database server using: /opt/local/lib/postgresql10/bin/pg_ctl -D /opt/local/var/db/postgresql10/defaultdb -l logfile start$ sudo su postgres -c '/opt/local/lib/postgresql10/bin/pg_upgrade --old-bindir /opt/local/lib/postgresql94/bin --old-datadir /opt/local/var/db/postgresql94/defaultdb --new-bindir /opt/local/lib/postgresql10/bin --new-datadir /opt/local/var/db/postgresql10/defaultdb'Performing Consistency Checks-----------------------------Checking cluster versions okChecking database user is the install user okChecking database connection settings okChecking for prepared transactions okChecking for reg* data types in user tables okChecking for contrib/isn with bigint-passing mismatch okChecking for invalid \"unknown\" user columns okChecking for roles starting with \"pg_\" okCreating dump of global objects okCreating dump of database schemas okChecking for presence of required libraries okChecking database user is the install user okChecking for prepared transactions okIf pg_upgrade fails after this point, you must re-initdb thenew cluster before continuing.Performing Upgrade------------------Analyzing all rows in the new cluster okFreezing all rows in the new cluster okDeleting files from new pg_xact okCopying old pg_clog to new server okSetting next transaction ID and epoch for new cluster okDeleting files from new pg_multixact/offsets okCopying old pg_multixact/offsets to new server okDeleting files from new pg_multixact/members okCopying old pg_multixact/members to new server okSetting next multixact ID and offset for new cluster okResetting WAL archives okSetting frozenxid and minmxid counters in new cluster okRestoring global objects in the new cluster okRestoring database schemas in the new cluster dm_test*failure*Consult the last few lines of \"pg_upgrade_dump_409041.log\" forthe probable cause of the failure.Failure, exiting----- snip -----And here is that pg_upgrade_dump_409041.log file:----- snip -----command: \"/opt/local/lib/postgresql10/bin/pg_dump\" --host /private/tmp --port 50432 --username postgres --schema-only --quote-all-identifiers --binary-upgrade --format=custom --file=\"pg_upgrade_dump_409041.custom\" 'dbname=dm_test' >> \"pg_upgrade_dump_409041.log\" 2>&1command: \"/opt/local/lib/postgresql10/bin/pg_restore\" --host /private/tmp --port 50432 --username postgres --exit-on-error --verbose --dbname 'dbname=dm_test' \"pg_upgrade_dump_409041.custom\" >> \"pg_upgrade_dump_409041.log\" 2>&1pg_restore: connecting to database for restorepg_restore: creating pg_largeobject \"pg_largeobject\"pg_restore: creating pg_largeobject_metadata \"pg_largeobject_metadata\"pg_restore: creating SCHEMA \"acl_admin\"pg_restore: creating COMMENT \"SCHEMA \"acl_admin\"\"pg_restore: creating SCHEMA \"public\"pg_restore: creating COMMENT \"SCHEMA \"public\"\"pg_restore: creating EXTENSION \"fuzzystrmatch\"pg_restore: creating COMMENT \"EXTENSION \"fuzzystrmatch\"\"pg_restore: creating EXTENSION \"hstore\"pg_restore: creating COMMENT \"EXTENSION \"hstore\"\"pg_restore: creating EXTENSION \"sslinfo\"pg_restore: creating COMMENT \"EXTENSION \"sslinfo\"\"pg_restore: creating TYPE \"public.dataimport_job_state\"pg_restore: creating TYPE \"public.dblink_pkey_results\"...[...]pg_restore: creating INDEX \"public.ix_semantic_mapping_lower_title\"pg_restore: creating INDEX \"public.ix_semantic_mapping_normalize_title\"pg_restore: [archiver (db)] Error while PROCESSING TOC:pg_restore: [archiver (db)] Error from TOC entry 5240; 1259 427215 INDEX ix_semantic_mapping_normalize_title dm_adminpg_restore: [archiver (db)] could not execute query: ERROR: function unaccent(text) does not existLINE 2: SELECT lower(unaccent(btrim(regexp_replace($1, '\\s+', ' ',... ^HINT: No function matches the given name and argument types. You might need to add explicit type casts.QUERY: SELECT lower(unaccent(btrim(regexp_replace($1, '\\s+', ' ', 'g'), ' \"')))CONTEXT: SQL function \"semantic_normalize\" during inlining Command was: -- For binary upgrade, must preserve pg_class oidsSELECT pg_catalog.binary_upgrade_set_next_index_pg_class_oid('427215'::pg_catalog.oid);CREATE INDEX \"ix_semantic_mapping_normalize_title\" ON \"public\".\"semantic_mapping\" USING \"btree\" (\"public\".\"semantic_normalize\"(\"title\"));----- snip -----These occurrences of unaccent at the end are the only ones in the file:$ sudo grep -n unaccent pg_upgrade_dump_409041.log g1713:pg_restore: [archiver (db)] could not execute query: ERROR: function unaccent(text) does not exist1714:LINE 2: SELECT lower(unaccent(btrim(regexp_replace($1, '\\s+', ' ',...1718: SELECT lower(unaccent(btrim(regexp_replace($1, '\\s+', ' ', 'g'), ' \"')))Here is the definition of that index in the old DB:\\d ix_semantic_mapping_normalize_title Index \"public.ix_semantic_mapping_normalize_title\" Column | Type | Definition--------------------+------+--------------------------- semantic_normalize | text | semantic_normalize(title)btree, for table \"public.semantic_mapping\"and the semantic_normalize function it uses:\\x \\df+ semantic_normalizeExpanded display is on.List of functions-[ RECORD 1 ]-------+---------------------------------------------------------------------------Schema | publicName | semantic_normalizeResult data type | textArgument data types | title textType | normalVolatility | immutableOwner | dm_adminSecurity | invokerAccess privileges |Language | sqlSource code | + | SELECT lower(unaccent(btrim(regexp_replace($1, '\\s+', ' ', 'g'), ' \"')))+ |Description |Any more information I can provide, to help troubleshoot this?Cheers,Gulli","msg_date":"Wed, 29 Aug 2018 18:09:53 +0000","msg_from":"Gunnlaugur Thor Briem <gunnlaugur@gmail.com>","msg_from_op":true,"msg_subject":"pg_upgrade fails saying function unaccent(text) doesn't exist"},{"msg_contents":"On Wed, Aug 29, 2018 at 06:09:53PM +0000, Gunnlaugur Thor Briem wrote:\n> Hi,\n> \n> I'm trying to upgrade an old PG 9.4 cluster to PG 10, and it's failing on\n> creating an index that uses the unaccent(text) function.\n> \n> That function is part of the unaccent extension, which is installed in the old\n> DB cluster. I expect pg_upgrade to create that extension as part of the\n> upgrade. It does create other extensions that are installed in the old DB\n> cluster. I don't get why this one isn't included.\n\nThis is caused by security changes made in PG 10.3 and other minor\nreleases. Please see this thread for an outline of the issue:\n\n https://www.postgresql.org/message-id/flat/152106914669.1223.5104148605998271987%40wrigleys.postgresql.org\n\nI think you have to change your index function to specify the schema\nname before the unacces function call, e.g.\n\n\tSELECT lower(public.unaccent(btrim(regexp_replace(\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n","msg_date":"Wed, 29 Aug 2018 15:06:35 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: pg_upgrade fails saying function unaccent(text) doesn't exist"},{"msg_contents":"Thank you! That got the pg_upgrade to completion. But then during\n./analyze_new_cluster.sh vacuum fails thus:\n\nvacuumdb: processing database \"dm_test\": Generating minimal optimizer\nstatistics (1 target)\nvacuumdb: vacuuming of database \"dm_test\" failed: ERROR: text search\ndictionary \"unaccent\" does not exist\nCONTEXT: SQL function \"semantic_normalize\" statement 1\n\nAnd yet a text search dictionary with that name does exist:\n\n$ psql -d dm_test -c '\\dFd+ unaccent'\n List of text search dictionaries\n Schema | Name | Template | Init options | Description\n--------+----------+-----------------+--------------------+-------------\n public | unaccent | public.unaccent | rules = 'unaccent' |\n(1 row)\n\nRunning VACUUM ANALYZE semantic_mapping in psql works:\n\n$ psql -d dm_test -c 'VACUUM ANALYZE semantic_mapping'\nVACUUM\nTime: 1231,767 ms (00:01,232)\n\nBut running it with the vacuumdb command doesn't:\n\nvacuumdb -z -t semantic_mapping dm_test\nvacuumdb: vacuuming database \"dm_test\"\nvacuumdb: vacuuming of table \"semantic_mapping\" in database \"dm_test\"\nfailed: ERROR: text search dictionary \"unaccent\" does not exist\nCONTEXT: SQL function \"semantic_normalize\" statement 1\n\nThis is presumably a similar search path problem, because I can reproduce\nthis in psql by setting the search path to exclude public:\n\nset search_path to \"$user\";\nvacuum analyze public.semantic_mapping;\nERROR: text search dictionary \"unaccent\" does not exist\nCONTEXT: SQL function \"semantic_normalize\" statement 1\nTime: 851,562 ms\n\nCan't find a place to poke the \"public.\" prefix in to work around this ...\nI can't even see where it's getting the link to the text search dictionary\nfrom. Is that in native code in the unaccent extension?\n\nThe unaccent definition looks like this:\n\n\\df+ public.unaccent\n\nList of functions\n Schema | Name | Result data type | Argument data types | Type |\nVolatility | Parallel | Owner | Security | Access privileges | Language |\nSource code | Description\n--------+----------+------------------+---------------------+--------+------------+----------+-------+----------+-------------------+----------+---------------+-------------\n public | unaccent | text | regdictionary, text | normal |\nstable | safe | gthb | invoker | | c |\nunaccent_dict |\n public | unaccent | text | text | normal |\nstable | safe | gthb | invoker | | c |\nunaccent_dict |\n(2 rows)\n\nAny tips?\n\nCheers,\nGulli\n\nOn Wed, Aug 29, 2018 at 7:06 PM Bruce Momjian <bruce@momjian.us> wrote:\n\n> On Wed, Aug 29, 2018 at 06:09:53PM +0000, Gunnlaugur Thor Briem wrote:\n> > Hi,\n> >\n> > I'm trying to upgrade an old PG 9.4 cluster to PG 10, and it's failing on\n> > creating an index that uses the unaccent(text) function.\n> >\n> > That function is part of the unaccent extension, which is installed in\n> the old\n> > DB cluster. I expect pg_upgrade to create that extension as part of the\n> > upgrade. It does create other extensions that are installed in the old DB\n> > cluster. I don't get why this one isn't included.\n>\n> This is caused by security changes made in PG 10.3 and other minor\n> releases. Please see this thread for an outline of the issue:\n>\n>\n> https://www.postgresql.org/message-id/flat/152106914669.1223.5104148605998271987%40wrigleys.postgresql.org\n>\n> I think you have to change your index function to specify the schema\n> name before the unacces function call, e.g.\n>\n> SELECT lower(public.unaccent(btrim(regexp_replace(\n>\n> --\n> Bruce Momjian <bruce@momjian.us> http://momjian.us\n> EnterpriseDB http://enterprisedb.com\n>\n> + As you are, so once was I. As I am, so you will be. +\n> + Ancient Roman grave inscription +\n>\n\nThank you! That got the pg_upgrade to completion. But then during ./analyze_new_cluster.sh vacuum fails thus:vacuumdb: processing database \"dm_test\": Generating minimal optimizer statistics (1 target)vacuumdb: vacuuming of database \"dm_test\" failed: ERROR: text search dictionary \"unaccent\" does not existCONTEXT: SQL function \"semantic_normalize\" statement 1And yet a text search dictionary with that name does exist:$ psql -d dm_test -c '\\dFd+ unaccent' List of text search dictionaries Schema | Name | Template | Init options | Description--------+----------+-----------------+--------------------+------------- public | unaccent | public.unaccent | rules = 'unaccent' |(1 row)Running VACUUM ANALYZE semantic_mapping in psql works:$ psql -d dm_test -c 'VACUUM ANALYZE semantic_mapping'VACUUMTime: 1231,767 ms (00:01,232)But running it with the vacuumdb command doesn't:vacuumdb -z -t semantic_mapping dm_testvacuumdb: vacuuming database \"dm_test\"vacuumdb: vacuuming of table \"semantic_mapping\" in database \"dm_test\" failed: ERROR: text search dictionary \"unaccent\" does not existCONTEXT: SQL function \"semantic_normalize\" statement 1This is presumably a similar search path problem, because I can reproduce this in psql by setting the search path to exclude public:set search_path to \"$user\";vacuum analyze public.semantic_mapping;ERROR: text search dictionary \"unaccent\" does not existCONTEXT: SQL function \"semantic_normalize\" statement 1Time: 851,562 msCan't find a place to poke the \"public.\" prefix in to work around this ... I can't even see where it's getting the link to the text search dictionary from. Is that in native code in the unaccent extension?The unaccent definition looks like this:\\df+ public.unaccent List of functions Schema | Name | Result data type | Argument data types | Type | Volatility | Parallel | Owner | Security | Access privileges | Language | Source code | Description--------+----------+------------------+---------------------+--------+------------+----------+-------+----------+-------------------+----------+---------------+------------- public | unaccent | text | regdictionary, text | normal | stable | safe | gthb | invoker | | c | unaccent_dict | public | unaccent | text | text | normal | stable | safe | gthb | invoker | | c | unaccent_dict |(2 rows)Any tips?Cheers,GulliOn Wed, Aug 29, 2018 at 7:06 PM Bruce Momjian <bruce@momjian.us> wrote:On Wed, Aug 29, 2018 at 06:09:53PM +0000, Gunnlaugur Thor Briem wrote:\n> Hi,\n> \n> I'm trying to upgrade an old PG 9.4 cluster to PG 10, and it's failing on\n> creating an index that uses the unaccent(text) function.\n> \n> That function is part of the unaccent extension, which is installed in the old\n> DB cluster. I expect pg_upgrade to create that extension as part of the\n> upgrade. It does create other extensions that are installed in the old DB\n> cluster. I don't get why this one isn't included.\n\nThis is caused by security changes made in PG 10.3 and other minor\nreleases. Please see this thread for an outline of the issue:\n\n https://www.postgresql.org/message-id/flat/152106914669.1223.5104148605998271987%40wrigleys.postgresql.org\n\nI think you have to change your index function to specify the schema\nname before the unacces function call, e.g.\n\n SELECT lower(public.unaccent(btrim(regexp_replace(\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +","msg_date":"Thu, 30 Aug 2018 09:24:13 +0000","msg_from":"Gunnlaugur Thor Briem <gunnlaugur@gmail.com>","msg_from_op":true,"msg_subject":"Re: pg_upgrade fails saying function unaccent(text) doesn't exist"},{"msg_contents":"On 08/30/2018 02:24 AM, Gunnlaugur Thor Briem wrote:\n> Thank you! That got the pg_upgrade to completion. But then during \n> ./analyze_new_cluster.sh vacuum fails thus:\n> \n> vacuumdb: processing database \"dm_test\": Generating minimal optimizer \n> statistics (1 target)\n> vacuumdb: vacuuming of database \"dm_test\" failed: ERROR: text search \n> dictionary \"unaccent\" does not exist\n> CONTEXT: SQL function \"semantic_normalize\" statement 1\n> \n> And yet a text search dictionary with that name does exist:\n> \n> $ psql -d dm_test -c '\\dFd+ unaccent'\n> List of text search dictionaries\n> Schema | Name | Template | Init options | Description\n> --------+----------+-----------------+--------------------+-------------\n> public | unaccent | public.unaccent | rules = 'unaccent' |\n> (1 row)\n> \n> Running VACUUM ANALYZE semantic_mapping in psql works:\n> \n> $ psql -d dm_test -c 'VACUUM ANALYZE semantic_mapping'\n> VACUUM\n> Time: 1231,767 ms (00:01,232)\n> \n> But running it with the vacuumdb command doesn't:\n> \n> vacuumdb -z -t semantic_mapping dm_test\n> vacuumdb: vacuuming database \"dm_test\"\n> vacuumdb: vacuuming of table \"semantic_mapping\" in database \"dm_test\" \n> failed: ERROR: text search dictionary \"unaccent\" does not exist\n> CONTEXT: SQL function \"semantic_normalize\" statement 1\n> \n> This is presumably a similar search path problem, because I can \n> reproduce this in psql by setting the search path to exclude public:\n> \n> set search_path to \"$user\";\n> vacuum analyze public.semantic_mapping;\n> ERROR: text search dictionary \"unaccent\" does not exist\n> CONTEXT: SQL function \"semantic_normalize\" statement 1\n> Time: 851,562 ms\n> \n> Can't find a place to poke the \"public.\" prefix in to work around this \n> ... I can't even see where it's getting the link to the text search \n> dictionary from. Is that in native code in the unaccent extension?\n\nSince the semantic_normalize function is tripping it and it uses \nunaccent I would say it is native to the extension.\n\nWhat does:\n\n\\dFd unaccent\n\nshow?\n\n> \n> The unaccent definition looks like this:\n> \n> \\df+ public.unaccent\n> List of functions\n> Schema | Name | Result data type | Argument data types | Type | \n> Volatility | Parallel | Owner | Security | Access privileges | Language \n> | Source code | Description\n> --------+----------+------------------+---------------------+--------+------------+----------+-------+----------+-------------------+----------+---------------+-------------\n> public | unaccent | text | regdictionary, text | normal | \n> stable | safe | gthb | invoker | | c \n> | unaccent_dict |\n> public | unaccent | text | text | normal | \n> stable | safe | gthb | invoker | | c \n> | unaccent_dict |\n> (2 rows)\n> \n> Any tips?\n> \n> Cheers,\n> Gulli\n> \n> On Wed, Aug 29, 2018 at 7:06 PM Bruce Momjian <bruce@momjian.us \n> <mailto:bruce@momjian.us>> wrote:\n> \n> On Wed, Aug 29, 2018 at 06:09:53PM +0000, Gunnlaugur Thor Briem wrote:\n> > Hi,\n> >\n> > I'm trying to upgrade an old PG 9.4 cluster to PG 10, and it's\n> failing on\n> > creating an index that uses the unaccent(text) function.\n> >\n> > That function is part of the unaccent extension, which is\n> installed in the old\n> > DB cluster. I expect pg_upgrade to create that extension as part\n> of the\n> > upgrade. It does create other extensions that are installed in\n> the old DB\n> > cluster. I don't get why this one isn't included.\n> \n> This is caused by security changes made in PG 10.3 and other minor\n> releases. Please see this thread for an outline of the issue:\n> \n> https://www.postgresql.org/message-id/flat/152106914669.1223.5104148605998271987%40wrigleys.postgresql.org\n> \n> I think you have to change your index function to specify the schema\n> name before the unacces function call, e.g.\n> \n> SELECT lower(public.unaccent(btrim(regexp_replace(\n> \n> -- \n> Bruce Momjian <bruce@momjian.us <mailto:bruce@momjian.us>>\n> http://momjian.us\n> EnterpriseDB http://enterprisedb.com\n> \n> + As you are, so once was I. As I am, so you will be. +\n> + Ancient Roman grave inscription +\n> \n\n\n-- \nAdrian Klaver\nadrian.klaver@aklaver.com\n\n","msg_date":"Thu, 30 Aug 2018 06:48:26 -0700","msg_from":"Adrian Klaver <adrian.klaver@aklaver.com>","msg_from_op":false,"msg_subject":"Re: pg_upgrade fails saying function unaccent(text) doesn't exist"},{"msg_contents":"\\dFd unaccent\nList of text search dictionaries\n Schema | Name | Description\n--------+----------+-------------\n public | unaccent |\n(1 row)\n\n\\dFd+ unaccent\n List of text search dictionaries\n Schema | Name | Template | Init options | Description\n--------+----------+-----------------+--------------------+-------------\n public | unaccent | public.unaccent | rules = 'unaccent' |\n(1 row)\n\nCheers,\nGulli\n\nOn Thu, Aug 30, 2018 at 1:48 PM Adrian Klaver <adrian.klaver@aklaver.com>\nwrote:\n\n> On 08/30/2018 02:24 AM, Gunnlaugur Thor Briem wrote:\n> > Thank you! That got the pg_upgrade to completion. But then during\n> > ./analyze_new_cluster.sh vacuum fails thus:\n> >\n> > vacuumdb: processing database \"dm_test\": Generating minimal optimizer\n> > statistics (1 target)\n> > vacuumdb: vacuuming of database \"dm_test\" failed: ERROR: text search\n> > dictionary \"unaccent\" does not exist\n> > CONTEXT: SQL function \"semantic_normalize\" statement 1\n> >\n> > And yet a text search dictionary with that name does exist:\n> >\n> > $ psql -d dm_test -c '\\dFd+ unaccent'\n> > List of text search dictionaries\n> > Schema | Name | Template | Init options | Description\n> > --------+----------+-----------------+--------------------+-------------\n> > public | unaccent | public.unaccent | rules = 'unaccent' |\n> > (1 row)\n> >\n> > Running VACUUM ANALYZE semantic_mapping in psql works:\n> >\n> > $ psql -d dm_test -c 'VACUUM ANALYZE semantic_mapping'\n> > VACUUM\n> > Time: 1231,767 ms (00:01,232)\n> >\n> > But running it with the vacuumdb command doesn't:\n> >\n> > vacuumdb -z -t semantic_mapping dm_test\n> > vacuumdb: vacuuming database \"dm_test\"\n> > vacuumdb: vacuuming of table \"semantic_mapping\" in database \"dm_test\"\n> > failed: ERROR: text search dictionary \"unaccent\" does not exist\n> > CONTEXT: SQL function \"semantic_normalize\" statement 1\n> >\n> > This is presumably a similar search path problem, because I can\n> > reproduce this in psql by setting the search path to exclude public:\n> >\n> > set search_path to \"$user\";\n> > vacuum analyze public.semantic_mapping;\n> > ERROR: text search dictionary \"unaccent\" does not exist\n> > CONTEXT: SQL function \"semantic_normalize\" statement 1\n> > Time: 851,562 ms\n> >\n> > Can't find a place to poke the \"public.\" prefix in to work around this\n> > ... I can't even see where it's getting the link to the text search\n> > dictionary from. Is that in native code in the unaccent extension?\n>\n> Since the semantic_normalize function is tripping it and it uses\n> unaccent I would say it is native to the extension.\n>\n> What does:\n>\n> \\dFd unaccent\n>\n> show?\n>\n> >\n> > The unaccent definition looks like this:\n> >\n> > \\df+ public.unaccent\n> >\n>\n> List of functions\n> > Schema | Name | Result data type | Argument data types | Type |\n> > Volatility | Parallel | Owner | Security | Access privileges | Language\n> > | Source code | Description\n> >\n> --------+----------+------------------+---------------------+--------+------------+----------+-------+----------+-------------------+----------+---------------+-------------\n> > public | unaccent | text | regdictionary, text | normal |\n> > stable | safe | gthb | invoker | | c\n> > | unaccent_dict |\n> > public | unaccent | text | text | normal |\n> > stable | safe | gthb | invoker | | c\n> > | unaccent_dict |\n> > (2 rows)\n> >\n> > Any tips?\n> >\n> > Cheers,\n> > Gulli\n> >\n> > On Wed, Aug 29, 2018 at 7:06 PM Bruce Momjian <bruce@momjian.us\n> > <mailto:bruce@momjian.us>> wrote:\n> >\n> > On Wed, Aug 29, 2018 at 06:09:53PM +0000, Gunnlaugur Thor Briem\n> wrote:\n> > > Hi,\n> > >\n> > > I'm trying to upgrade an old PG 9.4 cluster to PG 10, and it's\n> > failing on\n> > > creating an index that uses the unaccent(text) function.\n> > >\n> > > That function is part of the unaccent extension, which is\n> > installed in the old\n> > > DB cluster. I expect pg_upgrade to create that extension as part\n> > of the\n> > > upgrade. It does create other extensions that are installed in\n> > the old DB\n> > > cluster. I don't get why this one isn't included.\n> >\n> > This is caused by security changes made in PG 10.3 and other minor\n> > releases. Please see this thread for an outline of the issue:\n> >\n> >\n> https://www.postgresql.org/message-id/flat/152106914669.1223.5104148605998271987%40wrigleys.postgresql.org\n> >\n> > I think you have to change your index function to specify the schema\n> > name before the unacces function call, e.g.\n> >\n> > SELECT lower(public.unaccent(btrim(regexp_replace(\n> >\n> > --\n> > Bruce Momjian <bruce@momjian.us <mailto:bruce@momjian.us>>\n> > http://momjian.us\n> > EnterpriseDB http://enterprisedb.com\n> >\n> > + As you are, so once was I. As I am, so you will be. +\n> > + Ancient Roman grave inscription +\n> >\n>\n>\n> --\n> Adrian Klaver\n> adrian.klaver@aklaver.com\n>\n>\n\n\\dFd unaccentList of text search dictionaries Schema | Name | Description--------+----------+------------- public | unaccent |(1 row)\\dFd+ unaccent List of text search dictionaries Schema | Name | Template | Init options | Description--------+----------+-----------------+--------------------+------------- public | unaccent | public.unaccent | rules = 'unaccent' |(1 row)Cheers,GulliOn Thu, Aug 30, 2018 at 1:48 PM Adrian Klaver <adrian.klaver@aklaver.com> wrote:On 08/30/2018 02:24 AM, Gunnlaugur Thor Briem wrote:\n> Thank you! That got the pg_upgrade to completion. But then during \n> ./analyze_new_cluster.sh vacuum fails thus:\n> \n> vacuumdb: processing database \"dm_test\": Generating minimal optimizer \n> statistics (1 target)\n> vacuumdb: vacuuming of database \"dm_test\" failed: ERROR: text search \n> dictionary \"unaccent\" does not exist\n> CONTEXT: SQL function \"semantic_normalize\" statement 1\n> \n> And yet a text search dictionary with that name does exist:\n> \n> $ psql -d dm_test -c '\\dFd+ unaccent'\n> List of text search dictionaries\n> Schema | Name | Template | Init options | Description\n> --------+----------+-----------------+--------------------+-------------\n> public | unaccent | public.unaccent | rules = 'unaccent' |\n> (1 row)\n> \n> Running VACUUM ANALYZE semantic_mapping in psql works:\n> \n> $ psql -d dm_test -c 'VACUUM ANALYZE semantic_mapping'\n> VACUUM\n> Time: 1231,767 ms (00:01,232)\n> \n> But running it with the vacuumdb command doesn't:\n> \n> vacuumdb -z -t semantic_mapping dm_test\n> vacuumdb: vacuuming database \"dm_test\"\n> vacuumdb: vacuuming of table \"semantic_mapping\" in database \"dm_test\" \n> failed: ERROR: text search dictionary \"unaccent\" does not exist\n> CONTEXT: SQL function \"semantic_normalize\" statement 1\n> \n> This is presumably a similar search path problem, because I can \n> reproduce this in psql by setting the search path to exclude public:\n> \n> set search_path to \"$user\";\n> vacuum analyze public.semantic_mapping;\n> ERROR: text search dictionary \"unaccent\" does not exist\n> CONTEXT: SQL function \"semantic_normalize\" statement 1\n> Time: 851,562 ms\n> \n> Can't find a place to poke the \"public.\" prefix in to work around this \n> ... I can't even see where it's getting the link to the text search \n> dictionary from. Is that in native code in the unaccent extension?\n\nSince the semantic_normalize function is tripping it and it uses \nunaccent I would say it is native to the extension.\n\nWhat does:\n\n\\dFd unaccent\n\nshow?\n\n> \n> The unaccent definition looks like this:\n> \n> \\df+ public.unaccent\n> List of functions\n> Schema | Name | Result data type | Argument data types | Type | \n> Volatility | Parallel | Owner | Security | Access privileges | Language \n> | Source code | Description\n> --------+----------+------------------+---------------------+--------+------------+----------+-------+----------+-------------------+----------+---------------+-------------\n> public | unaccent | text | regdictionary, text | normal | \n> stable | safe | gthb | invoker | | c \n> | unaccent_dict |\n> public | unaccent | text | text | normal | \n> stable | safe | gthb | invoker | | c \n> | unaccent_dict |\n> (2 rows)\n> \n> Any tips?\n> \n> Cheers,\n> Gulli\n> \n> On Wed, Aug 29, 2018 at 7:06 PM Bruce Momjian <bruce@momjian.us \n> <mailto:bruce@momjian.us>> wrote:\n> \n> On Wed, Aug 29, 2018 at 06:09:53PM +0000, Gunnlaugur Thor Briem wrote:\n> > Hi,\n> >\n> > I'm trying to upgrade an old PG 9.4 cluster to PG 10, and it's\n> failing on\n> > creating an index that uses the unaccent(text) function.\n> >\n> > That function is part of the unaccent extension, which is\n> installed in the old\n> > DB cluster. I expect pg_upgrade to create that extension as part\n> of the\n> > upgrade. It does create other extensions that are installed in\n> the old DB\n> > cluster. I don't get why this one isn't included.\n> \n> This is caused by security changes made in PG 10.3 and other minor\n> releases. Please see this thread for an outline of the issue:\n> \n> https://www.postgresql.org/message-id/flat/152106914669.1223.5104148605998271987%40wrigleys.postgresql.org\n> \n> I think you have to change your index function to specify the schema\n> name before the unacces function call, e.g.\n> \n> SELECT lower(public.unaccent(btrim(regexp_replace(\n> \n> -- \n> Bruce Momjian <bruce@momjian.us <mailto:bruce@momjian.us>>\n> http://momjian.us\n> EnterpriseDB http://enterprisedb.com\n> \n> + As you are, so once was I. As I am, so you will be. +\n> + Ancient Roman grave inscription +\n> \n\n\n-- \nAdrian Klaver\nadrian.klaver@aklaver.com","msg_date":"Thu, 30 Aug 2018 16:03:11 +0000","msg_from":"Gunnlaugur Thor Briem <gunnlaugur@gmail.com>","msg_from_op":true,"msg_subject":"Re: pg_upgrade fails saying function unaccent(text) doesn't exist"},{"msg_contents":"On 08/30/2018 09:03 AM, Gunnlaugur Thor Briem wrote:\n> \\dFd unaccent\n> List of text search dictionaries\n> Schema | Name | Description\n> --------+----------+-------------\n> public | unaccent |\n> (1 row)\n> \n> \\dFd+ unaccent\n> List of text search dictionaries\n> Schema | Name | Template | Init options | Description\n> --------+----------+-----------------+--------------------+-------------\n> public | unaccent | public.unaccent | rules = 'unaccent' |\n> (1 row)\n> \n> Cheers,\n> Gulli\n> \n\nI could not replicate with simple case:\n\nselect version();\n version \n\n------------------------------------------------------------------------------------\n PostgreSQL 10.5 on x86_64-pc-linux-gnu, compiled by gcc (SUSE Linux) \n4.8.5, 64-bit\n\nCREATE OR REPLACE FUNCTION public.semantic_normalize(title text)\n RETURNS text\n LANGUAGE sql\nAS $function$\nSELECT lower(unaccent(btrim(regexp_replace($1,\n'\\s+', ' ', 'g'), ' \"')))\n$function$\n\nCREATE TABLE unaccent_test(title text);\n\nINSERT INTO unaccent_test values ('Hukić'), ('Böttcher'), ('ÀÁÂÃÄÅ'), \n('électro');\n\nCREATE INDEX ix_semantic_normalize_title on unaccent_test(title);\n\nVACUUM ANALYZE unaccent_test;\nVACUUM\n\nvacuumdb -U postgres -z -t unaccent_test test\nvacuumdb: vacuuming database \"test\"\n\n\nThe only thing I can think of is that you have an older version of \nvacuumdb that is not aware of the schema specification changes in the \nnewer versions of Postgrse.\n\n\n\n-- \nAdrian Klaver\nadrian.klaver@aklaver.com\n\n","msg_date":"Fri, 31 Aug 2018 07:14:18 -0700","msg_from":"Adrian Klaver <adrian.klaver@aklaver.com>","msg_from_op":false,"msg_subject":"Re: pg_upgrade fails saying function unaccent(text) doesn't exist"},{"msg_contents":"On 08/30/2018 09:03 AM, Gunnlaugur Thor Briem wrote:\n> \\dFd unaccent\n> List of text search dictionaries\n> Schema | Name | Description\n> --------+----------+-------------\n> public | unaccent |\n> (1 row)\n> \n> \\dFd+ unaccent\n> List of text search dictionaries\n> Schema | Name | Template | Init options | Description\n> --------+----------+-----------------+--------------------+-------------\n> public | unaccent | public.unaccent | rules = 'unaccent' |\n> (1 row)\n> \n> Cheers,\n> Gulli\n> \n\nForgot to add to previous post:\n\n\\dx unaccent\n List of installed extensions\n Name | Version | Schema | Description\n----------+---------+--------+---------------------------------------------\n unaccent | 1.1 | public | text search dictionary that removes accents\n\n\n\n-- \nAdrian Klaver\nadrian.klaver@aklaver.com\n\n","msg_date":"Fri, 31 Aug 2018 07:17:06 -0700","msg_from":"Adrian Klaver <adrian.klaver@aklaver.com>","msg_from_op":false,"msg_subject":"Re: pg_upgrade fails saying function unaccent(text) doesn't exist"},{"msg_contents":"OK, I found the cause of the unaccent dictionary problem, and a workaround.\n\nIt's not the vacuumdb version, not the unaccent version, and it's not even\na pg_upgrade problem: I get this error also with PG 9.4.18 running on the\nold cluster, with both the 10.5 vacuumdb and the 9.4.18 vacuumdb, and I get\nthe same error in both.\n\nAnd it's not strictly a vacuumdb problem, though vacuumdb triggers it.\n\nHere's a very minimal test case, unrelated to my DB, that you ought to be\nable to reproduce:\n\nSET search_path = \"$user\"; SELECT public.unaccent('fóö');\nSET\nERROR: text search dictionary \"unaccent\" does not exist\n\nand here's a workaround:\n\nSET search_path = \"$user\"; SELECT public.unaccent(tsdict.oid, 'fóö')\nFROM pg_catalog.pg_ts_dict tsdict WHERE dictname='unaccent';\nSET\n unaccent\n----------\n foo\n(1 row)\n\nThe workaround avoids the OID lookup of the dictionary ... that lookup (in\nthe single-argument unaccent function) is done by unqualified name:\n\nhttps://github.com/postgres/postgres/blob/fb8697b31aaeebe6170c572739867dcaa01053c6/contrib/unaccent/unaccent.c#L377\n\n dictOid = get_ts_dict_oid(stringToQualifiedNameList(\"unaccent\"),\nfalse);\n\nand that fails if the search path doesn't include public.\n\nSo it is indeed triggered by the security changes that Bruce mentioned;\nthose were backported into 9.4.17:\nhttps://www.postgresql.org/docs/9.4/static/release-9-4-17.html ... and so\ngot pulled in by my Macports upgrades. So nothing to do with pg_upgrade.\n\nSo the workaround for my vacuumdb/function-index problem is to give\nunaccent the OID of the text search dictionary, so that the search path\nisn't in play:\n\nCREATE OR REPLACE FUNCTION public.semantic_normalize(title text)\nRETURNS text\n LANGUAGE sql\n IMMUTABLE STRICT\nAS $function$\n SELECT lower(public.unaccent(16603, btrim(regexp_replace($1, '\\s+', ' ',\n'g'), ' \"')))\n$function$;\n\nand that makes vacuumdb -z work in both 9.4.18 and 10.5, and makes\n./analyze_new_cluster.sh complete without problems.\n\nThe proper fix is, I suppose, to make the single-argument unaccent function\nexplicitly look up the dictionary in the same schema as the function itself\nis in.\n\nCheers,\nGulli\n\nOK, I found the cause of the unaccent dictionary problem, and a workaround.It's not the vacuumdb version, not the unaccent version, and it's not even a pg_upgrade problem: I get this error also with PG 9.4.18 running on the old cluster, with both the 10.5 vacuumdb and the 9.4.18 vacuumdb, and I get the same error in both.And it's not strictly a vacuumdb problem, though vacuumdb triggers it.Here's a very minimal test case, unrelated to my DB, that you ought to be able to reproduce:SET search_path = \"$user\"; SELECT public.unaccent('fóö');SETERROR: text search dictionary \"unaccent\" does not existand here's a workaround:SET search_path = \"$user\"; SELECT public.unaccent(tsdict.oid, 'fóö') FROM pg_catalog.pg_ts_dict tsdict WHERE dictname='unaccent';SET unaccent---------- foo(1 row)The workaround avoids the OID lookup of the dictionary ... that lookup (in the single-argument unaccent function) is done by unqualified name:https://github.com/postgres/postgres/blob/fb8697b31aaeebe6170c572739867dcaa01053c6/contrib/unaccent/unaccent.c#L377 dictOid = get_ts_dict_oid(stringToQualifiedNameList(\"unaccent\"), false);and that fails if the search path doesn't include public.So it is indeed triggered by the security changes that Bruce mentioned; those were backported into 9.4.17: https://www.postgresql.org/docs/9.4/static/release-9-4-17.html ... and so got pulled in by my Macports upgrades. So nothing to do with pg_upgrade.So the workaround for my vacuumdb/function-index problem is to give unaccent the OID of the text search dictionary, so that the search path isn't in play:CREATE OR REPLACE FUNCTION public.semantic_normalize(title text)RETURNS text LANGUAGE sql IMMUTABLE STRICTAS $function$ SELECT lower(public.unaccent(16603, btrim(regexp_replace($1, '\\s+', ' ', 'g'), ' \"')))$function$;and that makes vacuumdb -z work in both 9.4.18 and 10.5, and makes ./analyze_new_cluster.sh complete without problems.The proper fix is, I suppose, to make the single-argument unaccent function explicitly look up the dictionary in the same schema as the function itself is in.Cheers,Gulli","msg_date":"Wed, 5 Sep 2018 08:49:50 +0000","msg_from":"Gunnlaugur Thor Briem <gunnlaugur@gmail.com>","msg_from_op":true,"msg_subject":"unaccent(text) fails depending on search_path (WAS: pg_upgrade fails\n saying function unaccent(text) doesn't exist)"},{"msg_contents":"On 09/05/2018 01:49 AM, Gunnlaugur Thor Briem wrote:\n> OK, I found the cause of the unaccent dictionary problem, and a workaround.\n> \n> It's not the vacuumdb version, not the unaccent version, and it's not \n> even a pg_upgrade problem: I get this error also with PG 9.4.18 running \n> on the old cluster, with both the 10.5 vacuumdb and the 9.4.18 vacuumdb, \n> and I get the same error in both.\n> \n> And it's not strictly a vacuumdb problem, though vacuumdb triggers it.\n> \n> Here's a very minimal test case, unrelated to my DB, that you ought to \n> be able to reproduce:\n> \n> SET search_path = \"$user\"; SELECT public.unaccent('fóö');\n> SET\n> ERROR: text search dictionary \"unaccent\" does not exist\n> \n> and here's a workaround:\n> \n> SET search_path = \"$user\"; SELECT public.unaccent(tsdict.oid, 'fóö')\n> FROM pg_catalog.pg_ts_dict tsdict WHERE dictname='unaccent';\n> SET\n> unaccent\n> ----------\n> foo\n> (1 row)\n> \n> The workaround avoids the OID lookup of the dictionary ... that lookup \n> (in the single-argument unaccent function) is done by unqualified name:\n> \n> https://github.com/postgres/postgres/blob/fb8697b31aaeebe6170c572739867dcaa01053c6/contrib/unaccent/unaccent.c#L377\n> \n> dictOid = \n> get_ts_dict_oid(stringToQualifiedNameList(\"unaccent\"), false);\n> \n> and that fails if the search path doesn't include public. >\n> So it is indeed triggered by the security changes that Bruce mentioned; \n> those were backported into 9.4.17: \n> https://www.postgresql.org/docs/9.4/static/release-9-4-17.html ... and \n> so got pulled in by my Macports upgrades. So nothing to do with pg_upgrade.\n> \n> So the workaround for my vacuumdb/function-index problem is to give \n> unaccent the OID of the text search dictionary, so that the search path \n> isn't in play:\n> \n> CREATE OR REPLACE FUNCTION public.semantic_normalize(title text)\n> RETURNS text\n> LANGUAGE sql\n> IMMUTABLE STRICT\n> AS $function$\n> SELECT lower(public.unaccent(16603, btrim(regexp_replace($1, '\\s+', ' \n> ', 'g'), ' \"')))\n> $function$;\n> \n> and that makes vacuumdb -z work in both 9.4.18 and 10.5, and makes \n> ./analyze_new_cluster.sh complete without problems.\n\n\nNice investigation. Working off the above, I offer a suggestion:\n\nSET search_path = \"$user\"; SELECT public.unaccent('unaccent', 'fóö');\nSET\nERROR: text search dictionary \"unaccent\" does not exist\nLINE 1: SELECT public.unaccent('unaccent', 'fóö');\n\n\nSET search_path = \"$user\"; SELECT public.unaccent('public.unaccent', 'fóö');\nSET\n unaccent\n----------\n foo\n\nThat eliminates hard wiring the OID.\n\n> \n> The proper fix is, I suppose, to make the single-argument unaccent \n> function explicitly look up the dictionary in the same schema as the \n> function itself is in.\n> \n> Cheers,\n> Gulli\n> \n\n\n-- \nAdrian Klaver\nadrian.klaver@aklaver.com\n\n","msg_date":"Wed, 5 Sep 2018 07:00:03 -0700","msg_from":"Adrian Klaver <adrian.klaver@aklaver.com>","msg_from_op":false,"msg_subject":"Re: unaccent(text) fails depending on search_path (WAS: pg_upgrade\n fails saying function unaccent(text) doesn't exist)"},{"msg_contents":"Yep, a neater workaround for sure!\n\nCheers,\nGulli\n\nOn Wed, Sep 5, 2018 at 2:00 PM Adrian Klaver <adrian.klaver@aklaver.com>\nwrote:\n\n> On 09/05/2018 01:49 AM, Gunnlaugur Thor Briem wrote:\n> > OK, I found the cause of the unaccent dictionary problem, and a\n> workaround.\n> >\n> > It's not the vacuumdb version, not the unaccent version, and it's not\n> > even a pg_upgrade problem: I get this error also with PG 9.4.18 running\n> > on the old cluster, with both the 10.5 vacuumdb and the 9.4.18 vacuumdb,\n> > and I get the same error in both.\n> >\n> > And it's not strictly a vacuumdb problem, though vacuumdb triggers it.\n> >\n> > Here's a very minimal test case, unrelated to my DB, that you ought to\n> > be able to reproduce:\n> >\n> > SET search_path = \"$user\"; SELECT public.unaccent('fóö');\n> > SET\n> > ERROR: text search dictionary \"unaccent\" does not exist\n> >\n> > and here's a workaround:\n> >\n> > SET search_path = \"$user\"; SELECT public.unaccent(tsdict.oid, 'fóö')\n> > FROM pg_catalog.pg_ts_dict tsdict WHERE dictname='unaccent';\n> > SET\n> > unaccent\n> > ----------\n> > foo\n> > (1 row)\n> >\n> > The workaround avoids the OID lookup of the dictionary ... that lookup\n> > (in the single-argument unaccent function) is done by unqualified name:\n> >\n> >\n> https://github.com/postgres/postgres/blob/fb8697b31aaeebe6170c572739867dcaa01053c6/contrib/unaccent/unaccent.c#L377\n> >\n> > dictOid =\n> > get_ts_dict_oid(stringToQualifiedNameList(\"unaccent\"), false);\n> >\n> > and that fails if the search path doesn't include public. >\n> > So it is indeed triggered by the security changes that Bruce mentioned;\n> > those were backported into 9.4.17:\n> > https://www.postgresql.org/docs/9.4/static/release-9-4-17.html ... and\n> > so got pulled in by my Macports upgrades. So nothing to do with\n> pg_upgrade.\n> >\n> > So the workaround for my vacuumdb/function-index problem is to give\n> > unaccent the OID of the text search dictionary, so that the search path\n> > isn't in play:\n> >\n> > CREATE OR REPLACE FUNCTION public.semantic_normalize(title text)\n> > RETURNS text\n> > LANGUAGE sql\n> > IMMUTABLE STRICT\n> > AS $function$\n> > SELECT lower(public.unaccent(16603, btrim(regexp_replace($1, '\\s+', '\n> > ', 'g'), ' \"')))\n> > $function$;\n> >\n> > and that makes vacuumdb -z work in both 9.4.18 and 10.5, and makes\n> > ./analyze_new_cluster.sh complete without problems.\n>\n>\n> Nice investigation. Working off the above, I offer a suggestion:\n>\n> SET search_path = \"$user\"; SELECT public.unaccent('unaccent', 'fóö');\n> SET\n> ERROR: text search dictionary \"unaccent\" does not exist\n> LINE 1: SELECT public.unaccent('unaccent', 'fóö');\n>\n>\n> SET search_path = \"$user\"; SELECT public.unaccent('public.unaccent',\n> 'fóö');\n> SET\n> unaccent\n> ----------\n> foo\n>\n> That eliminates hard wiring the OID.\n>\n> >\n> > The proper fix is, I suppose, to make the single-argument unaccent\n> > function explicitly look up the dictionary in the same schema as the\n> > function itself is in.\n> >\n> > Cheers,\n> > Gulli\n> >\n>\n>\n> --\n> Adrian Klaver\n> adrian.klaver@aklaver.com\n>\n\nYep, a neater workaround for sure!Cheers,GulliOn Wed, Sep 5, 2018 at 2:00 PM Adrian Klaver <adrian.klaver@aklaver.com> wrote:On 09/05/2018 01:49 AM, Gunnlaugur Thor Briem wrote:\n> OK, I found the cause of the unaccent dictionary problem, and a workaround.\n> \n> It's not the vacuumdb version, not the unaccent version, and it's not \n> even a pg_upgrade problem: I get this error also with PG 9.4.18 running \n> on the old cluster, with both the 10.5 vacuumdb and the 9.4.18 vacuumdb, \n> and I get the same error in both.\n> \n> And it's not strictly a vacuumdb problem, though vacuumdb triggers it.\n> \n> Here's a very minimal test case, unrelated to my DB, that you ought to \n> be able to reproduce:\n> \n> SET search_path = \"$user\"; SELECT public.unaccent('fóö');\n> SET\n> ERROR: text search dictionary \"unaccent\" does not exist\n> \n> and here's a workaround:\n> \n> SET search_path = \"$user\"; SELECT public.unaccent(tsdict.oid, 'fóö')\n> FROM pg_catalog.pg_ts_dict tsdict WHERE dictname='unaccent';\n> SET\n> unaccent\n> ----------\n> foo\n> (1 row)\n> \n> The workaround avoids the OID lookup of the dictionary ... that lookup \n> (in the single-argument unaccent function) is done by unqualified name:\n> \n> https://github.com/postgres/postgres/blob/fb8697b31aaeebe6170c572739867dcaa01053c6/contrib/unaccent/unaccent.c#L377\n> \n> dictOid = \n> get_ts_dict_oid(stringToQualifiedNameList(\"unaccent\"), false);\n> \n> and that fails if the search path doesn't include public. >\n> So it is indeed triggered by the security changes that Bruce mentioned; \n> those were backported into 9.4.17: \n> https://www.postgresql.org/docs/9.4/static/release-9-4-17.html ... and \n> so got pulled in by my Macports upgrades. So nothing to do with pg_upgrade.\n> \n> So the workaround for my vacuumdb/function-index problem is to give \n> unaccent the OID of the text search dictionary, so that the search path \n> isn't in play:\n> \n> CREATE OR REPLACE FUNCTION public.semantic_normalize(title text)\n> RETURNS text\n> LANGUAGE sql\n> IMMUTABLE STRICT\n> AS $function$\n> SELECT lower(public.unaccent(16603, btrim(regexp_replace($1, '\\s+', ' \n> ', 'g'), ' \"')))\n> $function$;\n> \n> and that makes vacuumdb -z work in both 9.4.18 and 10.5, and makes \n> ./analyze_new_cluster.sh complete without problems.\n\n\nNice investigation. Working off the above, I offer a suggestion:\n\nSET search_path = \"$user\"; SELECT public.unaccent('unaccent', 'fóö');\nSET\nERROR: text search dictionary \"unaccent\" does not exist\nLINE 1: SELECT public.unaccent('unaccent', 'fóö');\n\n\nSET search_path = \"$user\"; SELECT public.unaccent('public.unaccent', 'fóö');\nSET\n unaccent\n----------\n foo\n\nThat eliminates hard wiring the OID.\n\n> \n> The proper fix is, I suppose, to make the single-argument unaccent \n> function explicitly look up the dictionary in the same schema as the \n> function itself is in.\n> \n> Cheers,\n> Gulli\n> \n\n\n-- \nAdrian Klaver\nadrian.klaver@aklaver.com","msg_date":"Wed, 5 Sep 2018 14:06:17 +0000","msg_from":"Gunnlaugur Thor Briem <gunnlaugur@gmail.com>","msg_from_op":true,"msg_subject":"Re: unaccent(text) fails depending on search_path (WAS: pg_upgrade\n fails saying function unaccent(text) doesn't exist)"},{"msg_contents":"Gunnlaugur Thor Briem <gunnlaugur@gmail.com> writes:\n> SET search_path = \"$user\"; SELECT public.unaccent('foo');\n> SET\n> ERROR: text search dictionary \"unaccent\" does not exist\n\nMeh. I think we need the attached, or something just about like it.\n\nIt's barely possible that there's somebody out there who's relying on\nsetting the search path to allow choosing among multiple \"unaccent\"\ndictionaries. But there are way more people whose functions are\nbroken due to the recent search-path-tightening changes.\n\n\t\t\tregards, tom lane","msg_date":"Wed, 05 Sep 2018 15:03:33 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: unaccent(text) fails depending on search_path (WAS: pg_upgrade\n fails saying function unaccent(text) doesn't exist)"},{"msg_contents":"[ redirecting to pgsql-hackers ]\n\nI wrote:\n> Gunnlaugur Thor Briem <gunnlaugur@gmail.com> writes:\n>> SET search_path = \"$user\"; SELECT public.unaccent('foo');\n>> SET\n>> ERROR: text search dictionary \"unaccent\" does not exist\n\n> Meh. I think we need the attached, or something just about like it.\n>\n> It's barely possible that there's somebody out there who's relying on\n> setting the search path to allow choosing among multiple \"unaccent\"\n> dictionaries. But there are way more people whose functions are\n> broken due to the recent search-path-tightening changes.\n\nHere's a slightly more efficient version.\n\n\t\t\tregards, tom lane","msg_date":"Wed, 05 Sep 2018 18:37:00 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: unaccent(text) fails depending on search_path (WAS: pg_upgrade\n fails saying function unaccent(text) doesn't exist)"},{"msg_contents":"On Wed, Sep 5, 2018 at 06:37:00PM -0400, Tom Lane wrote:\n> [ redirecting to pgsql-hackers ]\n> \n> I wrote:\n> > Gunnlaugur Thor Briem <gunnlaugur@gmail.com> writes:\n> >> SET search_path = \"$user\"; SELECT public.unaccent('foo');\n> >> SET\n> >> ERROR: text search dictionary \"unaccent\" does not exist\n> \n> > Meh. I think we need the attached, or something just about like it.\n> >\n> > It's barely possible that there's somebody out there who's relying on\n> > setting the search path to allow choosing among multiple \"unaccent\"\n> > dictionaries. But there are way more people whose functions are\n> > broken due to the recent search-path-tightening changes.\n> \n> Here's a slightly more efficient version.\n\nIf we are going down this route, is there any thought of handling\nearchdistance the same way?\n\n\thttps://www.postgresql.org/message-id/20180330205229.GS8476@momjian.us\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n","msg_date":"Fri, 7 Sep 2018 18:32:47 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: unaccent(text) fails depending on search_path (WAS: pg_upgrade\n fails saying function unaccent(text) doesn't exist)"},{"msg_contents":"Bruce Momjian <bruce@momjian.us> writes:\n> If we are going down this route, is there any thought of handling\n> earchdistance the same way?\n> \thttps://www.postgresql.org/message-id/20180330205229.GS8476@momjian.us\n\nAFAICS there are no internal-to-the-C-code search path dependencies\nin earthdistance.c, so it's not the same problem.\n\n\t\t\tregards, tom lane\n\n","msg_date":"Fri, 07 Sep 2018 18:43:52 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: unaccent(text) fails depending on search_path (WAS: pg_upgrade\n fails saying function unaccent(text) doesn't exist)"},{"msg_contents":"On Fri, Sep 7, 2018 at 06:43:52PM -0400, Tom Lane wrote:\n> Bruce Momjian <bruce@momjian.us> writes:\n> > If we are going down this route, is there any thought of handling\n> > earchdistance the same way?\n> > \thttps://www.postgresql.org/message-id/20180330205229.GS8476@momjian.us\n> \n> AFAICS there are no internal-to-the-C-code search path dependencies\n> in earthdistance.c, so it's not the same problem.\n\nUh, there is an SQL function that calls functions from the module that\nfail. It would be a CREATE FUNCTION patch, I think, but I thought the\nissue was the same.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n","msg_date":"Fri, 7 Sep 2018 18:56:43 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: unaccent(text) fails depending on search_path (WAS: pg_upgrade\n fails saying function unaccent(text) doesn't exist)"},{"msg_contents":"Bruce Momjian <bruce@momjian.us> writes:\n> On Fri, Sep 7, 2018 at 06:43:52PM -0400, Tom Lane wrote:\n>> AFAICS there are no internal-to-the-C-code search path dependencies\n>> in earthdistance.c, so it's not the same problem.\n\n> Uh, there is an SQL function that calls functions from the module that\n> fail. It would be a CREATE FUNCTION patch, I think, but I thought the\n> issue was the same.\n\nNot really. You could either interpolate @extschema@ into the text\nof the referencing function, or (though much inferior for performance)\nhave it SET SEARCH_PATH FROM CURRENT. Either of those changes would\ninvolve an extension version bump since they're changing the extension\nscript. What's more of a problem is that we could no longer claim\nthe extension is relocatable. My unaccent fix dodged that by looking\nup the C function's current schema, but I don't think there's any\nequivalent functionality available at SQL level.\n\n\t\t\tregards, tom lane\n\n","msg_date":"Fri, 07 Sep 2018 19:05:39 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: unaccent(text) fails depending on search_path (WAS: pg_upgrade\n fails saying function unaccent(text) doesn't exist)"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nI'm trying to upgrade an old PG 9.4 cluster to PG 10, and it's failing on\\ncreating an index that uses the unaccent(text) function.\\n\\nThat function is part of the unaccent extension, which is installed in the\\nold DB cluster. I expect pg_upgrade to create that extension as part of the\\nupgrade. It does create other extensions that are installed in the old DB\\ncluster. I don't get why this one isn't included.\\n\\nHere are the commands I run, and their output:\\n\\n----- snip -----\\n$ sudo rm -rf /opt/local/var/db/postgresql10/defaultdb\\n\\n$ sudo mkdir -p /opt/local/var/db/postgresql10/defaultdb\\n\\n$ sudo chown postgres:postgres /opt/local/var/db/postgresql10/defaultdb\\n\\n$ sudo su postgres -c '/opt/local/lib/postgresql10/bin/initdb -D\\n/opt/local/var/db/postgresql10/defaultdb --locale en_US.UTF-8'\\nThe files belonging to this database system will be owned by user\\n\\\"postgres\\\".\\nThis user must also own the server process.\\n\\nThe database cluster will be initialized with locale \\\"en_US.UTF-8\\\".\\nThe default database encoding has accordingly been set to \\\"UTF8\\\".\\nThe default text search configuration will be set to \\\"english\\\".\\n\\nData page checksums are disabled.\\n\\nfixing permissions on existing directory\\n/opt/local/var/db/postgresql10/defaultdb ... ok\\ncreating subdirectories ... ok\\nselecting default max_connections ... 100\\nselecting default shared_buffers ... 128MB\\nselecting dynamic shared memory implementation ... posix\\ncreating configuration files ... ok\\nrunning bootstrap script ... ok\\nperforming post-bootstrap initialization ... ok\\nsyncing data to disk ... ok\\n\\nWARNING: enabling \\\"trust\\\" authentication for local connections\\nYou can change this by editing pg_hba.conf or using the option -A, or\\n--auth-local and --auth-host, the next time you run initdb.\\n\\nSuccess. You can now start the database server using:\\n\\n /opt/local/lib/postgresql10/bin/pg_ctl -D\\n/opt/local/var/db/postgresql10/defaultdb -l logfile start\\n\\n$ sudo su postgres -c '/opt/local/lib/postgresql10/bin/pg_upgrade\\n--old-bindir /opt/local/lib/postgresql94/bin --old-datadir\\n/opt/local/var/db/postgresql94/defaultdb --new-bindir\\n/opt/local/lib/postgresql10/bin --new-datadir\\n/opt/local/var/db/postgresql10/defaultdb'Performing Consistency Checks\\n-----------------------------\\nChecking cluster versions ok\\nChecking database user is the install user ok\\nChecking database connection settings ok\\nChecking for prepared transactions ok\\nChecking for reg* data types in user tables ok\\nChecking for contrib/isn with bigint-passing mismatch ok\\nChecking for invalid \\\"unknown\\\" user columns ok\\nChecking for roles starting with \\\"pg_\\\" ok\\nCreating dump of global objects ok\\nCreating dump of database schemas\\n ok\\nChecking for presence of required libraries ok\\nChecking database user is the install user ok\\nChecking for prepared transactions ok\\n\\nIf pg_upgrade fails after this point, you must re-initdb the\\nnew cluster before continuing.\\n\\nPerforming Upgrade\\n------------------\\nAnalyzing all rows in the new cluster ok\\nFreezing all rows in the new cluster ok\\nDeleting files from new pg_xact ok\\nCopying old pg_clog to new server ok\\nSetting next transaction ID and epoch for new cluster ok\\nDeleting files from new pg_multixact/offsets ok\\nCopying old pg_multixact/offsets to new server ok\\nDeleting files from new pg_multixact/members ok\\nCopying old pg_multixact/members to new server ok\\nSetting next multixact ID and offset for new cluster ok\\nResetting WAL archives ok\\nSetting frozenxid and minmxid counters in new cluster ok\\nRestoring global objects in the new cluster ok\\nRestoring database schemas in the new cluster\\n dm_test\\n*failure*\\n\\nConsult the last few lines of \\\"pg_upgrade_dump_409041.log\\\" for\\nthe probable cause of the failure.\\nFailure, exiting\\n----- snip -----\\n\\nAnd here is that pg_upgrade_dump_409041.log file:\\n\\n----- snip -----\\ncommand: \\\"/opt/local/lib/postgresql10/bin/pg_dump\\\" --host /private/tmp\\n--port 50432 --username postgres --schema-only --quote-all-identifiers\\n--binary-upgrade --format=custom --file=\\\"pg_upgrade_dump_409041.custom\\\"\\n'dbname=dm_test' >> \\\"pg_upgrade_dump_409041.log\\\" 2>&1\\n\\n\\ncommand: \\\"/opt/local/lib/postgresql10/bin/pg_restore\\\" --host /private/tmp\\n--port 50432 --username postgres --exit-on-error --verbose --dbname\\n'dbname=dm_test' \\\"pg_upgrade_dump_409041.custom\\\" >>\\n\\\"pg_upgrade_dump_409041.log\\\" 2>&1\\npg_restore: connecting to database for restore\\npg_restore: creating pg_largeobject \\\"pg_largeobject\\\"\\npg_restore: creating pg_largeobject_metadata \\\"pg_largeobject_metadata\\\"\\npg_restore: creating SCHEMA \\\"acl_admin\\\"\\npg_restore: creating COMMENT \\\"SCHEMA \\\"acl_admin\\\"\\\"\\npg_restore: creating SCHEMA \\\"public\\\"\\npg_restore: creating COMMENT \\\"SCHEMA \\\"public\\\"\\\"\\npg_restore: creating EXTENSION \\\"fuzzystrmatch\\\"\\npg_restore: creating COMMENT \\\"EXTENSION \\\"fuzzystrmatch\\\"\\\"\\npg_restore: creating EXTENSION \\\"hstore\\\"\\npg_restore: creating COMMENT \\\"EXTENSION \\\"hstore\\\"\\\"\\npg_restore: creating EXTENSION \\\"sslinfo\\\"\\npg_restore: creating COMMENT \\\"EXTENSION \\\"sslinfo\\\"\\\"\\npg_restore: creating TYPE \\\"public.dataimport_job_state\\\"\\npg_restore: creating TYPE \\\"public.dblink_pkey_results\\\"...\\n[...]\\npg_restore: creating INDEX \\\"public.ix_semantic_mapping_lower_title\\\"\\npg_restore: creating INDEX \\\"public.ix_semantic_mapping_normalize_title\\\"\\npg_restore: [archiver (db)] Error while PROCESSING TOC:\\npg_restore: [archiver (db)] Error from TOC entry 5240; 1259 427215 INDEX\\nix_semantic_mapping_normalize_title dm_admin\\npg_restore: [archiver (db)] could not execute query: ERROR: function\\nunaccent(text) does not exist\\nLINE 2: SELECT lower(unaccent(btrim(regexp_replace($1, '\\\\s+', ' ',...\\n ^\\nHINT: No function matches the given name and argument types. You might\\nneed to add explicit type casts.\\nQUERY:\\n SELECT lower(unaccent(btrim(regexp_replace($1, '\\\\s+', ' ', 'g'), ' \\\"')))\\n\\nCONTEXT: SQL function \\\"semantic_normalize\\\" during inlining\\n Command was:\\n-- For binary upgrade, must preserve pg_class oids\\nSELECT\\npg_catalog.binary_upgrade_set_next_index_pg_class_oid('427215'::pg_catalog.oid);\\n\\nCREATE INDEX \\\"ix_semantic_mapping_normalize_title\\\" ON\\n\\\"public\\\".\\\"semantic_mapping\\\" USING \\\"btree\\\"\\n(\\\"public\\\".\\\"semantic_normalize\\\"(\\\"title\\\"));\\n----- snip -----\\n\\nThese occurrences of unaccent at the end are the only ones in the file:\\n\\n$ sudo grep -n unaccent pg_upgrade_dump_409041.log g\\n1713:pg_restore: [archiver (db)] could not execute query: ERROR: function\\nunaccent(text) does not exist\\n1714:LINE 2: SELECT lower(unaccent(btrim(regexp_replace($1, '\\\\s+', ' ',...\\n1718: SELECT lower(unaccent(btrim(regexp_replace($1, '\\\\s+', ' ', 'g'), '\\n\\\"')))\\n\\nHere is the definition of that index in the old DB:\\n\\n\\\\d ix_semantic_mapping_normalize_title\\n Index \\\"public.ix_semantic_mapping_normalize_title\\\"\\n Column | Type | Definition\\n--------------------+------+---------------------------\\n semantic_normalize | text | semantic_normalize(title)\\nbtree, for table \\\"public.semantic_mapping\\\"\\n\\nand the semantic_normalize function it uses:\\n\\n\\\\x \\\\df+ semantic_normalize\\nExpanded display is on.\\nList of functions\\n-[ RECORD 1\\n]-------+---------------------------------------------------------------------------\\nSchema | public\\nName | semantic_normalize\\nResult data type | text\\nArgument data types | title text\\nType | normal\\nVolatility | immutable\\nOwner | dm_admin\\nSecurity | invoker\\nAccess privileges |\\nLanguage | sql\\nSource code\\n|\\n+\\n | SELECT lower(unaccent(btrim(regexp_replace($1,\\n'\\\\s+', ' ', 'g'), ' \\\"')))+\\n |\\nDescription |\\n\\nAny more information I can provide, to help troubleshoot this?\\n\\nCheers,\\nGulli\\n\\nHi,I'm trying to upgrade an old PG 9.4 cluster to PG 10, and it's failing on creating an index that uses the unaccent(text) function.That function is part of the unaccent extension, which is installed in the old DB cluster. I expect pg_upgrade to create that extension as part of the upgrade. It does create other extensions that are installed in the old DB cluster. I don't get why this one isn't included.Here are the commands I run, and their output:----- snip -----$ sudo rm -rf /opt/local/var/db/postgresql10/defaultdb$ sudo mkdir -p /opt/local/var/db/postgresql10/defaultdb$ sudo chown postgres:postgres /opt/local/var/db/postgresql10/defaultdb$ sudo su postgres -c '/opt/local/lib/postgresql10/bin/initdb -D /opt/local/var/db/postgresql10/defaultdb --locale en_US.UTF-8'The files belonging to this database system will be owned by user \\\"postgres\\\".This user must also own the server process.The database cluster will be initialized with locale \\\"en_US.UTF-8\\\".The default database encoding has accordingly been set to \\\"UTF8\\\".The default text search configuration will be set to \\\"english\\\".Data page checksums are disabled.fixing permissions on existing directory /opt/local/var/db/postgresql10/defaultdb ... okcreating subdirectories ... okselecting default max_connections ... 100selecting default shared_buffers ... 128MBselecting dynamic shared memory implementation ... posixcreating configuration files ... okrunning bootstrap script ... okperforming post-bootstrap initialization ... oksyncing data to disk ... okWARNING: enabling \\\"trust\\\" authentication for local connectionsYou can change this by editing pg_hba.conf or using the option -A, or--auth-local and --auth-host, the next time you run initdb.Success. You can now start the database server using: /opt/local/lib/postgresql10/bin/pg_ctl -D /opt/local/var/db/postgresql10/defaultdb -l logfile start$ sudo su postgres -c '/opt/local/lib/postgresql10/bin/pg_upgrade --old-bindir /opt/local/lib/postgresql94/bin --old-datadir /opt/local/var/db/postgresql94/defaultdb --new-bindir /opt/local/lib/postgresql10/bin --new-datadir /opt/local/var/db/postgresql10/defaultdb'Performing Consistency Checks-----------------------------Checking cluster versions okChecking database user is the install user okChecking database connection settings okChecking for prepared transactions okChecking for reg* data types in user tables okChecking for contrib/isn with bigint-passing mismatch okChecking for invalid \\\"unknown\\\" user columns okChecking for roles starting with \\\"pg_\\\" okCreating dump of global objects okCreating dump of database schemas okChecking for presence of required libraries okChecking database user is the install user okChecking for prepared transactions okIf pg_upgrade fails after this point, you must re-initdb thenew cluster before continuing.Performing Upgrade------------------Analyzing all rows in the new cluster okFreezing all rows in the new cluster okDeleting files from new pg_xact okCopying old pg_clog to new server okSetting next transaction ID and epoch for new cluster okDeleting files from new pg_multixact/offsets okCopying old pg_multixact/offsets to new server okDeleting files from new pg_multixact/members okCopying old pg_multixact/members to new server okSetting next multixact ID and offset for new cluster okResetting WAL archives okSetting frozenxid and minmxid counters in new cluster okRestoring global objects in the new cluster okRestoring database schemas in the new cluster dm_test*failure*Consult the last few lines of \\\"pg_upgrade_dump_409041.log\\\" forthe probable cause of the failure.Failure, exiting----- snip -----And here is that pg_upgrade_dump_409041.log file:----- snip -----command: \\\"/opt/local/lib/postgresql10/bin/pg_dump\\\" --host /private/tmp --port 50432 --username postgres --schema-only --quote-all-identifiers --binary-upgrade --format=custom --file=\\\"pg_upgrade_dump_409041.custom\\\" 'dbname=dm_test' >> \\\"pg_upgrade_dump_409041.log\\\" 2>&1command: \\\"/opt/local/lib/postgresql10/bin/pg_restore\\\" --host /private/tmp --port 50432 --username postgres --exit-on-error --verbose --dbname 'dbname=dm_test' \\\"pg_upgrade_dump_409041.custom\\\" >> \\\"pg_upgrade_dump_409041.log\\\" 2>&1pg_restore: connecting to database for restorepg_restore: creating pg_largeobject \\\"pg_largeobject\\\"pg_restore: creating pg_largeobject_metadata \\\"pg_largeobject_metadata\\\"pg_restore: creating SCHEMA \\\"acl_admin\\\"pg_restore: creating COMMENT \\\"SCHEMA \\\"acl_admin\\\"\\\"pg_restore: creating SCHEMA \\\"public\\\"pg_restore: creating COMMENT \\\"SCHEMA \\\"public\\\"\\\"pg_restore: creating EXTENSION \\\"fuzzystrmatch\\\"pg_restore: creating COMMENT \\\"EXTENSION \\\"fuzzystrmatch\\\"\\\"pg_restore: creating EXTENSION \\\"hstore\\\"pg_restore: creating COMMENT \\\"EXTENSION \\\"hstore\\\"\\\"pg_restore: creating EXTENSION \\\"sslinfo\\\"pg_restore: creating COMMENT \\\"EXTENSION \\\"sslinfo\\\"\\\"pg_restore: creating TYPE \\\"public.dataimport_job_state\\\"pg_restore: creating TYPE \\\"public.dblink_pkey_results\\\"...[...]pg_restore: creating INDEX \\\"public.ix_semantic_mapping_lower_title\\\"pg_restore: creating INDEX \\\"public.ix_semantic_mapping_normalize_title\\\"pg_restore: [archiver (db)] Error while PROCESSING TOC:pg_restore: [archiver (db)] Error from TOC entry 5240; 1259 427215 INDEX ix_semantic_mapping_normalize_title dm_adminpg_restore: [archiver (db)] could not execute query: ERROR: function unaccent(text) does not existLINE 2: SELECT lower(unaccent(btrim(regexp_replace($1, '\\\\s+', ' ',... ^HINT: No function matches the given name and argument types. You might need to add explicit type casts.QUERY: SELECT lower(unaccent(btrim(regexp_replace($1, '\\\\s+', ' ', 'g'), ' \\\"')))CONTEXT: SQL function \\\"semantic_normalize\\\" during inlining Command was: -- For binary upgrade, must preserve pg_class oidsSELECT pg_catalog.binary_upgrade_set_next_index_pg_class_oid('427215'::pg_catalog.oid);CREATE INDEX \\\"ix_semantic_mapping_normalize_title\\\" ON \\\"public\\\".\\\"semantic_mapping\\\" USING \\\"btree\\\" (\\\"public\\\".\\\"semantic_normalize\\\"(\\\"title\\\"));----- snip -----These occurrences of unaccent at the end are the only ones in the file:$ sudo grep -n unaccent pg_upgrade_dump_409041.log g1713:pg_restore: [archiver (db)] could not execute query: ERROR: function unaccent(text) does not exist1714:LINE 2: SELECT lower(unaccent(btrim(regexp_replace($1, '\\\\s+', ' ',...1718: SELECT lower(unaccent(btrim(regexp_replace($1, '\\\\s+', ' ', 'g'), ' \\\"')))Here is the definition of that index in the old DB:\\\\d ix_semantic_mapping_normalize_title Index \\\"public.ix_semantic_mapping_normalize_title\\\" Column | Type | Definition--------------------+------+--------------------------- semantic_normalize | text | semantic_normalize(title)btree, for table \\\"public.semantic_mapping\\\"and the semantic_normalize function it uses:\\\\x \\\\df+ semantic_normalizeExpanded display is on.List of functions-[ RECORD 1 ]-------+---------------------------------------------------------------------------Schema | publicName | semantic_normalizeResult data type | textArgument data types | title textType | normalVolatility | immutableOwner | dm_adminSecurity | invokerAccess privileges |Language | sqlSource code | + | SELECT lower(unaccent(btrim(regexp_replace($1, '\\\\s+', ' ', 'g'), ' \\\"')))+ |Description |Any more information I can provide, to help troubleshoot this?Cheers,Gulli\",\n \"msg_date\": \"Wed, 29 Aug 2018 18:09:53 +0000\",\n \"msg_from\": \"Gunnlaugur Thor Briem <gunnlaugur@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"pg_upgrade fails saying function unaccent(text) doesn't exist\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 29, 2018 at 06:09:53PM +0000, Gunnlaugur Thor Briem wrote:\\n> Hi,\\n> \\n> I'm trying to upgrade an old PG 9.4 cluster to PG 10, and it's failing on\\n> creating an index that uses the unaccent(text) function.\\n> \\n> That function is part of the unaccent extension, which is installed in the old\\n> DB cluster. I expect pg_upgrade to create that extension as part of the\\n> upgrade. It does create other extensions that are installed in the old DB\\n> cluster. I don't get why this one isn't included.\\n\\nThis is caused by security changes made in PG 10.3 and other minor\\nreleases. Please see this thread for an outline of the issue:\\n\\n https://www.postgresql.org/message-id/flat/152106914669.1223.5104148605998271987%40wrigleys.postgresql.org\\n\\nI think you have to change your index function to specify the schema\\nname before the unacces function call, e.g.\\n\\n\\tSELECT lower(public.unaccent(btrim(regexp_replace(\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\",\n \"msg_date\": \"Wed, 29 Aug 2018 15:06:35 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_upgrade fails saying function unaccent(text) doesn't exist\"\n },\n {\n \"msg_contents\": \"Thank you! That got the pg_upgrade to completion. But then during\\n./analyze_new_cluster.sh vacuum fails thus:\\n\\nvacuumdb: processing database \\\"dm_test\\\": Generating minimal optimizer\\nstatistics (1 target)\\nvacuumdb: vacuuming of database \\\"dm_test\\\" failed: ERROR: text search\\ndictionary \\\"unaccent\\\" does not exist\\nCONTEXT: SQL function \\\"semantic_normalize\\\" statement 1\\n\\nAnd yet a text search dictionary with that name does exist:\\n\\n$ psql -d dm_test -c '\\\\dFd+ unaccent'\\n List of text search dictionaries\\n Schema | Name | Template | Init options | Description\\n--------+----------+-----------------+--------------------+-------------\\n public | unaccent | public.unaccent | rules = 'unaccent' |\\n(1 row)\\n\\nRunning VACUUM ANALYZE semantic_mapping in psql works:\\n\\n$ psql -d dm_test -c 'VACUUM ANALYZE semantic_mapping'\\nVACUUM\\nTime: 1231,767 ms (00:01,232)\\n\\nBut running it with the vacuumdb command doesn't:\\n\\nvacuumdb -z -t semantic_mapping dm_test\\nvacuumdb: vacuuming database \\\"dm_test\\\"\\nvacuumdb: vacuuming of table \\\"semantic_mapping\\\" in database \\\"dm_test\\\"\\nfailed: ERROR: text search dictionary \\\"unaccent\\\" does not exist\\nCONTEXT: SQL function \\\"semantic_normalize\\\" statement 1\\n\\nThis is presumably a similar search path problem, because I can reproduce\\nthis in psql by setting the search path to exclude public:\\n\\nset search_path to \\\"$user\\\";\\nvacuum analyze public.semantic_mapping;\\nERROR: text search dictionary \\\"unaccent\\\" does not exist\\nCONTEXT: SQL function \\\"semantic_normalize\\\" statement 1\\nTime: 851,562 ms\\n\\nCan't find a place to poke the \\\"public.\\\" prefix in to work around this ...\\nI can't even see where it's getting the link to the text search dictionary\\nfrom. Is that in native code in the unaccent extension?\\n\\nThe unaccent definition looks like this:\\n\\n\\\\df+ public.unaccent\\n\\nList of functions\\n Schema | Name | Result data type | Argument data types | Type |\\nVolatility | Parallel | Owner | Security | Access privileges | Language |\\nSource code | Description\\n--------+----------+------------------+---------------------+--------+------------+----------+-------+----------+-------------------+----------+---------------+-------------\\n public | unaccent | text | regdictionary, text | normal |\\nstable | safe | gthb | invoker | | c |\\nunaccent_dict |\\n public | unaccent | text | text | normal |\\nstable | safe | gthb | invoker | | c |\\nunaccent_dict |\\n(2 rows)\\n\\nAny tips?\\n\\nCheers,\\nGulli\\n\\nOn Wed, Aug 29, 2018 at 7:06 PM Bruce Momjian <bruce@momjian.us> wrote:\\n\\n> On Wed, Aug 29, 2018 at 06:09:53PM +0000, Gunnlaugur Thor Briem wrote:\\n> > Hi,\\n> >\\n> > I'm trying to upgrade an old PG 9.4 cluster to PG 10, and it's failing on\\n> > creating an index that uses the unaccent(text) function.\\n> >\\n> > That function is part of the unaccent extension, which is installed in\\n> the old\\n> > DB cluster. I expect pg_upgrade to create that extension as part of the\\n> > upgrade. It does create other extensions that are installed in the old DB\\n> > cluster. I don't get why this one isn't included.\\n>\\n> This is caused by security changes made in PG 10.3 and other minor\\n> releases. Please see this thread for an outline of the issue:\\n>\\n>\\n> https://www.postgresql.org/message-id/flat/152106914669.1223.5104148605998271987%40wrigleys.postgresql.org\\n>\\n> I think you have to change your index function to specify the schema\\n> name before the unacces function call, e.g.\\n>\\n> SELECT lower(public.unaccent(btrim(regexp_replace(\\n>\\n> --\\n> Bruce Momjian <bruce@momjian.us> http://momjian.us\\n> EnterpriseDB http://enterprisedb.com\\n>\\n> + As you are, so once was I. As I am, so you will be. +\\n> + Ancient Roman grave inscription +\\n>\\n\\nThank you! That got the pg_upgrade to completion. But then during ./analyze_new_cluster.sh vacuum fails thus:vacuumdb: processing database \\\"dm_test\\\": Generating minimal optimizer statistics (1 target)vacuumdb: vacuuming of database \\\"dm_test\\\" failed: ERROR: text search dictionary \\\"unaccent\\\" does not existCONTEXT: SQL function \\\"semantic_normalize\\\" statement 1And yet a text search dictionary with that name does exist:$ psql -d dm_test -c '\\\\dFd+ unaccent' List of text search dictionaries Schema | Name | Template | Init options | Description--------+----------+-----------------+--------------------+------------- public | unaccent | public.unaccent | rules = 'unaccent' |(1 row)Running VACUUM ANALYZE semantic_mapping in psql works:$ psql -d dm_test -c 'VACUUM ANALYZE semantic_mapping'VACUUMTime: 1231,767 ms (00:01,232)But running it with the vacuumdb command doesn't:vacuumdb -z -t semantic_mapping dm_testvacuumdb: vacuuming database \\\"dm_test\\\"vacuumdb: vacuuming of table \\\"semantic_mapping\\\" in database \\\"dm_test\\\" failed: ERROR: text search dictionary \\\"unaccent\\\" does not existCONTEXT: SQL function \\\"semantic_normalize\\\" statement 1This is presumably a similar search path problem, because I can reproduce this in psql by setting the search path to exclude public:set search_path to \\\"$user\\\";vacuum analyze public.semantic_mapping;ERROR: text search dictionary \\\"unaccent\\\" does not existCONTEXT: SQL function \\\"semantic_normalize\\\" statement 1Time: 851,562 msCan't find a place to poke the \\\"public.\\\" prefix in to work around this ... I can't even see where it's getting the link to the text search dictionary from. Is that in native code in the unaccent extension?The unaccent definition looks like this:\\\\df+ public.unaccent List of functions Schema | Name | Result data type | Argument data types | Type | Volatility | Parallel | Owner | Security | Access privileges | Language | Source code | Description--------+----------+------------------+---------------------+--------+------------+----------+-------+----------+-------------------+----------+---------------+------------- public | unaccent | text | regdictionary, text | normal | stable | safe | gthb | invoker | | c | unaccent_dict | public | unaccent | text | text | normal | stable | safe | gthb | invoker | | c | unaccent_dict |(2 rows)Any tips?Cheers,GulliOn Wed, Aug 29, 2018 at 7:06 PM Bruce Momjian <bruce@momjian.us> wrote:On Wed, Aug 29, 2018 at 06:09:53PM +0000, Gunnlaugur Thor Briem wrote:\\n> Hi,\\n> \\n> I'm trying to upgrade an old PG 9.4 cluster to PG 10, and it's failing on\\n> creating an index that uses the unaccent(text) function.\\n> \\n> That function is part of the unaccent extension, which is installed in the old\\n> DB cluster. I expect pg_upgrade to create that extension as part of the\\n> upgrade. It does create other extensions that are installed in the old DB\\n> cluster. I don't get why this one isn't included.\\n\\nThis is caused by security changes made in PG 10.3 and other minor\\nreleases. Please see this thread for an outline of the issue:\\n\\n https://www.postgresql.org/message-id/flat/152106914669.1223.5104148605998271987%40wrigleys.postgresql.org\\n\\nI think you have to change your index function to specify the schema\\nname before the unacces function call, e.g.\\n\\n SELECT lower(public.unaccent(btrim(regexp_replace(\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\",\n \"msg_date\": \"Thu, 30 Aug 2018 09:24:13 +0000\",\n \"msg_from\": \"Gunnlaugur Thor Briem <gunnlaugur@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_upgrade fails saying function unaccent(text) doesn't exist\"\n },\n {\n \"msg_contents\": \"On 08/30/2018 02:24 AM, Gunnlaugur Thor Briem wrote:\\n> Thank you! That got the pg_upgrade to completion. But then during \\n> ./analyze_new_cluster.sh vacuum fails thus:\\n> \\n> vacuumdb: processing database \\\"dm_test\\\": Generating minimal optimizer \\n> statistics (1 target)\\n> vacuumdb: vacuuming of database \\\"dm_test\\\" failed: ERROR: text search \\n> dictionary \\\"unaccent\\\" does not exist\\n> CONTEXT: SQL function \\\"semantic_normalize\\\" statement 1\\n> \\n> And yet a text search dictionary with that name does exist:\\n> \\n> $ psql -d dm_test -c '\\\\dFd+ unaccent'\\n> List of text search dictionaries\\n> Schema | Name | Template | Init options | Description\\n> --------+----------+-----------------+--------------------+-------------\\n> public | unaccent | public.unaccent | rules = 'unaccent' |\\n> (1 row)\\n> \\n> Running VACUUM ANALYZE semantic_mapping in psql works:\\n> \\n> $ psql -d dm_test -c 'VACUUM ANALYZE semantic_mapping'\\n> VACUUM\\n> Time: 1231,767 ms (00:01,232)\\n> \\n> But running it with the vacuumdb command doesn't:\\n> \\n> vacuumdb -z -t semantic_mapping dm_test\\n> vacuumdb: vacuuming database \\\"dm_test\\\"\\n> vacuumdb: vacuuming of table \\\"semantic_mapping\\\" in database \\\"dm_test\\\" \\n> failed: ERROR: text search dictionary \\\"unaccent\\\" does not exist\\n> CONTEXT: SQL function \\\"semantic_normalize\\\" statement 1\\n> \\n> This is presumably a similar search path problem, because I can \\n> reproduce this in psql by setting the search path to exclude public:\\n> \\n> set search_path to \\\"$user\\\";\\n> vacuum analyze public.semantic_mapping;\\n> ERROR: text search dictionary \\\"unaccent\\\" does not exist\\n> CONTEXT: SQL function \\\"semantic_normalize\\\" statement 1\\n> Time: 851,562 ms\\n> \\n> Can't find a place to poke the \\\"public.\\\" prefix in to work around this \\n> ... I can't even see where it's getting the link to the text search \\n> dictionary from. Is that in native code in the unaccent extension?\\n\\nSince the semantic_normalize function is tripping it and it uses \\nunaccent I would say it is native to the extension.\\n\\nWhat does:\\n\\n\\\\dFd unaccent\\n\\nshow?\\n\\n> \\n> The unaccent definition looks like this:\\n> \\n> \\\\df+ public.unaccent\\n> List of functions\\n> Schema | Name | Result data type | Argument data types | Type | \\n> Volatility | Parallel | Owner | Security | Access privileges | Language \\n> | Source code | Description\\n> --------+----------+------------------+---------------------+--------+------------+----------+-------+----------+-------------------+----------+---------------+-------------\\n> public | unaccent | text | regdictionary, text | normal | \\n> stable | safe | gthb | invoker | | c \\n> | unaccent_dict |\\n> public | unaccent | text | text | normal | \\n> stable | safe | gthb | invoker | | c \\n> | unaccent_dict |\\n> (2 rows)\\n> \\n> Any tips?\\n> \\n> Cheers,\\n> Gulli\\n> \\n> On Wed, Aug 29, 2018 at 7:06 PM Bruce Momjian <bruce@momjian.us \\n> <mailto:bruce@momjian.us>> wrote:\\n> \\n> On Wed, Aug 29, 2018 at 06:09:53PM +0000, Gunnlaugur Thor Briem wrote:\\n> > Hi,\\n> >\\n> > I'm trying to upgrade an old PG 9.4 cluster to PG 10, and it's\\n> failing on\\n> > creating an index that uses the unaccent(text) function.\\n> >\\n> > That function is part of the unaccent extension, which is\\n> installed in the old\\n> > DB cluster. I expect pg_upgrade to create that extension as part\\n> of the\\n> > upgrade. It does create other extensions that are installed in\\n> the old DB\\n> > cluster. I don't get why this one isn't included.\\n> \\n> This is caused by security changes made in PG 10.3 and other minor\\n> releases. Please see this thread for an outline of the issue:\\n> \\n> https://www.postgresql.org/message-id/flat/152106914669.1223.5104148605998271987%40wrigleys.postgresql.org\\n> \\n> I think you have to change your index function to specify the schema\\n> name before the unacces function call, e.g.\\n> \\n> SELECT lower(public.unaccent(btrim(regexp_replace(\\n> \\n> -- \\n> Bruce Momjian <bruce@momjian.us <mailto:bruce@momjian.us>>\\n> http://momjian.us\\n> EnterpriseDB http://enterprisedb.com\\n> \\n> + As you are, so once was I. As I am, so you will be. +\\n> + Ancient Roman grave inscription +\\n> \\n\\n\\n-- \\nAdrian Klaver\\nadrian.klaver@aklaver.com\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 06:48:26 -0700\",\n \"msg_from\": \"Adrian Klaver <adrian.klaver@aklaver.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_upgrade fails saying function unaccent(text) doesn't exist\"\n },\n {\n \"msg_contents\": \"\\\\dFd unaccent\\nList of text search dictionaries\\n Schema | Name | Description\\n--------+----------+-------------\\n public | unaccent |\\n(1 row)\\n\\n\\\\dFd+ unaccent\\n List of text search dictionaries\\n Schema | Name | Template | Init options | Description\\n--------+----------+-----------------+--------------------+-------------\\n public | unaccent | public.unaccent | rules = 'unaccent' |\\n(1 row)\\n\\nCheers,\\nGulli\\n\\nOn Thu, Aug 30, 2018 at 1:48 PM Adrian Klaver <adrian.klaver@aklaver.com>\\nwrote:\\n\\n> On 08/30/2018 02:24 AM, Gunnlaugur Thor Briem wrote:\\n> > Thank you! That got the pg_upgrade to completion. But then during\\n> > ./analyze_new_cluster.sh vacuum fails thus:\\n> >\\n> > vacuumdb: processing database \\\"dm_test\\\": Generating minimal optimizer\\n> > statistics (1 target)\\n> > vacuumdb: vacuuming of database \\\"dm_test\\\" failed: ERROR: text search\\n> > dictionary \\\"unaccent\\\" does not exist\\n> > CONTEXT: SQL function \\\"semantic_normalize\\\" statement 1\\n> >\\n> > And yet a text search dictionary with that name does exist:\\n> >\\n> > $ psql -d dm_test -c '\\\\dFd+ unaccent'\\n> > List of text search dictionaries\\n> > Schema | Name | Template | Init options | Description\\n> > --------+----------+-----------------+--------------------+-------------\\n> > public | unaccent | public.unaccent | rules = 'unaccent' |\\n> > (1 row)\\n> >\\n> > Running VACUUM ANALYZE semantic_mapping in psql works:\\n> >\\n> > $ psql -d dm_test -c 'VACUUM ANALYZE semantic_mapping'\\n> > VACUUM\\n> > Time: 1231,767 ms (00:01,232)\\n> >\\n> > But running it with the vacuumdb command doesn't:\\n> >\\n> > vacuumdb -z -t semantic_mapping dm_test\\n> > vacuumdb: vacuuming database \\\"dm_test\\\"\\n> > vacuumdb: vacuuming of table \\\"semantic_mapping\\\" in database \\\"dm_test\\\"\\n> > failed: ERROR: text search dictionary \\\"unaccent\\\" does not exist\\n> > CONTEXT: SQL function \\\"semantic_normalize\\\" statement 1\\n> >\\n> > This is presumably a similar search path problem, because I can\\n> > reproduce this in psql by setting the search path to exclude public:\\n> >\\n> > set search_path to \\\"$user\\\";\\n> > vacuum analyze public.semantic_mapping;\\n> > ERROR: text search dictionary \\\"unaccent\\\" does not exist\\n> > CONTEXT: SQL function \\\"semantic_normalize\\\" statement 1\\n> > Time: 851,562 ms\\n> >\\n> > Can't find a place to poke the \\\"public.\\\" prefix in to work around this\\n> > ... I can't even see where it's getting the link to the text search\\n> > dictionary from. Is that in native code in the unaccent extension?\\n>\\n> Since the semantic_normalize function is tripping it and it uses\\n> unaccent I would say it is native to the extension.\\n>\\n> What does:\\n>\\n> \\\\dFd unaccent\\n>\\n> show?\\n>\\n> >\\n> > The unaccent definition looks like this:\\n> >\\n> > \\\\df+ public.unaccent\\n> >\\n>\\n> List of functions\\n> > Schema | Name | Result data type | Argument data types | Type |\\n> > Volatility | Parallel | Owner | Security | Access privileges | Language\\n> > | Source code | Description\\n> >\\n> --------+----------+------------------+---------------------+--------+------------+----------+-------+----------+-------------------+----------+---------------+-------------\\n> > public | unaccent | text | regdictionary, text | normal |\\n> > stable | safe | gthb | invoker | | c\\n> > | unaccent_dict |\\n> > public | unaccent | text | text | normal |\\n> > stable | safe | gthb | invoker | | c\\n> > | unaccent_dict |\\n> > (2 rows)\\n> >\\n> > Any tips?\\n> >\\n> > Cheers,\\n> > Gulli\\n> >\\n> > On Wed, Aug 29, 2018 at 7:06 PM Bruce Momjian <bruce@momjian.us\\n> > <mailto:bruce@momjian.us>> wrote:\\n> >\\n> > On Wed, Aug 29, 2018 at 06:09:53PM +0000, Gunnlaugur Thor Briem\\n> wrote:\\n> > > Hi,\\n> > >\\n> > > I'm trying to upgrade an old PG 9.4 cluster to PG 10, and it's\\n> > failing on\\n> > > creating an index that uses the unaccent(text) function.\\n> > >\\n> > > That function is part of the unaccent extension, which is\\n> > installed in the old\\n> > > DB cluster. I expect pg_upgrade to create that extension as part\\n> > of the\\n> > > upgrade. It does create other extensions that are installed in\\n> > the old DB\\n> > > cluster. I don't get why this one isn't included.\\n> >\\n> > This is caused by security changes made in PG 10.3 and other minor\\n> > releases. Please see this thread for an outline of the issue:\\n> >\\n> >\\n> https://www.postgresql.org/message-id/flat/152106914669.1223.5104148605998271987%40wrigleys.postgresql.org\\n> >\\n> > I think you have to change your index function to specify the schema\\n> > name before the unacces function call, e.g.\\n> >\\n> > SELECT lower(public.unaccent(btrim(regexp_replace(\\n> >\\n> > --\\n> > Bruce Momjian <bruce@momjian.us <mailto:bruce@momjian.us>>\\n> > http://momjian.us\\n> > EnterpriseDB http://enterprisedb.com\\n> >\\n> > + As you are, so once was I. As I am, so you will be. +\\n> > + Ancient Roman grave inscription +\\n> >\\n>\\n>\\n> --\\n> Adrian Klaver\\n> adrian.klaver@aklaver.com\\n>\\n>\\n\\n\\\\dFd unaccentList of text search dictionaries Schema | Name | Description--------+----------+------------- public | unaccent |(1 row)\\\\dFd+ unaccent List of text search dictionaries Schema | Name | Template | Init options | Description--------+----------+-----------------+--------------------+------------- public | unaccent | public.unaccent | rules = 'unaccent' |(1 row)Cheers,GulliOn Thu, Aug 30, 2018 at 1:48 PM Adrian Klaver <adrian.klaver@aklaver.com> wrote:On 08/30/2018 02:24 AM, Gunnlaugur Thor Briem wrote:\\n> Thank you! That got the pg_upgrade to completion. But then during \\n> ./analyze_new_cluster.sh vacuum fails thus:\\n> \\n> vacuumdb: processing database \\\"dm_test\\\": Generating minimal optimizer \\n> statistics (1 target)\\n> vacuumdb: vacuuming of database \\\"dm_test\\\" failed: ERROR: text search \\n> dictionary \\\"unaccent\\\" does not exist\\n> CONTEXT: SQL function \\\"semantic_normalize\\\" statement 1\\n> \\n> And yet a text search dictionary with that name does exist:\\n> \\n> $ psql -d dm_test -c '\\\\dFd+ unaccent'\\n> List of text search dictionaries\\n> Schema | Name | Template | Init options | Description\\n> --------+----------+-----------------+--------------------+-------------\\n> public | unaccent | public.unaccent | rules = 'unaccent' |\\n> (1 row)\\n> \\n> Running VACUUM ANALYZE semantic_mapping in psql works:\\n> \\n> $ psql -d dm_test -c 'VACUUM ANALYZE semantic_mapping'\\n> VACUUM\\n> Time: 1231,767 ms (00:01,232)\\n> \\n> But running it with the vacuumdb command doesn't:\\n> \\n> vacuumdb -z -t semantic_mapping dm_test\\n> vacuumdb: vacuuming database \\\"dm_test\\\"\\n> vacuumdb: vacuuming of table \\\"semantic_mapping\\\" in database \\\"dm_test\\\" \\n> failed: ERROR: text search dictionary \\\"unaccent\\\" does not exist\\n> CONTEXT: SQL function \\\"semantic_normalize\\\" statement 1\\n> \\n> This is presumably a similar search path problem, because I can \\n> reproduce this in psql by setting the search path to exclude public:\\n> \\n> set search_path to \\\"$user\\\";\\n> vacuum analyze public.semantic_mapping;\\n> ERROR: text search dictionary \\\"unaccent\\\" does not exist\\n> CONTEXT: SQL function \\\"semantic_normalize\\\" statement 1\\n> Time: 851,562 ms\\n> \\n> Can't find a place to poke the \\\"public.\\\" prefix in to work around this \\n> ... I can't even see where it's getting the link to the text search \\n> dictionary from. Is that in native code in the unaccent extension?\\n\\nSince the semantic_normalize function is tripping it and it uses \\nunaccent I would say it is native to the extension.\\n\\nWhat does:\\n\\n\\\\dFd unaccent\\n\\nshow?\\n\\n> \\n> The unaccent definition looks like this:\\n> \\n> \\\\df+ public.unaccent\\n> List of functions\\n> Schema | Name | Result data type | Argument data types | Type | \\n> Volatility | Parallel | Owner | Security | Access privileges | Language \\n> | Source code | Description\\n> --------+----------+------------------+---------------------+--------+------------+----------+-------+----------+-------------------+----------+---------------+-------------\\n> public | unaccent | text | regdictionary, text | normal | \\n> stable | safe | gthb | invoker | | c \\n> | unaccent_dict |\\n> public | unaccent | text | text | normal | \\n> stable | safe | gthb | invoker | | c \\n> | unaccent_dict |\\n> (2 rows)\\n> \\n> Any tips?\\n> \\n> Cheers,\\n> Gulli\\n> \\n> On Wed, Aug 29, 2018 at 7:06 PM Bruce Momjian <bruce@momjian.us \\n> <mailto:bruce@momjian.us>> wrote:\\n> \\n> On Wed, Aug 29, 2018 at 06:09:53PM +0000, Gunnlaugur Thor Briem wrote:\\n> > Hi,\\n> >\\n> > I'm trying to upgrade an old PG 9.4 cluster to PG 10, and it's\\n> failing on\\n> > creating an index that uses the unaccent(text) function.\\n> >\\n> > That function is part of the unaccent extension, which is\\n> installed in the old\\n> > DB cluster. I expect pg_upgrade to create that extension as part\\n> of the\\n> > upgrade. It does create other extensions that are installed in\\n> the old DB\\n> > cluster. I don't get why this one isn't included.\\n> \\n> This is caused by security changes made in PG 10.3 and other minor\\n> releases. Please see this thread for an outline of the issue:\\n> \\n> https://www.postgresql.org/message-id/flat/152106914669.1223.5104148605998271987%40wrigleys.postgresql.org\\n> \\n> I think you have to change your index function to specify the schema\\n> name before the unacces function call, e.g.\\n> \\n> SELECT lower(public.unaccent(btrim(regexp_replace(\\n> \\n> -- \\n> Bruce Momjian <bruce@momjian.us <mailto:bruce@momjian.us>>\\n> http://momjian.us\\n> EnterpriseDB http://enterprisedb.com\\n> \\n> + As you are, so once was I. As I am, so you will be. +\\n> + Ancient Roman grave inscription +\\n> \\n\\n\\n-- \\nAdrian Klaver\\nadrian.klaver@aklaver.com\",\n \"msg_date\": \"Thu, 30 Aug 2018 16:03:11 +0000\",\n \"msg_from\": \"Gunnlaugur Thor Briem <gunnlaugur@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_upgrade fails saying function unaccent(text) doesn't exist\"\n },\n {\n \"msg_contents\": \"On 08/30/2018 09:03 AM, Gunnlaugur Thor Briem wrote:\\n> \\\\dFd unaccent\\n> List of text search dictionaries\\n> Schema | Name | Description\\n> --------+----------+-------------\\n> public | unaccent |\\n> (1 row)\\n> \\n> \\\\dFd+ unaccent\\n> List of text search dictionaries\\n> Schema | Name | Template | Init options | Description\\n> --------+----------+-----------------+--------------------+-------------\\n> public | unaccent | public.unaccent | rules = 'unaccent' |\\n> (1 row)\\n> \\n> Cheers,\\n> Gulli\\n> \\n\\nI could not replicate with simple case:\\n\\nselect version();\\n version \\n\\n------------------------------------------------------------------------------------\\n PostgreSQL 10.5 on x86_64-pc-linux-gnu, compiled by gcc (SUSE Linux) \\n4.8.5, 64-bit\\n\\nCREATE OR REPLACE FUNCTION public.semantic_normalize(title text)\\n RETURNS text\\n LANGUAGE sql\\nAS $function$\\nSELECT lower(unaccent(btrim(regexp_replace($1,\\n'\\\\s+', ' ', 'g'), ' \\\"')))\\n$function$\\n\\nCREATE TABLE unaccent_test(title text);\\n\\nINSERT INTO unaccent_test values ('Hukić'), ('Böttcher'), ('ÀÁÂÃÄÅ'), \\n('électro');\\n\\nCREATE INDEX ix_semantic_normalize_title on unaccent_test(title);\\n\\nVACUUM ANALYZE unaccent_test;\\nVACUUM\\n\\nvacuumdb -U postgres -z -t unaccent_test test\\nvacuumdb: vacuuming database \\\"test\\\"\\n\\n\\nThe only thing I can think of is that you have an older version of \\nvacuumdb that is not aware of the schema specification changes in the \\nnewer versions of Postgrse.\\n\\n\\n\\n-- \\nAdrian Klaver\\nadrian.klaver@aklaver.com\\n\\n\",\n \"msg_date\": \"Fri, 31 Aug 2018 07:14:18 -0700\",\n \"msg_from\": \"Adrian Klaver <adrian.klaver@aklaver.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_upgrade fails saying function unaccent(text) doesn't exist\"\n },\n {\n \"msg_contents\": \"On 08/30/2018 09:03 AM, Gunnlaugur Thor Briem wrote:\\n> \\\\dFd unaccent\\n> List of text search dictionaries\\n> Schema | Name | Description\\n> --------+----------+-------------\\n> public | unaccent |\\n> (1 row)\\n> \\n> \\\\dFd+ unaccent\\n> List of text search dictionaries\\n> Schema | Name | Template | Init options | Description\\n> --------+----------+-----------------+--------------------+-------------\\n> public | unaccent | public.unaccent | rules = 'unaccent' |\\n> (1 row)\\n> \\n> Cheers,\\n> Gulli\\n> \\n\\nForgot to add to previous post:\\n\\n\\\\dx unaccent\\n List of installed extensions\\n Name | Version | Schema | Description\\n----------+---------+--------+---------------------------------------------\\n unaccent | 1.1 | public | text search dictionary that removes accents\\n\\n\\n\\n-- \\nAdrian Klaver\\nadrian.klaver@aklaver.com\\n\\n\",\n \"msg_date\": \"Fri, 31 Aug 2018 07:17:06 -0700\",\n \"msg_from\": \"Adrian Klaver <adrian.klaver@aklaver.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_upgrade fails saying function unaccent(text) doesn't exist\"\n },\n {\n \"msg_contents\": \"OK, I found the cause of the unaccent dictionary problem, and a workaround.\\n\\nIt's not the vacuumdb version, not the unaccent version, and it's not even\\na pg_upgrade problem: I get this error also with PG 9.4.18 running on the\\nold cluster, with both the 10.5 vacuumdb and the 9.4.18 vacuumdb, and I get\\nthe same error in both.\\n\\nAnd it's not strictly a vacuumdb problem, though vacuumdb triggers it.\\n\\nHere's a very minimal test case, unrelated to my DB, that you ought to be\\nable to reproduce:\\n\\nSET search_path = \\\"$user\\\"; SELECT public.unaccent('fóö');\\nSET\\nERROR: text search dictionary \\\"unaccent\\\" does not exist\\n\\nand here's a workaround:\\n\\nSET search_path = \\\"$user\\\"; SELECT public.unaccent(tsdict.oid, 'fóö')\\nFROM pg_catalog.pg_ts_dict tsdict WHERE dictname='unaccent';\\nSET\\n unaccent\\n----------\\n foo\\n(1 row)\\n\\nThe workaround avoids the OID lookup of the dictionary ... that lookup (in\\nthe single-argument unaccent function) is done by unqualified name:\\n\\nhttps://github.com/postgres/postgres/blob/fb8697b31aaeebe6170c572739867dcaa01053c6/contrib/unaccent/unaccent.c#L377\\n\\n dictOid = get_ts_dict_oid(stringToQualifiedNameList(\\\"unaccent\\\"),\\nfalse);\\n\\nand that fails if the search path doesn't include public.\\n\\nSo it is indeed triggered by the security changes that Bruce mentioned;\\nthose were backported into 9.4.17:\\nhttps://www.postgresql.org/docs/9.4/static/release-9-4-17.html ... and so\\ngot pulled in by my Macports upgrades. So nothing to do with pg_upgrade.\\n\\nSo the workaround for my vacuumdb/function-index problem is to give\\nunaccent the OID of the text search dictionary, so that the search path\\nisn't in play:\\n\\nCREATE OR REPLACE FUNCTION public.semantic_normalize(title text)\\nRETURNS text\\n LANGUAGE sql\\n IMMUTABLE STRICT\\nAS $function$\\n SELECT lower(public.unaccent(16603, btrim(regexp_replace($1, '\\\\s+', ' ',\\n'g'), ' \\\"')))\\n$function$;\\n\\nand that makes vacuumdb -z work in both 9.4.18 and 10.5, and makes\\n./analyze_new_cluster.sh complete without problems.\\n\\nThe proper fix is, I suppose, to make the single-argument unaccent function\\nexplicitly look up the dictionary in the same schema as the function itself\\nis in.\\n\\nCheers,\\nGulli\\n\\nOK, I found the cause of the unaccent dictionary problem, and a workaround.It's not the vacuumdb version, not the unaccent version, and it's not even a pg_upgrade problem: I get this error also with PG 9.4.18 running on the old cluster, with both the 10.5 vacuumdb and the 9.4.18 vacuumdb, and I get the same error in both.And it's not strictly a vacuumdb problem, though vacuumdb triggers it.Here's a very minimal test case, unrelated to my DB, that you ought to be able to reproduce:SET search_path = \\\"$user\\\"; SELECT public.unaccent('fóö');SETERROR: text search dictionary \\\"unaccent\\\" does not existand here's a workaround:SET search_path = \\\"$user\\\"; SELECT public.unaccent(tsdict.oid, 'fóö') FROM pg_catalog.pg_ts_dict tsdict WHERE dictname='unaccent';SET unaccent---------- foo(1 row)The workaround avoids the OID lookup of the dictionary ... that lookup (in the single-argument unaccent function) is done by unqualified name:https://github.com/postgres/postgres/blob/fb8697b31aaeebe6170c572739867dcaa01053c6/contrib/unaccent/unaccent.c#L377 dictOid = get_ts_dict_oid(stringToQualifiedNameList(\\\"unaccent\\\"), false);and that fails if the search path doesn't include public.So it is indeed triggered by the security changes that Bruce mentioned; those were backported into 9.4.17: https://www.postgresql.org/docs/9.4/static/release-9-4-17.html ... and so got pulled in by my Macports upgrades. So nothing to do with pg_upgrade.So the workaround for my vacuumdb/function-index problem is to give unaccent the OID of the text search dictionary, so that the search path isn't in play:CREATE OR REPLACE FUNCTION public.semantic_normalize(title text)RETURNS text LANGUAGE sql IMMUTABLE STRICTAS $function$ SELECT lower(public.unaccent(16603, btrim(regexp_replace($1, '\\\\s+', ' ', 'g'), ' \\\"')))$function$;and that makes vacuumdb -z work in both 9.4.18 and 10.5, and makes ./analyze_new_cluster.sh complete without problems.The proper fix is, I suppose, to make the single-argument unaccent function explicitly look up the dictionary in the same schema as the function itself is in.Cheers,Gulli\",\n \"msg_date\": \"Wed, 5 Sep 2018 08:49:50 +0000\",\n \"msg_from\": \"Gunnlaugur Thor Briem <gunnlaugur@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"unaccent(text) fails depending on search_path (WAS: pg_upgrade fails\\n saying function unaccent(text) doesn't exist)\"\n },\n {\n \"msg_contents\": \"On 09/05/2018 01:49 AM, Gunnlaugur Thor Briem wrote:\\n> OK, I found the cause of the unaccent dictionary problem, and a workaround.\\n> \\n> It's not the vacuumdb version, not the unaccent version, and it's not \\n> even a pg_upgrade problem: I get this error also with PG 9.4.18 running \\n> on the old cluster, with both the 10.5 vacuumdb and the 9.4.18 vacuumdb, \\n> and I get the same error in both.\\n> \\n> And it's not strictly a vacuumdb problem, though vacuumdb triggers it.\\n> \\n> Here's a very minimal test case, unrelated to my DB, that you ought to \\n> be able to reproduce:\\n> \\n> SET search_path = \\\"$user\\\"; SELECT public.unaccent('fóö');\\n> SET\\n> ERROR: text search dictionary \\\"unaccent\\\" does not exist\\n> \\n> and here's a workaround:\\n> \\n> SET search_path = \\\"$user\\\"; SELECT public.unaccent(tsdict.oid, 'fóö')\\n> FROM pg_catalog.pg_ts_dict tsdict WHERE dictname='unaccent';\\n> SET\\n> unaccent\\n> ----------\\n> foo\\n> (1 row)\\n> \\n> The workaround avoids the OID lookup of the dictionary ... that lookup \\n> (in the single-argument unaccent function) is done by unqualified name:\\n> \\n> https://github.com/postgres/postgres/blob/fb8697b31aaeebe6170c572739867dcaa01053c6/contrib/unaccent/unaccent.c#L377\\n> \\n> dictOid = \\n> get_ts_dict_oid(stringToQualifiedNameList(\\\"unaccent\\\"), false);\\n> \\n> and that fails if the search path doesn't include public. >\\n> So it is indeed triggered by the security changes that Bruce mentioned; \\n> those were backported into 9.4.17: \\n> https://www.postgresql.org/docs/9.4/static/release-9-4-17.html ... and \\n> so got pulled in by my Macports upgrades. So nothing to do with pg_upgrade.\\n> \\n> So the workaround for my vacuumdb/function-index problem is to give \\n> unaccent the OID of the text search dictionary, so that the search path \\n> isn't in play:\\n> \\n> CREATE OR REPLACE FUNCTION public.semantic_normalize(title text)\\n> RETURNS text\\n> LANGUAGE sql\\n> IMMUTABLE STRICT\\n> AS $function$\\n> SELECT lower(public.unaccent(16603, btrim(regexp_replace($1, '\\\\s+', ' \\n> ', 'g'), ' \\\"')))\\n> $function$;\\n> \\n> and that makes vacuumdb -z work in both 9.4.18 and 10.5, and makes \\n> ./analyze_new_cluster.sh complete without problems.\\n\\n\\nNice investigation. Working off the above, I offer a suggestion:\\n\\nSET search_path = \\\"$user\\\"; SELECT public.unaccent('unaccent', 'fóö');\\nSET\\nERROR: text search dictionary \\\"unaccent\\\" does not exist\\nLINE 1: SELECT public.unaccent('unaccent', 'fóö');\\n\\n\\nSET search_path = \\\"$user\\\"; SELECT public.unaccent('public.unaccent', 'fóö');\\nSET\\n unaccent\\n----------\\n foo\\n\\nThat eliminates hard wiring the OID.\\n\\n> \\n> The proper fix is, I suppose, to make the single-argument unaccent \\n> function explicitly look up the dictionary in the same schema as the \\n> function itself is in.\\n> \\n> Cheers,\\n> Gulli\\n> \\n\\n\\n-- \\nAdrian Klaver\\nadrian.klaver@aklaver.com\\n\\n\",\n \"msg_date\": \"Wed, 5 Sep 2018 07:00:03 -0700\",\n \"msg_from\": \"Adrian Klaver <adrian.klaver@aklaver.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: unaccent(text) fails depending on search_path (WAS: pg_upgrade\\n fails saying function unaccent(text) doesn't exist)\"\n },\n {\n \"msg_contents\": \"Yep, a neater workaround for sure!\\n\\nCheers,\\nGulli\\n\\nOn Wed, Sep 5, 2018 at 2:00 PM Adrian Klaver <adrian.klaver@aklaver.com>\\nwrote:\\n\\n> On 09/05/2018 01:49 AM, Gunnlaugur Thor Briem wrote:\\n> > OK, I found the cause of the unaccent dictionary problem, and a\\n> workaround.\\n> >\\n> > It's not the vacuumdb version, not the unaccent version, and it's not\\n> > even a pg_upgrade problem: I get this error also with PG 9.4.18 running\\n> > on the old cluster, with both the 10.5 vacuumdb and the 9.4.18 vacuumdb,\\n> > and I get the same error in both.\\n> >\\n> > And it's not strictly a vacuumdb problem, though vacuumdb triggers it.\\n> >\\n> > Here's a very minimal test case, unrelated to my DB, that you ought to\\n> > be able to reproduce:\\n> >\\n> > SET search_path = \\\"$user\\\"; SELECT public.unaccent('fóö');\\n> > SET\\n> > ERROR: text search dictionary \\\"unaccent\\\" does not exist\\n> >\\n> > and here's a workaround:\\n> >\\n> > SET search_path = \\\"$user\\\"; SELECT public.unaccent(tsdict.oid, 'fóö')\\n> > FROM pg_catalog.pg_ts_dict tsdict WHERE dictname='unaccent';\\n> > SET\\n> > unaccent\\n> > ----------\\n> > foo\\n> > (1 row)\\n> >\\n> > The workaround avoids the OID lookup of the dictionary ... that lookup\\n> > (in the single-argument unaccent function) is done by unqualified name:\\n> >\\n> >\\n> https://github.com/postgres/postgres/blob/fb8697b31aaeebe6170c572739867dcaa01053c6/contrib/unaccent/unaccent.c#L377\\n> >\\n> > dictOid =\\n> > get_ts_dict_oid(stringToQualifiedNameList(\\\"unaccent\\\"), false);\\n> >\\n> > and that fails if the search path doesn't include public. >\\n> > So it is indeed triggered by the security changes that Bruce mentioned;\\n> > those were backported into 9.4.17:\\n> > https://www.postgresql.org/docs/9.4/static/release-9-4-17.html ... and\\n> > so got pulled in by my Macports upgrades. So nothing to do with\\n> pg_upgrade.\\n> >\\n> > So the workaround for my vacuumdb/function-index problem is to give\\n> > unaccent the OID of the text search dictionary, so that the search path\\n> > isn't in play:\\n> >\\n> > CREATE OR REPLACE FUNCTION public.semantic_normalize(title text)\\n> > RETURNS text\\n> > LANGUAGE sql\\n> > IMMUTABLE STRICT\\n> > AS $function$\\n> > SELECT lower(public.unaccent(16603, btrim(regexp_replace($1, '\\\\s+', '\\n> > ', 'g'), ' \\\"')))\\n> > $function$;\\n> >\\n> > and that makes vacuumdb -z work in both 9.4.18 and 10.5, and makes\\n> > ./analyze_new_cluster.sh complete without problems.\\n>\\n>\\n> Nice investigation. Working off the above, I offer a suggestion:\\n>\\n> SET search_path = \\\"$user\\\"; SELECT public.unaccent('unaccent', 'fóö');\\n> SET\\n> ERROR: text search dictionary \\\"unaccent\\\" does not exist\\n> LINE 1: SELECT public.unaccent('unaccent', 'fóö');\\n>\\n>\\n> SET search_path = \\\"$user\\\"; SELECT public.unaccent('public.unaccent',\\n> 'fóö');\\n> SET\\n> unaccent\\n> ----------\\n> foo\\n>\\n> That eliminates hard wiring the OID.\\n>\\n> >\\n> > The proper fix is, I suppose, to make the single-argument unaccent\\n> > function explicitly look up the dictionary in the same schema as the\\n> > function itself is in.\\n> >\\n> > Cheers,\\n> > Gulli\\n> >\\n>\\n>\\n> --\\n> Adrian Klaver\\n> adrian.klaver@aklaver.com\\n>\\n\\nYep, a neater workaround for sure!Cheers,GulliOn Wed, Sep 5, 2018 at 2:00 PM Adrian Klaver <adrian.klaver@aklaver.com> wrote:On 09/05/2018 01:49 AM, Gunnlaugur Thor Briem wrote:\\n> OK, I found the cause of the unaccent dictionary problem, and a workaround.\\n> \\n> It's not the vacuumdb version, not the unaccent version, and it's not \\n> even a pg_upgrade problem: I get this error also with PG 9.4.18 running \\n> on the old cluster, with both the 10.5 vacuumdb and the 9.4.18 vacuumdb, \\n> and I get the same error in both.\\n> \\n> And it's not strictly a vacuumdb problem, though vacuumdb triggers it.\\n> \\n> Here's a very minimal test case, unrelated to my DB, that you ought to \\n> be able to reproduce:\\n> \\n> SET search_path = \\\"$user\\\"; SELECT public.unaccent('fóö');\\n> SET\\n> ERROR: text search dictionary \\\"unaccent\\\" does not exist\\n> \\n> and here's a workaround:\\n> \\n> SET search_path = \\\"$user\\\"; SELECT public.unaccent(tsdict.oid, 'fóö')\\n> FROM pg_catalog.pg_ts_dict tsdict WHERE dictname='unaccent';\\n> SET\\n> unaccent\\n> ----------\\n> foo\\n> (1 row)\\n> \\n> The workaround avoids the OID lookup of the dictionary ... that lookup \\n> (in the single-argument unaccent function) is done by unqualified name:\\n> \\n> https://github.com/postgres/postgres/blob/fb8697b31aaeebe6170c572739867dcaa01053c6/contrib/unaccent/unaccent.c#L377\\n> \\n> dictOid = \\n> get_ts_dict_oid(stringToQualifiedNameList(\\\"unaccent\\\"), false);\\n> \\n> and that fails if the search path doesn't include public. >\\n> So it is indeed triggered by the security changes that Bruce mentioned; \\n> those were backported into 9.4.17: \\n> https://www.postgresql.org/docs/9.4/static/release-9-4-17.html ... and \\n> so got pulled in by my Macports upgrades. So nothing to do with pg_upgrade.\\n> \\n> So the workaround for my vacuumdb/function-index problem is to give \\n> unaccent the OID of the text search dictionary, so that the search path \\n> isn't in play:\\n> \\n> CREATE OR REPLACE FUNCTION public.semantic_normalize(title text)\\n> RETURNS text\\n> LANGUAGE sql\\n> IMMUTABLE STRICT\\n> AS $function$\\n> SELECT lower(public.unaccent(16603, btrim(regexp_replace($1, '\\\\s+', ' \\n> ', 'g'), ' \\\"')))\\n> $function$;\\n> \\n> and that makes vacuumdb -z work in both 9.4.18 and 10.5, and makes \\n> ./analyze_new_cluster.sh complete without problems.\\n\\n\\nNice investigation. Working off the above, I offer a suggestion:\\n\\nSET search_path = \\\"$user\\\"; SELECT public.unaccent('unaccent', 'fóö');\\nSET\\nERROR: text search dictionary \\\"unaccent\\\" does not exist\\nLINE 1: SELECT public.unaccent('unaccent', 'fóö');\\n\\n\\nSET search_path = \\\"$user\\\"; SELECT public.unaccent('public.unaccent', 'fóö');\\nSET\\n unaccent\\n----------\\n foo\\n\\nThat eliminates hard wiring the OID.\\n\\n> \\n> The proper fix is, I suppose, to make the single-argument unaccent \\n> function explicitly look up the dictionary in the same schema as the \\n> function itself is in.\\n> \\n> Cheers,\\n> Gulli\\n> \\n\\n\\n-- \\nAdrian Klaver\\nadrian.klaver@aklaver.com\",\n \"msg_date\": \"Wed, 5 Sep 2018 14:06:17 +0000\",\n \"msg_from\": \"Gunnlaugur Thor Briem <gunnlaugur@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: unaccent(text) fails depending on search_path (WAS: pg_upgrade\\n fails saying function unaccent(text) doesn't exist)\"\n },\n {\n \"msg_contents\": \"Gunnlaugur Thor Briem <gunnlaugur@gmail.com> writes:\\n> SET search_path = \\\"$user\\\"; SELECT public.unaccent('foo');\\n> SET\\n> ERROR: text search dictionary \\\"unaccent\\\" does not exist\\n\\nMeh. I think we need the attached, or something just about like it.\\n\\nIt's barely possible that there's somebody out there who's relying on\\nsetting the search path to allow choosing among multiple \\\"unaccent\\\"\\ndictionaries. But there are way more people whose functions are\\nbroken due to the recent search-path-tightening changes.\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Wed, 05 Sep 2018 15:03:33 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: unaccent(text) fails depending on search_path (WAS: pg_upgrade\\n fails saying function unaccent(text) doesn't exist)\"\n },\n {\n \"msg_contents\": \"[ redirecting to pgsql-hackers ]\\n\\nI wrote:\\n> Gunnlaugur Thor Briem <gunnlaugur@gmail.com> writes:\\n>> SET search_path = \\\"$user\\\"; SELECT public.unaccent('foo');\\n>> SET\\n>> ERROR: text search dictionary \\\"unaccent\\\" does not exist\\n\\n> Meh. I think we need the attached, or something just about like it.\\n>\\n> It's barely possible that there's somebody out there who's relying on\\n> setting the search path to allow choosing among multiple \\\"unaccent\\\"\\n> dictionaries. But there are way more people whose functions are\\n> broken due to the recent search-path-tightening changes.\\n\\nHere's a slightly more efficient version.\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Wed, 05 Sep 2018 18:37:00 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: unaccent(text) fails depending on search_path (WAS: pg_upgrade\\n fails saying function unaccent(text) doesn't exist)\"\n },\n {\n \"msg_contents\": \"On Wed, Sep 5, 2018 at 06:37:00PM -0400, Tom Lane wrote:\\n> [ redirecting to pgsql-hackers ]\\n> \\n> I wrote:\\n> > Gunnlaugur Thor Briem <gunnlaugur@gmail.com> writes:\\n> >> SET search_path = \\\"$user\\\"; SELECT public.unaccent('foo');\\n> >> SET\\n> >> ERROR: text search dictionary \\\"unaccent\\\" does not exist\\n> \\n> > Meh. I think we need the attached, or something just about like it.\\n> >\\n> > It's barely possible that there's somebody out there who's relying on\\n> > setting the search path to allow choosing among multiple \\\"unaccent\\\"\\n> > dictionaries. But there are way more people whose functions are\\n> > broken due to the recent search-path-tightening changes.\\n> \\n> Here's a slightly more efficient version.\\n\\nIf we are going down this route, is there any thought of handling\\nearchdistance the same way?\\n\\n\\thttps://www.postgresql.org/message-id/20180330205229.GS8476@momjian.us\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\",\n \"msg_date\": \"Fri, 7 Sep 2018 18:32:47 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: unaccent(text) fails depending on search_path (WAS: pg_upgrade\\n fails saying function unaccent(text) doesn't exist)\"\n },\n {\n \"msg_contents\": \"Bruce Momjian <bruce@momjian.us> writes:\\n> If we are going down this route, is there any thought of handling\\n> earchdistance the same way?\\n> \\thttps://www.postgresql.org/message-id/20180330205229.GS8476@momjian.us\\n\\nAFAICS there are no internal-to-the-C-code search path dependencies\\nin earthdistance.c, so it's not the same problem.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Fri, 07 Sep 2018 18:43:52 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: unaccent(text) fails depending on search_path (WAS: pg_upgrade\\n fails saying function unaccent(text) doesn't exist)\"\n },\n {\n \"msg_contents\": \"On Fri, Sep 7, 2018 at 06:43:52PM -0400, Tom Lane wrote:\\n> Bruce Momjian <bruce@momjian.us> writes:\\n> > If we are going down this route, is there any thought of handling\\n> > earchdistance the same way?\\n> > \\thttps://www.postgresql.org/message-id/20180330205229.GS8476@momjian.us\\n> \\n> AFAICS there are no internal-to-the-C-code search path dependencies\\n> in earthdistance.c, so it's not the same problem.\\n\\nUh, there is an SQL function that calls functions from the module that\\nfail. It would be a CREATE FUNCTION patch, I think, but I thought the\\nissue was the same.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\",\n \"msg_date\": \"Fri, 7 Sep 2018 18:56:43 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: unaccent(text) fails depending on search_path (WAS: pg_upgrade\\n fails saying function unaccent(text) doesn't exist)\"\n },\n {\n \"msg_contents\": \"Bruce Momjian <bruce@momjian.us> writes:\\n> On Fri, Sep 7, 2018 at 06:43:52PM -0400, Tom Lane wrote:\\n>> AFAICS there are no internal-to-the-C-code search path dependencies\\n>> in earthdistance.c, so it's not the same problem.\\n\\n> Uh, there is an SQL function that calls functions from the module that\\n> fail. It would be a CREATE FUNCTION patch, I think, but I thought the\\n> issue was the same.\\n\\nNot really. You could either interpolate @extschema@ into the text\\nof the referencing function, or (though much inferior for performance)\\nhave it SET SEARCH_PATH FROM CURRENT. Either of those changes would\\ninvolve an extension version bump since they're changing the extension\\nscript. What's more of a problem is that we could no longer claim\\nthe extension is relocatable. My unaccent fix dodged that by looking\\nup the C function's current schema, but I don't think there's any\\nequivalent functionality available at SQL level.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Fri, 07 Sep 2018 19:05:39 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: unaccent(text) fails depending on search_path (WAS: pg_upgrade\\n fails saying function unaccent(text) doesn't exist)\"\n }\n]","isConversational":false}}},{"rowIdx":47193,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Stumble upon this formatting issue while reading the pg_upgrade docs.\n\nMartín","msg_date":"Wed, 29 Aug 2018 20:02:54 -0700","msg_from":"=?UTF-8?q?Mart=C3=ADn_Fern=C3=A1ndez?= <fmartin91@gmail.com>","msg_from_op":true,"msg_subject":"[PATCH] Fix formatting in pg_upgrade manpage doc"},{"msg_contents":"On Wed, Aug 29, 2018 at 08:02:54PM -0700, Mart�n Fern�ndez wrote:\n> *\n> Stumble upon this formatting issue while reading the pg_upgrade docs.\n\nI am confused. What is the value of moving \"without\" to the next line\nin the sgml file?\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n","msg_date":"Wed, 29 Aug 2018 23:34:01 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Fix formatting in pg_upgrade manpage doc"},{"msg_contents":"I spotted the formatting issue in an older version and the change I did confused me thinking that it was fixing the issue. \n\nSorry.\n\nMartín\n\nOn Thu, Aug 30, 2018 at 12:34 AM Bruce Momjian < Bruce Momjian ( Bruce Momjian <bruce@momjian.us> ) > wrote:\n\n> \n> \n> \n> On Wed, Aug 29, 2018 at 08:02:54PM -0700, Martín Fernández wrote:\n> > *\n> > Stumble upon this formatting issue while reading the pg_upgrade docs.\n> \n> I am confused. What is the value of moving \"without\" to the next line\n> in the sgml file?\n> \n> --\n> Bruce Momjian < bruce@momjian.us > http://momjian.us\n> EnterpriseDB http://enterprisedb.com\n> \n> + As you are, so once was I. As I am, so you will be. +\n> + Ancient Roman grave inscription +\n> \n> \n>\n\n\nI spotted the formatting issue in an older version and the change I did confused me thinking that it was fixing the issue. Sorry.\n\nMartín\nOn Thu, Aug 30, 2018 at 12:34 AM Bruce Momjian <Bruce Momjian > wrote:On Wed, Aug 29, 2018 at 08:02:54PM -0700, Martín Fernández wrote:\n> *\n> Stumble upon this formatting issue while reading the pg_upgrade docs.\n\nI am confused. What is the value of moving \"without\" to the next line\nin the sgml file?\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +","msg_date":"Wed, 29 Aug 2018 21:31:25 -0700","msg_from":"=?UTF-8?q?Mart=C3=ADn_Fern=C3=A1ndez?= <fmartin91@gmail.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Fix formatting in pg_upgrade manpage doc"}],"string":"[\n {\n \"msg_contents\": \"Stumble upon this formatting issue while reading the pg_upgrade docs.\\n\\nMartín\",\n \"msg_date\": \"Wed, 29 Aug 2018 20:02:54 -0700\",\n \"msg_from\": \"=?UTF-8?q?Mart=C3=ADn_Fern=C3=A1ndez?= <fmartin91@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"[PATCH] Fix formatting in pg_upgrade manpage doc\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 29, 2018 at 08:02:54PM -0700, Mart�n Fern�ndez wrote:\\n> *\\n> Stumble upon this formatting issue while reading the pg_upgrade docs.\\n\\nI am confused. What is the value of moving \\\"without\\\" to the next line\\nin the sgml file?\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\",\n \"msg_date\": \"Wed, 29 Aug 2018 23:34:01 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Fix formatting in pg_upgrade manpage doc\"\n },\n {\n \"msg_contents\": \"I spotted the formatting issue in an older version and the change I did confused me thinking that it was fixing the issue. \\n\\nSorry.\\n\\nMartín\\n\\nOn Thu, Aug 30, 2018 at 12:34 AM Bruce Momjian < Bruce Momjian ( Bruce Momjian <bruce@momjian.us> ) > wrote:\\n\\n> \\n> \\n> \\n> On Wed, Aug 29, 2018 at 08:02:54PM -0700, Martín Fernández wrote:\\n> > *\\n> > Stumble upon this formatting issue while reading the pg_upgrade docs.\\n> \\n> I am confused. What is the value of moving \\\"without\\\" to the next line\\n> in the sgml file?\\n> \\n> --\\n> Bruce Momjian < bruce@momjian.us > http://momjian.us\\n> EnterpriseDB http://enterprisedb.com\\n> \\n> + As you are, so once was I. As I am, so you will be. +\\n> + Ancient Roman grave inscription +\\n> \\n> \\n>\\n\\n\\nI spotted the formatting issue in an older version and the change I did confused me thinking that it was fixing the issue. Sorry.\\n\\nMartín\\nOn Thu, Aug 30, 2018 at 12:34 AM Bruce Momjian <Bruce Momjian > wrote:On Wed, Aug 29, 2018 at 08:02:54PM -0700, Martín Fernández wrote:\\n> *\\n> Stumble upon this formatting issue while reading the pg_upgrade docs.\\n\\nI am confused. What is the value of moving \\\"without\\\" to the next line\\nin the sgml file?\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\",\n \"msg_date\": \"Wed, 29 Aug 2018 21:31:25 -0700\",\n \"msg_from\": \"=?UTF-8?q?Mart=C3=ADn_Fern=C3=A1ndez?= <fmartin91@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Fix formatting in pg_upgrade manpage doc\"\n }\n]","isConversational":false}}},{"rowIdx":47194,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nAttached is a patch to fix comments of IndexInfo. ii_KeyAttrNumbers was\nrenamed to ii_IndexAttrNumbers and ii_Am was added but these are\nnot reflected to the comment.\n\nRegards,\n-- \nYugo Nagata <nagata@sraoss.co.jp>","msg_date":"Thu, 30 Aug 2018 13:48:31 +0900","msg_from":"Yugo Nagata <nagata@sraoss.co.jp>","msg_from_op":true,"msg_subject":"Fix comments of IndexInfo"},{"msg_contents":"On 30/08/18 07:48, Yugo Nagata wrote:\n> Attached is a patch to fix comments of IndexInfo. ii_KeyAttrNumbers was\n> renamed to ii_IndexAttrNumbers and ii_Am was added but these are\n> not reflected to the comment.\n\nApplied, thanks!\n\n- Heikki\n\n","msg_date":"Thu, 30 Aug 2018 09:14:26 +0300","msg_from":"Heikki Linnakangas <hlinnaka@iki.fi>","msg_from_op":false,"msg_subject":"Re: Fix comments of IndexInfo"},{"msg_contents":"On Thu, 30 Aug 2018 09:14:26 +0300\nHeikki Linnakangas <hlinnaka@iki.fi> wrote:\n\n> On 30/08/18 07:48, Yugo Nagata wrote:\n> > Attached is a patch to fix comments of IndexInfo. ii_KeyAttrNumbers was\n> > renamed to ii_IndexAttrNumbers and ii_Am was added but these are\n> > not reflected to the comment.\n> \n> Applied, thanks!\n\nThanks!\n-- \nYugo Nagata <nagata@sraoss.co.jp>\n\n","msg_date":"Thu, 30 Aug 2018 15:18:09 +0900","msg_from":"Yugo Nagata <nagata@sraoss.co.jp>","msg_from_op":true,"msg_subject":"Re: Fix comments of IndexInfo"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nAttached is a patch to fix comments of IndexInfo. ii_KeyAttrNumbers was\\nrenamed to ii_IndexAttrNumbers and ii_Am was added but these are\\nnot reflected to the comment.\\n\\nRegards,\\n-- \\nYugo Nagata <nagata@sraoss.co.jp>\",\n \"msg_date\": \"Thu, 30 Aug 2018 13:48:31 +0900\",\n \"msg_from\": \"Yugo Nagata <nagata@sraoss.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Fix comments of IndexInfo\"\n },\n {\n \"msg_contents\": \"On 30/08/18 07:48, Yugo Nagata wrote:\\n> Attached is a patch to fix comments of IndexInfo. ii_KeyAttrNumbers was\\n> renamed to ii_IndexAttrNumbers and ii_Am was added but these are\\n> not reflected to the comment.\\n\\nApplied, thanks!\\n\\n- Heikki\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 09:14:26 +0300\",\n \"msg_from\": \"Heikki Linnakangas <hlinnaka@iki.fi>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Fix comments of IndexInfo\"\n },\n {\n \"msg_contents\": \"On Thu, 30 Aug 2018 09:14:26 +0300\\nHeikki Linnakangas <hlinnaka@iki.fi> wrote:\\n\\n> On 30/08/18 07:48, Yugo Nagata wrote:\\n> > Attached is a patch to fix comments of IndexInfo. ii_KeyAttrNumbers was\\n> > renamed to ii_IndexAttrNumbers and ii_Am was added but these are\\n> > not reflected to the comment.\\n> \\n> Applied, thanks!\\n\\nThanks!\\n-- \\nYugo Nagata <nagata@sraoss.co.jp>\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 15:18:09 +0900\",\n \"msg_from\": \"Yugo Nagata <nagata@sraoss.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Fix comments of IndexInfo\"\n }\n]","isConversational":false}}},{"rowIdx":47195,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi hackers,\n\nI am really confused. If my conclusions are correct, then nobody ever \ntried to use replication slots at Windows!\nThe function RestoreSlotFromDisk in slot.c contains the following code:\n\nstatic void\nRestoreSlotFromDisk(const char *name)\n{\n ReplicationSlotOnDisk cp;\n int i;\n char path[MAXPGPATH + 22];\n int fd;\n bool restored = false;\n int readBytes;\n pg_crc32c checksum;\n\n /* no need to lock here, no concurrent access allowed yet */\n\n /* delete temp file if it exists */\n sprintf(path, \"pg_replslot/%s/state.tmp\", name);\n if (unlink(path) < 0 && errno != ENOENT)\n ereport(PANIC,\n (errcode_for_file_access(),\n errmsg(\"could not remove file \\\"%s\\\": %m\", path)));\n\n sprintf(path, \"pg_replslot/%s/state\", name);\n\n elog(DEBUG1, \"restoring replication slot from \\\"%s\\\"\", path);\n\n fd = OpenTransientFile(path, O_RDWR | PG_BINARY);\n\n /*\n * We do not need to handle this as we are rename()ing the \ndirectory into\n * place only after we fsync()ed the state file.\n */\n if (fd < 0)\n ereport(PANIC,\n (errcode_for_file_access(),\n errmsg(\"could not open file \\\"%s\\\": %m\", path)));\n\n /*\n * Sync state file before we're reading from it. We might have crashed\n * while it wasn't synced yet and we shouldn't continue on that basis.\n */\npgstat_report_wait_start(WAIT_EVENT_REPLICATION_SLOT_RESTORE_SYNC);\n if (pg_fsync(fd) != 0)\n {\n int save_errno = errno;\n\n CloseTransientFile(fd);\n errno = save_errno;\n ereport(PANIC,\n (errcode_for_file_access(),\n errmsg(\"could not fsync file \\\"%s\\\": %m\",\n path)));\n }\n pgstat_report_wait_end();\n\n /* Also sync the parent directory */\n START_CRIT_SECTION();\n fsync_fname(path, true);\n END_CRIT_SECTION();\n\n ...\n\nPlease notice that fsync_fname with comment \"also sync parent directory\" \nis called for path of the file!\nfsync_fname in turn does the following:\n\n\n /*\n * fsync_fname -- Try to fsync a file or directory\n *\n * Ignores errors trying to open unreadable files, or trying to fsync\n * directories on systems where that isn't allowed/required. Reports\n * other errors non-fatally.\n */\nint\nfsync_fname(const char *fname, bool isdir, const char *progname)\n{\n int fd;\n int flags;\n int returncode;\n\n /*\n * Some OSs require directories to be opened read-only whereas other\n * systems don't allow us to fsync files opened read-only; so we \nneed both\n * cases here. Using O_RDWR will cause us to fail to fsync files \nthat are\n * not writable by our userid, but we assume that's OK.\n */\n flags = PG_BINARY;\n if (!isdir)\n flags |= O_RDWR;\n else\n flags |= O_RDONLY;\n\n /*\n * Open the file, silently ignoring errors about unreadable files (or\n * unsupported operations, e.g. opening a directory under Windows), and\n * logging others.\n */\n fd = open(fname, flags);\n if (fd < 0)\n {\n if (errno == EACCES || (isdir && errno == EISDIR))\n return 0;\n fprintf(stderr, _(\"%s: could not open file \\\"%s\\\": %s\\n\"),\n progname, fname, strerror(errno));\n return -1;\n }\n\n returncode = fsync(fd);\n\n\nSo if \"isdir\" is true (and it is true in this case), it sets O_RDONLY flag.\nThen fsync_fname successfully opens slot file in readonly mode and calls \nfsync() which at windows\nis substituted with _commit() which in turn is wrapper for FlushFileBuffers.\nFinally FlushFileBuffers returns ERROR_ACCESS_DENINED which cause \nassertion failure in _commit:\n\nif ( !FlushFileBuffers((HANDLE)_get_osfhandle(filedes)) ) {\n retval = GetLastError();\n }\n else {\n retval = 0; /* return success */\n }\n\n /* map the OS return code to C errno value and return code */\n if (retval == 0)\n goto good;\n\n _doserrno = retval;\n\n }\n\n errno = EBADF;\n retval = -1;\n\n _ASSERTE((\"Invalid file descriptor. File possibly closed by a different thread\",0));\n\n\nI think that the problem happen only with debug version of postgres.\nRelease version will just return error in this case which is silently ignored by RestoreSlotFromDisk function.\n\nI think that bug fix is trivial: we just need to use fsync_parent_path instead of fsync_fname in RestoreSlotFromDisk.\n\n-- \nKonstantin Knizhnik\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres Company\n\n\n","msg_date":"Thu, 30 Aug 2018 11:00:43 +0300","msg_from":"Konstantin Knizhnik <k.knizhnik@postgrespro.ru>","msg_from_op":true,"msg_subject":"Bug in slot.c and are replication slots ever used at Window?"},{"msg_contents":"On Thu, Aug 30, 2018 at 11:00:43AM +0300, Konstantin Knizhnik wrote:\n> So if \"isdir\" is true (and it is true in this case), it sets O_RDONLY\n> flag. Then fsync_fname successfully opens slot file in readonly mode\n> and calls fsync() which at windows is substituted with _commit() which\n> in turn is wrapper for FlushFileBuffers. \n\nIt seems to me that you are right here, \"path\" points to\npg_replslot/$SLOTNAME/state which is a file so the fsync is incorrect.\nI am not sure that we'd want to publish fsync_parent_path out of fd.c\nthough, so perhaps we could just save the slot path in a different\nvariable and use it?\n--\nMichael","msg_date":"Thu, 30 Aug 2018 11:27:31 -0700","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Bug in slot.c and are replication slots ever used at Window?"},{"msg_contents":"Hi Konstantin,\n\nOn Thu, Aug 30, 2018 at 11:27:31AM -0700, Michael Paquier wrote:\n> It seems to me that you are right here, \"path\" points to\n> pg_replslot/$SLOTNAME/state which is a file so the fsync is incorrect.\n> I am not sure that we'd want to publish fsync_parent_path out of fd.c\n> though, so perhaps we could just save the slot path in a different\n> variable and use it?\n\nI have spent more time on this bug, and the code path you have pointed\nat is the only one having such an issue. Attached is a patch to fix the\nproblem. It includes the sanity checks I have used to check all code\npaths calling fsync_fname() for both the frontend and the backend code.\n\nThe checks will not be included in the final fix, still they look useful\nso I am planning to start a new thread on the matter as perhaps other\nfolks have more and/or better ideas.\n--\nMichael","msg_date":"Fri, 31 Aug 2018 11:46:47 -0700","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Bug in slot.c and are replication slots ever used at Window?"},{"msg_contents":"On Fri, Aug 31, 2018 at 11:46:47AM -0700, Michael Paquier wrote:\n> The checks will not be included in the final fix, still they look useful\n> so I am planning to start a new thread on the matter as perhaps other\n> folks have more and/or better ideas.\n\nPushed the fix down to 9.4, without the extra sanity checks.\n--\nMichael","msg_date":"Sun, 2 Sep 2018 12:44:31 -0700","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Bug in slot.c and are replication slots ever used at Window?"}],"string":"[\n {\n \"msg_contents\": \"Hi hackers,\\n\\nI am really confused. If my conclusions are correct, then nobody ever \\ntried to use replication slots at Windows!\\nThe function RestoreSlotFromDisk in slot.c contains the following code:\\n\\nstatic void\\nRestoreSlotFromDisk(const char *name)\\n{\\n ReplicationSlotOnDisk cp;\\n int i;\\n char path[MAXPGPATH + 22];\\n int fd;\\n bool restored = false;\\n int readBytes;\\n pg_crc32c checksum;\\n\\n /* no need to lock here, no concurrent access allowed yet */\\n\\n /* delete temp file if it exists */\\n sprintf(path, \\\"pg_replslot/%s/state.tmp\\\", name);\\n if (unlink(path) < 0 && errno != ENOENT)\\n ereport(PANIC,\\n (errcode_for_file_access(),\\n errmsg(\\\"could not remove file \\\\\\\"%s\\\\\\\": %m\\\", path)));\\n\\n sprintf(path, \\\"pg_replslot/%s/state\\\", name);\\n\\n elog(DEBUG1, \\\"restoring replication slot from \\\\\\\"%s\\\\\\\"\\\", path);\\n\\n fd = OpenTransientFile(path, O_RDWR | PG_BINARY);\\n\\n /*\\n * We do not need to handle this as we are rename()ing the \\ndirectory into\\n * place only after we fsync()ed the state file.\\n */\\n if (fd < 0)\\n ereport(PANIC,\\n (errcode_for_file_access(),\\n errmsg(\\\"could not open file \\\\\\\"%s\\\\\\\": %m\\\", path)));\\n\\n /*\\n * Sync state file before we're reading from it. We might have crashed\\n * while it wasn't synced yet and we shouldn't continue on that basis.\\n */\\npgstat_report_wait_start(WAIT_EVENT_REPLICATION_SLOT_RESTORE_SYNC);\\n if (pg_fsync(fd) != 0)\\n {\\n int save_errno = errno;\\n\\n CloseTransientFile(fd);\\n errno = save_errno;\\n ereport(PANIC,\\n (errcode_for_file_access(),\\n errmsg(\\\"could not fsync file \\\\\\\"%s\\\\\\\": %m\\\",\\n path)));\\n }\\n pgstat_report_wait_end();\\n\\n /* Also sync the parent directory */\\n START_CRIT_SECTION();\\n fsync_fname(path, true);\\n END_CRIT_SECTION();\\n\\n ...\\n\\nPlease notice that fsync_fname with comment \\\"also sync parent directory\\\" \\nis called for path of the file!\\nfsync_fname in turn does the following:\\n\\n\\n /*\\n * fsync_fname -- Try to fsync a file or directory\\n *\\n * Ignores errors trying to open unreadable files, or trying to fsync\\n * directories on systems where that isn't allowed/required. Reports\\n * other errors non-fatally.\\n */\\nint\\nfsync_fname(const char *fname, bool isdir, const char *progname)\\n{\\n int fd;\\n int flags;\\n int returncode;\\n\\n /*\\n * Some OSs require directories to be opened read-only whereas other\\n * systems don't allow us to fsync files opened read-only; so we \\nneed both\\n * cases here. Using O_RDWR will cause us to fail to fsync files \\nthat are\\n * not writable by our userid, but we assume that's OK.\\n */\\n flags = PG_BINARY;\\n if (!isdir)\\n flags |= O_RDWR;\\n else\\n flags |= O_RDONLY;\\n\\n /*\\n * Open the file, silently ignoring errors about unreadable files (or\\n * unsupported operations, e.g. opening a directory under Windows), and\\n * logging others.\\n */\\n fd = open(fname, flags);\\n if (fd < 0)\\n {\\n if (errno == EACCES || (isdir && errno == EISDIR))\\n return 0;\\n fprintf(stderr, _(\\\"%s: could not open file \\\\\\\"%s\\\\\\\": %s\\\\n\\\"),\\n progname, fname, strerror(errno));\\n return -1;\\n }\\n\\n returncode = fsync(fd);\\n\\n\\nSo if \\\"isdir\\\" is true (and it is true in this case), it sets O_RDONLY flag.\\nThen fsync_fname successfully opens slot file in readonly mode and calls \\nfsync() which at windows\\nis substituted with _commit() which in turn is wrapper for FlushFileBuffers.\\nFinally FlushFileBuffers returns ERROR_ACCESS_DENINED which cause \\nassertion failure in _commit:\\n\\nif ( !FlushFileBuffers((HANDLE)_get_osfhandle(filedes)) ) {\\n retval = GetLastError();\\n }\\n else {\\n retval = 0; /* return success */\\n }\\n\\n /* map the OS return code to C errno value and return code */\\n if (retval == 0)\\n goto good;\\n\\n _doserrno = retval;\\n\\n }\\n\\n errno = EBADF;\\n retval = -1;\\n\\n _ASSERTE((\\\"Invalid file descriptor. File possibly closed by a different thread\\\",0));\\n\\n\\nI think that the problem happen only with debug version of postgres.\\nRelease version will just return error in this case which is silently ignored by RestoreSlotFromDisk function.\\n\\nI think that bug fix is trivial: we just need to use fsync_parent_path instead of fsync_fname in RestoreSlotFromDisk.\\n\\n-- \\nKonstantin Knizhnik\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 11:00:43 +0300\",\n \"msg_from\": \"Konstantin Knizhnik <k.knizhnik@postgrespro.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Bug in slot.c and are replication slots ever used at Window?\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 30, 2018 at 11:00:43AM +0300, Konstantin Knizhnik wrote:\\n> So if \\\"isdir\\\" is true (and it is true in this case), it sets O_RDONLY\\n> flag. Then fsync_fname successfully opens slot file in readonly mode\\n> and calls fsync() which at windows is substituted with _commit() which\\n> in turn is wrapper for FlushFileBuffers. \\n\\nIt seems to me that you are right here, \\\"path\\\" points to\\npg_replslot/$SLOTNAME/state which is a file so the fsync is incorrect.\\nI am not sure that we'd want to publish fsync_parent_path out of fd.c\\nthough, so perhaps we could just save the slot path in a different\\nvariable and use it?\\n--\\nMichael\",\n \"msg_date\": \"Thu, 30 Aug 2018 11:27:31 -0700\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Bug in slot.c and are replication slots ever used at Window?\"\n },\n {\n \"msg_contents\": \"Hi Konstantin,\\n\\nOn Thu, Aug 30, 2018 at 11:27:31AM -0700, Michael Paquier wrote:\\n> It seems to me that you are right here, \\\"path\\\" points to\\n> pg_replslot/$SLOTNAME/state which is a file so the fsync is incorrect.\\n> I am not sure that we'd want to publish fsync_parent_path out of fd.c\\n> though, so perhaps we could just save the slot path in a different\\n> variable and use it?\\n\\nI have spent more time on this bug, and the code path you have pointed\\nat is the only one having such an issue. Attached is a patch to fix the\\nproblem. It includes the sanity checks I have used to check all code\\npaths calling fsync_fname() for both the frontend and the backend code.\\n\\nThe checks will not be included in the final fix, still they look useful\\nso I am planning to start a new thread on the matter as perhaps other\\nfolks have more and/or better ideas.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 31 Aug 2018 11:46:47 -0700\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Bug in slot.c and are replication slots ever used at Window?\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 31, 2018 at 11:46:47AM -0700, Michael Paquier wrote:\\n> The checks will not be included in the final fix, still they look useful\\n> so I am planning to start a new thread on the matter as perhaps other\\n> folks have more and/or better ideas.\\n\\nPushed the fix down to 9.4, without the extra sanity checks.\\n--\\nMichael\",\n \"msg_date\": \"Sun, 2 Sep 2018 12:44:31 -0700\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Bug in slot.c and are replication slots ever used at Window?\"\n }\n]","isConversational":false}}},{"rowIdx":47196,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nI noticed that pg_verify_checksums computes bogus checksums if I compile\nit with '-O2 -Wall' but without -fno-strict-aliasing. Also I am getting\na compile warning then:\n\n|src/bin/pg_verify_checksums$ make CFLAGS='-g -O2 -Wall'\n[...]\n|gcc -g -O2 -Wall -I../../../src/include \n| -I/home/mba/Projekte/OSS/PostgreSQL/git/postgresql/build/../src/include\n| -D_GNU_SOURCE -c -o pg_verify_checksums.o\n| /home/mba/Projekte/OSS/PostgreSQL/git/postgresql/build/../src/bin/pg_verify_checksums/pg_verify_checksums.c\n| -MMD -MP -MF .deps/pg_verify_checksums.Po\n|/home/mba/Projekte/OSS/PostgreSQL/git/postgresql/build/../src/bin/pg_verify_checksums/pg_verify_checksums.c: \n| In function ‘scan_file’:\n|/home/mba/Projekte/OSS/PostgreSQL/git/postgresql/build/../src/bin/pg_verify_checksums/pg_verify_checksums.c:112:3: \n| warning: dereferencing type-punned pointer will break strict-aliasing\n| rules [-Wstrict-aliasing]\n| if (PageIsNew(buf))\n| ^~\n[...]\n|src/bin/pg_verify_checksums$ ./pg_verify_checksums -D ../../test/regress/tmp_check/data\n|pg_verify_checksums: checksum verification failed in file \"../../test/regress/tmp_check/data/global/1233\", block 0: calculated checksum FC8A but expected F857\n|pg_verify_checksums: checksum verification failed in file \"../../test/regress/tmp_check/data/global/1233\", block 1: calculated checksum F340 but expected A68D\n[...]\n|pg_verify_checksums: checksum verification failed in file \"../../test/regress/tmp_check/data/base/12368/2659\", block 5: calculated checksum 954D but expected D2ED\n|pg_verify_checksums: checksum verification failed in file \"../../test/regress/tmp_check/data/base/12368/2659\", block 6: calculated checksum 361 but expected E7F7\n|pg_verify_checksums: checksum verification failed in file \"../../test/regress/tmp_check/data/base/12368/2659\", block 7: calculated checksum 4C0D but expected 4B10\n|pg_verify_checksums: checksum verification failed in file \"../../test/regress/tmp_check/data/base/12368/2659\", block 8: calculated checksum 2B9A but expected 71E3\n|pg_verify_checksums: checksum verification failed in file \"../../test/regress/tmp_check/data/base/12368/2659\", block 9: calculated checksum 19CD but expected 541C\n|Checksum scan completed\n|Data checksum version: 1\n|Files scanned: 932\n|Blocks scanned: 2789\n|Bad checksums: 2789\n\nIf I add -fno-strict-aliasing to $CFLAGS, the problem goes away.\n\nIs this something to worry about, or just pilot error cause I am not\nusing the same $CFLAGS as for the rest of the build? I originally\nnoticed this problem with my external fork of pg_verify_checksums.\n\n\nMichael\n\n-- \nMichael Banck\nProjektleiter / Senior Berater\nTel.: +49 2166 9901-171\nFax: +49 2166 9901-100\nEmail: michael.banck@credativ.de\n\ncredativ GmbH, HRB Mönchengladbach 12080\nUSt-ID-Nummer: DE204566209\nTrompeterallee 108, 41189 Mönchengladbach\nGeschäftsführung: Dr. Michael Meskes, Jörg Folz, Sascha Heuer\n\nUnser Umgang mit personenbezogenen Daten unterliegt\nfolgenden Bestimmungen: https://www.credativ.de/datenschutz\n\n","msg_date":"Thu, 30 Aug 2018 10:35:00 +0200","msg_from":"Michael Banck <michael.banck@credativ.de>","msg_from_op":true,"msg_subject":"pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"\n> I noticed that pg_verify_checksums computes bogus checksums if I compile\n> it with '-O2 -Wall' but without -fno-strict-aliasing. Also I am getting\n> a compile warning then:\n>\n> [...]\n>\n> If I add -fno-strict-aliasing to $CFLAGS, the problem goes away.\n>\n> Is this something to worry about, or just pilot error cause I am not\n> using the same $CFLAGS as for the rest of the build? I originally\n> noticed this problem with my external fork of pg_verify_checksums.\n\nIt looks like the code is using aliasing where the standard says it should \nnot, which breaks compiler optimization, and the added options tells the \ncompiler to not assume that the code conforms to the standard...\n\nAs PostgreSQL source is expected to conform to some C standard (unsure \nwhich one right now, possibly c89 but maybe it is beginning to switch to \nc99, a young 19 years old standard), I'd suggest that the right fix is \nrather to actually remove the aliasing issue.\n\n-- \nFabien.\n\n","msg_date":"Thu, 30 Aug 2018 13:25:14 +0200 (CEST)","msg_from":"Fabien COELHO <coelho@cri.ensmp.fr>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"Fabien COELHO <coelho@cri.ensmp.fr> writes:\n>> If I add -fno-strict-aliasing to $CFLAGS, the problem goes away.\n>> Is this something to worry about, or just pilot error cause I am not\n>> using the same $CFLAGS as for the rest of the build? I originally\n>> noticed this problem with my external fork of pg_verify_checksums.\n\n> It looks like the code is using aliasing where the standard says it should \n> not, which breaks compiler optimization, and the added options tells the \n> compiler to not assume that the code conforms to the standard...\n\nActually, this code is just plain broken:\n\n\tchar\t\tbuf[BLCKSZ];\n\tPageHeader\theader = (PageHeader) buf;\n\nThere is no guarantee that a local char[] array is aligned on anything\nstronger than a byte boundary, and if it isn't, word-sized accesses to\n*header are going to fail on machines with strict alignment rules.\nI suppose that that is really what Michael's compiler is moaning about.\n\nI rather suspect that this hasn't been tested on anything but Intel\nhardware, which is famously misalignment-tolerant. The lack of any\napparent regression test infrastructure for it isn't leaving a warm\nfeeling about how much the buildfarm is testing it.\n\n(The right fix, of course, is to malloc the work buffer rather than\nput it on the stack.)\n\n\t\t\tregards, tom lane\n\n","msg_date":"Thu, 30 Aug 2018 10:39:26 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"On 2018-Aug-30, Fabien COELHO wrote:\n\n> As PostgreSQL source is expected to conform to some C standard (unsure which\n> one right now, possibly c89 but maybe it is beginning to switch to c99, a\n> young 19 years old standard), I'd suggest that the right fix is rather to\n> actually remove the aliasing issue.\n\nYeah, type aliasing is quite a rampant issue in our code and I don't\nthink there's an easy fix.\n\n-- \n�lvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n","msg_date":"Thu, 30 Aug 2018 11:45:47 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"On Thu, Aug 30, 2018 at 10:39:26AM -0400, Tom Lane wrote:\n> I rather suspect that this hasn't been tested on anything but Intel\n> hardware, which is famously misalignment-tolerant. The lack of any\n> apparent regression test infrastructure for it isn't leaving a warm\n> feeling about how much the buildfarm is testing it.\n> \n> (The right fix, of course, is to malloc the work buffer rather than\n> put it on the stack.)\n\npg_upgrade/file.c is careful about that (5afcd2a), and has a comment on\nthe matter, as does pg_standby.c.\n\nNow, grepping around for \"BLCKSZ]\", some garbage may have accumulated?\n- entrySplitPage in ginentrypage.c\n- rewind_copy_file_range in file.c\n- _hash_alloc_buckets in hashpage.c\n- pg_prewarm.c\n- blinsert.c\n- pg_waldump.c\n- logtape.c\n--\nMichael","msg_date":"Thu, 30 Aug 2018 10:56:52 -0700","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"On Thu, Aug 30, 2018 at 4:39 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n\n> Fabien COELHO <coelho@cri.ensmp.fr> writes:\n> >> If I add -fno-strict-aliasing to $CFLAGS, the problem goes away.\n> >> Is this something to worry about, or just pilot error cause I am not\n> >> using the same $CFLAGS as for the rest of the build? I originally\n> >> noticed this problem with my external fork of pg_verify_checksums.\n>\n> > It looks like the code is using aliasing where the standard says it\n> should\n> > not, which breaks compiler optimization, and the added options tells the\n> > compiler to not assume that the code conforms to the standard...\n>\n> Actually, this code is just plain broken:\n>\n> char buf[BLCKSZ];\n> PageHeader header = (PageHeader) buf;\n>\n> There is no guarantee that a local char[] array is aligned on anything\n> stronger than a byte boundary, and if it isn't, word-sized accesses to\n> *header are going to fail on machines with strict alignment rules.\n> I suppose that that is really what Michael's compiler is moaning about.\n>\n> I rather suspect that this hasn't been tested on anything but Intel\n> hardware, which is famously misalignment-tolerant. The lack of any\n> apparent regression test infrastructure for it isn't leaving a warm\n> feeling about how much the buildfarm is testing it.\n>\n\nI know I certainly didn't test it on non-intel.\n\nWe did have that in the online verify checksum patch, but it came out as\npart of the revert of that part.\n\nGiven that we also recently found bugs in the actual hash index\nimplementation that would've gotten caught by this, perhaps we should add a\nTAP test for this.\n\nShould we make it a separate test in pg_verify_checksums, or should we\npiggyback on the pg_basebackup tests (which AFAICT is the only ones that\ncreate a cluster with checksums enabled at all, and thus is the only\ncodepath that actually uses the backend checksum code at all, which I think\nis an even worse thing to have no tests of)\n\n\n(The right fix, of course, is to malloc the work buffer rather than\n> put it on the stack.)\n\n\nSo if I get you right, you're saying the attached patch should be all\nthat's needed?\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/ <http://www.hagander.net/>\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>","msg_date":"Thu, 30 Aug 2018 21:35:33 +0200","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"On Thu, Aug 30, 2018 at 9:35 PM, Magnus Hagander <magnus@hagander.net>\nwrote:\n\n> On Thu, Aug 30, 2018 at 4:39 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n>> Fabien COELHO <coelho@cri.ensmp.fr> writes:\n>> >> If I add -fno-strict-aliasing to $CFLAGS, the problem goes away.\n>> >> Is this something to worry about, or just pilot error cause I am not\n>> >> using the same $CFLAGS as for the rest of the build? I originally\n>> >> noticed this problem with my external fork of pg_verify_checksums.\n>>\n>> > It looks like the code is using aliasing where the standard says it\n>> should\n>> > not, which breaks compiler optimization, and the added options tells\n>> the\n>> > compiler to not assume that the code conforms to the standard...\n>>\n>> Actually, this code is just plain broken:\n>>\n>> char buf[BLCKSZ];\n>> PageHeader header = (PageHeader) buf;\n>>\n>> There is no guarantee that a local char[] array is aligned on anything\n>> stronger than a byte boundary, and if it isn't, word-sized accesses to\n>> *header are going to fail on machines with strict alignment rules.\n>> I suppose that that is really what Michael's compiler is moaning about.\n>>\n>> I rather suspect that this hasn't been tested on anything but Intel\n>> hardware, which is famously misalignment-tolerant. The lack of any\n>> apparent regression test infrastructure for it isn't leaving a warm\n>> feeling about how much the buildfarm is testing it.\n>>\n>\n> I know I certainly didn't test it on non-intel.\n>\n> We did have that in the online verify checksum patch, but it came out as\n> part of the revert of that part.\n>\n> Given that we also recently found bugs in the actual hash index\n> implementation that would've gotten caught by this, perhaps we should add a\n> TAP test for this.\n>\n> Should we make it a separate test in pg_verify_checksums, or should we\n> piggyback on the pg_basebackup tests (which AFAICT is the only ones that\n> create a cluster with checksums enabled at all, and thus is the only\n> codepath that actually uses the backend checksum code at all, which I think\n> is an even worse thing to have no tests of)\n>\n\nPFA some *very* basic tests for pg_verify_checksums, which should at least\nbe enough to catch the kind of errors we had now in the tool itself.\n\nDoes not at this point try to solve the fact that the backend checksum code\nis not tested.\n\n\n(The right fix, of course, is to malloc the work buffer rather than\n>> put it on the stack.)\n>\n>\n> So if I get you right, you're saying the attached patch should be all\n> that's needed?\n>\n>\n-- \n Magnus Hagander\n Me: https://www.hagander.net/ <http://www.hagander.net/>\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>","msg_date":"Thu, 30 Aug 2018 22:02:13 +0200","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"On Thu, Aug 30, 2018 at 09:35:33PM +0200, Magnus Hagander wrote:\n> Should we make it a separate test in pg_verify_checksums, or should we\n> piggyback on the pg_basebackup tests (which AFAICT is the only ones that\n> create a cluster with checksums enabled at all, and thus is the only\n> codepath that actually uses the backend checksum code at all, which I think\n> is an even worse thing to have no tests of)\n\nThis should be a separate suite. And FWIW, we only use pg_regress to\nmake sure that an already-initialized data folder has the correct,\nsecure authentication set. So creating a node with checksums enabled is\njust that:\n$node->init(extra => ['--data-checksums']);\n\n[... 20 minutes later ...]\nAttached is a basic test suite ;)\n--\nMichael","msg_date":"Thu, 30 Aug 2018 13:02:58 -0700","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"On Thu, Aug 30, 2018 at 10:02 PM, Michael Paquier <michael@paquier.xyz>\nwrote:\n\n> On Thu, Aug 30, 2018 at 09:35:33PM +0200, Magnus Hagander wrote:\n> > Should we make it a separate test in pg_verify_checksums, or should we\n> > piggyback on the pg_basebackup tests (which AFAICT is the only ones that\n> > create a cluster with checksums enabled at all, and thus is the only\n> > codepath that actually uses the backend checksum code at all, which I\n> think\n> > is an even worse thing to have no tests of)\n>\n> This should be a separate suite. And FWIW, we only use pg_regress to\n> make sure that an already-initialized data folder has the correct,\n> secure authentication set. So creating a node with checksums enabled is\n> just that:\n> $node->init(extra => ['--data-checksums']);\n>\n> [... 20 minutes later ...]\n> Attached is a basic test suite ;)\n>\n\nHaha, nice timing :)\n\nI wonder if your tests that pg_control has picked things up belong more in\nthe tests of initdb itself?\n\nDo you think there is value in testing against a non-checksum cluster? I\nguess there's some point to it. I think testing actual corruption (like my\nversion of the tests) is more valuable, but perhaps we should just do both?\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/ <http://www.hagander.net/>\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>\n\nOn Thu, Aug 30, 2018 at 10:02 PM, Michael Paquier <michael@paquier.xyz> wrote:On Thu, Aug 30, 2018 at 09:35:33PM +0200, Magnus Hagander wrote:\n> Should we make it a separate test in pg_verify_checksums, or should we\n> piggyback on the pg_basebackup tests (which AFAICT is the only ones that\n> create a cluster with checksums enabled at all, and thus is the only\n> codepath that actually uses the backend checksum code at all, which I think\n> is an even worse thing to have no tests of)\n\nThis should be a separate suite. And FWIW, we only use pg_regress to\nmake sure that an already-initialized data folder has the correct,\nsecure authentication set. So creating a node with checksums enabled is\njust that:\n$node->init(extra => ['--data-checksums']);\n\n[... 20 minutes later ...]\nAttached is a basic test suite ;)Haha, nice timing :) I wonder if your tests that pg_control has picked things up belong more in the tests of initdb itself?Do you think there is value in testing against a non-checksum cluster? I guess there's some point to it. I think testing actual corruption (like my version of the tests) is more valuable, but perhaps we should just do both?-- Magnus Hagander Me: https://www.hagander.net/ Work: https://www.redpill-linpro.com/","msg_date":"Thu, 30 Aug 2018 22:07:38 +0200","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"On 2018-08-30 10:39:26 -0400, Tom Lane wrote:\n> \tchar\t\tbuf[BLCKSZ];\n> \tPageHeader\theader = (PageHeader) buf;\n\n> (The right fix, of course, is to malloc the work buffer rather than\n> put it on the stack.)\n\nOr alternatively, for places where such allocations could be a problem\nfor performance, using a union can be used to force the required\nalignment. I'm fairly certain that just allocating is more than OK\nhere, to be clear.\n\nGreetings,\n\nAndres Freund\n\n","msg_date":"Thu, 30 Aug 2018 13:11:45 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"Hi,\n\nAm Donnerstag, den 30.08.2018, 22:02 +0200 schrieb Magnus Hagander:\n> PFA some *very* basic tests for pg_verify_checksums, which should at\n> least be enough to catch the kind of errors we had now in the tool\n> itself.\n\nI proposed something similar for pg_basebackup back then and IIRC Peter\n(rightfully) complained that using a system catalog for that wasn't so\ngreat - also, are you sure the relfilenode will always be stable? Then\nagain, this is obviously a throw-away data directory so it should be no\ngeneral issue, just a matter of taste.\n\nhttps://github.com/credativ/pg_checksums/blob/master/t/001_checksums.pl\nhas some similar tests at the end but creates a table for corruption,\nadds some safeguards for block size and also uses command_checks_all()\nto parse the output. It's PGDG licensed so you're welcome to take some\nor all of that.\n\n\nMichael\n\n-- \nMichael Banck\nProjektleiter / Senior Berater\nTel.: +49 2166 9901-171\nFax: +49 2166 9901-100\nEmail: michael.banck@credativ.de\n\ncredativ GmbH, HRB Mönchengladbach 12080\nUSt-ID-Nummer: DE204566209\nTrompeterallee 108, 41189 Mönchengladbach\nGeschäftsführung: Dr. Michael Meskes, Jörg Folz, Sascha Heuer\n\nUnser Umgang mit personenbezogenen Daten unterliegt\nfolgenden Bestimmungen: https://www.credativ.de/datenschutz\n\n","msg_date":"Thu, 30 Aug 2018 22:22:24 +0200","msg_from":"Michael Banck <michael.banck@credativ.de>","msg_from_op":true,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"Hi,\n\nAm Donnerstag, den 30.08.2018, 21:35 +0200 schrieb Magnus Hagander:\n> So if I get you right, you're saying the attached patch should be all\n> that's needed? \n\nI tried to do some similar changes but neither what you proposed nor\nwhat I came up with actually fixes the checksum failures, though they\nsilence the warning at -Wall level.\n\nCould well be I'm doing something wrong, so it would be cool if somebody\ncould reproduce this first. In principle, it should be enough to run\n'make clean && make CLFAGS=-O2' in the src/bin/pg_verify_checksums\nsubdirectory in order to get a broken executable.\n\n\nMichael\n\n-- \nMichael Banck\nProjektleiter / Senior Berater\nTel.: +49 2166 9901-171\nFax: +49 2166 9901-100\nEmail: michael.banck@credativ.de\n\ncredativ GmbH, HRB Mönchengladbach 12080\nUSt-ID-Nummer: DE204566209\nTrompeterallee 108, 41189 Mönchengladbach\nGeschäftsführung: Dr. Michael Meskes, Jörg Folz, Sascha Heuer\n\nUnser Umgang mit personenbezogenen Daten unterliegt\nfolgenden Bestimmungen: https://www.credativ.de/datenschutz\n\n","msg_date":"Thu, 30 Aug 2018 22:28:21 +0200","msg_from":"Michael Banck <michael.banck@credativ.de>","msg_from_op":true,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"Michael Banck <michael.banck@credativ.de> writes:\n> Could well be I'm doing something wrong, so it would be cool if somebody\n> could reproduce this first. In principle, it should be enough to run\n> 'make clean && make CLFAGS=-O2' in the src/bin/pg_verify_checksums\n> subdirectory in order to get a broken executable.\n\nI can reproduce it on my Fedora 28 machine (gcc 8.1.1) by compiling\npg_verify_checksums with all the normal flags except -fno-strict-aliasing.\nI don't see any warning, not even with -Wstrict-aliasing=2 :-(, but the\ngenerated code malfunctions as Michael describes. As best I can tell,\nthe problem is in the inlined code from checksum_impl.h rather than\nin pg_verify_checksums.c itself. I think what is happening is gcc\nis concluding that it can optimize away the code that temporarily\nsets the page's checksum field to zero.\n\nAlthough, as Alvaro mentioned upthread, there's basically no hope of\ngetting away from -fno-strict-aliasing for Postgres-at-large, I think it'd\nbe a good thing if we could get checksum_impl.h to not depend on it. The\nwhole point of that header, according to its own self-description, is to\nexport the checksum calculation for use by external programs --- and we\ncan't really control what compiler flags will be used by those. The fact\nthat Michael even ran into this problem seems to be an instance of that.\n\nSo, I've been fooling around trying to get it to work without\n-fno-strict-aliasing, but with little luck so far.\n\n\t\t\tregards, tom lane\n\n","msg_date":"Thu, 30 Aug 2018 17:19:28 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"Hi,\n\nOn 2018-08-30 17:19:28 -0400, Tom Lane wrote:\n> So, I've been fooling around trying to get it to work without\n> -fno-strict-aliasing, but with little luck so far.\n\nThe problem presumably is that pg_checksum_block() accesses the relevant\nfields as an uint32, whereas pg_checksum_page() accesses it as a\nPageHeader. That's an aliasing violation. *One* cast from char* to\neither type is fine, it's accessing under both those types that's\nproblematic.\n\nOne way to fix it would be to memcpy in/out the modified PageHeader, or\njust do offset math and memcpy to that offset.\n\nGreetings,\n\nAndres Freund\n\n","msg_date":"Thu, 30 Aug 2018 14:46:06 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"Magnus Hagander <magnus@hagander.net> writes:\n> On Thu, Aug 30, 2018 at 4:39 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> (The right fix, of course, is to malloc the work buffer rather than\n>> put it on the stack.)\n\n> So if I get you right, you're saying the attached patch should be all\n> that's needed?\n\nWell, that's some of what's needed. I notice this code is also being\nsloppy about whether block numbers are signed or unsigned, which means\nit'd probably misbehave on relations exceeding 2^31 blocks. I have\na patch in progress to clean all that up, though I'm still working\non the strict-aliasing part.\n\n\t\t\tregards, tom lane\n\n","msg_date":"Thu, 30 Aug 2018 17:49:04 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"On 2018-08-30 14:46:06 -0700, Andres Freund wrote:\n> Hi,\n> \n> On 2018-08-30 17:19:28 -0400, Tom Lane wrote:\n> > So, I've been fooling around trying to get it to work without\n> > -fno-strict-aliasing, but with little luck so far.\n> \n> The problem presumably is that pg_checksum_block() accesses the relevant\n> fields as an uint32, whereas pg_checksum_page() accesses it as a\n> PageHeader. That's an aliasing violation. *One* cast from char* to\n> either type is fine, it's accessing under both those types that's\n> problematic.\n> \n> One way to fix it would be to memcpy in/out the modified PageHeader, or\n> just do offset math and memcpy to that offset.\n\nIt took me a bit to reproduce the issue (due to sheer stupidity on my\npart: no, changing the flags passed to gcc to link pg_verify_checksums\ndoesn't do the trick), but the above indeed fixes the issue for me.\n\nThe attached is just for demonstration that the approach works.\n\nGreetings,\n\nAndres Freund","msg_date":"Thu, 30 Aug 2018 15:04:09 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> On 2018-08-30 14:46:06 -0700, Andres Freund wrote:\n>> One way to fix it would be to memcpy in/out the modified PageHeader, or\n>> just do offset math and memcpy to that offset.\n\n> It took me a bit to reproduce the issue (due to sheer stupidity on my\n> part: no, changing the flags passed to gcc to link pg_verify_checksums\n> doesn't do the trick), but the above indeed fixes the issue for me.\n\nI suspect people will complain about the added cost of doing that.\n\nI've been AFK all afternoon, but what I was intending to try next was\nthe union approach, specifically union'ing PageHeaderData with the uint32\narray representation needed by pg_checksum_block(). That might also\nend up giving us code less unreadable than this:\n\n\tuint32\t\t(*dataArr)[N_SUMS] = (uint32 (*)[N_SUMS]) data;\n\n\nBTW, not to mention the elephant in the room, but: is it *really* OK\nthat pg_checksum_page scribbles on the page buffer, even temporarily?\nIt's certainly quite horrid that there aren't large warnings about\nthat in the function's API comment.\n\n\t\t\tregards, tom lane\n\n","msg_date":"Thu, 30 Aug 2018 18:11:40 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"On 2018-08-30 18:11:40 -0400, Tom Lane wrote:\n> Andres Freund <andres@anarazel.de> writes:\n> > On 2018-08-30 14:46:06 -0700, Andres Freund wrote:\n> >> One way to fix it would be to memcpy in/out the modified PageHeader, or\n> >> just do offset math and memcpy to that offset.\n> \n> > It took me a bit to reproduce the issue (due to sheer stupidity on my\n> > part: no, changing the flags passed to gcc to link pg_verify_checksums\n> > doesn't do the trick), but the above indeed fixes the issue for me.\n> \n> I suspect people will complain about the added cost of doing that.\n\nI think the compiler will just optimize it away. But if we're concerned\nwe could write it as\n\n\tmemcpy(&save_checksum, page + offsetof(PageHeaderData, pd_checksum), sizeof(save_checksum));\n\tmemset(page + offsetof(PageHeaderData, pd_checksum), 0, sizeof(save_checksum));\n\n\tchecksum = pg_checksum_block(page, BLCKSZ);\n\n\tmemcpy(page + offsetof(PageHeaderData, pd_checksum), &save_checksum, sizeof(save_checksum));\n\nworks, but still not exceedingly pretty :/. The code generated looks\nreasonable:\n\n194\t\tmemcpy(&save_checksum, page + offsetof(PageHeaderData, pd_checksum), sizeof(save_checksum));\n 0x00000000000035d0 <+0>:\tpush %r12\n 0x00000000000035d2 <+2>:\txor %eax,%eax\n 0x00000000000035d4 <+4>:\tmovzwl 0x8(%rdi),%r12d\n\n195\t\tmemset(page + offsetof(PageHeaderData, pd_checksum), 0, sizeof(save_checksum));\n 0x00000000000035d9 <+9>:\tpush %rbp\n 0x00000000000035da <+10>:\tmov %ax,0x8(%rdi)\n\n(the pushes are just interspersed stuff, yay latency aware instruction\nscheduling)\n\n\n> I've been AFK all afternoon, but what I was intending to try next was\n> the union approach, specifically union'ing PageHeaderData with the uint32\n> array representation needed by pg_checksum_block(). That might also\n> end up giving us code less unreadable than this:\n> \n> \tuint32\t\t(*dataArr)[N_SUMS] = (uint32 (*)[N_SUMS]) data;\n\nHm.\n\n\n> BTW, not to mention the elephant in the room, but: is it *really* OK\n> that pg_checksum_page scribbles on the page buffer, even temporarily?\n> It's certainly quite horrid that there aren't large warnings about\n> that in the function's API comment.\n\nIt certainly should be warned about. Practically I don't think it's a\nproblem, because we pretty much always operate on a copy of the page\nwhen writing out, as otherwise concurrently set hint bits would be\ntroublesome.\n\nGreetings,\n\nAndres Freund\n\n","msg_date":"Thu, 30 Aug 2018 15:33:15 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"On Thu, Aug 30, 2018 at 10:07:38PM +0200, Magnus Hagander wrote:\n> I wonder if your tests that pg_control has picked things up belong more in\n> the tests of initdb itself?\n\nFor the case where checksums are disabled, moving there the check on\ncontrol data makes sense.\n\n> Do you think there is value in testing against a non-checksum cluster? I\n> guess there's some point to it. I think testing actual corruption (like my\n> version of the tests) is more valuable, but perhaps we should just do both?\n\nYeah, let's do stuff on a single cluster which has them only enabled,\nas initializing a node is one of the most costly operations in TAP\ntests. Checking that the server is stopped is definitely a must in my\nopinion, and your addition about emulating corrupted blocks is a good\nidea. I would personally vote for keeping a control file check within\nthe tests of pg_verify_checksums as that's cheap.\n--\nMichael","msg_date":"Thu, 30 Aug 2018 15:44:58 -0700","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> On 2018-08-30 18:11:40 -0400, Tom Lane wrote:\n>> I suspect people will complain about the added cost of doing that.\n\n> I think the compiler will just optimize it away.\n\nMaybe. In any case, the attached version avoids any need for memcpy\nand is, I think, cleaner code anyhow. It fixes the problem for me\nwith Fedora's gcc, though I'm not sure that it'd be enough to get rid\nof the warning Michael sees (I wonder what compiler he's using).\n\n>> BTW, not to mention the elephant in the room, but: is it *really* OK\n>> that pg_checksum_page scribbles on the page buffer, even temporarily?\n>> It's certainly quite horrid that there aren't large warnings about\n>> that in the function's API comment.\n\n> It certainly should be warned about. Practically I don't think it's a\n> problem, because we pretty much always operate on a copy of the page\n> when writing out, as otherwise concurrently set hint bits would be\n> troublesome.\n\nThe write end of things is not a problem IMO, since presumably the caller\nwould be about to overwrite the checksum field anyway. The issue is for\nreading and/or verifying, where it's much less obvious that clobbering\nthe page image is safe.\n\n\t\t\tregards, tom lane","msg_date":"Thu, 30 Aug 2018 19:02:15 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"On 2018-08-30 19:02:15 -0400, Tom Lane wrote:\n> Andres Freund <andres@anarazel.de> writes:\n> > It certainly should be warned about. Practically I don't think it's a\n> > problem, because we pretty much always operate on a copy of the page\n> > when writing out, as otherwise concurrently set hint bits would be\n> > troublesome.\n> \n> The write end of things is not a problem IMO, since presumably the caller\n> would be about to overwrite the checksum field anyway. The issue is for\n> reading and/or verifying, where it's much less obvious that clobbering\n> the page image is safe.\n\nWith \"reading\" you mean putting the page into the buffercache? If so,\nthat should be ok, the page isn't yet accessible (BM_VALID isn't set).\n\nI don't think it's possible to do verification from the page in s_b,\nbecause we don't keep the checksum current, so there should never be\ncalls to do so.\n\nBut we probably should still add a comment making clear that the\nfunction modifies the buffer temporarily.\n\nGreetings,\n\nAndres Freund\n\n","msg_date":"Thu, 30 Aug 2018 16:12:14 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> On Thu, Aug 30, 2018 at 10:39:26AM -0400, Tom Lane wrote:\n>> (The right fix, of course, is to malloc the work buffer rather than\n>> put it on the stack.)\n\n> pg_upgrade/file.c is careful about that (5afcd2a), and has a comment on\n> the matter, as does pg_standby.c.\n\n> Now, grepping around for \"BLCKSZ]\", some garbage may have accumulated?\n\nAh, that's a cute idea for finding trouble spots.\n\n> - entrySplitPage in ginentrypage.c\n> - rewind_copy_file_range in file.c\n> - _hash_alloc_buckets in hashpage.c\n> - pg_prewarm.c\n> - blinsert.c\n> - pg_waldump.c\n> - logtape.c\n\nSome of these are safe, I think, because the buffers are only used as\ntargets for read() and write(). But some are definitely broken.\nMy own list of files that seem to have issues is\n\nblinsert.c\ngeneric_xlog.c\nginentrypage.c\nhashpage.c\npg_verify_checksums.c\npg_waldump.c\nxloginsert.c\n\nThe fact that some of these are pretty old and we've not noticed is\nnot good. It suggests that we don't currently have any compilers in the\nbuildfarm that under-align char[] arrays on the stack, which seems like\na gotcha waiting to bite us. I wonder if there is any way to persuade\nsome compiler on a non-Intel box to do that.\n\nAnyway, I'll work on a patch for that, unless you were on it already?\n\n\t\t\tregards, tom lane\n\n","msg_date":"Thu, 30 Aug 2018 19:37:37 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"On Thu, Aug 30, 2018 at 07:37:37PM -0400, Tom Lane wrote:\n> Some of these are safe, I think, because the buffers are only used as\n> targets for read() and write(). But some are definitely broken.\n\nYes, I have not spent more than a couple of minutes on this issue. I\nnoticed some of them easily though.\n\n> My own list of files that seem to have issues is\n> \n> blinsert.c\n> generic_xlog.c\n> ginentrypage.c\n> hashpage.c\n> pg_verify_checksums.c\n> pg_waldump.c\n> xloginsert.c\n> \n> The fact that some of these are pretty old and we've not noticed is\n> not good. It suggests that we don't currently have any compilers in the\n> buildfarm that under-align char[] arrays on the stack, which seems like\n> a gotcha waiting to bite us. I wonder if there is any way to persuade\n> some compiler on a non-Intel box to do that.\n\nAgreed, that's annoying.\n\n> Anyway, I'll work on a patch for that, unless you were on it already?\n\nI have begun working on that and am less than halfway through it as I\nneeded a fresh problem, however I am not sure I would be able to finish\nit today, perhaps tomorrow... If you have time and want to press on as\n11 would release soon, of course feel free to wrap up more quickly than\nI can.\n--\nMichael","msg_date":"Thu, 30 Aug 2018 17:06:00 -0700","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"Hi,\n\nAm Donnerstag, den 30.08.2018, 19:02 -0400 schrieb Tom Lane:\n> Andres Freund <andres@anarazel.de> writes:\n> > On 2018-08-30 18:11:40 -0400, Tom Lane wrote:\n> > > I suspect people will complain about the added cost of doing that.\n> > I think the compiler will just optimize it away.\n> \n> Maybe. In any case, the attached version avoids any need for memcpy\n> and is, I think, cleaner code anyhow. It fixes the problem for me\n> with Fedora's gcc, though I'm not sure that it'd be enough to get rid\n> of the warning Michael sees (I wonder what compiler he's using).\n\nThis fixes the bogus checksums, thanks!\n\nI am on Debian stable, i.e. 'gcc version 6.3.0 20170516 (Debian 6.3.0-\n18+deb9u1)'. \n\nThe warning shows up only if I add -Wall *and* -O2 to CFLAGS, I think I\ndid not mention that explictly before.\n\nAs it is in pg_verify_checksums.c I need Magnus' patch as well to make\nit go away. But the warning is independent of the runtime issue so less\nof an issue (but probably worth fixing).\n\n\nMichael \n\n-- \nMichael Banck\nProjektleiter / Senior Berater\nTel.: +49 2166 9901-171\nFax: +49 2166 9901-100\nEmail: michael.banck@credativ.de\n\ncredativ GmbH, HRB Mönchengladbach 12080\nUSt-ID-Nummer: DE204566209\nTrompeterallee 108, 41189 Mönchengladbach\nGeschäftsführung: Dr. Michael Meskes, Jörg Folz, Sascha Heuer\n\nUnser Umgang mit personenbezogenen Daten unterliegt\nfolgenden Bestimmungen: https://www.credativ.de/datenschutz\n\n","msg_date":"Fri, 31 Aug 2018 10:11:07 +0200","msg_from":"Michael Banck <michael.banck@credativ.de>","msg_from_op":true,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"On 31/08/2018 01:37, Tom Lane wrote:\n> The fact that some of these are pretty old and we've not noticed is\n> not good. It suggests that we don't currently have any compilers in the\n> buildfarm that under-align char[] arrays on the stack, which seems like\n> a gotcha waiting to bite us. I wonder if there is any way to persuade\n> some compiler on a non-Intel box to do that.\n\n-Wcast-align=strict (requires gcc-8) warns about the issue in\npg_verify_checksums.c, but also about a handful^W^W15211 other things,\nso it wouldn't be immediately useful.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n","msg_date":"Fri, 31 Aug 2018 13:10:08 +0200","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> On Thu, Aug 30, 2018 at 07:37:37PM -0400, Tom Lane wrote:\n>> Anyway, I'll work on a patch for that, unless you were on it already?\n\n> I have begun working on that and am less than halfway through it as I\n> needed a fresh problem, however I am not sure I would be able to finish\n> it today, perhaps tomorrow... If you have time and want to press on as\n> 11 would release soon, of course feel free to wrap up more quickly than\n> I can.\n\nSome of these places might be performance-critical enough that adding\na palloc/pfree cycle would not be nice. What I was considering doing\nwas inventing something like\n\ntypedef union PGAlignedBuffer\n{\n\tchar\tdata[BLCKSZ];\n\tdouble\tforce_align;\n} PGAlignedBuffer;\n\nand replacing plain char[] variables with that as appropriate.\nWe might need one for XLOG_BLCKSZ too.\n\nSince this'd be used in frontend as well as backend, it'd likely\nhave to end up in c.h, which is slightly annoying but not really\na big deal as long as we pick a nonconflicting type name.\n\n\t\t\tregards, tom lane\n\n","msg_date":"Fri, 31 Aug 2018 10:55:06 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"Michael Banck <michael.banck@credativ.de> writes:\n> Am Donnerstag, den 30.08.2018, 19:02 -0400 schrieb Tom Lane:\n>> Maybe. In any case, the attached version avoids any need for memcpy\n>> and is, I think, cleaner code anyhow. It fixes the problem for me\n>> with Fedora's gcc, though I'm not sure that it'd be enough to get rid\n>> of the warning Michael sees (I wonder what compiler he's using).\n\n> This fixes the bogus checksums, thanks!\n> I am on Debian stable, i.e. 'gcc version 6.3.0 20170516 (Debian 6.3.0-\n> 18+deb9u1)'.\n\nAh-hah. I still can't replicate any warning with gcc 8.1.1, but\nI do have a Raspbian installation at hand, with\ngcc version 6.3.0 20170516 (Raspbian 6.3.0-18+rpi1+deb9u1) \nand on that I get\n\n$ gcc -Wall -Wmissing-prototypes -Wpointer-arith -Wdeclaration-after-statement -Werror=vla -Wendif-labels -Wmissing-format-attribute -Wformat-security -fwrapv -fexcess-precision=standard -g -O2 -I../../../src/include -D_GNU_SOURCE -c -o pg_verify_checksums.o pg_verify_checksums.c\npg_verify_checksums.c: In function 'scan_file':\npg_verify_checksums.c:112:3: warning: dereferencing type-punned pointer will break strict-aliasing rules [-Wstrict-aliasing]\n if (PageIsNew(buf))\n ^~\n\nAdding -Wstrict-aliasing=2 makes it slightly more verbose:\n\n$ gcc -Wstrict-aliasing=2 -Wall -Wmissing-prototypes -Wpointer-arith -Wdeclaration-after-statement -Werror=vla -Wendif-labels -Wmissing-format-attribute -Wformat-security -fwrapv -fexcess-precision=standard -g -O2 -I../../../src/include -D_GNU_SOURCE -c -o pg_verify_checksums.o pg_verify_checksums.c\npg_verify_checksums.c: In function 'scan_file':\npg_verify_checksums.c:82:2: warning: dereferencing type-punned pointer will break strict-aliasing rules [-Wstrict-aliasing]\n PageHeader header = (PageHeader) buf;\n ^~~~~~~~~~\npg_verify_checksums.c:112:3: warning: dereferencing type-punned pointer will break strict-aliasing rules [-Wstrict-aliasing]\n if (PageIsNew(buf))\n ^~\n\nCompiling this way also causes pg_verify_checksums to compute wrong\nchecksums. Installing the patched version of checksum_impl.h fixes\nthat, but doesn't change the warnings, meaning they have absolutely\nnothing to do with the actual problem :-(\n\nSo, even aside from the difficulty of getting rid of aliasing-unsafe\ncode in Postgres, the state of the compiler warning technology is\nstill years short of where we could sanely consider dispensing with\n-fno-strict-aliasing in general. Still, as I said before, it'd be\na good idea for checksum_impl.h to not depend on that. I'll go\nahead and push the fix.\n\n\t\t\tregards, tom lane\n\n","msg_date":"Fri, 31 Aug 2018 12:02:04 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"I wrote:\n> Some of these places might be performance-critical enough that adding\n> a palloc/pfree cycle would not be nice. What I was considering doing\n> was inventing something like\n>\n> typedef union PGAlignedBuffer\n> {\n> \tchar\tdata[BLCKSZ];\n> \tdouble\tforce_align;\n> } PGAlignedBuffer;\n\nHere's a proposed patch using that approach.\n\nAlthough some of the places that were using \"char buf[BLCKSZ]\" variables\nweren't doing anything that really requires better alignment, I made\nalmost all of them use PGAlignedBuffer variables anyway, on the grounds\nthat you typically get better memcpy speed for aligned than unaligned\ntransfers.\n\nI also fixed a few places that were using the palloc solution, and\none that was actually doing hand-alignment of the pointer (ugh).\nI didn't try to be thorough about getting rid of such pallocs,\njust fix places that seemed likely to be performance-critical.\n\nThis needs to be back-patched as relevant, of course.\n\n\t\t\tregards, tom lane","msg_date":"Fri, 31 Aug 2018 15:55:51 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"On Fri, Aug 31, 2018 at 03:55:51PM -0400, Tom Lane wrote:\n> I wrote:\n>> Some of these places might be performance-critical enough that adding\n>> a palloc/pfree cycle would not be nice. What I was considering doing\n>> was inventing something like\n>>\n>> typedef union PGAlignedBuffer\n>> {\n>> \tchar\tdata[BLCKSZ];\n>> \tdouble\tforce_align;\n>> } PGAlignedBuffer;\n> \n> Here's a proposed patch using that approach.\n\nThis solution is smarter than using malloc/palloc to enforce alignment.\nI was digging into the GIN and bloom code with this pattern, but except\nif I used TopTransactionContext for a simple approach, which is not\nacceptable by the way, I was finishing with ugly memory contexts\nused... I am still not sure if that was completely correct either, this\nneeded more time ;)\n\n> Although some of the places that were using \"char buf[BLCKSZ]\" variables\n> weren't doing anything that really requires better alignment, I made\n> almost all of them use PGAlignedBuffer variables anyway, on the grounds\n> that you typically get better memcpy speed for aligned than unaligned\n> transfers.\n\nMakes sense.\n\n> I also fixed a few places that were using the palloc solution, and\n> one that was actually doing hand-alignment of the pointer (ugh).\n> I didn't try to be thorough about getting rid of such pallocs,\n> just fix places that seemed likely to be performance-critical.\n\nOkay, for the memo replay_image_masked and master_image_masked\nin xlog.c could make use of the new structure. SetWALSegSize in\npg_standby.c and WriteEmptyXLOG in pg_resetwal.c, and pg_upgrade's\nfile.c could also be patched.\n\nwalmethods.c could also use some static buffers, not worth the\ncomplication perhaps.. \n\n> +typedef union PGAlignedBuffer\n> +{\n> +\tchar\t\tdata[BLCKSZ];\n> +\tdouble\t\tforce_align_d;\n> +\tint64\t\tforce_align_i64;\n> +} PGAlignedBuffer;\n> +\n> +/* Same, but for an XLOG_BLCKSZ-sized buffer */\n> +typedef union PGAlignedXLogBuffer\n> +{\n> +\tchar\t\tdata[XLOG_BLCKSZ];\n> +\tdouble\t\tforce_align_d;\n> +\tint64\t\tforce_align_i64;\n> +} PGAlignedXLogBuffer;\n\nOne complain I have is about the name of those structures. Perhaps\nPGAlignedBlock and PGAlignedXlogBlock make more sense?\n--\nMichael","msg_date":"Fri, 31 Aug 2018 15:21:19 -0700","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> On Fri, Aug 31, 2018 at 03:55:51PM -0400, Tom Lane wrote:\n>> I also fixed a few places that were using the palloc solution, and\n>> one that was actually doing hand-alignment of the pointer (ugh).\n>> I didn't try to be thorough about getting rid of such pallocs,\n>> just fix places that seemed likely to be performance-critical.\n\n> Okay, for the memo replay_image_masked and master_image_masked\n> in xlog.c could make use of the new structure. SetWALSegSize in\n> pg_standby.c and WriteEmptyXLOG in pg_resetwal.c, and pg_upgrade's\n> file.c could also be patched.\n\nI intentionally didn't change replay_image_masked/master_image_masked\nto use statically-allocated buffers. Since, AFAICS, those aren't\nneeded in most backend processes, they'd just be eating 16KB of\nper-process data space to no purpose.\n\nThe others you mention could be changed, probably, but I didn't\nbother as they didn't seem performance-critical.\n\n>> +typedef union PGAlignedBuffer\n\n> One complain I have is about the name of those structures. Perhaps\n> PGAlignedBlock and PGAlignedXlogBlock make more sense?\n\nDon't have a strong preference, anybody else have an opinion?\n\n(I also wondered whether to use \"WAL\" instead of \"XLog\" in that\nstruct name, but it seems like we've mostly stuck with \"xlog\"\nin internal C names.)\n\n\t\t\tregards, tom lane\n\n","msg_date":"Fri, 31 Aug 2018 19:59:58 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"On Fri, Aug 31, 2018 at 07:59:58PM -0400, Tom Lane wrote:\n> The others you mention could be changed, probably, but I didn't\n> bother as they didn't seem performance-critical.\n\nIt is not really critical indeed. There is an argument to change them\nso as other folks get used to it though.\n\n> (I also wondered whether to use \"WAL\" instead of \"XLog\" in that\n> struct name, but it seems like we've mostly stuck with \"xlog\"\n> in internal C names.)\n\nXLOG_BLCKSZ is used, which makes me think that XLog is better than WAL\nhere. A matter of taste of course.\n--\nMichael","msg_date":"Fri, 31 Aug 2018 17:33:03 -0700","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"\nHello,\n\n>> Okay, for the memo replay_image_masked and master_image_masked\n>> in xlog.c could make use of the new structure. SetWALSegSize in\n>> pg_standby.c and WriteEmptyXLOG in pg_resetwal.c, and pg_upgrade's\n>> file.c could also be patched.\n>\n> I intentionally didn't change replay_image_masked/master_image_masked\n> to use statically-allocated buffers. Since, AFAICS, those aren't\n> needed in most backend processes, they'd just be eating 16KB of\n> per-process data space to no purpose.\n>\n> The others you mention could be changed, probably, but I didn't\n> bother as they didn't seem performance-critical.\n\nI'd go for having just one same approach everywhere, for code base \nhomogeneity.\n\n>>> +typedef union PGAlignedBuffer\n>\n>> One complain I have is about the name of those structures. Perhaps\n>> PGAlignedBlock and PGAlignedXlogBlock make more sense?\n>\n> Don't have a strong preference, anybody else have an opinion?\n\nI like \"Block\" better, because it's more precise.\n\n> (I also wondered whether to use \"WAL\" instead of \"XLog\" in that\n> struct name, but it seems like we've mostly stuck with \"xlog\"\n> in internal C names.)\n\nBest to blend with the surrounding code in the header file?\n\n-- \nFabien.\n\n","msg_date":"Sat, 1 Sep 2018 07:24:38 +0200 (CEST)","msg_from":"Fabien COELHO <coelho@cri.ensmp.fr>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> On Fri, Aug 31, 2018 at 07:59:58PM -0400, Tom Lane wrote:\n>> The others you mention could be changed, probably, but I didn't\n>> bother as they didn't seem performance-critical.\n\n> It is not really critical indeed. There is an argument to change them\n> so as other folks get used to it though.\n\nFair enough. I renamed the types as suggested, changed a few more\nplaces for consistency's sake, and pushed.\n\nThere still remain some places where palloc(BLCKSZ) or equivalent is used,\nbut there's no matching pfree. In a lot of them the buffer is returned\nto the caller so there's no choice. It's likely that some are just\nleaking the storage transiently and we could convert them to using a\nPGAlignedBlock local variable, but I didn't bother trying to do the\nanalysis.\n\n\t\t\tregards, tom lane\n\n","msg_date":"Sat, 01 Sep 2018 15:32:10 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"},{"msg_contents":"On Sat, Sep 01, 2018 at 03:32:10PM -0400, Tom Lane wrote:\n> Fair enough. I renamed the types as suggested, changed a few more\n> places for consistency's sake, and pushed.\n\nThanks!\n\n> There still remain some places where palloc(BLCKSZ) or equivalent is used,\n> but there's no matching pfree. In a lot of them the buffer is returned\n> to the caller so there's no choice. It's likely that some are just\n> leaking the storage transiently and we could convert them to using a\n> PGAlignedBlock local variable, but I didn't bother trying to do the\n> analysis.\n\nAt quick glance, I am not seeing anything critical. So the result looks\ngood to me.\n--\nMichael","msg_date":"Sat, 1 Sep 2018 16:19:00 -0700","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_verify_checksums and -fno-strict-aliasing"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nI noticed that pg_verify_checksums computes bogus checksums if I compile\\nit with '-O2 -Wall' but without -fno-strict-aliasing. Also I am getting\\na compile warning then:\\n\\n|src/bin/pg_verify_checksums$ make CFLAGS='-g -O2 -Wall'\\n[...]\\n|gcc -g -O2 -Wall -I../../../src/include \\n| -I/home/mba/Projekte/OSS/PostgreSQL/git/postgresql/build/../src/include\\n| -D_GNU_SOURCE -c -o pg_verify_checksums.o\\n| /home/mba/Projekte/OSS/PostgreSQL/git/postgresql/build/../src/bin/pg_verify_checksums/pg_verify_checksums.c\\n| -MMD -MP -MF .deps/pg_verify_checksums.Po\\n|/home/mba/Projekte/OSS/PostgreSQL/git/postgresql/build/../src/bin/pg_verify_checksums/pg_verify_checksums.c: \\n| In function ‘scan_file’:\\n|/home/mba/Projekte/OSS/PostgreSQL/git/postgresql/build/../src/bin/pg_verify_checksums/pg_verify_checksums.c:112:3: \\n| warning: dereferencing type-punned pointer will break strict-aliasing\\n| rules [-Wstrict-aliasing]\\n| if (PageIsNew(buf))\\n| ^~\\n[...]\\n|src/bin/pg_verify_checksums$ ./pg_verify_checksums -D ../../test/regress/tmp_check/data\\n|pg_verify_checksums: checksum verification failed in file \\\"../../test/regress/tmp_check/data/global/1233\\\", block 0: calculated checksum FC8A but expected F857\\n|pg_verify_checksums: checksum verification failed in file \\\"../../test/regress/tmp_check/data/global/1233\\\", block 1: calculated checksum F340 but expected A68D\\n[...]\\n|pg_verify_checksums: checksum verification failed in file \\\"../../test/regress/tmp_check/data/base/12368/2659\\\", block 5: calculated checksum 954D but expected D2ED\\n|pg_verify_checksums: checksum verification failed in file \\\"../../test/regress/tmp_check/data/base/12368/2659\\\", block 6: calculated checksum 361 but expected E7F7\\n|pg_verify_checksums: checksum verification failed in file \\\"../../test/regress/tmp_check/data/base/12368/2659\\\", block 7: calculated checksum 4C0D but expected 4B10\\n|pg_verify_checksums: checksum verification failed in file \\\"../../test/regress/tmp_check/data/base/12368/2659\\\", block 8: calculated checksum 2B9A but expected 71E3\\n|pg_verify_checksums: checksum verification failed in file \\\"../../test/regress/tmp_check/data/base/12368/2659\\\", block 9: calculated checksum 19CD but expected 541C\\n|Checksum scan completed\\n|Data checksum version: 1\\n|Files scanned: 932\\n|Blocks scanned: 2789\\n|Bad checksums: 2789\\n\\nIf I add -fno-strict-aliasing to $CFLAGS, the problem goes away.\\n\\nIs this something to worry about, or just pilot error cause I am not\\nusing the same $CFLAGS as for the rest of the build? I originally\\nnoticed this problem with my external fork of pg_verify_checksums.\\n\\n\\nMichael\\n\\n-- \\nMichael Banck\\nProjektleiter / Senior Berater\\nTel.: +49 2166 9901-171\\nFax: +49 2166 9901-100\\nEmail: michael.banck@credativ.de\\n\\ncredativ GmbH, HRB Mönchengladbach 12080\\nUSt-ID-Nummer: DE204566209\\nTrompeterallee 108, 41189 Mönchengladbach\\nGeschäftsführung: Dr. Michael Meskes, Jörg Folz, Sascha Heuer\\n\\nUnser Umgang mit personenbezogenen Daten unterliegt\\nfolgenden Bestimmungen: https://www.credativ.de/datenschutz\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 10:35:00 +0200\",\n \"msg_from\": \"Michael Banck <michael.banck@credativ.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"\\n> I noticed that pg_verify_checksums computes bogus checksums if I compile\\n> it with '-O2 -Wall' but without -fno-strict-aliasing. Also I am getting\\n> a compile warning then:\\n>\\n> [...]\\n>\\n> If I add -fno-strict-aliasing to $CFLAGS, the problem goes away.\\n>\\n> Is this something to worry about, or just pilot error cause I am not\\n> using the same $CFLAGS as for the rest of the build? I originally\\n> noticed this problem with my external fork of pg_verify_checksums.\\n\\nIt looks like the code is using aliasing where the standard says it should \\nnot, which breaks compiler optimization, and the added options tells the \\ncompiler to not assume that the code conforms to the standard...\\n\\nAs PostgreSQL source is expected to conform to some C standard (unsure \\nwhich one right now, possibly c89 but maybe it is beginning to switch to \\nc99, a young 19 years old standard), I'd suggest that the right fix is \\nrather to actually remove the aliasing issue.\\n\\n-- \\nFabien.\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 13:25:14 +0200 (CEST)\",\n \"msg_from\": \"Fabien COELHO <coelho@cri.ensmp.fr>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"Fabien COELHO <coelho@cri.ensmp.fr> writes:\\n>> If I add -fno-strict-aliasing to $CFLAGS, the problem goes away.\\n>> Is this something to worry about, or just pilot error cause I am not\\n>> using the same $CFLAGS as for the rest of the build? I originally\\n>> noticed this problem with my external fork of pg_verify_checksums.\\n\\n> It looks like the code is using aliasing where the standard says it should \\n> not, which breaks compiler optimization, and the added options tells the \\n> compiler to not assume that the code conforms to the standard...\\n\\nActually, this code is just plain broken:\\n\\n\\tchar\\t\\tbuf[BLCKSZ];\\n\\tPageHeader\\theader = (PageHeader) buf;\\n\\nThere is no guarantee that a local char[] array is aligned on anything\\nstronger than a byte boundary, and if it isn't, word-sized accesses to\\n*header are going to fail on machines with strict alignment rules.\\nI suppose that that is really what Michael's compiler is moaning about.\\n\\nI rather suspect that this hasn't been tested on anything but Intel\\nhardware, which is famously misalignment-tolerant. The lack of any\\napparent regression test infrastructure for it isn't leaving a warm\\nfeeling about how much the buildfarm is testing it.\\n\\n(The right fix, of course, is to malloc the work buffer rather than\\nput it on the stack.)\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 10:39:26 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"On 2018-Aug-30, Fabien COELHO wrote:\\n\\n> As PostgreSQL source is expected to conform to some C standard (unsure which\\n> one right now, possibly c89 but maybe it is beginning to switch to c99, a\\n> young 19 years old standard), I'd suggest that the right fix is rather to\\n> actually remove the aliasing issue.\\n\\nYeah, type aliasing is quite a rampant issue in our code and I don't\\nthink there's an easy fix.\\n\\n-- \\n�lvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 11:45:47 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 30, 2018 at 10:39:26AM -0400, Tom Lane wrote:\\n> I rather suspect that this hasn't been tested on anything but Intel\\n> hardware, which is famously misalignment-tolerant. The lack of any\\n> apparent regression test infrastructure for it isn't leaving a warm\\n> feeling about how much the buildfarm is testing it.\\n> \\n> (The right fix, of course, is to malloc the work buffer rather than\\n> put it on the stack.)\\n\\npg_upgrade/file.c is careful about that (5afcd2a), and has a comment on\\nthe matter, as does pg_standby.c.\\n\\nNow, grepping around for \\\"BLCKSZ]\\\", some garbage may have accumulated?\\n- entrySplitPage in ginentrypage.c\\n- rewind_copy_file_range in file.c\\n- _hash_alloc_buckets in hashpage.c\\n- pg_prewarm.c\\n- blinsert.c\\n- pg_waldump.c\\n- logtape.c\\n--\\nMichael\",\n \"msg_date\": \"Thu, 30 Aug 2018 10:56:52 -0700\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 30, 2018 at 4:39 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n\\n> Fabien COELHO <coelho@cri.ensmp.fr> writes:\\n> >> If I add -fno-strict-aliasing to $CFLAGS, the problem goes away.\\n> >> Is this something to worry about, or just pilot error cause I am not\\n> >> using the same $CFLAGS as for the rest of the build? I originally\\n> >> noticed this problem with my external fork of pg_verify_checksums.\\n>\\n> > It looks like the code is using aliasing where the standard says it\\n> should\\n> > not, which breaks compiler optimization, and the added options tells the\\n> > compiler to not assume that the code conforms to the standard...\\n>\\n> Actually, this code is just plain broken:\\n>\\n> char buf[BLCKSZ];\\n> PageHeader header = (PageHeader) buf;\\n>\\n> There is no guarantee that a local char[] array is aligned on anything\\n> stronger than a byte boundary, and if it isn't, word-sized accesses to\\n> *header are going to fail on machines with strict alignment rules.\\n> I suppose that that is really what Michael's compiler is moaning about.\\n>\\n> I rather suspect that this hasn't been tested on anything but Intel\\n> hardware, which is famously misalignment-tolerant. The lack of any\\n> apparent regression test infrastructure for it isn't leaving a warm\\n> feeling about how much the buildfarm is testing it.\\n>\\n\\nI know I certainly didn't test it on non-intel.\\n\\nWe did have that in the online verify checksum patch, but it came out as\\npart of the revert of that part.\\n\\nGiven that we also recently found bugs in the actual hash index\\nimplementation that would've gotten caught by this, perhaps we should add a\\nTAP test for this.\\n\\nShould we make it a separate test in pg_verify_checksums, or should we\\npiggyback on the pg_basebackup tests (which AFAICT is the only ones that\\ncreate a cluster with checksums enabled at all, and thus is the only\\ncodepath that actually uses the backend checksum code at all, which I think\\nis an even worse thing to have no tests of)\\n\\n\\n(The right fix, of course, is to malloc the work buffer rather than\\n> put it on the stack.)\\n\\n\\nSo if I get you right, you're saying the attached patch should be all\\nthat's needed?\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/ <http://www.hagander.net/>\\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>\",\n \"msg_date\": \"Thu, 30 Aug 2018 21:35:33 +0200\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 30, 2018 at 9:35 PM, Magnus Hagander <magnus@hagander.net>\\nwrote:\\n\\n> On Thu, Aug 30, 2018 at 4:39 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n>> Fabien COELHO <coelho@cri.ensmp.fr> writes:\\n>> >> If I add -fno-strict-aliasing to $CFLAGS, the problem goes away.\\n>> >> Is this something to worry about, or just pilot error cause I am not\\n>> >> using the same $CFLAGS as for the rest of the build? I originally\\n>> >> noticed this problem with my external fork of pg_verify_checksums.\\n>>\\n>> > It looks like the code is using aliasing where the standard says it\\n>> should\\n>> > not, which breaks compiler optimization, and the added options tells\\n>> the\\n>> > compiler to not assume that the code conforms to the standard...\\n>>\\n>> Actually, this code is just plain broken:\\n>>\\n>> char buf[BLCKSZ];\\n>> PageHeader header = (PageHeader) buf;\\n>>\\n>> There is no guarantee that a local char[] array is aligned on anything\\n>> stronger than a byte boundary, and if it isn't, word-sized accesses to\\n>> *header are going to fail on machines with strict alignment rules.\\n>> I suppose that that is really what Michael's compiler is moaning about.\\n>>\\n>> I rather suspect that this hasn't been tested on anything but Intel\\n>> hardware, which is famously misalignment-tolerant. The lack of any\\n>> apparent regression test infrastructure for it isn't leaving a warm\\n>> feeling about how much the buildfarm is testing it.\\n>>\\n>\\n> I know I certainly didn't test it on non-intel.\\n>\\n> We did have that in the online verify checksum patch, but it came out as\\n> part of the revert of that part.\\n>\\n> Given that we also recently found bugs in the actual hash index\\n> implementation that would've gotten caught by this, perhaps we should add a\\n> TAP test for this.\\n>\\n> Should we make it a separate test in pg_verify_checksums, or should we\\n> piggyback on the pg_basebackup tests (which AFAICT is the only ones that\\n> create a cluster with checksums enabled at all, and thus is the only\\n> codepath that actually uses the backend checksum code at all, which I think\\n> is an even worse thing to have no tests of)\\n>\\n\\nPFA some *very* basic tests for pg_verify_checksums, which should at least\\nbe enough to catch the kind of errors we had now in the tool itself.\\n\\nDoes not at this point try to solve the fact that the backend checksum code\\nis not tested.\\n\\n\\n(The right fix, of course, is to malloc the work buffer rather than\\n>> put it on the stack.)\\n>\\n>\\n> So if I get you right, you're saying the attached patch should be all\\n> that's needed?\\n>\\n>\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/ <http://www.hagander.net/>\\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>\",\n \"msg_date\": \"Thu, 30 Aug 2018 22:02:13 +0200\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 30, 2018 at 09:35:33PM +0200, Magnus Hagander wrote:\\n> Should we make it a separate test in pg_verify_checksums, or should we\\n> piggyback on the pg_basebackup tests (which AFAICT is the only ones that\\n> create a cluster with checksums enabled at all, and thus is the only\\n> codepath that actually uses the backend checksum code at all, which I think\\n> is an even worse thing to have no tests of)\\n\\nThis should be a separate suite. And FWIW, we only use pg_regress to\\nmake sure that an already-initialized data folder has the correct,\\nsecure authentication set. So creating a node with checksums enabled is\\njust that:\\n$node->init(extra => ['--data-checksums']);\\n\\n[... 20 minutes later ...]\\nAttached is a basic test suite ;)\\n--\\nMichael\",\n \"msg_date\": \"Thu, 30 Aug 2018 13:02:58 -0700\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 30, 2018 at 10:02 PM, Michael Paquier <michael@paquier.xyz>\\nwrote:\\n\\n> On Thu, Aug 30, 2018 at 09:35:33PM +0200, Magnus Hagander wrote:\\n> > Should we make it a separate test in pg_verify_checksums, or should we\\n> > piggyback on the pg_basebackup tests (which AFAICT is the only ones that\\n> > create a cluster with checksums enabled at all, and thus is the only\\n> > codepath that actually uses the backend checksum code at all, which I\\n> think\\n> > is an even worse thing to have no tests of)\\n>\\n> This should be a separate suite. And FWIW, we only use pg_regress to\\n> make sure that an already-initialized data folder has the correct,\\n> secure authentication set. So creating a node with checksums enabled is\\n> just that:\\n> $node->init(extra => ['--data-checksums']);\\n>\\n> [... 20 minutes later ...]\\n> Attached is a basic test suite ;)\\n>\\n\\nHaha, nice timing :)\\n\\nI wonder if your tests that pg_control has picked things up belong more in\\nthe tests of initdb itself?\\n\\nDo you think there is value in testing against a non-checksum cluster? I\\nguess there's some point to it. I think testing actual corruption (like my\\nversion of the tests) is more valuable, but perhaps we should just do both?\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/ <http://www.hagander.net/>\\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>\\n\\nOn Thu, Aug 30, 2018 at 10:02 PM, Michael Paquier <michael@paquier.xyz> wrote:On Thu, Aug 30, 2018 at 09:35:33PM +0200, Magnus Hagander wrote:\\n> Should we make it a separate test in pg_verify_checksums, or should we\\n> piggyback on the pg_basebackup tests (which AFAICT is the only ones that\\n> create a cluster with checksums enabled at all, and thus is the only\\n> codepath that actually uses the backend checksum code at all, which I think\\n> is an even worse thing to have no tests of)\\n\\nThis should be a separate suite. And FWIW, we only use pg_regress to\\nmake sure that an already-initialized data folder has the correct,\\nsecure authentication set. So creating a node with checksums enabled is\\njust that:\\n$node->init(extra => ['--data-checksums']);\\n\\n[... 20 minutes later ...]\\nAttached is a basic test suite ;)Haha, nice timing :) I wonder if your tests that pg_control has picked things up belong more in the tests of initdb itself?Do you think there is value in testing against a non-checksum cluster? I guess there's some point to it. I think testing actual corruption (like my version of the tests) is more valuable, but perhaps we should just do both?-- Magnus Hagander Me: https://www.hagander.net/ Work: https://www.redpill-linpro.com/\",\n \"msg_date\": \"Thu, 30 Aug 2018 22:07:38 +0200\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"On 2018-08-30 10:39:26 -0400, Tom Lane wrote:\\n> \\tchar\\t\\tbuf[BLCKSZ];\\n> \\tPageHeader\\theader = (PageHeader) buf;\\n\\n> (The right fix, of course, is to malloc the work buffer rather than\\n> put it on the stack.)\\n\\nOr alternatively, for places where such allocations could be a problem\\nfor performance, using a union can be used to force the required\\nalignment. I'm fairly certain that just allocating is more than OK\\nhere, to be clear.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 13:11:45 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nAm Donnerstag, den 30.08.2018, 22:02 +0200 schrieb Magnus Hagander:\\n> PFA some *very* basic tests for pg_verify_checksums, which should at\\n> least be enough to catch the kind of errors we had now in the tool\\n> itself.\\n\\nI proposed something similar for pg_basebackup back then and IIRC Peter\\n(rightfully) complained that using a system catalog for that wasn't so\\ngreat - also, are you sure the relfilenode will always be stable? Then\\nagain, this is obviously a throw-away data directory so it should be no\\ngeneral issue, just a matter of taste.\\n\\nhttps://github.com/credativ/pg_checksums/blob/master/t/001_checksums.pl\\nhas some similar tests at the end but creates a table for corruption,\\nadds some safeguards for block size and also uses command_checks_all()\\nto parse the output. It's PGDG licensed so you're welcome to take some\\nor all of that.\\n\\n\\nMichael\\n\\n-- \\nMichael Banck\\nProjektleiter / Senior Berater\\nTel.: +49 2166 9901-171\\nFax: +49 2166 9901-100\\nEmail: michael.banck@credativ.de\\n\\ncredativ GmbH, HRB Mönchengladbach 12080\\nUSt-ID-Nummer: DE204566209\\nTrompeterallee 108, 41189 Mönchengladbach\\nGeschäftsführung: Dr. Michael Meskes, Jörg Folz, Sascha Heuer\\n\\nUnser Umgang mit personenbezogenen Daten unterliegt\\nfolgenden Bestimmungen: https://www.credativ.de/datenschutz\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 22:22:24 +0200\",\n \"msg_from\": \"Michael Banck <michael.banck@credativ.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nAm Donnerstag, den 30.08.2018, 21:35 +0200 schrieb Magnus Hagander:\\n> So if I get you right, you're saying the attached patch should be all\\n> that's needed? \\n\\nI tried to do some similar changes but neither what you proposed nor\\nwhat I came up with actually fixes the checksum failures, though they\\nsilence the warning at -Wall level.\\n\\nCould well be I'm doing something wrong, so it would be cool if somebody\\ncould reproduce this first. In principle, it should be enough to run\\n'make clean && make CLFAGS=-O2' in the src/bin/pg_verify_checksums\\nsubdirectory in order to get a broken executable.\\n\\n\\nMichael\\n\\n-- \\nMichael Banck\\nProjektleiter / Senior Berater\\nTel.: +49 2166 9901-171\\nFax: +49 2166 9901-100\\nEmail: michael.banck@credativ.de\\n\\ncredativ GmbH, HRB Mönchengladbach 12080\\nUSt-ID-Nummer: DE204566209\\nTrompeterallee 108, 41189 Mönchengladbach\\nGeschäftsführung: Dr. Michael Meskes, Jörg Folz, Sascha Heuer\\n\\nUnser Umgang mit personenbezogenen Daten unterliegt\\nfolgenden Bestimmungen: https://www.credativ.de/datenschutz\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 22:28:21 +0200\",\n \"msg_from\": \"Michael Banck <michael.banck@credativ.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"Michael Banck <michael.banck@credativ.de> writes:\\n> Could well be I'm doing something wrong, so it would be cool if somebody\\n> could reproduce this first. In principle, it should be enough to run\\n> 'make clean && make CLFAGS=-O2' in the src/bin/pg_verify_checksums\\n> subdirectory in order to get a broken executable.\\n\\nI can reproduce it on my Fedora 28 machine (gcc 8.1.1) by compiling\\npg_verify_checksums with all the normal flags except -fno-strict-aliasing.\\nI don't see any warning, not even with -Wstrict-aliasing=2 :-(, but the\\ngenerated code malfunctions as Michael describes. As best I can tell,\\nthe problem is in the inlined code from checksum_impl.h rather than\\nin pg_verify_checksums.c itself. I think what is happening is gcc\\nis concluding that it can optimize away the code that temporarily\\nsets the page's checksum field to zero.\\n\\nAlthough, as Alvaro mentioned upthread, there's basically no hope of\\ngetting away from -fno-strict-aliasing for Postgres-at-large, I think it'd\\nbe a good thing if we could get checksum_impl.h to not depend on it. The\\nwhole point of that header, according to its own self-description, is to\\nexport the checksum calculation for use by external programs --- and we\\ncan't really control what compiler flags will be used by those. The fact\\nthat Michael even ran into this problem seems to be an instance of that.\\n\\nSo, I've been fooling around trying to get it to work without\\n-fno-strict-aliasing, but with little luck so far.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 17:19:28 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2018-08-30 17:19:28 -0400, Tom Lane wrote:\\n> So, I've been fooling around trying to get it to work without\\n> -fno-strict-aliasing, but with little luck so far.\\n\\nThe problem presumably is that pg_checksum_block() accesses the relevant\\nfields as an uint32, whereas pg_checksum_page() accesses it as a\\nPageHeader. That's an aliasing violation. *One* cast from char* to\\neither type is fine, it's accessing under both those types that's\\nproblematic.\\n\\nOne way to fix it would be to memcpy in/out the modified PageHeader, or\\njust do offset math and memcpy to that offset.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 14:46:06 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"Magnus Hagander <magnus@hagander.net> writes:\\n> On Thu, Aug 30, 2018 at 4:39 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> (The right fix, of course, is to malloc the work buffer rather than\\n>> put it on the stack.)\\n\\n> So if I get you right, you're saying the attached patch should be all\\n> that's needed?\\n\\nWell, that's some of what's needed. I notice this code is also being\\nsloppy about whether block numbers are signed or unsigned, which means\\nit'd probably misbehave on relations exceeding 2^31 blocks. I have\\na patch in progress to clean all that up, though I'm still working\\non the strict-aliasing part.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 17:49:04 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"On 2018-08-30 14:46:06 -0700, Andres Freund wrote:\\n> Hi,\\n> \\n> On 2018-08-30 17:19:28 -0400, Tom Lane wrote:\\n> > So, I've been fooling around trying to get it to work without\\n> > -fno-strict-aliasing, but with little luck so far.\\n> \\n> The problem presumably is that pg_checksum_block() accesses the relevant\\n> fields as an uint32, whereas pg_checksum_page() accesses it as a\\n> PageHeader. That's an aliasing violation. *One* cast from char* to\\n> either type is fine, it's accessing under both those types that's\\n> problematic.\\n> \\n> One way to fix it would be to memcpy in/out the modified PageHeader, or\\n> just do offset math and memcpy to that offset.\\n\\nIt took me a bit to reproduce the issue (due to sheer stupidity on my\\npart: no, changing the flags passed to gcc to link pg_verify_checksums\\ndoesn't do the trick), but the above indeed fixes the issue for me.\\n\\nThe attached is just for demonstration that the approach works.\\n\\nGreetings,\\n\\nAndres Freund\",\n \"msg_date\": \"Thu, 30 Aug 2018 15:04:09 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> On 2018-08-30 14:46:06 -0700, Andres Freund wrote:\\n>> One way to fix it would be to memcpy in/out the modified PageHeader, or\\n>> just do offset math and memcpy to that offset.\\n\\n> It took me a bit to reproduce the issue (due to sheer stupidity on my\\n> part: no, changing the flags passed to gcc to link pg_verify_checksums\\n> doesn't do the trick), but the above indeed fixes the issue for me.\\n\\nI suspect people will complain about the added cost of doing that.\\n\\nI've been AFK all afternoon, but what I was intending to try next was\\nthe union approach, specifically union'ing PageHeaderData with the uint32\\narray representation needed by pg_checksum_block(). That might also\\nend up giving us code less unreadable than this:\\n\\n\\tuint32\\t\\t(*dataArr)[N_SUMS] = (uint32 (*)[N_SUMS]) data;\\n\\n\\nBTW, not to mention the elephant in the room, but: is it *really* OK\\nthat pg_checksum_page scribbles on the page buffer, even temporarily?\\nIt's certainly quite horrid that there aren't large warnings about\\nthat in the function's API comment.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 18:11:40 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"On 2018-08-30 18:11:40 -0400, Tom Lane wrote:\\n> Andres Freund <andres@anarazel.de> writes:\\n> > On 2018-08-30 14:46:06 -0700, Andres Freund wrote:\\n> >> One way to fix it would be to memcpy in/out the modified PageHeader, or\\n> >> just do offset math and memcpy to that offset.\\n> \\n> > It took me a bit to reproduce the issue (due to sheer stupidity on my\\n> > part: no, changing the flags passed to gcc to link pg_verify_checksums\\n> > doesn't do the trick), but the above indeed fixes the issue for me.\\n> \\n> I suspect people will complain about the added cost of doing that.\\n\\nI think the compiler will just optimize it away. But if we're concerned\\nwe could write it as\\n\\n\\tmemcpy(&save_checksum, page + offsetof(PageHeaderData, pd_checksum), sizeof(save_checksum));\\n\\tmemset(page + offsetof(PageHeaderData, pd_checksum), 0, sizeof(save_checksum));\\n\\n\\tchecksum = pg_checksum_block(page, BLCKSZ);\\n\\n\\tmemcpy(page + offsetof(PageHeaderData, pd_checksum), &save_checksum, sizeof(save_checksum));\\n\\nworks, but still not exceedingly pretty :/. The code generated looks\\nreasonable:\\n\\n194\\t\\tmemcpy(&save_checksum, page + offsetof(PageHeaderData, pd_checksum), sizeof(save_checksum));\\n 0x00000000000035d0 <+0>:\\tpush %r12\\n 0x00000000000035d2 <+2>:\\txor %eax,%eax\\n 0x00000000000035d4 <+4>:\\tmovzwl 0x8(%rdi),%r12d\\n\\n195\\t\\tmemset(page + offsetof(PageHeaderData, pd_checksum), 0, sizeof(save_checksum));\\n 0x00000000000035d9 <+9>:\\tpush %rbp\\n 0x00000000000035da <+10>:\\tmov %ax,0x8(%rdi)\\n\\n(the pushes are just interspersed stuff, yay latency aware instruction\\nscheduling)\\n\\n\\n> I've been AFK all afternoon, but what I was intending to try next was\\n> the union approach, specifically union'ing PageHeaderData with the uint32\\n> array representation needed by pg_checksum_block(). That might also\\n> end up giving us code less unreadable than this:\\n> \\n> \\tuint32\\t\\t(*dataArr)[N_SUMS] = (uint32 (*)[N_SUMS]) data;\\n\\nHm.\\n\\n\\n> BTW, not to mention the elephant in the room, but: is it *really* OK\\n> that pg_checksum_page scribbles on the page buffer, even temporarily?\\n> It's certainly quite horrid that there aren't large warnings about\\n> that in the function's API comment.\\n\\nIt certainly should be warned about. Practically I don't think it's a\\nproblem, because we pretty much always operate on a copy of the page\\nwhen writing out, as otherwise concurrently set hint bits would be\\ntroublesome.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 15:33:15 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 30, 2018 at 10:07:38PM +0200, Magnus Hagander wrote:\\n> I wonder if your tests that pg_control has picked things up belong more in\\n> the tests of initdb itself?\\n\\nFor the case where checksums are disabled, moving there the check on\\ncontrol data makes sense.\\n\\n> Do you think there is value in testing against a non-checksum cluster? I\\n> guess there's some point to it. I think testing actual corruption (like my\\n> version of the tests) is more valuable, but perhaps we should just do both?\\n\\nYeah, let's do stuff on a single cluster which has them only enabled,\\nas initializing a node is one of the most costly operations in TAP\\ntests. Checking that the server is stopped is definitely a must in my\\nopinion, and your addition about emulating corrupted blocks is a good\\nidea. I would personally vote for keeping a control file check within\\nthe tests of pg_verify_checksums as that's cheap.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 30 Aug 2018 15:44:58 -0700\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> On 2018-08-30 18:11:40 -0400, Tom Lane wrote:\\n>> I suspect people will complain about the added cost of doing that.\\n\\n> I think the compiler will just optimize it away.\\n\\nMaybe. In any case, the attached version avoids any need for memcpy\\nand is, I think, cleaner code anyhow. It fixes the problem for me\\nwith Fedora's gcc, though I'm not sure that it'd be enough to get rid\\nof the warning Michael sees (I wonder what compiler he's using).\\n\\n>> BTW, not to mention the elephant in the room, but: is it *really* OK\\n>> that pg_checksum_page scribbles on the page buffer, even temporarily?\\n>> It's certainly quite horrid that there aren't large warnings about\\n>> that in the function's API comment.\\n\\n> It certainly should be warned about. Practically I don't think it's a\\n> problem, because we pretty much always operate on a copy of the page\\n> when writing out, as otherwise concurrently set hint bits would be\\n> troublesome.\\n\\nThe write end of things is not a problem IMO, since presumably the caller\\nwould be about to overwrite the checksum field anyway. The issue is for\\nreading and/or verifying, where it's much less obvious that clobbering\\nthe page image is safe.\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Thu, 30 Aug 2018 19:02:15 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"On 2018-08-30 19:02:15 -0400, Tom Lane wrote:\\n> Andres Freund <andres@anarazel.de> writes:\\n> > It certainly should be warned about. Practically I don't think it's a\\n> > problem, because we pretty much always operate on a copy of the page\\n> > when writing out, as otherwise concurrently set hint bits would be\\n> > troublesome.\\n> \\n> The write end of things is not a problem IMO, since presumably the caller\\n> would be about to overwrite the checksum field anyway. The issue is for\\n> reading and/or verifying, where it's much less obvious that clobbering\\n> the page image is safe.\\n\\nWith \\\"reading\\\" you mean putting the page into the buffercache? If so,\\nthat should be ok, the page isn't yet accessible (BM_VALID isn't set).\\n\\nI don't think it's possible to do verification from the page in s_b,\\nbecause we don't keep the checksum current, so there should never be\\ncalls to do so.\\n\\nBut we probably should still add a comment making clear that the\\nfunction modifies the buffer temporarily.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 16:12:14 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> On Thu, Aug 30, 2018 at 10:39:26AM -0400, Tom Lane wrote:\\n>> (The right fix, of course, is to malloc the work buffer rather than\\n>> put it on the stack.)\\n\\n> pg_upgrade/file.c is careful about that (5afcd2a), and has a comment on\\n> the matter, as does pg_standby.c.\\n\\n> Now, grepping around for \\\"BLCKSZ]\\\", some garbage may have accumulated?\\n\\nAh, that's a cute idea for finding trouble spots.\\n\\n> - entrySplitPage in ginentrypage.c\\n> - rewind_copy_file_range in file.c\\n> - _hash_alloc_buckets in hashpage.c\\n> - pg_prewarm.c\\n> - blinsert.c\\n> - pg_waldump.c\\n> - logtape.c\\n\\nSome of these are safe, I think, because the buffers are only used as\\ntargets for read() and write(). But some are definitely broken.\\nMy own list of files that seem to have issues is\\n\\nblinsert.c\\ngeneric_xlog.c\\nginentrypage.c\\nhashpage.c\\npg_verify_checksums.c\\npg_waldump.c\\nxloginsert.c\\n\\nThe fact that some of these are pretty old and we've not noticed is\\nnot good. It suggests that we don't currently have any compilers in the\\nbuildfarm that under-align char[] arrays on the stack, which seems like\\na gotcha waiting to bite us. I wonder if there is any way to persuade\\nsome compiler on a non-Intel box to do that.\\n\\nAnyway, I'll work on a patch for that, unless you were on it already?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 19:37:37 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 30, 2018 at 07:37:37PM -0400, Tom Lane wrote:\\n> Some of these are safe, I think, because the buffers are only used as\\n> targets for read() and write(). But some are definitely broken.\\n\\nYes, I have not spent more than a couple of minutes on this issue. I\\nnoticed some of them easily though.\\n\\n> My own list of files that seem to have issues is\\n> \\n> blinsert.c\\n> generic_xlog.c\\n> ginentrypage.c\\n> hashpage.c\\n> pg_verify_checksums.c\\n> pg_waldump.c\\n> xloginsert.c\\n> \\n> The fact that some of these are pretty old and we've not noticed is\\n> not good. It suggests that we don't currently have any compilers in the\\n> buildfarm that under-align char[] arrays on the stack, which seems like\\n> a gotcha waiting to bite us. I wonder if there is any way to persuade\\n> some compiler on a non-Intel box to do that.\\n\\nAgreed, that's annoying.\\n\\n> Anyway, I'll work on a patch for that, unless you were on it already?\\n\\nI have begun working on that and am less than halfway through it as I\\nneeded a fresh problem, however I am not sure I would be able to finish\\nit today, perhaps tomorrow... If you have time and want to press on as\\n11 would release soon, of course feel free to wrap up more quickly than\\nI can.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 30 Aug 2018 17:06:00 -0700\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nAm Donnerstag, den 30.08.2018, 19:02 -0400 schrieb Tom Lane:\\n> Andres Freund <andres@anarazel.de> writes:\\n> > On 2018-08-30 18:11:40 -0400, Tom Lane wrote:\\n> > > I suspect people will complain about the added cost of doing that.\\n> > I think the compiler will just optimize it away.\\n> \\n> Maybe. In any case, the attached version avoids any need for memcpy\\n> and is, I think, cleaner code anyhow. It fixes the problem for me\\n> with Fedora's gcc, though I'm not sure that it'd be enough to get rid\\n> of the warning Michael sees (I wonder what compiler he's using).\\n\\nThis fixes the bogus checksums, thanks!\\n\\nI am on Debian stable, i.e. 'gcc version 6.3.0 20170516 (Debian 6.3.0-\\n18+deb9u1)'. \\n\\nThe warning shows up only if I add -Wall *and* -O2 to CFLAGS, I think I\\ndid not mention that explictly before.\\n\\nAs it is in pg_verify_checksums.c I need Magnus' patch as well to make\\nit go away. But the warning is independent of the runtime issue so less\\nof an issue (but probably worth fixing).\\n\\n\\nMichael \\n\\n-- \\nMichael Banck\\nProjektleiter / Senior Berater\\nTel.: +49 2166 9901-171\\nFax: +49 2166 9901-100\\nEmail: michael.banck@credativ.de\\n\\ncredativ GmbH, HRB Mönchengladbach 12080\\nUSt-ID-Nummer: DE204566209\\nTrompeterallee 108, 41189 Mönchengladbach\\nGeschäftsführung: Dr. Michael Meskes, Jörg Folz, Sascha Heuer\\n\\nUnser Umgang mit personenbezogenen Daten unterliegt\\nfolgenden Bestimmungen: https://www.credativ.de/datenschutz\\n\\n\",\n \"msg_date\": \"Fri, 31 Aug 2018 10:11:07 +0200\",\n \"msg_from\": \"Michael Banck <michael.banck@credativ.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"On 31/08/2018 01:37, Tom Lane wrote:\\n> The fact that some of these are pretty old and we've not noticed is\\n> not good. It suggests that we don't currently have any compilers in the\\n> buildfarm that under-align char[] arrays on the stack, which seems like\\n> a gotcha waiting to bite us. I wonder if there is any way to persuade\\n> some compiler on a non-Intel box to do that.\\n\\n-Wcast-align=strict (requires gcc-8) warns about the issue in\\npg_verify_checksums.c, but also about a handful^W^W15211 other things,\\nso it wouldn't be immediately useful.\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\",\n \"msg_date\": \"Fri, 31 Aug 2018 13:10:08 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> On Thu, Aug 30, 2018 at 07:37:37PM -0400, Tom Lane wrote:\\n>> Anyway, I'll work on a patch for that, unless you were on it already?\\n\\n> I have begun working on that and am less than halfway through it as I\\n> needed a fresh problem, however I am not sure I would be able to finish\\n> it today, perhaps tomorrow... If you have time and want to press on as\\n> 11 would release soon, of course feel free to wrap up more quickly than\\n> I can.\\n\\nSome of these places might be performance-critical enough that adding\\na palloc/pfree cycle would not be nice. What I was considering doing\\nwas inventing something like\\n\\ntypedef union PGAlignedBuffer\\n{\\n\\tchar\\tdata[BLCKSZ];\\n\\tdouble\\tforce_align;\\n} PGAlignedBuffer;\\n\\nand replacing plain char[] variables with that as appropriate.\\nWe might need one for XLOG_BLCKSZ too.\\n\\nSince this'd be used in frontend as well as backend, it'd likely\\nhave to end up in c.h, which is slightly annoying but not really\\na big deal as long as we pick a nonconflicting type name.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Fri, 31 Aug 2018 10:55:06 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"Michael Banck <michael.banck@credativ.de> writes:\\n> Am Donnerstag, den 30.08.2018, 19:02 -0400 schrieb Tom Lane:\\n>> Maybe. In any case, the attached version avoids any need for memcpy\\n>> and is, I think, cleaner code anyhow. It fixes the problem for me\\n>> with Fedora's gcc, though I'm not sure that it'd be enough to get rid\\n>> of the warning Michael sees (I wonder what compiler he's using).\\n\\n> This fixes the bogus checksums, thanks!\\n> I am on Debian stable, i.e. 'gcc version 6.3.0 20170516 (Debian 6.3.0-\\n> 18+deb9u1)'.\\n\\nAh-hah. I still can't replicate any warning with gcc 8.1.1, but\\nI do have a Raspbian installation at hand, with\\ngcc version 6.3.0 20170516 (Raspbian 6.3.0-18+rpi1+deb9u1) \\nand on that I get\\n\\n$ gcc -Wall -Wmissing-prototypes -Wpointer-arith -Wdeclaration-after-statement -Werror=vla -Wendif-labels -Wmissing-format-attribute -Wformat-security -fwrapv -fexcess-precision=standard -g -O2 -I../../../src/include -D_GNU_SOURCE -c -o pg_verify_checksums.o pg_verify_checksums.c\\npg_verify_checksums.c: In function 'scan_file':\\npg_verify_checksums.c:112:3: warning: dereferencing type-punned pointer will break strict-aliasing rules [-Wstrict-aliasing]\\n if (PageIsNew(buf))\\n ^~\\n\\nAdding -Wstrict-aliasing=2 makes it slightly more verbose:\\n\\n$ gcc -Wstrict-aliasing=2 -Wall -Wmissing-prototypes -Wpointer-arith -Wdeclaration-after-statement -Werror=vla -Wendif-labels -Wmissing-format-attribute -Wformat-security -fwrapv -fexcess-precision=standard -g -O2 -I../../../src/include -D_GNU_SOURCE -c -o pg_verify_checksums.o pg_verify_checksums.c\\npg_verify_checksums.c: In function 'scan_file':\\npg_verify_checksums.c:82:2: warning: dereferencing type-punned pointer will break strict-aliasing rules [-Wstrict-aliasing]\\n PageHeader header = (PageHeader) buf;\\n ^~~~~~~~~~\\npg_verify_checksums.c:112:3: warning: dereferencing type-punned pointer will break strict-aliasing rules [-Wstrict-aliasing]\\n if (PageIsNew(buf))\\n ^~\\n\\nCompiling this way also causes pg_verify_checksums to compute wrong\\nchecksums. Installing the patched version of checksum_impl.h fixes\\nthat, but doesn't change the warnings, meaning they have absolutely\\nnothing to do with the actual problem :-(\\n\\nSo, even aside from the difficulty of getting rid of aliasing-unsafe\\ncode in Postgres, the state of the compiler warning technology is\\nstill years short of where we could sanely consider dispensing with\\n-fno-strict-aliasing in general. Still, as I said before, it'd be\\na good idea for checksum_impl.h to not depend on that. I'll go\\nahead and push the fix.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Fri, 31 Aug 2018 12:02:04 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> Some of these places might be performance-critical enough that adding\\n> a palloc/pfree cycle would not be nice. What I was considering doing\\n> was inventing something like\\n>\\n> typedef union PGAlignedBuffer\\n> {\\n> \\tchar\\tdata[BLCKSZ];\\n> \\tdouble\\tforce_align;\\n> } PGAlignedBuffer;\\n\\nHere's a proposed patch using that approach.\\n\\nAlthough some of the places that were using \\\"char buf[BLCKSZ]\\\" variables\\nweren't doing anything that really requires better alignment, I made\\nalmost all of them use PGAlignedBuffer variables anyway, on the grounds\\nthat you typically get better memcpy speed for aligned than unaligned\\ntransfers.\\n\\nI also fixed a few places that were using the palloc solution, and\\none that was actually doing hand-alignment of the pointer (ugh).\\nI didn't try to be thorough about getting rid of such pallocs,\\njust fix places that seemed likely to be performance-critical.\\n\\nThis needs to be back-patched as relevant, of course.\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Fri, 31 Aug 2018 15:55:51 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 31, 2018 at 03:55:51PM -0400, Tom Lane wrote:\\n> I wrote:\\n>> Some of these places might be performance-critical enough that adding\\n>> a palloc/pfree cycle would not be nice. What I was considering doing\\n>> was inventing something like\\n>>\\n>> typedef union PGAlignedBuffer\\n>> {\\n>> \\tchar\\tdata[BLCKSZ];\\n>> \\tdouble\\tforce_align;\\n>> } PGAlignedBuffer;\\n> \\n> Here's a proposed patch using that approach.\\n\\nThis solution is smarter than using malloc/palloc to enforce alignment.\\nI was digging into the GIN and bloom code with this pattern, but except\\nif I used TopTransactionContext for a simple approach, which is not\\nacceptable by the way, I was finishing with ugly memory contexts\\nused... I am still not sure if that was completely correct either, this\\nneeded more time ;)\\n\\n> Although some of the places that were using \\\"char buf[BLCKSZ]\\\" variables\\n> weren't doing anything that really requires better alignment, I made\\n> almost all of them use PGAlignedBuffer variables anyway, on the grounds\\n> that you typically get better memcpy speed for aligned than unaligned\\n> transfers.\\n\\nMakes sense.\\n\\n> I also fixed a few places that were using the palloc solution, and\\n> one that was actually doing hand-alignment of the pointer (ugh).\\n> I didn't try to be thorough about getting rid of such pallocs,\\n> just fix places that seemed likely to be performance-critical.\\n\\nOkay, for the memo replay_image_masked and master_image_masked\\nin xlog.c could make use of the new structure. SetWALSegSize in\\npg_standby.c and WriteEmptyXLOG in pg_resetwal.c, and pg_upgrade's\\nfile.c could also be patched.\\n\\nwalmethods.c could also use some static buffers, not worth the\\ncomplication perhaps.. \\n\\n> +typedef union PGAlignedBuffer\\n> +{\\n> +\\tchar\\t\\tdata[BLCKSZ];\\n> +\\tdouble\\t\\tforce_align_d;\\n> +\\tint64\\t\\tforce_align_i64;\\n> +} PGAlignedBuffer;\\n> +\\n> +/* Same, but for an XLOG_BLCKSZ-sized buffer */\\n> +typedef union PGAlignedXLogBuffer\\n> +{\\n> +\\tchar\\t\\tdata[XLOG_BLCKSZ];\\n> +\\tdouble\\t\\tforce_align_d;\\n> +\\tint64\\t\\tforce_align_i64;\\n> +} PGAlignedXLogBuffer;\\n\\nOne complain I have is about the name of those structures. Perhaps\\nPGAlignedBlock and PGAlignedXlogBlock make more sense?\\n--\\nMichael\",\n \"msg_date\": \"Fri, 31 Aug 2018 15:21:19 -0700\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> On Fri, Aug 31, 2018 at 03:55:51PM -0400, Tom Lane wrote:\\n>> I also fixed a few places that were using the palloc solution, and\\n>> one that was actually doing hand-alignment of the pointer (ugh).\\n>> I didn't try to be thorough about getting rid of such pallocs,\\n>> just fix places that seemed likely to be performance-critical.\\n\\n> Okay, for the memo replay_image_masked and master_image_masked\\n> in xlog.c could make use of the new structure. SetWALSegSize in\\n> pg_standby.c and WriteEmptyXLOG in pg_resetwal.c, and pg_upgrade's\\n> file.c could also be patched.\\n\\nI intentionally didn't change replay_image_masked/master_image_masked\\nto use statically-allocated buffers. Since, AFAICS, those aren't\\nneeded in most backend processes, they'd just be eating 16KB of\\nper-process data space to no purpose.\\n\\nThe others you mention could be changed, probably, but I didn't\\nbother as they didn't seem performance-critical.\\n\\n>> +typedef union PGAlignedBuffer\\n\\n> One complain I have is about the name of those structures. Perhaps\\n> PGAlignedBlock and PGAlignedXlogBlock make more sense?\\n\\nDon't have a strong preference, anybody else have an opinion?\\n\\n(I also wondered whether to use \\\"WAL\\\" instead of \\\"XLog\\\" in that\\nstruct name, but it seems like we've mostly stuck with \\\"xlog\\\"\\nin internal C names.)\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Fri, 31 Aug 2018 19:59:58 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 31, 2018 at 07:59:58PM -0400, Tom Lane wrote:\\n> The others you mention could be changed, probably, but I didn't\\n> bother as they didn't seem performance-critical.\\n\\nIt is not really critical indeed. There is an argument to change them\\nso as other folks get used to it though.\\n\\n> (I also wondered whether to use \\\"WAL\\\" instead of \\\"XLog\\\" in that\\n> struct name, but it seems like we've mostly stuck with \\\"xlog\\\"\\n> in internal C names.)\\n\\nXLOG_BLCKSZ is used, which makes me think that XLog is better than WAL\\nhere. A matter of taste of course.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 31 Aug 2018 17:33:03 -0700\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"\\nHello,\\n\\n>> Okay, for the memo replay_image_masked and master_image_masked\\n>> in xlog.c could make use of the new structure. SetWALSegSize in\\n>> pg_standby.c and WriteEmptyXLOG in pg_resetwal.c, and pg_upgrade's\\n>> file.c could also be patched.\\n>\\n> I intentionally didn't change replay_image_masked/master_image_masked\\n> to use statically-allocated buffers. Since, AFAICS, those aren't\\n> needed in most backend processes, they'd just be eating 16KB of\\n> per-process data space to no purpose.\\n>\\n> The others you mention could be changed, probably, but I didn't\\n> bother as they didn't seem performance-critical.\\n\\nI'd go for having just one same approach everywhere, for code base \\nhomogeneity.\\n\\n>>> +typedef union PGAlignedBuffer\\n>\\n>> One complain I have is about the name of those structures. Perhaps\\n>> PGAlignedBlock and PGAlignedXlogBlock make more sense?\\n>\\n> Don't have a strong preference, anybody else have an opinion?\\n\\nI like \\\"Block\\\" better, because it's more precise.\\n\\n> (I also wondered whether to use \\\"WAL\\\" instead of \\\"XLog\\\" in that\\n> struct name, but it seems like we've mostly stuck with \\\"xlog\\\"\\n> in internal C names.)\\n\\nBest to blend with the surrounding code in the header file?\\n\\n-- \\nFabien.\\n\\n\",\n \"msg_date\": \"Sat, 1 Sep 2018 07:24:38 +0200 (CEST)\",\n \"msg_from\": \"Fabien COELHO <coelho@cri.ensmp.fr>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> On Fri, Aug 31, 2018 at 07:59:58PM -0400, Tom Lane wrote:\\n>> The others you mention could be changed, probably, but I didn't\\n>> bother as they didn't seem performance-critical.\\n\\n> It is not really critical indeed. There is an argument to change them\\n> so as other folks get used to it though.\\n\\nFair enough. I renamed the types as suggested, changed a few more\\nplaces for consistency's sake, and pushed.\\n\\nThere still remain some places where palloc(BLCKSZ) or equivalent is used,\\nbut there's no matching pfree. In a lot of them the buffer is returned\\nto the caller so there's no choice. It's likely that some are just\\nleaking the storage transiently and we could convert them to using a\\nPGAlignedBlock local variable, but I didn't bother trying to do the\\nanalysis.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Sat, 01 Sep 2018 15:32:10 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n },\n {\n \"msg_contents\": \"On Sat, Sep 01, 2018 at 03:32:10PM -0400, Tom Lane wrote:\\n> Fair enough. I renamed the types as suggested, changed a few more\\n> places for consistency's sake, and pushed.\\n\\nThanks!\\n\\n> There still remain some places where palloc(BLCKSZ) or equivalent is used,\\n> but there's no matching pfree. In a lot of them the buffer is returned\\n> to the caller so there's no choice. It's likely that some are just\\n> leaking the storage transiently and we could convert them to using a\\n> PGAlignedBlock local variable, but I didn't bother trying to do the\\n> analysis.\\n\\nAt quick glance, I am not seeing anything critical. So the result looks\\ngood to me.\\n--\\nMichael\",\n \"msg_date\": \"Sat, 1 Sep 2018 16:19:00 -0700\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_verify_checksums and -fno-strict-aliasing\"\n }\n]","isConversational":false}}},{"rowIdx":47197,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nHere is a small patch to remove $SUBJECT: s/has contains/contains/\n\nBest regards,\nEtsuro Fujita","msg_date":"Thu, 30 Aug 2018 20:27:52 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Extra word in src/backend/optimizer/README"},{"msg_contents":"On Thu, Aug 30, 2018 at 1:27 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\nwrote:\n\n> Hi,\n>\n> Here is a small patch to remove $SUBJECT: s/has contains/contains/\n>\n\nDefinitely looks correct. A good first test to verify your own commit/push\nprivileges, perhaps?\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/ <http://www.hagander.net/>\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>\n\nOn Thu, Aug 30, 2018 at 1:27 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp> wrote:Hi,\n\nHere is a small patch to remove $SUBJECT: s/has contains/contains/Definitely looks correct. A good first test to verify your own commit/push privileges, perhaps? -- Magnus Hagander Me: https://www.hagander.net/ Work: https://www.redpill-linpro.com/","msg_date":"Thu, 30 Aug 2018 21:37:55 +0200","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":false,"msg_subject":"Re: Extra word in src/backend/optimizer/README"},{"msg_contents":"(2018/08/31 4:37), Magnus Hagander wrote:\n> On Thu, Aug 30, 2018 at 1:27 PM, Etsuro Fujita\n> <fujita.etsuro@lab.ntt.co.jp <mailto:fujita.etsuro@lab.ntt.co.jp>> wrote:\n> Here is a small patch to remove $SUBJECT: s/has contains/contains/\n\n> Definitely looks correct. A good first test to verify your own\n> commit/push privileges, perhaps?\n\nYeah, this is a good exercise for me, so I committed the patch.\n\nThanks!\n\nBest regards,\nEtsuro Fujita\n\n","msg_date":"Fri, 31 Aug 2018 16:58:58 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Re: Extra word in src/backend/optimizer/README"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nHere is a small patch to remove $SUBJECT: s/has contains/contains/\\n\\nBest regards,\\nEtsuro Fujita\",\n \"msg_date\": \"Thu, 30 Aug 2018 20:27:52 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Extra word in src/backend/optimizer/README\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 30, 2018 at 1:27 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\\nwrote:\\n\\n> Hi,\\n>\\n> Here is a small patch to remove $SUBJECT: s/has contains/contains/\\n>\\n\\nDefinitely looks correct. A good first test to verify your own commit/push\\nprivileges, perhaps?\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/ <http://www.hagander.net/>\\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>\\n\\nOn Thu, Aug 30, 2018 at 1:27 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp> wrote:Hi,\\n\\nHere is a small patch to remove $SUBJECT: s/has contains/contains/Definitely looks correct. A good first test to verify your own commit/push privileges, perhaps? -- Magnus Hagander Me: https://www.hagander.net/ Work: https://www.redpill-linpro.com/\",\n \"msg_date\": \"Thu, 30 Aug 2018 21:37:55 +0200\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extra word in src/backend/optimizer/README\"\n },\n {\n \"msg_contents\": \"(2018/08/31 4:37), Magnus Hagander wrote:\\n> On Thu, Aug 30, 2018 at 1:27 PM, Etsuro Fujita\\n> <fujita.etsuro@lab.ntt.co.jp <mailto:fujita.etsuro@lab.ntt.co.jp>> wrote:\\n> Here is a small patch to remove $SUBJECT: s/has contains/contains/\\n\\n> Definitely looks correct. A good first test to verify your own\\n> commit/push privileges, perhaps?\\n\\nYeah, this is a good exercise for me, so I committed the patch.\\n\\nThanks!\\n\\nBest regards,\\nEtsuro Fujita\\n\\n\",\n \"msg_date\": \"Fri, 31 Aug 2018 16:58:58 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Extra word in src/backend/optimizer/README\"\n }\n]","isConversational":false}}},{"rowIdx":47198,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi!\n\nOur customer have a bad plan problem, which could be reduced to the\nfollowing example.\n\ncreate table t1 (id int primary key, k int);\ncreate table t2 (id int);\n\ninsert into t1 (select i, i from generate_series(1,1000000) i);\ninsert into t2 (select 0 from generate_series(1,100)i);\ninsert into t2 values (500000);\n\nFor the following query the plan is OK despite number of resulting\nrows is very much overestimated.\n\n# explain select * from\nt1 join t2 x1 on x1.id = t1.k\njoin t2 x2 on x2.id = t1.k\njoin t2 x3 on x3.id = t1.k\njoin t2 x4 on x4.id = t1.k\njoin t2 x5 on x5.id = t1.k\njoin t2 x6 on x6.id = t1.k\njoin t2 x7 on x7.id = t1.k\njoin t2 x8 on x8.id = t1.k\njoin t2 x9 on x9.id = t1.k\nwhere t1.id = 500000;\n\n QUERY PLAN\n------------------------------------------------------------------------------------------------------------------------------------------------------------\n Hash Join (cost=265.98..104071348014760.11 rows=9240411823550620 width=44)\n Hash Cond: (x1.id = x8.id)\n -> Hash Join (cost=22.80..25.20 rows=942425865760 width=36)\n Hash Cond: (x1.id = t1.k)\n -> Seq Scan on t2 x1 (cost=0.00..2.01 rows=101 width=4)\n -> Hash (cost=22.79..22.79 rows=1 width=32)\n -> Hash Join (cost=20.39..22.79 rows=1 width=32)\n Hash Cond: (x7.id = t1.k)\n -> Seq Scan on t2 x7 (cost=0.00..2.01 rows=101 width=4)\n -> Hash (cost=20.38..20.38 rows=1 width=28)\n -> Hash Join (cost=17.98..20.38 rows=1 width=28)\n Hash Cond: (x6.id = t1.k)\n -> Seq Scan on t2 x6\n(cost=0.00..2.01 rows=101 width=4)\n -> Hash (cost=17.96..17.96 rows=1 width=24)\n -> Hash Join\n(cost=15.57..17.96 rows=1 width=24)\n Hash Cond: (x5.id = t1.k)\n -> Seq Scan on t2 x5\n(cost=0.00..2.01 rows=101 width=4)\n -> Hash\n(cost=15.55..15.55 rows=1 width=20)\n -> Hash Join\n(cost=13.15..15.55 rows=1 width=20)\n Hash Cond:\n(x4.id = t1.k)\n -> Seq Scan\non t2 x4 (cost=0.00..2.01 rows=101 width=4)\n -> Hash\n(cost=13.14..13.14 rows=1 width=16)\n ->\nHash Join (cost=10.74..13.14 rows=1 width=16)\n\nHash Cond: (x3.id = t1.k)\n\n-> Seq Scan on t2 x3 (cost=0.00..2.01 rows=101 width=4)\n\n-> Hash (cost=10.73..10.73 rows=1 width=12)\n\n -> Hash Join (cost=8.46..10.73 rows=1 width=12)\n\n Hash Cond: (x2.id = t1.k)\n\n -> Seq Scan on t2 x2 (cost=0.00..2.01 rows=101 width=4)\n\n -> Hash (cost=8.44..8.44 rows=1 width=8)\n\n -> Index Scan using t1_pkey on t1 (cost=0.42..8.44\nrows=1 width=8)\n\n Index Cond: (id = 500000)\n -> Hash (cost=118.17..118.17 rows=10001 width=8)\n -> Hash Join (cost=3.27..118.17 rows=10001 width=8)\n Hash Cond: (x8.id = x9.id)\n -> Seq Scan on t2 x8 (cost=0.00..2.01 rows=101 width=4)\n -> Hash (cost=2.01..2.01 rows=101 width=4)\n -> Seq Scan on t2 x9 (cost=0.00..2.01 rows=101 width=4)\n(38 rows)\n\nBut when you add LIMIT clause, index scan on t1 turns out into sequential scan.\n\n# explain select * from\nt1 join t2 x1 on x1.id = t1.k\njoin t2 x2 on x2.id = t1.k\njoin t2 x3 on x3.id = t1.k\njoin t2 x4 on x4.id = t1.k\njoin t2 x5 on x5.id = t1.k\njoin t2 x6 on x6.id = t1.k\njoin t2 x7 on x7.id = t1.k\njoin t2 x8 on x8.id = t1.k\njoin t2 x9 on x9.id = t1.k\nwhere t1.id = 500000 limit 100;\n QUERY PLAN\n-------------------------------------------------------------------------------------------------------------------\n Limit (cost=0.00..1.55 rows=100 width=44)\n -> Nested Loop (cost=0.00..142805795187416.44\nrows=9240411823550620 width=44)\n Join Filter: (x1.id = x9.id)\n -> Nested Loop (cost=0.00..1427775203576.57\nrows=93318825071840 width=40)\n Join Filter: (x1.id = x8.id)\n -> Nested Loop (cost=0.00..16947.91 rows=942425865760 width=36)\n Join Filter: (t1.k = x1.id)\n -> Nested Loop (cost=0.00..16944.63 rows=1 width=32)\n Join Filter: (t1.k = x7.id)\n -> Nested Loop (cost=0.00..16941.36\nrows=1 width=28)\n Join Filter: (t1.k = x6.id)\n -> Nested Loop (cost=0.00..16938.09\nrows=1 width=24)\n Join Filter: (t1.k = x5.id)\n -> Nested Loop\n(cost=0.00..16934.82 rows=1 width=20)\n Join Filter: (t1.k = x4.id)\n -> Nested Loop\n(cost=0.00..16931.54 rows=1 width=16)\n Join Filter: (t1.k = x3.id)\n -> Nested Loop\n(cost=0.00..16928.27 rows=1 width=12)\n Join Filter:\n(t1.k = x2.id)\n -> Seq Scan\non t1 (cost=0.00..16925.00 rows=1 width=8)\n Filter:\n(id = 500000)\n -> Seq Scan\non t2 x2 (cost=0.00..2.01 rows=101 width=4)\n -> Seq Scan on t2\nx3 (cost=0.00..2.01 rows=101 width=4)\n -> Seq Scan on t2 x4\n(cost=0.00..2.01 rows=101 width=4)\n -> Seq Scan on t2 x5\n(cost=0.00..2.01 rows=101 width=4)\n -> Seq Scan on t2 x6\n(cost=0.00..2.01 rows=101 width=4)\n -> Seq Scan on t2 x7 (cost=0.00..2.01\nrows=101 width=4)\n -> Seq Scan on t2 x1 (cost=0.00..2.01 rows=101 width=4)\n -> Materialize (cost=0.00..2.51 rows=101 width=4)\n -> Seq Scan on t2 x8 (cost=0.00..2.01 rows=101 width=4)\n -> Materialize (cost=0.00..2.51 rows=101 width=4)\n -> Seq Scan on t2 x9 (cost=0.00..2.01 rows=101 width=4)\n(32 rows)\n\nObviously, that happens because selectivity for join with t2 is\noverestimated and multiplied many times. And improvements in this\narea seems quite hard for me.\n\nBut I think there is another issue in sequential scan cost. We have\nzero startup cost for sequential scan. But why? As I get sequential\nscan should estimate amount of resources we need to spend in order to\nstart returning rows. So, in our example when we expect finding 1 row\nfrom the table, in average we have to scan half of table before find\nthis row. Thus, startup_cost should be about half of total cost.\nAttached POC patch implements that. In the given example sequential\nscan turns out back to index scan.\n\nAny thoughts?\n\n------\nAlexander Korotkov\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres Company","msg_date":"Thu, 30 Aug 2018 16:56:53 +0300","msg_from":"Alexander Korotkov <a.korotkov@postgrespro.ru>","msg_from_op":true,"msg_subject":"Startup cost of sequential scan"},{"msg_contents":"Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\n> But I think there is another issue in sequential scan cost. We have\n> zero startup cost for sequential scan. But why?\n\nBecause it's what the mental model of startup cost says it should be.\nAlso, I do not think we can change that without a whole lot of unpleasant\nside effects on cases that work well today. Have you even checked to\nsee how much of the regression tests break with this change?\n\n\t\t\tregards, tom lane\n\n","msg_date":"Thu, 30 Aug 2018 10:04:55 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Startup cost of sequential scan"},{"msg_contents":"On Thu, Aug 30, 2018 at 5:05 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\n> > But I think there is another issue in sequential scan cost. We have\n> > zero startup cost for sequential scan. But why?\n>\n> Because it's what the mental model of startup cost says it should be.\n\nRight. So as I understand. The model is that we start sequential\nscan immediately, and then find one row uniformly distributed over the\nwhole table.\n\n From this model we make a conclusion that we're starting getting rows\nfrom sequential scan sooner than from index scan. And this conclusion\ndoesn't reflect reality. If even estimates for join with t2 wouldn't\nbe wrong, then our plan for LIMIT query would be still bad, because it\nwould be still faster to get that one row using index scan rather than\nsequential scan.\n\nSo, if understand the mental model correctly, our cost estimate for\nLIMIT query is based on the idea that we need to fetch only *fraction*\nof row from t1 in order to get 100 resulting rows. This is obviously\nwrong, because we're always dealing with whole rows :) But I can't\nimagine how we can take care of it without redesigning our whole\ncosting model...\n\n> Also, I do not think we can change that without a whole lot of unpleasant\n> side effects on cases that work well today. Have you even checked to\n> see how much of the regression tests break with this change?\n\nI though it's to early to discuss this. But yet, I've checked this.\nIt appears to be surprisingly few broken tests. Just some reordering\nof tables in partition_join, select_parallel.------\nAlexander Korotkov\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres Company","msg_date":"Thu, 30 Aug 2018 17:31:36 +0300","msg_from":"Alexander Korotkov <a.korotkov@postgrespro.ru>","msg_from_op":true,"msg_subject":"Re: Startup cost of sequential scan"},{"msg_contents":"Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\n> On Thu, Aug 30, 2018 at 5:05 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> Because it's what the mental model of startup cost says it should be.\n\n> From this model we make a conclusion that we're starting getting rows\n> from sequential scan sooner than from index scan. And this conclusion\n> doesn't reflect reality.\n\nNo, startup cost is not the \"time to find the first row\". It's overhead\npaid before you even get to start examining rows.\n\nI'm disinclined to consider fundamental changes to our costing model\non the basis of this example. The fact that the rowcount estimates are\nso far off reality means that you're basically looking at \"garbage in,\ngarbage out\" for the cost calculations --- and applying a small LIMIT\njust magnifies that.\n\nIt'd be more useful to think first about how to make the selectivity\nestimates better; after that, we might or might not still think there's\na costing issue.\n\n\t\t\tregards, tom lane\n\n","msg_date":"Thu, 30 Aug 2018 10:58:38 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Startup cost of sequential scan"},{"msg_contents":"\n\nOn 30.08.2018 17:58, Tom Lane wrote:\n> Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\n>> On Thu, Aug 30, 2018 at 5:05 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>>> Because it's what the mental model of startup cost says it should be.\n>> From this model we make a conclusion that we're starting getting rows\n>> from sequential scan sooner than from index scan. And this conclusion\n>> doesn't reflect reality.\n> No, startup cost is not the \"time to find the first row\". It's overhead\n> paid before you even get to start examining rows.\nBut it seems to me that calculation of cost in LIMIT node contradicts \nwith this statement:\n\n pathnode->path.startup_cost +=\n (subpath->total_cost - subpath->startup_cost)\n * offset_rows / subpath->rows;\n\n\n\n>\n> I'm disinclined to consider fundamental changes to our costing model\n> on the basis of this example. The fact that the rowcount estimates are\n> so far off reality means that you're basically looking at \"garbage in,\n> garbage out\" for the cost calculations --- and applying a small LIMIT\n> just magnifies that.\n>\n> It'd be more useful to think first about how to make the selectivity\n> estimates better; after that, we might or might not still think there's\n> a costing issue.\n>\n> \t\t\tregards, tom lane\n>\n\n-- \nKonstantin Knizhnik\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres Company\n\n\n","msg_date":"Thu, 30 Aug 2018 18:08:00 +0300","msg_from":"Konstantin Knizhnik <k.knizhnik@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: Startup cost of sequential scan"},{"msg_contents":"On Thu, Aug 30, 2018 at 6:08 PM Konstantin Knizhnik\n<k.knizhnik@postgrespro.ru> wrote:\n> On 30.08.2018 17:58, Tom Lane wrote:\n> > Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\n> >> On Thu, Aug 30, 2018 at 5:05 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> >>> Because it's what the mental model of startup cost says it should be.\n> >> From this model we make a conclusion that we're starting getting rows\n> >> from sequential scan sooner than from index scan. And this conclusion\n> >> doesn't reflect reality.\n> > No, startup cost is not the \"time to find the first row\". It's overhead\n> > paid before you even get to start examining rows.\n> But it seems to me that calculation of cost in LIMIT node contradicts\n> with this statement:\n>\n> pathnode->path.startup_cost +=\n> (subpath->total_cost - subpath->startup_cost)\n> * offset_rows / subpath->rows;\n\nWhy does it contradict? It just assumes that skipping OFFSET rows to\nbe preliminary work before returning results rows...\n\n------\nAlexander Korotkov\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres Company\n\n","msg_date":"Thu, 30 Aug 2018 18:23:30 +0300","msg_from":"Alexander Korotkov <a.korotkov@postgrespro.ru>","msg_from_op":true,"msg_subject":"Re: Startup cost of sequential scan"},{"msg_contents":">>>>> \"Konstantin\" == Konstantin Knizhnik <k.knizhnik@postgrespro.ru> writes:\n\n >> No, startup cost is not the \"time to find the first row\". It's\n >> overhead paid before you even get to start examining rows.\n\n Konstantin> But it seems to me that calculation of cost in LIMIT node\n Konstantin> contradicts with this statement:\n\nThe model (assuming I understand it rightly) is that what we're actually\ntracking is a startup cost and a per-output-row cost, but for comparison\npurposes we actually store the rows and the computed total, rather than\njust the per-row cost:\n\nrows\nstartup_cost\ntotal_cost = startup_cost + (rows * per_row_cost)\n\nSo what Limit is doing the for the offset count is recovering the\nsubpath's per_row_cost from (total_cost - startup_cost)/rows, and then\nscaling that by the number of rows in the offset (which are being\ndiscarded), and adding that to the startup cost. So this is saying: the\nstartup cost for OFFSET N is the startup cost of the subplan, plus the\ncost of fetching N rows from the subplan. (And after fetching those N\nrows, we still haven't found the first row that we will actually\nreturn.)\n\nFor LIMIT N, we instead replace the old total cost with a new one\ncalculated from the startup cost plus N times the subplan's per-row\ncost.\n\n-- \nAndrew (irc:RhodiumToad)\n\n","msg_date":"Thu, 30 Aug 2018 16:23:44 +0100","msg_from":"Andrew Gierth <andrew@tao11.riddles.org.uk>","msg_from_op":false,"msg_subject":"Re: Startup cost of sequential scan"},{"msg_contents":"Andrew Gierth <andrew@tao11.riddles.org.uk> writes:\n> The model (assuming I understand it rightly) is that what we're actually\n> tracking is a startup cost and a per-output-row cost, but for comparison\n> purposes we actually store the rows and the computed total, rather than\n> just the per-row cost:\n\n> rows\n> startup_cost\n> total_cost = startup_cost + (rows * per_row_cost)\n\nRight. I tend to think of it differently: the cost to retrieve the\nfirst K rows out of a total of N is\n\tstartup_cost + (total_cost - startup_cost) * K/N\nbut that comes out to the same thing as what Andrew said.\n\n\t\t\tregards, tom lane\n\n","msg_date":"Thu, 30 Aug 2018 11:33:19 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Startup cost of sequential scan"},{"msg_contents":"On Thu, Aug 30, 2018 at 5:58 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\n> > On Thu, Aug 30, 2018 at 5:05 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> >> Because it's what the mental model of startup cost says it should be.\n>\n> > From this model we make a conclusion that we're starting getting rows\n> > from sequential scan sooner than from index scan. And this conclusion\n> > doesn't reflect reality.\n>\n> No, startup cost is not the \"time to find the first row\". It's overhead\n> paid before you even get to start examining rows.\n>\n> I'm disinclined to consider fundamental changes to our costing model\n> on the basis of this example. The fact that the rowcount estimates are\n> so far off reality means that you're basically looking at \"garbage in,\n> garbage out\" for the cost calculations --- and applying a small LIMIT\n> just magnifies that.\n>\n> It'd be more useful to think first about how to make the selectivity\n> estimates better; after that, we might or might not still think there's\n> a costing issue.\n\nI understand that startup cost is not \"time to find the first row\".\nBut I think this example highlight not one but two issues.\n1) Row count estimates for joins are wrong.\n2) Rows are assumed to be continuous while in reality they are\ndiscrete. So, if we reverse the assumptions made in LIMIT clause\nestimation, we may say that it's basically assuming that we need to\nfetch only fraction of row from the sequential scan node. And in the\ncase we really fetch 101 rows in each join with t2, this logic would\nstill bring us to the bad plan. And now I'm not proposing go rush\nredesigning planner to fix that. I just think it's probably something\nworth discussion.\n\n------\nAlexander Korotkov\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres Company\n\n","msg_date":"Thu, 30 Aug 2018 18:38:04 +0300","msg_from":"Alexander Korotkov <a.korotkov@postgrespro.ru>","msg_from_op":true,"msg_subject":"Re: Startup cost of sequential scan"},{"msg_contents":"Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\n> I understand that startup cost is not \"time to find the first row\".\n> But I think this example highlight not one but two issues.\n> 1) Row count estimates for joins are wrong.\n\nYup.\n\n> 2) Rows are assumed to be continuous while in reality they are\n> discrete.\n\nWhere do you get that from?\n\nThe problem this example is highlighting is mostly just the bad\njoin size estimate, imo. It's not at all clear that the planner\nwould still do the wrong thing if that were fixed. In fact,\nif I replace the contents of t2 with some other distribution,\nsuch as\n\tinsert into t2 (select i from generate_series(1,100)i);\nI get a much better join size estimate *and* a sane plan, even\nin the LIMIT case. So the problem here is just with the join\nsize estimate.\n\n\t\t\tregards, tom lane\n\n","msg_date":"Thu, 30 Aug 2018 11:55:27 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Startup cost of sequential scan"},{"msg_contents":"On Thu, Aug 30, 2018 at 10:04 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\n>> But I think there is another issue in sequential scan cost. We have\n>> zero startup cost for sequential scan. But why?\n>\n> Because it's what the mental model of startup cost says it should be.\n\nWhose mental model? I guess the Tom Lane mind is the canonical one\nfor this project, but I'm not sure that it entirely agrees with mine.\nIIRC, it was previously proposed that we ought to charge\nrandom_page_cost for the first block of a sequential scan, because at\npresent the cost of fetching 1 block differs depending on whether we\nare fetching it from a heap or an index, which seems unprincipled.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n","msg_date":"Thu, 30 Aug 2018 15:43:59 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Startup cost of sequential scan"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> Whose mental model? I guess the Tom Lane mind is the canonical one\n> for this project, but I'm not sure that it entirely agrees with mine.\n\nSince the fact that we have a notion of startup cost at all is entirely my\nfault, I don't feel shy about claiming to have the authoritative view of\nwhat it means.\n\n(Whether that's adequately documented is another question :-()\n\n> IIRC, it was previously proposed that we ought to charge\n> random_page_cost for the first block of a sequential scan, because at\n> present the cost of fetching 1 block differs depending on whether we\n> are fetching it from a heap or an index, which seems unprincipled.\n\nThat might well be a sane thing to do ... but it'd still be part of run\ncost not startup cost.\n\n\t\t\tregards, tom lane\n\n","msg_date":"Thu, 30 Aug 2018 17:38:36 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Startup cost of sequential scan"},{"msg_contents":"On Thu, Aug 30, 2018 at 6:55 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\n> > I understand that startup cost is not \"time to find the first row\".\n> > But I think this example highlight not one but two issues.\n> > 1) Row count estimates for joins are wrong.\n>\n> Yup.\n>\n> > 2) Rows are assumed to be continuous while in reality they are\n> > discrete.\n>\n> Where do you get that from?\n\nI made a similar example, where estimates are good. It's quite\nartificial, because it selects small limit from enormous virtual table\nconstructed by cross joins. But it illustrates how our model,\nassuming tuples to be continuous, can differ from reality.\n\ncreate table t1 (id int primary key, k int);\ncreate table t2 (id int);\n\ninsert into t1 (select i, i from generate_series(1,1000000) i);\ninsert into t2 (select 0 from generate_series(1,100) i);\nvacuum analyze t1, t2;\n\nBy default, the query is processed using sequential scan for t1.\n\n# explain analyze select * from t1,t2 x1,t2 x2,t2 x3,t2 x4,t2 x5 where\nt1.id = 500000 limit 100;\n QUERY PLAN\n--------------------------------------------------------------------------------------------------------------------------------------\n Limit (cost=0.00..1.26 rows=100 width=28) (actual\ntime=32.879..32.926 rows=100 loops=1)\n -> Nested Loop (cost=0.00..126279562.00 rows=10000000000\nwidth=28) (actual time=32.878..32.912 rows=100 loops=1)\n -> Nested Loop (cost=0.00..1279559.75 rows=100000000\nwidth=24) (actual time=32.873..32.873 rows=1 loops=1)\n -> Nested Loop (cost=0.00..29557.50 rows=1000000\nwidth=20) (actual time=32.868..32.868 rows=1 loops=1)\n -> Nested Loop (cost=0.00..17055.25 rows=10000\nwidth=16) (actual time=32.864..32.864 rows=1 loops=1)\n -> Nested Loop (cost=0.00..16928.00\nrows=100 width=12) (actual time=32.856..32.856 rows=1 loops=1)\n -> Seq Scan on t1\n(cost=0.00..16925.00 rows=1 width=8) (actual time=32.841..32.841\nrows=1 loops=1)\n Filter: (id = 500000)\n Rows Removed by Filter: 501983\n -> Seq Scan on t2 x1\n(cost=0.00..2.00 rows=100 width=4) (actual time=0.011..0.011 rows=1\nloops=1)\n -> Materialize (cost=0.00..2.50 rows=100\nwidth=4) (actual time=0.007..0.007 rows=1 loops=1)\n -> Seq Scan on t2 x2\n(cost=0.00..2.00 rows=100 width=4) (actual time=0.003..0.003 rows=1\nloops=1)\n -> Materialize (cost=0.00..2.50 rows=100\nwidth=4) (actual time=0.003..0.003 rows=1 loops=1)\n -> Seq Scan on t2 x3 (cost=0.00..2.00\nrows=100 width=4) (actual time=0.002..0.003 rows=1 loops=1)\n -> Materialize (cost=0.00..2.50 rows=100 width=4)\n(actual time=0.003..0.003 rows=1 loops=1)\n -> Seq Scan on t2 x4 (cost=0.00..2.00 rows=100\nwidth=4) (actual time=0.002..0.003 rows=1 loops=1)\n -> Materialize (cost=0.00..2.50 rows=100 width=4) (actual\ntime=0.004..0.024 rows=100 loops=1)\n -> Seq Scan on t2 x5 (cost=0.00..2.00 rows=100\nwidth=4) (actual time=0.003..0.010 rows=100 loops=1)\n Planning Time: 0.372 ms\n Execution Time: 32.987 ms\n(20 rows)\n\nBut if we disable sequential scan and bitmap scan then it would be\nindex scan, which is much faster.\n# set enable_seqscan = off;\n# set enable_bitmapscan = off;\n\n# explain analyze select * from t1,t2 x1,t2 x2,t2 x3,t2 x4,t2 x5 where\nt1.id = 500000 limit 100;\n\nQUERY PLAN\n---------------------------------------------------------------------------------------------------------------------------------------------------------\n Limit (cost=50000000000.43..50000000001.69 rows=100 width=28)\n(actual time=0.041..0.092 rows=100 loops=1)\n -> Nested Loop (cost=50000000000.43..50126262645.44\nrows=10000000000 width=28) (actual time=0.040..0.079 rows=100 loops=1)\n -> Nested Loop (cost=40000000000.43..40001262643.19\nrows=100000000 width=24) (actual time=0.036..0.036 rows=1 loops=1)\n -> Nested Loop (cost=30000000000.42..30000012640.94\nrows=1000000 width=20) (actual time=0.033..0.033 rows=1 loops=1)\n -> Nested Loop\n(cost=20000000000.42..20000000138.69 rows=10000 width=16) (actual\ntime=0.029..0.029 rows=1 loops=1)\n -> Nested Loop\n(cost=10000000000.42..10000000011.44 rows=100 width=12) (actual\ntime=0.023..0.023 rows=1 loops=1)\n -> Index Scan using t1_pkey on t1\n(cost=0.42..8.44 rows=1 width=8) (actual time=0.015..0.015 rows=1\nloops=1)\n Index Cond: (id = 500000)\n -> Seq Scan on t2 x1\n(cost=10000000000.00..10000000002.00 rows=100 width=4) (actual\ntime=0.007..0.007 rows=1 loops=1)\n -> Materialize\n(cost=10000000000.00..10000000002.50 rows=100 width=4) (actual\ntime=0.004..0.005 rows=1 loops=1)\n -> Seq Scan on t2 x2\n(cost=10000000000.00..10000000002.00 rows=100 width=4) (actual\ntime=0.003..0.003 rows=1 loops=1)\n -> Materialize\n(cost=10000000000.00..10000000002.50 rows=100 width=4) (actual\ntime=0.003..0.003 rows=1 loops=1)\n -> Seq Scan on t2 x3\n(cost=10000000000.00..10000000002.00 rows=100 width=4) (actual\ntime=0.003..0.003 rows=1 loops=1)\n -> Materialize (cost=10000000000.00..10000000002.50\nrows=100 width=4) (actual time=0.003..0.003 rows=1 loops=1)\n -> Seq Scan on t2 x4\n(cost=10000000000.00..10000000002.00 rows=100 width=4) (actual\ntime=0.002..0.003 rows=1 loops=1)\n -> Materialize (cost=10000000000.00..10000000002.50\nrows=100 width=4) (actual time=0.003..0.026 rows=100 loops=1)\n -> Seq Scan on t2 x5\n(cost=10000000000.00..10000000002.00 rows=100 width=4) (actual\ntime=0.003..0.010 rows=100 loops=1)\n Planning Time: 0.274 ms\n Execution Time: 0.150 ms\n(19 rows)\n\n From this example, I get that there is a distinct issue that we assume\nrows to be continuous.\n\n------\nAlexander Korotkov\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres Company\n\n","msg_date":"Fri, 31 Aug 2018 13:50:32 +0300","msg_from":"Alexander Korotkov <a.korotkov@postgrespro.ru>","msg_from_op":true,"msg_subject":"Re: Startup cost of sequential scan"}],"string":"[\n {\n \"msg_contents\": \"Hi!\\n\\nOur customer have a bad plan problem, which could be reduced to the\\nfollowing example.\\n\\ncreate table t1 (id int primary key, k int);\\ncreate table t2 (id int);\\n\\ninsert into t1 (select i, i from generate_series(1,1000000) i);\\ninsert into t2 (select 0 from generate_series(1,100)i);\\ninsert into t2 values (500000);\\n\\nFor the following query the plan is OK despite number of resulting\\nrows is very much overestimated.\\n\\n# explain select * from\\nt1 join t2 x1 on x1.id = t1.k\\njoin t2 x2 on x2.id = t1.k\\njoin t2 x3 on x3.id = t1.k\\njoin t2 x4 on x4.id = t1.k\\njoin t2 x5 on x5.id = t1.k\\njoin t2 x6 on x6.id = t1.k\\njoin t2 x7 on x7.id = t1.k\\njoin t2 x8 on x8.id = t1.k\\njoin t2 x9 on x9.id = t1.k\\nwhere t1.id = 500000;\\n\\n QUERY PLAN\\n------------------------------------------------------------------------------------------------------------------------------------------------------------\\n Hash Join (cost=265.98..104071348014760.11 rows=9240411823550620 width=44)\\n Hash Cond: (x1.id = x8.id)\\n -> Hash Join (cost=22.80..25.20 rows=942425865760 width=36)\\n Hash Cond: (x1.id = t1.k)\\n -> Seq Scan on t2 x1 (cost=0.00..2.01 rows=101 width=4)\\n -> Hash (cost=22.79..22.79 rows=1 width=32)\\n -> Hash Join (cost=20.39..22.79 rows=1 width=32)\\n Hash Cond: (x7.id = t1.k)\\n -> Seq Scan on t2 x7 (cost=0.00..2.01 rows=101 width=4)\\n -> Hash (cost=20.38..20.38 rows=1 width=28)\\n -> Hash Join (cost=17.98..20.38 rows=1 width=28)\\n Hash Cond: (x6.id = t1.k)\\n -> Seq Scan on t2 x6\\n(cost=0.00..2.01 rows=101 width=4)\\n -> Hash (cost=17.96..17.96 rows=1 width=24)\\n -> Hash Join\\n(cost=15.57..17.96 rows=1 width=24)\\n Hash Cond: (x5.id = t1.k)\\n -> Seq Scan on t2 x5\\n(cost=0.00..2.01 rows=101 width=4)\\n -> Hash\\n(cost=15.55..15.55 rows=1 width=20)\\n -> Hash Join\\n(cost=13.15..15.55 rows=1 width=20)\\n Hash Cond:\\n(x4.id = t1.k)\\n -> Seq Scan\\non t2 x4 (cost=0.00..2.01 rows=101 width=4)\\n -> Hash\\n(cost=13.14..13.14 rows=1 width=16)\\n ->\\nHash Join (cost=10.74..13.14 rows=1 width=16)\\n\\nHash Cond: (x3.id = t1.k)\\n\\n-> Seq Scan on t2 x3 (cost=0.00..2.01 rows=101 width=4)\\n\\n-> Hash (cost=10.73..10.73 rows=1 width=12)\\n\\n -> Hash Join (cost=8.46..10.73 rows=1 width=12)\\n\\n Hash Cond: (x2.id = t1.k)\\n\\n -> Seq Scan on t2 x2 (cost=0.00..2.01 rows=101 width=4)\\n\\n -> Hash (cost=8.44..8.44 rows=1 width=8)\\n\\n -> Index Scan using t1_pkey on t1 (cost=0.42..8.44\\nrows=1 width=8)\\n\\n Index Cond: (id = 500000)\\n -> Hash (cost=118.17..118.17 rows=10001 width=8)\\n -> Hash Join (cost=3.27..118.17 rows=10001 width=8)\\n Hash Cond: (x8.id = x9.id)\\n -> Seq Scan on t2 x8 (cost=0.00..2.01 rows=101 width=4)\\n -> Hash (cost=2.01..2.01 rows=101 width=4)\\n -> Seq Scan on t2 x9 (cost=0.00..2.01 rows=101 width=4)\\n(38 rows)\\n\\nBut when you add LIMIT clause, index scan on t1 turns out into sequential scan.\\n\\n# explain select * from\\nt1 join t2 x1 on x1.id = t1.k\\njoin t2 x2 on x2.id = t1.k\\njoin t2 x3 on x3.id = t1.k\\njoin t2 x4 on x4.id = t1.k\\njoin t2 x5 on x5.id = t1.k\\njoin t2 x6 on x6.id = t1.k\\njoin t2 x7 on x7.id = t1.k\\njoin t2 x8 on x8.id = t1.k\\njoin t2 x9 on x9.id = t1.k\\nwhere t1.id = 500000 limit 100;\\n QUERY PLAN\\n-------------------------------------------------------------------------------------------------------------------\\n Limit (cost=0.00..1.55 rows=100 width=44)\\n -> Nested Loop (cost=0.00..142805795187416.44\\nrows=9240411823550620 width=44)\\n Join Filter: (x1.id = x9.id)\\n -> Nested Loop (cost=0.00..1427775203576.57\\nrows=93318825071840 width=40)\\n Join Filter: (x1.id = x8.id)\\n -> Nested Loop (cost=0.00..16947.91 rows=942425865760 width=36)\\n Join Filter: (t1.k = x1.id)\\n -> Nested Loop (cost=0.00..16944.63 rows=1 width=32)\\n Join Filter: (t1.k = x7.id)\\n -> Nested Loop (cost=0.00..16941.36\\nrows=1 width=28)\\n Join Filter: (t1.k = x6.id)\\n -> Nested Loop (cost=0.00..16938.09\\nrows=1 width=24)\\n Join Filter: (t1.k = x5.id)\\n -> Nested Loop\\n(cost=0.00..16934.82 rows=1 width=20)\\n Join Filter: (t1.k = x4.id)\\n -> Nested Loop\\n(cost=0.00..16931.54 rows=1 width=16)\\n Join Filter: (t1.k = x3.id)\\n -> Nested Loop\\n(cost=0.00..16928.27 rows=1 width=12)\\n Join Filter:\\n(t1.k = x2.id)\\n -> Seq Scan\\non t1 (cost=0.00..16925.00 rows=1 width=8)\\n Filter:\\n(id = 500000)\\n -> Seq Scan\\non t2 x2 (cost=0.00..2.01 rows=101 width=4)\\n -> Seq Scan on t2\\nx3 (cost=0.00..2.01 rows=101 width=4)\\n -> Seq Scan on t2 x4\\n(cost=0.00..2.01 rows=101 width=4)\\n -> Seq Scan on t2 x5\\n(cost=0.00..2.01 rows=101 width=4)\\n -> Seq Scan on t2 x6\\n(cost=0.00..2.01 rows=101 width=4)\\n -> Seq Scan on t2 x7 (cost=0.00..2.01\\nrows=101 width=4)\\n -> Seq Scan on t2 x1 (cost=0.00..2.01 rows=101 width=4)\\n -> Materialize (cost=0.00..2.51 rows=101 width=4)\\n -> Seq Scan on t2 x8 (cost=0.00..2.01 rows=101 width=4)\\n -> Materialize (cost=0.00..2.51 rows=101 width=4)\\n -> Seq Scan on t2 x9 (cost=0.00..2.01 rows=101 width=4)\\n(32 rows)\\n\\nObviously, that happens because selectivity for join with t2 is\\noverestimated and multiplied many times. And improvements in this\\narea seems quite hard for me.\\n\\nBut I think there is another issue in sequential scan cost. We have\\nzero startup cost for sequential scan. But why? As I get sequential\\nscan should estimate amount of resources we need to spend in order to\\nstart returning rows. So, in our example when we expect finding 1 row\\nfrom the table, in average we have to scan half of table before find\\nthis row. Thus, startup_cost should be about half of total cost.\\nAttached POC patch implements that. In the given example sequential\\nscan turns out back to index scan.\\n\\nAny thoughts?\\n\\n------\\nAlexander Korotkov\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres Company\",\n \"msg_date\": \"Thu, 30 Aug 2018 16:56:53 +0300\",\n \"msg_from\": \"Alexander Korotkov <a.korotkov@postgrespro.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Startup cost of sequential scan\"\n },\n {\n \"msg_contents\": \"Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\\n> But I think there is another issue in sequential scan cost. We have\\n> zero startup cost for sequential scan. But why?\\n\\nBecause it's what the mental model of startup cost says it should be.\\nAlso, I do not think we can change that without a whole lot of unpleasant\\nside effects on cases that work well today. Have you even checked to\\nsee how much of the regression tests break with this change?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 10:04:55 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Startup cost of sequential scan\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 30, 2018 at 5:05 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\\n> > But I think there is another issue in sequential scan cost. We have\\n> > zero startup cost for sequential scan. But why?\\n>\\n> Because it's what the mental model of startup cost says it should be.\\n\\nRight. So as I understand. The model is that we start sequential\\nscan immediately, and then find one row uniformly distributed over the\\nwhole table.\\n\\n From this model we make a conclusion that we're starting getting rows\\nfrom sequential scan sooner than from index scan. And this conclusion\\ndoesn't reflect reality. If even estimates for join with t2 wouldn't\\nbe wrong, then our plan for LIMIT query would be still bad, because it\\nwould be still faster to get that one row using index scan rather than\\nsequential scan.\\n\\nSo, if understand the mental model correctly, our cost estimate for\\nLIMIT query is based on the idea that we need to fetch only *fraction*\\nof row from t1 in order to get 100 resulting rows. This is obviously\\nwrong, because we're always dealing with whole rows :) But I can't\\nimagine how we can take care of it without redesigning our whole\\ncosting model...\\n\\n> Also, I do not think we can change that without a whole lot of unpleasant\\n> side effects on cases that work well today. Have you even checked to\\n> see how much of the regression tests break with this change?\\n\\nI though it's to early to discuss this. But yet, I've checked this.\\nIt appears to be surprisingly few broken tests. Just some reordering\\nof tables in partition_join, select_parallel.------\\nAlexander Korotkov\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres Company\",\n \"msg_date\": \"Thu, 30 Aug 2018 17:31:36 +0300\",\n \"msg_from\": \"Alexander Korotkov <a.korotkov@postgrespro.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Startup cost of sequential scan\"\n },\n {\n \"msg_contents\": \"Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\\n> On Thu, Aug 30, 2018 at 5:05 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> Because it's what the mental model of startup cost says it should be.\\n\\n> From this model we make a conclusion that we're starting getting rows\\n> from sequential scan sooner than from index scan. And this conclusion\\n> doesn't reflect reality.\\n\\nNo, startup cost is not the \\\"time to find the first row\\\". It's overhead\\npaid before you even get to start examining rows.\\n\\nI'm disinclined to consider fundamental changes to our costing model\\non the basis of this example. The fact that the rowcount estimates are\\nso far off reality means that you're basically looking at \\\"garbage in,\\ngarbage out\\\" for the cost calculations --- and applying a small LIMIT\\njust magnifies that.\\n\\nIt'd be more useful to think first about how to make the selectivity\\nestimates better; after that, we might or might not still think there's\\na costing issue.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 10:58:38 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Startup cost of sequential scan\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 30.08.2018 17:58, Tom Lane wrote:\\n> Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\\n>> On Thu, Aug 30, 2018 at 5:05 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>>> Because it's what the mental model of startup cost says it should be.\\n>> From this model we make a conclusion that we're starting getting rows\\n>> from sequential scan sooner than from index scan. And this conclusion\\n>> doesn't reflect reality.\\n> No, startup cost is not the \\\"time to find the first row\\\". It's overhead\\n> paid before you even get to start examining rows.\\nBut it seems to me that calculation of cost in LIMIT node contradicts \\nwith this statement:\\n\\n pathnode->path.startup_cost +=\\n (subpath->total_cost - subpath->startup_cost)\\n * offset_rows / subpath->rows;\\n\\n\\n\\n>\\n> I'm disinclined to consider fundamental changes to our costing model\\n> on the basis of this example. The fact that the rowcount estimates are\\n> so far off reality means that you're basically looking at \\\"garbage in,\\n> garbage out\\\" for the cost calculations --- and applying a small LIMIT\\n> just magnifies that.\\n>\\n> It'd be more useful to think first about how to make the selectivity\\n> estimates better; after that, we might or might not still think there's\\n> a costing issue.\\n>\\n> \\t\\t\\tregards, tom lane\\n>\\n\\n-- \\nKonstantin Knizhnik\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 18:08:00 +0300\",\n \"msg_from\": \"Konstantin Knizhnik <k.knizhnik@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Startup cost of sequential scan\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 30, 2018 at 6:08 PM Konstantin Knizhnik\\n<k.knizhnik@postgrespro.ru> wrote:\\n> On 30.08.2018 17:58, Tom Lane wrote:\\n> > Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\\n> >> On Thu, Aug 30, 2018 at 5:05 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> >>> Because it's what the mental model of startup cost says it should be.\\n> >> From this model we make a conclusion that we're starting getting rows\\n> >> from sequential scan sooner than from index scan. And this conclusion\\n> >> doesn't reflect reality.\\n> > No, startup cost is not the \\\"time to find the first row\\\". It's overhead\\n> > paid before you even get to start examining rows.\\n> But it seems to me that calculation of cost in LIMIT node contradicts\\n> with this statement:\\n>\\n> pathnode->path.startup_cost +=\\n> (subpath->total_cost - subpath->startup_cost)\\n> * offset_rows / subpath->rows;\\n\\nWhy does it contradict? It just assumes that skipping OFFSET rows to\\nbe preliminary work before returning results rows...\\n\\n------\\nAlexander Korotkov\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres Company\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 18:23:30 +0300\",\n \"msg_from\": \"Alexander Korotkov <a.korotkov@postgrespro.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Startup cost of sequential scan\"\n },\n {\n \"msg_contents\": \">>>>> \\\"Konstantin\\\" == Konstantin Knizhnik <k.knizhnik@postgrespro.ru> writes:\\n\\n >> No, startup cost is not the \\\"time to find the first row\\\". It's\\n >> overhead paid before you even get to start examining rows.\\n\\n Konstantin> But it seems to me that calculation of cost in LIMIT node\\n Konstantin> contradicts with this statement:\\n\\nThe model (assuming I understand it rightly) is that what we're actually\\ntracking is a startup cost and a per-output-row cost, but for comparison\\npurposes we actually store the rows and the computed total, rather than\\njust the per-row cost:\\n\\nrows\\nstartup_cost\\ntotal_cost = startup_cost + (rows * per_row_cost)\\n\\nSo what Limit is doing the for the offset count is recovering the\\nsubpath's per_row_cost from (total_cost - startup_cost)/rows, and then\\nscaling that by the number of rows in the offset (which are being\\ndiscarded), and adding that to the startup cost. So this is saying: the\\nstartup cost for OFFSET N is the startup cost of the subplan, plus the\\ncost of fetching N rows from the subplan. (And after fetching those N\\nrows, we still haven't found the first row that we will actually\\nreturn.)\\n\\nFor LIMIT N, we instead replace the old total cost with a new one\\ncalculated from the startup cost plus N times the subplan's per-row\\ncost.\\n\\n-- \\nAndrew (irc:RhodiumToad)\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 16:23:44 +0100\",\n \"msg_from\": \"Andrew Gierth <andrew@tao11.riddles.org.uk>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Startup cost of sequential scan\"\n },\n {\n \"msg_contents\": \"Andrew Gierth <andrew@tao11.riddles.org.uk> writes:\\n> The model (assuming I understand it rightly) is that what we're actually\\n> tracking is a startup cost and a per-output-row cost, but for comparison\\n> purposes we actually store the rows and the computed total, rather than\\n> just the per-row cost:\\n\\n> rows\\n> startup_cost\\n> total_cost = startup_cost + (rows * per_row_cost)\\n\\nRight. I tend to think of it differently: the cost to retrieve the\\nfirst K rows out of a total of N is\\n\\tstartup_cost + (total_cost - startup_cost) * K/N\\nbut that comes out to the same thing as what Andrew said.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 11:33:19 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Startup cost of sequential scan\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 30, 2018 at 5:58 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\\n> > On Thu, Aug 30, 2018 at 5:05 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> >> Because it's what the mental model of startup cost says it should be.\\n>\\n> > From this model we make a conclusion that we're starting getting rows\\n> > from sequential scan sooner than from index scan. And this conclusion\\n> > doesn't reflect reality.\\n>\\n> No, startup cost is not the \\\"time to find the first row\\\". It's overhead\\n> paid before you even get to start examining rows.\\n>\\n> I'm disinclined to consider fundamental changes to our costing model\\n> on the basis of this example. The fact that the rowcount estimates are\\n> so far off reality means that you're basically looking at \\\"garbage in,\\n> garbage out\\\" for the cost calculations --- and applying a small LIMIT\\n> just magnifies that.\\n>\\n> It'd be more useful to think first about how to make the selectivity\\n> estimates better; after that, we might or might not still think there's\\n> a costing issue.\\n\\nI understand that startup cost is not \\\"time to find the first row\\\".\\nBut I think this example highlight not one but two issues.\\n1) Row count estimates for joins are wrong.\\n2) Rows are assumed to be continuous while in reality they are\\ndiscrete. So, if we reverse the assumptions made in LIMIT clause\\nestimation, we may say that it's basically assuming that we need to\\nfetch only fraction of row from the sequential scan node. And in the\\ncase we really fetch 101 rows in each join with t2, this logic would\\nstill bring us to the bad plan. And now I'm not proposing go rush\\nredesigning planner to fix that. I just think it's probably something\\nworth discussion.\\n\\n------\\nAlexander Korotkov\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres Company\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 18:38:04 +0300\",\n \"msg_from\": \"Alexander Korotkov <a.korotkov@postgrespro.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Startup cost of sequential scan\"\n },\n {\n \"msg_contents\": \"Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\\n> I understand that startup cost is not \\\"time to find the first row\\\".\\n> But I think this example highlight not one but two issues.\\n> 1) Row count estimates for joins are wrong.\\n\\nYup.\\n\\n> 2) Rows are assumed to be continuous while in reality they are\\n> discrete.\\n\\nWhere do you get that from?\\n\\nThe problem this example is highlighting is mostly just the bad\\njoin size estimate, imo. It's not at all clear that the planner\\nwould still do the wrong thing if that were fixed. In fact,\\nif I replace the contents of t2 with some other distribution,\\nsuch as\\n\\tinsert into t2 (select i from generate_series(1,100)i);\\nI get a much better join size estimate *and* a sane plan, even\\nin the LIMIT case. So the problem here is just with the join\\nsize estimate.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 11:55:27 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Startup cost of sequential scan\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 30, 2018 at 10:04 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\\n>> But I think there is another issue in sequential scan cost. We have\\n>> zero startup cost for sequential scan. But why?\\n>\\n> Because it's what the mental model of startup cost says it should be.\\n\\nWhose mental model? I guess the Tom Lane mind is the canonical one\\nfor this project, but I'm not sure that it entirely agrees with mine.\\nIIRC, it was previously proposed that we ought to charge\\nrandom_page_cost for the first block of a sequential scan, because at\\npresent the cost of fetching 1 block differs depending on whether we\\nare fetching it from a heap or an index, which seems unprincipled.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 15:43:59 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Startup cost of sequential scan\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> Whose mental model? I guess the Tom Lane mind is the canonical one\\n> for this project, but I'm not sure that it entirely agrees with mine.\\n\\nSince the fact that we have a notion of startup cost at all is entirely my\\nfault, I don't feel shy about claiming to have the authoritative view of\\nwhat it means.\\n\\n(Whether that's adequately documented is another question :-()\\n\\n> IIRC, it was previously proposed that we ought to charge\\n> random_page_cost for the first block of a sequential scan, because at\\n> present the cost of fetching 1 block differs depending on whether we\\n> are fetching it from a heap or an index, which seems unprincipled.\\n\\nThat might well be a sane thing to do ... but it'd still be part of run\\ncost not startup cost.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 17:38:36 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Startup cost of sequential scan\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 30, 2018 at 6:55 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Alexander Korotkov <a.korotkov@postgrespro.ru> writes:\\n> > I understand that startup cost is not \\\"time to find the first row\\\".\\n> > But I think this example highlight not one but two issues.\\n> > 1) Row count estimates for joins are wrong.\\n>\\n> Yup.\\n>\\n> > 2) Rows are assumed to be continuous while in reality they are\\n> > discrete.\\n>\\n> Where do you get that from?\\n\\nI made a similar example, where estimates are good. It's quite\\nartificial, because it selects small limit from enormous virtual table\\nconstructed by cross joins. But it illustrates how our model,\\nassuming tuples to be continuous, can differ from reality.\\n\\ncreate table t1 (id int primary key, k int);\\ncreate table t2 (id int);\\n\\ninsert into t1 (select i, i from generate_series(1,1000000) i);\\ninsert into t2 (select 0 from generate_series(1,100) i);\\nvacuum analyze t1, t2;\\n\\nBy default, the query is processed using sequential scan for t1.\\n\\n# explain analyze select * from t1,t2 x1,t2 x2,t2 x3,t2 x4,t2 x5 where\\nt1.id = 500000 limit 100;\\n QUERY PLAN\\n--------------------------------------------------------------------------------------------------------------------------------------\\n Limit (cost=0.00..1.26 rows=100 width=28) (actual\\ntime=32.879..32.926 rows=100 loops=1)\\n -> Nested Loop (cost=0.00..126279562.00 rows=10000000000\\nwidth=28) (actual time=32.878..32.912 rows=100 loops=1)\\n -> Nested Loop (cost=0.00..1279559.75 rows=100000000\\nwidth=24) (actual time=32.873..32.873 rows=1 loops=1)\\n -> Nested Loop (cost=0.00..29557.50 rows=1000000\\nwidth=20) (actual time=32.868..32.868 rows=1 loops=1)\\n -> Nested Loop (cost=0.00..17055.25 rows=10000\\nwidth=16) (actual time=32.864..32.864 rows=1 loops=1)\\n -> Nested Loop (cost=0.00..16928.00\\nrows=100 width=12) (actual time=32.856..32.856 rows=1 loops=1)\\n -> Seq Scan on t1\\n(cost=0.00..16925.00 rows=1 width=8) (actual time=32.841..32.841\\nrows=1 loops=1)\\n Filter: (id = 500000)\\n Rows Removed by Filter: 501983\\n -> Seq Scan on t2 x1\\n(cost=0.00..2.00 rows=100 width=4) (actual time=0.011..0.011 rows=1\\nloops=1)\\n -> Materialize (cost=0.00..2.50 rows=100\\nwidth=4) (actual time=0.007..0.007 rows=1 loops=1)\\n -> Seq Scan on t2 x2\\n(cost=0.00..2.00 rows=100 width=4) (actual time=0.003..0.003 rows=1\\nloops=1)\\n -> Materialize (cost=0.00..2.50 rows=100\\nwidth=4) (actual time=0.003..0.003 rows=1 loops=1)\\n -> Seq Scan on t2 x3 (cost=0.00..2.00\\nrows=100 width=4) (actual time=0.002..0.003 rows=1 loops=1)\\n -> Materialize (cost=0.00..2.50 rows=100 width=4)\\n(actual time=0.003..0.003 rows=1 loops=1)\\n -> Seq Scan on t2 x4 (cost=0.00..2.00 rows=100\\nwidth=4) (actual time=0.002..0.003 rows=1 loops=1)\\n -> Materialize (cost=0.00..2.50 rows=100 width=4) (actual\\ntime=0.004..0.024 rows=100 loops=1)\\n -> Seq Scan on t2 x5 (cost=0.00..2.00 rows=100\\nwidth=4) (actual time=0.003..0.010 rows=100 loops=1)\\n Planning Time: 0.372 ms\\n Execution Time: 32.987 ms\\n(20 rows)\\n\\nBut if we disable sequential scan and bitmap scan then it would be\\nindex scan, which is much faster.\\n# set enable_seqscan = off;\\n# set enable_bitmapscan = off;\\n\\n# explain analyze select * from t1,t2 x1,t2 x2,t2 x3,t2 x4,t2 x5 where\\nt1.id = 500000 limit 100;\\n\\nQUERY PLAN\\n---------------------------------------------------------------------------------------------------------------------------------------------------------\\n Limit (cost=50000000000.43..50000000001.69 rows=100 width=28)\\n(actual time=0.041..0.092 rows=100 loops=1)\\n -> Nested Loop (cost=50000000000.43..50126262645.44\\nrows=10000000000 width=28) (actual time=0.040..0.079 rows=100 loops=1)\\n -> Nested Loop (cost=40000000000.43..40001262643.19\\nrows=100000000 width=24) (actual time=0.036..0.036 rows=1 loops=1)\\n -> Nested Loop (cost=30000000000.42..30000012640.94\\nrows=1000000 width=20) (actual time=0.033..0.033 rows=1 loops=1)\\n -> Nested Loop\\n(cost=20000000000.42..20000000138.69 rows=10000 width=16) (actual\\ntime=0.029..0.029 rows=1 loops=1)\\n -> Nested Loop\\n(cost=10000000000.42..10000000011.44 rows=100 width=12) (actual\\ntime=0.023..0.023 rows=1 loops=1)\\n -> Index Scan using t1_pkey on t1\\n(cost=0.42..8.44 rows=1 width=8) (actual time=0.015..0.015 rows=1\\nloops=1)\\n Index Cond: (id = 500000)\\n -> Seq Scan on t2 x1\\n(cost=10000000000.00..10000000002.00 rows=100 width=4) (actual\\ntime=0.007..0.007 rows=1 loops=1)\\n -> Materialize\\n(cost=10000000000.00..10000000002.50 rows=100 width=4) (actual\\ntime=0.004..0.005 rows=1 loops=1)\\n -> Seq Scan on t2 x2\\n(cost=10000000000.00..10000000002.00 rows=100 width=4) (actual\\ntime=0.003..0.003 rows=1 loops=1)\\n -> Materialize\\n(cost=10000000000.00..10000000002.50 rows=100 width=4) (actual\\ntime=0.003..0.003 rows=1 loops=1)\\n -> Seq Scan on t2 x3\\n(cost=10000000000.00..10000000002.00 rows=100 width=4) (actual\\ntime=0.003..0.003 rows=1 loops=1)\\n -> Materialize (cost=10000000000.00..10000000002.50\\nrows=100 width=4) (actual time=0.003..0.003 rows=1 loops=1)\\n -> Seq Scan on t2 x4\\n(cost=10000000000.00..10000000002.00 rows=100 width=4) (actual\\ntime=0.002..0.003 rows=1 loops=1)\\n -> Materialize (cost=10000000000.00..10000000002.50\\nrows=100 width=4) (actual time=0.003..0.026 rows=100 loops=1)\\n -> Seq Scan on t2 x5\\n(cost=10000000000.00..10000000002.00 rows=100 width=4) (actual\\ntime=0.003..0.010 rows=100 loops=1)\\n Planning Time: 0.274 ms\\n Execution Time: 0.150 ms\\n(19 rows)\\n\\n From this example, I get that there is a distinct issue that we assume\\nrows to be continuous.\\n\\n------\\nAlexander Korotkov\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres Company\\n\\n\",\n \"msg_date\": \"Fri, 31 Aug 2018 13:50:32 +0300\",\n \"msg_from\": \"Alexander Korotkov <a.korotkov@postgrespro.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Startup cost of sequential scan\"\n }\n]","isConversational":false}}},{"rowIdx":47199,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi all,\nWe want to introduce disk quota feature into Postgres.\n\n*Why disk quota*\n*In a multi-tenant environment, there is a requirement to limit the disk\nquota that database/schema/table can be written or a user can consume for\ndifferent organizations.*\n*Meanwhile, other databases such as Oracle, Teradata, DB2 have already\nsupported disk quota feature.*\n\n*Heikki has already implemented disk quota feature in Postgres as an\nextension pg_quota <https://github.com/hlinnaka/pg_quota>. We plan to\nenhance disk quota feature based on Heikki's implementation.*\n\n*Scope*\n\nThe scope of disk quota feature is to limit the disk usage of\ndatabase/schema/table objects and users.\n\nHere table means heap table, ao table, index table, partition table and\nassociated table( toast table, visible table, large object etc.). Schema\ndisk quota is the disk quota of all the tables in this schema. Database\ndisk quota is the disk quota of all the tables in this database.\n\nUser's quota is the size of all the tables whose owner are this user.\nOut of Scope: Note that spill files, xlogs, clogs and logs are not\nconsidered for database object level disk quota at this stage.\n\n\n*Design*\nWe propose disk quota with the following components:\n\n1. Quota Setting Store is where the disk quota setting to be stored and\naccessed. DBA or object owner uses SQL queries to configure the disk quota\nfor each database objects.\n\n\n\n*2. Quota Change Detector is the monitor of size change of database objects\nin Postgres. It will write change information to shared memory to notify\nQuota Size Checker. The implementation of Change Detector could be hooks in\nsmgr_extend/smgr_unlink/smgr_truncate when modifying the size of a heap\ntable. The hooks will write to shared memory in a batched way to reduce the\nimpact on OLTP performance.3. Quota Size Checker is implemented as a worker\nprocess. It maintains the current disk usage for each database objects and\nusers, and compare them with settings in Quota Setting Store. If it detects\nthe disk usage hit the quota redzone(either upon or below), it will notify\nQuota Enforcement Operator. 4. Quota Enforcement Operator has two roles:\none is to check the disk quota permission before queries are\nexecuted(QueryBeforeRun Check), the other is to cancel the running queries\nwhen it reaches the disk quota limit dynamically(QueryRunning Check). Quota\nEnforcement Operator uses shared memory to store the enforcement\ninformation to guarantee a quick check. *\n\n\n*To implement the right proof of concept, we want to receive feedback from\nthe community from the following aspects: *\nQ1. User Interface: when setting a disk quota,\n Option 1 is to use *insert into quota.config(user1, 10G)*\n Option 2 is to use UDF *select set_quota(\"role\",\"user1\",10G)*\n Option 3 is to use native SQL syntax *create disk quota on ROLE user1\n10G, or create disk quota on SCHEMA s1 25G;*\nQ2. Quota Setting Store using user table or catalog?\n Option 1 is to create a schema called quota for each database and write\nquota settings into quota.config table, only DBA could modify it. This\ncorresponds to Q1.option1\n Option 2 is to store quota settings into the catalog. For Schema and\nTable write them to database level catalog. For database or user, write\nthem to either database level or global catalog.\n\nI personally prefer Q1's option3 and Q2's option2, since it makes disk\nquota more like a native feature. We could support the quota worker process\nimplementation as an extension for now, but in the long term, disk quota is\nlike a fundamental feature of a database and should be a native feature\njust like other databases. So store quota conf into catalog and supply\nnative syntax is better.\n\nOpen Problem\nWe prepare to implement Quota Size Checker as a worker process. Worker\nprocess needs to connect to a database to build the disk usage map and\nquota map(e.g. a user/shcema's disk usage on a given database) But one\nworker process can only bind to one database(InitPostgres(dbname)) at the\ninitialization stage. It results in we need a separate worker process for\neach database. The solution to this problem is not straightforward, here\nare some ideas:\n1 To make worker process could retrieve and cache information from all the\ndatabases. As Tom Lane pointed out that it needs to flush all the database\nspecific thing, like relcache syscache etc.\n2 Keep one worker process for each database. But using a parent/global\nquota worker process to manage the lifecycle of database level worker\nprocesses. It could handle the newly created database(avoid restart\ndatabase) and save resource when a database is not used. But this needs to\nchange worker process to be hierarchical. Postmaster becomes the grandfather\n of database level worker processes in this case.\n\nAny better ideas on it?\n\n\n\n-- \nThanks\n\nHubert Zhang\n\nHi all,We want to introduce disk quota feature into Postgres.Why disk quotaIn a multi-tenant environment, there is a requirement to limit the disk quota that database/schema/table can be written or a user can consume for different organizations.Meanwhile, other databases such as Oracle, Teradata, DB2 have already supported disk quota feature.Heikki has already implemented disk quota feature in Postgres as an extension pg_quota. We plan to enhance disk quota feature based on Heikki's implementation.ScopeThe scope of disk quota feature is to limit the disk usage of database/schema/table objects and users.Here table means heap table, ao table, index table, partition table and associated table( toast table, visible table, large object etc.). Schema disk quota is the disk quota of all the tables in this schema. Database disk quota is the disk quota of all the tables in this database.User's quota is the size of all the tables whose owner are this user.Out of Scope: Note that spill files, xlogs, clogs and logs are not considered for database object level disk quota at this stage.DesignWe propose disk quota with the following components:1. Quota Setting Store is where the disk quota setting to be stored and accessed. DBA or object owner uses SQL queries to configure the disk quota for each database objects. 2. Quota Change Detector is the monitor of size change of database objects in Postgres. It will write change information to shared memory to notify Quota Size Checker. The implementation of Change Detector could be hooks in smgr_extend/smgr_unlink/smgr_truncate when modifying the size of a heap table. The hooks will write to shared memory in a batched way to reduce the impact on OLTP performance.3. Quota Size Checker is implemented as a worker process. It maintains the current disk usage for each database objects and users, and compare them with settings in Quota Setting Store. If it detects the disk usage hit the quota redzone(either upon or below), it will notify Quota Enforcement Operator. 4. Quota Enforcement Operator has two roles: one is to check the disk quota permission before queries are executed(QueryBeforeRun Check), the other is to cancel the running queries when it reaches the disk quota limit dynamically(QueryRunning Check). Quota Enforcement Operator uses shared memory to store the enforcement information to guarantee a quick check. To implement the right proof of concept, we want to receive feedback from the community from the following aspects: Q1. User Interface: when setting a disk quota, Option 1 is to use insert into quota.config(user1, 10G) Option 2 is to use UDF select set_quota(\"role\",\"user1\",10G) Option 3 is to use native SQL syntax create disk quota on ROLE user1 10G, or create disk quota on SCHEMA s1 25G;Q2. Quota Setting Store using user table or catalog? Option 1 is to create a schema called quota for each database and write quota settings into quota.config table, only DBA could modify it. This corresponds to Q1.option1 Option 2 is to store quota settings into the catalog. For Schema and Table write them to database level catalog. For database or user, write them to either database level or global catalog. I personally prefer Q1's option3 and Q2's option2, since it makes disk quota more like a native feature. We could support the quota worker process implementation as an extension for now, but in the long term, disk quota is like a fundamental feature of a database and should be a native feature just like other databases. So store quota conf into catalog and supply native syntax is better.Open ProblemWe prepare to implement Quota Size Checker as a worker process. Worker process needs to connect to a database to build the disk usage map and quota map(e.g. a user/shcema's disk usage on a given database) But one worker process can only bind to one database(InitPostgres(dbname)) at the initialization stage. It results in we need a separate worker process for each database. The solution to this problem is not straightforward, here are some ideas:1 To make worker process could retrieve and cache information from all the databases. As Tom Lane pointed out that it needs to flush all the database specific thing, like relcache syscache etc.2 Keep one worker process for each database. But using a parent/global quota worker process to manage the lifecycle of database level worker processes. It could handle the newly created database(avoid restart database) and save resource when a database is not used. But this needs to change worker process to be hierarchical. Postmaster becomes the grandfather of database level worker processes in this case.Any better ideas on it?-- ThanksHubert Zhang","msg_date":"Thu, 30 Aug 2018 21:57:41 +0800","msg_from":"Hubert Zhang <hzhang@pivotal.io>","msg_from_op":true,"msg_subject":"Proposal for disk quota feature"},{"msg_contents":"On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n\n> 2 Keep one worker process for each database. But using a parent/global\n> quota worker process to manage the lifecycle of database level worker\n> processes. It could handle the newly created database(avoid restart\n> database) and save resource when a database is not used. But this needs to\n> change worker process to be hierarchical. Postmaster becomes the grandfather\n> of database level worker processes in this case.\n\nI am using background workers this way in 9.5 at $work.\n\nIn my case, one worker lives forever, wakes up on a set period, and\nstarts a short-lived worker for every database, waiting for each\none before starting the next.\n\nIt was straightforward to implement. Looking back over the code,\nI see the global worker assigns its own PID to worker.bgw_notify_pid\nof each of its children, and also obtains a handle for each child\nfrom RegisterDynamicBackgroundWorker().\n\nI imagine the global quota worker would prefer to start workers\nfor every database and then just wait for notifications from any\nof them, but that seems equally straightforward at first glance.\n\n-Chap\n\n","msg_date":"Thu, 30 Aug 2018 10:22:37 -0400","msg_from":"Chapman Flack <chap@anastigmatix.net>","msg_from_op":false,"msg_subject":"Re: Proposal for disk quota feature"},{"msg_contents":"2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\n\n> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n>\n> > 2 Keep one worker process for each database. But using a parent/global\n> > quota worker process to manage the lifecycle of database level worker\n> > processes. It could handle the newly created database(avoid restart\n> > database) and save resource when a database is not used. But this needs\n> to\n> > change worker process to be hierarchical. Postmaster becomes the\n> grandfather\n> > of database level worker processes in this case.\n>\n> I am using background workers this way in 9.5 at $work.\n>\n> In my case, one worker lives forever, wakes up on a set period, and\n> starts a short-lived worker for every database, waiting for each\n> one before starting the next.\n>\n> It was straightforward to implement. Looking back over the code,\n> I see the global worker assigns its own PID to worker.bgw_notify_pid\n> of each of its children, and also obtains a handle for each child\n> from RegisterDynamicBackgroundWorker().\n>\n> I imagine the global quota worker would prefer to start workers\n> for every database and then just wait for notifications from any\n> of them, but that seems equally straightforward at first glance.\n>\n\nThere are servers with thousands databases. Worker per database is not good\nidea.\n\nIt should to share ideas, code with autovacuum process.\n\nNot sure, how to effective implementation based on bg workers can be. On\nservers with large set of databases, large set of tables it can identify\ntoo big table too late.\n\nIsn't better to implement some quotas on storage level?\n\nRegards\n\nPavel\n\n\n\n> -Chap\n>\n>\n\n2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n\n> 2 Keep one worker process for each database. But using a parent/global\n> quota worker process to manage the lifecycle of database level worker\n> processes. It could handle the newly created database(avoid restart\n> database) and save resource when a database is not used. But this needs to\n> change worker process to be hierarchical. Postmaster becomes the grandfather\n> of database level worker processes in this case.\n\nI am using background workers this way in 9.5 at $work.\n\nIn my case, one worker lives forever, wakes up on a set period, and\nstarts a short-lived worker for every database, waiting for each\none before starting the next.\n\nIt was straightforward to implement. Looking back over the code,\nI see the global worker assigns its own PID to worker.bgw_notify_pid\nof each of its children, and also obtains a handle for each child\nfrom RegisterDynamicBackgroundWorker().\n\nI imagine the global quota worker would prefer to start workers\nfor every database and then just wait for notifications from any\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\n\n-Chap","msg_date":"Thu, 30 Aug 2018 21:32:58 +0200","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":false,"msg_subject":"Re: Proposal for disk quota feature"},{"msg_contents":"Thanks Chapman.\n@Pavel, could you please explain more about your second suggestion \"implement\nsome quotas on storage level?\"\nWe will not keep the long-lived processes attach to all databases(just like\nyou mentioned servers with thousands of databases)\nAnd you are right, we could share ideas with autovacuum process, fork\nworker processes in need.\n\"autovacuum checks for tables that have had a large number of inserted,\nupdated or deleted tuples. These checks use the statistics collection\nfacility\"\ndiskquota process is similar to autovacuum at caring about insert, but the\ndifference is that it also care about vucuum full, truncate and drop. While\nupdate and delete may not be interested since no file change happens. So a\nseparate diskquota process is preferred.\n\nSo if we implemented disk quota as a full native feature, and in the first\ninitial version I prefer to implement the following features:\n1 Fork diskquota launcher process under Postmaster serverloop, which is\nlong-lived.\n2 Diskquota launcher process is responsible for creating diskquota worker\nprocess for every database.\n3 DIskquota setting is stored in a separate catalog table for each database.\n4 Initialization stage, Diskquota launcher process creates diskquota worker\nprocess for all the databases(traverse like autovacuum). Worker process\ncalculates disk usage of db objects and their diskquota setting. If any\ndb object exceeds its quota limit, put them into the blacklist in the\nshared memory, which will later be used by enforcement operator. Worker\nprocess exits when works are done.\n5 Running stage, Diskquota launcher process creates diskquota worker\nprocess for the database with a large number of insert, copy, truncate,\ndrop etc. or create disk quota statement. Worker process updates the file\nsize for db objects containing the result relation, and compare with the\ndiskquota setting. Again, if exceeds quota limit, put them into blacklist,\nremove from blacklist vice versa. Worker process exits when works are done\nand a GUC could control the frequency of worker process restart to a\nspecific database. As you know, this GUC also controls the delay when we do\nenforcement.\n6 Enforcement. When postgres backend executes queries, check the blacklist\nin shared memory to determine whether the query is allowed(before execute)\nor need rollback(is executing)?\n\nIf we implemented disk quota as an extension, we could just use background\nworker to start diskquota launcher process and use\nRegisterDynamicBackgroundWorker() to fork child diskquota worker processes\nby the launcher process as suggested by @Chapman. Diskquota setting could\nbe stored in user table in a separate schema for each database(Schema and\ntable created by create extension statement) just like what Heikki has done\nin pg_quota project. But in this case, we need to create extension for each\ndatabase before diskquota worker process can be set up for that database.\n\nAny comments on the above design and which is preferred, native feature or\nextension as the POC?\n\n\n-- Hubert\n\n\n\nOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com>\nwrote:\n\n>\n>\n> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\n>\n>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n>>\n>> > 2 Keep one worker process for each database. But using a parent/global\n>> > quota worker process to manage the lifecycle of database level worker\n>> > processes. It could handle the newly created database(avoid restart\n>> > database) and save resource when a database is not used. But this needs\n>> to\n>> > change worker process to be hierarchical. Postmaster becomes the\n>> grandfather\n>> > of database level worker processes in this case.\n>>\n>> I am using background workers this way in 9.5 at $work.\n>>\n>> In my case, one worker lives forever, wakes up on a set period, and\n>> starts a short-lived worker for every database, waiting for each\n>> one before starting the next.\n>>\n>> It was straightforward to implement. Looking back over the code,\n>> I see the global worker assigns its own PID to worker.bgw_notify_pid\n>> of each of its children, and also obtains a handle for each child\n>> from RegisterDynamicBackgroundWorker().\n>>\n>> I imagine the global quota worker would prefer to start workers\n>> for every database and then just wait for notifications from any\n>> of them, but that seems equally straightforward at first glance.\n>>\n>\n> There are servers with thousands databases. Worker per database is not\n> good idea.\n>\n> It should to share ideas, code with autovacuum process.\n>\n> Not sure, how to effective implementation based on bg workers can be. On\n> servers with large set of databases, large set of tables it can identify\n> too big table too late.\n>\n> Isn't better to implement some quotas on storage level?\n>\n> Regards\n>\n> Pavel\n>\n>\n>\n>> -Chap\n>>\n>>\n>\n\n\n-- \nThanks\n\nHubert Zhang\n\nThanks Chapman.@Pavel, could you please explain more about your second suggestion \"implement some quotas on storage level?\"We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n\n> 2 Keep one worker process for each database. But using a parent/global\n> quota worker process to manage the lifecycle of database level worker\n> processes. It could handle the newly created database(avoid restart\n> database) and save resource when a database is not used. But this needs to\n> change worker process to be hierarchical. Postmaster becomes the grandfather\n> of database level worker processes in this case.\n\nI am using background workers this way in 9.5 at $work.\n\nIn my case, one worker lives forever, wakes up on a set period, and\nstarts a short-lived worker for every database, waiting for each\none before starting the next.\n\nIt was straightforward to implement. Looking back over the code,\nI see the global worker assigns its own PID to worker.bgw_notify_pid\nof each of its children, and also obtains a handle for each child\nfrom RegisterDynamicBackgroundWorker().\n\nI imagine the global quota worker would prefer to start workers\nfor every database and then just wait for notifications from any\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\n\n-Chap\n\n\n-- ThanksHubert Zhang","msg_date":"Sun, 2 Sep 2018 20:18:56 +0800","msg_from":"Hubert Zhang <hzhang@pivotal.io>","msg_from_op":true,"msg_subject":"Re: Proposal for disk quota feature"},{"msg_contents":"Hi\n\n2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n\n> Thanks Chapman.\n> @Pavel, could you please explain more about your second suggestion \"implement\n> some quotas on storage level?\"\n>\n\nSee attached patch - it is very simple - and good enough for our purposes.\n\nRegards\n\nPavel\n\n\n\n> We will not keep the long-lived processes attach to all databases(just\n> like you mentioned servers with thousands of databases)\n> And you are right, we could share ideas with autovacuum process, fork\n> worker processes in need.\n> \"autovacuum checks for tables that have had a large number of inserted,\n> updated or deleted tuples. These checks use the statistics collection\n> facility\"\n> diskquota process is similar to autovacuum at caring about insert, but the\n> difference is that it also care about vucuum full, truncate and drop. While\n> update and delete may not be interested since no file change happens. So a\n> separate diskquota process is preferred.\n>\n> So if we implemented disk quota as a full native feature, and in the\n> first initial version I prefer to implement the following features:\n> 1 Fork diskquota launcher process under Postmaster serverloop, which is\n> long-lived.\n> 2 Diskquota launcher process is responsible for creating diskquota worker\n> process for every database.\n> 3 DIskquota setting is stored in a separate catalog table for each\n> database.\n> 4 Initialization stage, Diskquota launcher process creates diskquota worker\n> process for all the databases(traverse like autovacuum). Worker process\n> calculates disk usage of db objects and their diskquota setting. If any\n> db object exceeds its quota limit, put them into the blacklist in the\n> shared memory, which will later be used by enforcement operator. Worker\n> process exits when works are done.\n> 5 Running stage, Diskquota launcher process creates diskquota worker\n> process for the database with a large number of insert, copy, truncate,\n> drop etc. or create disk quota statement. Worker process updates the file\n> size for db objects containing the result relation, and compare with the\n> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\n> remove from blacklist vice versa. Worker process exits when works are\n> done and a GUC could control the frequency of worker process restart to a\n> specific database. As you know, this GUC also controls the delay when we do\n> enforcement.\n> 6 Enforcement. When postgres backend executes queries, check the blacklist\n> in shared memory to determine whether the query is allowed(before execute)\n> or need rollback(is executing)?\n>\n> If we implemented disk quota as an extension, we could just use background\n> worker to start diskquota launcher process and use\n> RegisterDynamicBackgroundWorker() to fork child diskquota worker\n> processes by the launcher process as suggested by @Chapman.\n> Diskquota setting could be stored in user table in a separate schema for\n> each database(Schema and table created by create extension statement) just\n> like what Heikki has done in pg_quota project. But in this case, we need to\n> create extension for each database before diskquota worker process can be\n> set up for that database.\n>\n> Any comments on the above design and which is preferred, native feature or\n> extension as the POC?\n>\n>\n> -- Hubert\n>\n>\n>\n> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com>\n> wrote:\n>\n>>\n>>\n>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\n>>\n>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n>>>\n>>> > 2 Keep one worker process for each database. But using a parent/global\n>>> > quota worker process to manage the lifecycle of database level worker\n>>> > processes. It could handle the newly created database(avoid restart\n>>> > database) and save resource when a database is not used. But this\n>>> needs to\n>>> > change worker process to be hierarchical. Postmaster becomes the\n>>> grandfather\n>>> > of database level worker processes in this case.\n>>>\n>>> I am using background workers this way in 9.5 at $work.\n>>>\n>>> In my case, one worker lives forever, wakes up on a set period, and\n>>> starts a short-lived worker for every database, waiting for each\n>>> one before starting the next.\n>>>\n>>> It was straightforward to implement. Looking back over the code,\n>>> I see the global worker assigns its own PID to worker.bgw_notify_pid\n>>> of each of its children, and also obtains a handle for each child\n>>> from RegisterDynamicBackgroundWorker().\n>>>\n>>> I imagine the global quota worker would prefer to start workers\n>>> for every database and then just wait for notifications from any\n>>> of them, but that seems equally straightforward at first glance.\n>>>\n>>\n>> There are servers with thousands databases. Worker per database is not\n>> good idea.\n>>\n>> It should to share ideas, code with autovacuum process.\n>>\n>> Not sure, how to effective implementation based on bg workers can be. On\n>> servers with large set of databases, large set of tables it can identify\n>> too big table too late.\n>>\n>> Isn't better to implement some quotas on storage level?\n>>\n>> Regards\n>>\n>> Pavel\n>>\n>>\n>>\n>>> -Chap\n>>>\n>>>\n>>\n>\n>\n> --\n> Thanks\n>\n> Hubert Zhang\n>","msg_date":"Sun, 2 Sep 2018 14:44:40 +0200","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":false,"msg_subject":"Re: Proposal for disk quota feature"},{"msg_contents":"Thanks Pavel.\nYour patch did enforcement on storage level(md.c or we could also use\nsmgr_extend). It's straight forward.\nBut I prefer to implement disk_quota as a feature with following objectives:\n1 set/alter disk quota setting on different database objects, e.g. user,\ndatabase, schema etc. not only a general GUC, but we could set separate\nquota limit for a specific objects.\n2 enforcement operator should work at two positions: before query is\nrunning and when query is running. The latter one's implementation maybe\nsimilar to your patch.\n\nOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com>\nwrote:\n\n> Hi\n>\n> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>\n>> Thanks Chapman.\n>> @Pavel, could you please explain more about your second suggestion \"implement\n>> some quotas on storage level?\"\n>>\n>\n> See attached patch - it is very simple - and good enough for our purposes.\n>\n> Regards\n>\n> Pavel\n>\n>\n>\n>> We will not keep the long-lived processes attach to all databases(just\n>> like you mentioned servers with thousands of databases)\n>> And you are right, we could share ideas with autovacuum process, fork\n>> worker processes in need.\n>> \"autovacuum checks for tables that have had a large number of inserted,\n>> updated or deleted tuples. These checks use the statistics collection\n>> facility\"\n>> diskquota process is similar to autovacuum at caring about insert, but\n>> the difference is that it also care about vucuum full, truncate and drop.\n>> While update and delete may not be interested since no file change happens.\n>> So a separate diskquota process is preferred.\n>>\n>> So if we implemented disk quota as a full native feature, and in the\n>> first initial version I prefer to implement the following features:\n>> 1 Fork diskquota launcher process under Postmaster serverloop, which is\n>> long-lived.\n>> 2 Diskquota launcher process is responsible for creating diskquota\n>> worker process for every database.\n>> 3 DIskquota setting is stored in a separate catalog table for each\n>> database.\n>> 4 Initialization stage, Diskquota launcher process creates diskquota worker\n>> process for all the databases(traverse like autovacuum). Worker process\n>> calculates disk usage of db objects and their diskquota setting. If any\n>> db object exceeds its quota limit, put them into the blacklist in the\n>> shared memory, which will later be used by enforcement operator. Worker\n>> process exits when works are done.\n>> 5 Running stage, Diskquota launcher process creates diskquota worker\n>> process for the database with a large number of insert, copy, truncate,\n>> drop etc. or create disk quota statement. Worker process updates the file\n>> size for db objects containing the result relation, and compare with the\n>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\n>> remove from blacklist vice versa. Worker process exits when works are\n>> done and a GUC could control the frequency of worker process restart to a\n>> specific database. As you know, this GUC also controls the delay when we do\n>> enforcement.\n>> 6 Enforcement. When postgres backend executes queries, check the\n>> blacklist in shared memory to determine whether the query is allowed(before\n>> execute) or need rollback(is executing)?\n>>\n>> If we implemented disk quota as an extension, we could just use\n>> background worker to start diskquota launcher process and use\n>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\n>> processes by the launcher process as suggested by @Chapman.\n>> Diskquota setting could be stored in user table in a separate schema for\n>> each database(Schema and table created by create extension statement) just\n>> like what Heikki has done in pg_quota project. But in this case, we need to\n>> create extension for each database before diskquota worker process can be\n>> set up for that database.\n>>\n>> Any comments on the above design and which is preferred, native feature\n>> or extension as the POC?\n>>\n>>\n>> -- Hubert\n>>\n>>\n>>\n>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com>\n>> wrote:\n>>\n>>>\n>>>\n>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\n>>>\n>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n>>>>\n>>>> > 2 Keep one worker process for each database. But using a parent/global\n>>>> > quota worker process to manage the lifecycle of database level worker\n>>>> > processes. It could handle the newly created database(avoid restart\n>>>> > database) and save resource when a database is not used. But this\n>>>> needs to\n>>>> > change worker process to be hierarchical. Postmaster becomes the\n>>>> grandfather\n>>>> > of database level worker processes in this case.\n>>>>\n>>>> I am using background workers this way in 9.5 at $work.\n>>>>\n>>>> In my case, one worker lives forever, wakes up on a set period, and\n>>>> starts a short-lived worker for every database, waiting for each\n>>>> one before starting the next.\n>>>>\n>>>> It was straightforward to implement. Looking back over the code,\n>>>> I see the global worker assigns its own PID to worker.bgw_notify_pid\n>>>> of each of its children, and also obtains a handle for each child\n>>>> from RegisterDynamicBackgroundWorker().\n>>>>\n>>>> I imagine the global quota worker would prefer to start workers\n>>>> for every database and then just wait for notifications from any\n>>>> of them, but that seems equally straightforward at first glance.\n>>>>\n>>>\n>>> There are servers with thousands databases. Worker per database is not\n>>> good idea.\n>>>\n>>> It should to share ideas, code with autovacuum process.\n>>>\n>>> Not sure, how to effective implementation based on bg workers can be. On\n>>> servers with large set of databases, large set of tables it can identify\n>>> too big table too late.\n>>>\n>>> Isn't better to implement some quotas on storage level?\n>>>\n>>> Regards\n>>>\n>>> Pavel\n>>>\n>>>\n>>>\n>>>> -Chap\n>>>>\n>>>>\n>>>\n>>\n>>\n>> --\n>> Thanks\n>>\n>> Hubert Zhang\n>>\n>\n>\n\n\n-- \nThanks\n\nHubert Zhang\n\nThanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \"implement some quotas on storage level?\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n\n> 2 Keep one worker process for each database. But using a parent/global\n> quota worker process to manage the lifecycle of database level worker\n> processes. It could handle the newly created database(avoid restart\n> database) and save resource when a database is not used. But this needs to\n> change worker process to be hierarchical. Postmaster becomes the grandfather\n> of database level worker processes in this case.\n\nI am using background workers this way in 9.5 at $work.\n\nIn my case, one worker lives forever, wakes up on a set period, and\nstarts a short-lived worker for every database, waiting for each\none before starting the next.\n\nIt was straightforward to implement. Looking back over the code,\nI see the global worker assigns its own PID to worker.bgw_notify_pid\nof each of its children, and also obtains a handle for each child\nfrom RegisterDynamicBackgroundWorker().\n\nI imagine the global quota worker would prefer to start workers\nfor every database and then just wait for notifications from any\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\n\n-Chap\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang","msg_date":"Mon, 3 Sep 2018 09:49:22 +0800","msg_from":"Hubert Zhang <hzhang@pivotal.io>","msg_from_op":true,"msg_subject":"Re: Proposal for disk quota feature"},{"msg_contents":"2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n\n> Thanks Pavel.\n> Your patch did enforcement on storage level(md.c or we could also use\n> smgr_extend). It's straight forward.\n> But I prefer to implement disk_quota as a feature with following\n> objectives:\n> 1 set/alter disk quota setting on different database objects, e.g. user,\n> database, schema etc. not only a general GUC, but we could set separate\n> quota limit for a specific objects.\n> 2 enforcement operator should work at two positions: before query is\n> running and when query is running. The latter one's implementation maybe\n> similar to your patch.\n>\n\nThe patch was just example. The resource quotes should be more complex -\nper partition, table, schema, database, user - so GUC are possible, but not\nvery user friendly.\n\nOur case is specific, but not too much. The servers are used for\nmultidimensional analyses - and some tables can grow too fast (COPY, INSERT\nSELECT). We need to solve limits immediately. The implementation is simple,\nso I did it. Same implementation on database level, or schema level needs\nsome more locks, so it will not be too effective. The resource management\ncan be complex very complex, and I expect so it will be hard work.\n\nRegards\n\nPavel\n\n\n> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com>\n> wrote:\n>\n>> Hi\n>>\n>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>\n>>> Thanks Chapman.\n>>> @Pavel, could you please explain more about your second suggestion \"implement\n>>> some quotas on storage level?\"\n>>>\n>>\n>> See attached patch - it is very simple - and good enough for our\n>> purposes.\n>>\n>> Regards\n>>\n>> Pavel\n>>\n>>\n>>\n>>> We will not keep the long-lived processes attach to all databases(just\n>>> like you mentioned servers with thousands of databases)\n>>> And you are right, we could share ideas with autovacuum process, fork\n>>> worker processes in need.\n>>> \"autovacuum checks for tables that have had a large number of inserted,\n>>> updated or deleted tuples. These checks use the statistics collection\n>>> facility\"\n>>> diskquota process is similar to autovacuum at caring about insert, but\n>>> the difference is that it also care about vucuum full, truncate and drop.\n>>> While update and delete may not be interested since no file change happens.\n>>> So a separate diskquota process is preferred.\n>>>\n>>> So if we implemented disk quota as a full native feature, and in the\n>>> first initial version I prefer to implement the following features:\n>>> 1 Fork diskquota launcher process under Postmaster serverloop, which is\n>>> long-lived.\n>>> 2 Diskquota launcher process is responsible for creating diskquota\n>>> worker process for every database.\n>>> 3 DIskquota setting is stored in a separate catalog table for each\n>>> database.\n>>> 4 Initialization stage, Diskquota launcher process creates diskquota worker\n>>> process for all the databases(traverse like autovacuum). Worker process\n>>> calculates disk usage of db objects and their diskquota setting. If any\n>>> db object exceeds its quota limit, put them into the blacklist in the\n>>> shared memory, which will later be used by enforcement operator. Worker\n>>> process exits when works are done.\n>>> 5 Running stage, Diskquota launcher process creates diskquota worker\n>>> process for the database with a large number of insert, copy, truncate,\n>>> drop etc. or create disk quota statement. Worker process updates the file\n>>> size for db objects containing the result relation, and compare with the\n>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\n>>> remove from blacklist vice versa. Worker process exits when works are\n>>> done and a GUC could control the frequency of worker process restart to a\n>>> specific database. As you know, this GUC also controls the delay when we do\n>>> enforcement.\n>>> 6 Enforcement. When postgres backend executes queries, check the\n>>> blacklist in shared memory to determine whether the query is allowed(before\n>>> execute) or need rollback(is executing)?\n>>>\n>>> If we implemented disk quota as an extension, we could just use\n>>> background worker to start diskquota launcher process and use\n>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\n>>> processes by the launcher process as suggested by @Chapman.\n>>> Diskquota setting could be stored in user table in a separate schema for\n>>> each database(Schema and table created by create extension statement) just\n>>> like what Heikki has done in pg_quota project. But in this case, we need to\n>>> create extension for each database before diskquota worker process can be\n>>> set up for that database.\n>>>\n>>> Any comments on the above design and which is preferred, native feature\n>>> or extension as the POC?\n>>>\n>>>\n>>> -- Hubert\n>>>\n>>>\n>>>\n>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com>\n>>> wrote:\n>>>\n>>>>\n>>>>\n>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\n>>>>\n>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n>>>>>\n>>>>> > 2 Keep one worker process for each database. But using a\n>>>>> parent/global\n>>>>> > quota worker process to manage the lifecycle of database level worker\n>>>>> > processes. It could handle the newly created database(avoid restart\n>>>>> > database) and save resource when a database is not used. But this\n>>>>> needs to\n>>>>> > change worker process to be hierarchical. Postmaster becomes the\n>>>>> grandfather\n>>>>> > of database level worker processes in this case.\n>>>>>\n>>>>> I am using background workers this way in 9.5 at $work.\n>>>>>\n>>>>> In my case, one worker lives forever, wakes up on a set period, and\n>>>>> starts a short-lived worker for every database, waiting for each\n>>>>> one before starting the next.\n>>>>>\n>>>>> It was straightforward to implement. Looking back over the code,\n>>>>> I see the global worker assigns its own PID to worker.bgw_notify_pid\n>>>>> of each of its children, and also obtains a handle for each child\n>>>>> from RegisterDynamicBackgroundWorker().\n>>>>>\n>>>>> I imagine the global quota worker would prefer to start workers\n>>>>> for every database and then just wait for notifications from any\n>>>>> of them, but that seems equally straightforward at first glance.\n>>>>>\n>>>>\n>>>> There are servers with thousands databases. Worker per database is not\n>>>> good idea.\n>>>>\n>>>> It should to share ideas, code with autovacuum process.\n>>>>\n>>>> Not sure, how to effective implementation based on bg workers can be.\n>>>> On servers with large set of databases, large set of tables it can identify\n>>>> too big table too late.\n>>>>\n>>>> Isn't better to implement some quotas on storage level?\n>>>>\n>>>> Regards\n>>>>\n>>>> Pavel\n>>>>\n>>>>\n>>>>\n>>>>> -Chap\n>>>>>\n>>>>>\n>>>>\n>>>\n>>>\n>>> --\n>>> Thanks\n>>>\n>>> Hubert Zhang\n>>>\n>>\n>>\n>\n>\n> --\n> Thanks\n>\n> Hubert Zhang\n>\n\n2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \"implement some quotas on storage level?\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n\n> 2 Keep one worker process for each database. But using a parent/global\n> quota worker process to manage the lifecycle of database level worker\n> processes. It could handle the newly created database(avoid restart\n> database) and save resource when a database is not used. But this needs to\n> change worker process to be hierarchical. Postmaster becomes the grandfather\n> of database level worker processes in this case.\n\nI am using background workers this way in 9.5 at $work.\n\nIn my case, one worker lives forever, wakes up on a set period, and\nstarts a short-lived worker for every database, waiting for each\none before starting the next.\n\nIt was straightforward to implement. Looking back over the code,\nI see the global worker assigns its own PID to worker.bgw_notify_pid\nof each of its children, and also obtains a handle for each child\nfrom RegisterDynamicBackgroundWorker().\n\nI imagine the global quota worker would prefer to start workers\nfor every database and then just wait for notifications from any\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\n\n-Chap\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang","msg_date":"Mon, 3 Sep 2018 06:05:40 +0200","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":false,"msg_subject":"Re: Proposal for disk quota feature"},{"msg_contents":"*Hi all,We redesign disk quota feature based on the comments from Pavel\nStehule and Chapman Flack. Here are the new design.OverviewBasically, disk\nquota feature is used to support multi-tenancy environment, different level\nof database objects could be set a quota limit to avoid over use of disk\nspace. A common case could be as follows: DBA could enable disk quota on a\nspecified database list. DBA could set disk quota limit for\ntables/schemas/roles in these databases. Separate disk quota worker process\nwill monitor the disk usage for these objects and detect the objects which\nexceed their quota limit. Queries loading data into these “out of disk\nquota” tables/schemas/roles will be cancelled.We are currently working at\ninit implementation stage. We would like to propose our idea firstly and\nget feedbacks from community to do quick iteration.SQL Syntax (How to use\ndisk quota)1 Specify the databases with disk quota enabled in GUC\n“diskquota_databases” in postgresql.conf and restart the database.2 DBA\ncould set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\nON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\n(quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\n‘3MB’);3 Simulate a schema out of quota limit case: suppose table a1 and\ntable a2 are both under schema s1.INSERT INTO a1 SELECT\ngenerate_series(1,1000);INSERT INTO a2 SELECT\ngenerate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\ngenerate_series(1,1000);ERROR: schema's disk space quota exceededDROP\nTABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\ngenerate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\nfollowing components.1. Quota Setting Store is where the disk quota setting\nto be stored and accessed. We plan to use catalog table pg_diskquota to\nstore these information. pg_diskquota is\nlike:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\ndiskquota name */ int16 quotatype; /* diskquota type name */ Oid\nquotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\ndiskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\nMB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\nchange of database objects. We plan to use stat collector to detect the\n‘active’ table list at initial stage. But stat collector has some\nlimitation on finding the active table which is in a running transaction.\nDetails see TODO section.3. Quota Size Checker is where to calculate the\nsize and compare with quota limit for database objects. According to\nPavel’s comment, autovacuum launcher and worker process could be a good\nreference to disk quota. So we plan to use a disk quota launcher daemon\nprocess and several disk quota worker process to finish this work. Launcher\nprocess is responsible for starting worker process based on a user defined\ndatabase list from GUC. Worker process will connect to its target database\nand monitor the disk usage for objects in this database. In init stage of\nworker process, it will call calculate_total_relation_size() to calculate\nthe size for each user table. After init stage, worker process will refresh\nthe disk model every N seconds. Refreshing will only recalculate the size\nof tables in ‘active’ table list, which is generated by Quata Change\nDetector to minimize the cost.4. Quota Enforcement Operator is where to\ncheck for the quota limitation at postgres backend side. We will firstly\nimplement it in ExecCheckRTPerms() as pre-running enforcement. It will\ncheck the disk quota of tables being inserted or updated, and report error\nif table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\nAs a native feature, we plan to add more checkpoint to do running query\nenforcement. For example, if a disk quota lefts 10MB quota, a query could\ninsert 1GB data. This query could be allowed in pre-running enforcement\ncheck, but will be cancelled in running query enforcement check. Therefore,\nit can improve the accurate of disk quota usage. To achieve this, we plan\nto add a checkpoint in lower API such as smgr_extened. Hence, the Quota\nEnforcement Operator will check the disk quota usage when smgr_extened is\ncalled. If the quota is over limited, current query will be cancelled.\nHighlight1. Native feature.Support native Create/Drop Disk Quota SQL\nstatement.New catalog table pg_diskquota to store disk quota setting.2.\nAuto DML/DDL detection. Table\ncreate/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\n Schema create/drop and Role create/drop will be detected by disk quota\nautomatically. 3. Low cost disk quota checker.Worker process of disk quota\nneed to refresh the disk usage model every N seconds. Since recalculate the\nfile size using stat() system call is expensive for a large number of\nfiles, we use an ‘active’ table list to reduce the real work at each\niteration. A basic experiment on our init stage implementation on database\nwith 20K tables shows that the refresh cost is 1% cpu usage and will be\nfinished within 50ms. Todo/LimitationBefore we propose our patch, we plan\nto enhance it with the following ideas:1. Setting database list with disk\nquota enabled dynamically without restart database. Since we have the disk\nquota launcher process, it could detect the new ‘diskquota_databases’ list\nand start/stop the corresponding disk quota worker process.2. Enforcement\nwhen query is running. Considering the case when there is 10MB quota left,\nbut next query will insert 10GB data. Current enforcement design will allow\nthis query to be executed. This is limited by the ‘active’ table detection\nis generated by stat collector. Postgres backend will only send table stat\ninformation to collector only when the transaction ends. We need a new way\nto detect the ‘active’ table even when this table is being modified inside\na running transaction.3. Monitor unlimited number of databases. Current we\nset the max number of disk quota worker process to be 10 to reduce the\naffection normal workload. But how about if we want to monitor the disk\nquota of more than 10 databases? Our solution is to let disk quota launcher\nto manage a queue of database need to be monitored. And disk quota worker\nprocess consuming the queue and refresh the disk usage/quota for this\ndatabase. After some periods, worker will return the database to the queue,\nand fetch the top database from queue to process. The period determine the\ndelay of detecting disk quota change. To implement this feature, we need to\nsupport a subprocess of postmaster to rebind to another database instead of\nthe database binded in InitPostgres().4. Support active table detection on\nvacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\nare not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\nFlack for the former comments on disk quota feature. Any comments on how to\nimprove disk quota feature are appreciated.*\n\n\nOn Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com>\nwrote:\n\n>\n>\n> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>\n>> Thanks Pavel.\n>> Your patch did enforcement on storage level(md.c or we could also use\n>> smgr_extend). It's straight forward.\n>> But I prefer to implement disk_quota as a feature with following\n>> objectives:\n>> 1 set/alter disk quota setting on different database objects, e.g. user,\n>> database, schema etc. not only a general GUC, but we could set separate\n>> quota limit for a specific objects.\n>> 2 enforcement operator should work at two positions: before query is\n>> running and when query is running. The latter one's implementation maybe\n>> similar to your patch.\n>>\n>\n> The patch was just example. The resource quotes should be more complex -\n> per partition, table, schema, database, user - so GUC are possible, but not\n> very user friendly.\n>\n> Our case is specific, but not too much. The servers are used for\n> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\n> SELECT). We need to solve limits immediately. The implementation is simple,\n> so I did it. Same implementation on database level, or schema level needs\n> some more locks, so it will not be too effective. The resource management\n> can be complex very complex, and I expect so it will be hard work.\n>\n> Regards\n>\n> Pavel\n>\n>\n>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com>\n>> wrote:\n>>\n>>> Hi\n>>>\n>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>>\n>>>> Thanks Chapman.\n>>>> @Pavel, could you please explain more about your second suggestion \"implement\n>>>> some quotas on storage level?\"\n>>>>\n>>>\n>>> See attached patch - it is very simple - and good enough for our\n>>> purposes.\n>>>\n>>> Regards\n>>>\n>>> Pavel\n>>>\n>>>\n>>>\n>>>> We will not keep the long-lived processes attach to all databases(just\n>>>> like you mentioned servers with thousands of databases)\n>>>> And you are right, we could share ideas with autovacuum process, fork\n>>>> worker processes in need.\n>>>> \"autovacuum checks for tables that have had a large number of\n>>>> inserted, updated or deleted tuples. These checks use the statistics\n>>>> collection facility\"\n>>>> diskquota process is similar to autovacuum at caring about insert, but\n>>>> the difference is that it also care about vucuum full, truncate and drop.\n>>>> While update and delete may not be interested since no file change happens.\n>>>> So a separate diskquota process is preferred.\n>>>>\n>>>> So if we implemented disk quota as a full native feature, and in the\n>>>> first initial version I prefer to implement the following features:\n>>>> 1 Fork diskquota launcher process under Postmaster serverloop, which is\n>>>> long-lived.\n>>>> 2 Diskquota launcher process is responsible for creating diskquota\n>>>> worker process for every database.\n>>>> 3 DIskquota setting is stored in a separate catalog table for each\n>>>> database.\n>>>> 4 Initialization stage, Diskquota launcher process creates diskquota worker\n>>>> process for all the databases(traverse like autovacuum). Worker process\n>>>> calculates disk usage of db objects and their diskquota setting. If any\n>>>> db object exceeds its quota limit, put them into the blacklist in the\n>>>> shared memory, which will later be used by enforcement operator. Worker\n>>>> process exits when works are done.\n>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\n>>>> process for the database with a large number of insert, copy, truncate,\n>>>> drop etc. or create disk quota statement. Worker process updates the file\n>>>> size for db objects containing the result relation, and compare with the\n>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\n>>>> remove from blacklist vice versa. Worker process exits when works are\n>>>> done and a GUC could control the frequency of worker process restart to a\n>>>> specific database. As you know, this GUC also controls the delay when we do\n>>>> enforcement.\n>>>> 6 Enforcement. When postgres backend executes queries, check the\n>>>> blacklist in shared memory to determine whether the query is allowed(before\n>>>> execute) or need rollback(is executing)?\n>>>>\n>>>> If we implemented disk quota as an extension, we could just use\n>>>> background worker to start diskquota launcher process and use\n>>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\n>>>> processes by the launcher process as suggested by @Chapman.\n>>>> Diskquota setting could be stored in user table in a separate schema for\n>>>> each database(Schema and table created by create extension statement) just\n>>>> like what Heikki has done in pg_quota project. But in this case, we need to\n>>>> create extension for each database before diskquota worker process can be\n>>>> set up for that database.\n>>>>\n>>>> Any comments on the above design and which is preferred, native feature\n>>>> or extension as the POC?\n>>>>\n>>>>\n>>>> -- Hubert\n>>>>\n>>>>\n>>>>\n>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com\n>>>> > wrote:\n>>>>\n>>>>>\n>>>>>\n>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\n>>>>>\n>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n>>>>>>\n>>>>>> > 2 Keep one worker process for each database. But using a\n>>>>>> parent/global\n>>>>>> > quota worker process to manage the lifecycle of database level\n>>>>>> worker\n>>>>>> > processes. It could handle the newly created database(avoid restart\n>>>>>> > database) and save resource when a database is not used. But this\n>>>>>> needs to\n>>>>>> > change worker process to be hierarchical. Postmaster becomes the\n>>>>>> grandfather\n>>>>>> > of database level worker processes in this case.\n>>>>>>\n>>>>>> I am using background workers this way in 9.5 at $work.\n>>>>>>\n>>>>>> In my case, one worker lives forever, wakes up on a set period, and\n>>>>>> starts a short-lived worker for every database, waiting for each\n>>>>>> one before starting the next.\n>>>>>>\n>>>>>> It was straightforward to implement. Looking back over the code,\n>>>>>> I see the global worker assigns its own PID to worker.bgw_notify_pid\n>>>>>> of each of its children, and also obtains a handle for each child\n>>>>>> from RegisterDynamicBackgroundWorker().\n>>>>>>\n>>>>>> I imagine the global quota worker would prefer to start workers\n>>>>>> for every database and then just wait for notifications from any\n>>>>>> of them, but that seems equally straightforward at first glance.\n>>>>>>\n>>>>>\n>>>>> There are servers with thousands databases. Worker per database is not\n>>>>> good idea.\n>>>>>\n>>>>> It should to share ideas, code with autovacuum process.\n>>>>>\n>>>>> Not sure, how to effective implementation based on bg workers can be.\n>>>>> On servers with large set of databases, large set of tables it can identify\n>>>>> too big table too late.\n>>>>>\n>>>>> Isn't better to implement some quotas on storage level?\n>>>>>\n>>>>> Regards\n>>>>>\n>>>>> Pavel\n>>>>>\n>>>>>\n>>>>>\n>>>>>> -Chap\n>>>>>>\n>>>>>>\n>>>>>\n>>>>\n>>>>\n>>>> --\n>>>> Thanks\n>>>>\n>>>> Hubert Zhang\n>>>>\n>>>\n>>>\n>>\n>>\n>> --\n>> Thanks\n>>\n>> Hubert Zhang\n>>\n>\n>\n\n\n-- \nThanks\n\nHubert Zhang\n\nHi all,We redesign disk quota feature based on the comments from Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk quota feature is used to support multi-tenancy environment, different level of database objects could be set a quota limit to avoid over use of disk space. A common case could be as follows: DBA could enable disk quota on a specified database list. DBA could set disk quota limit for tables/schemas/roles in these databases. Separate disk quota worker process will monitor the disk usage for these objects and detect the objects which exceed their quota limit. Queries loading data into these “out of disk quota” tables/schemas/roles will be cancelled.We are currently working at init implementation stage. We would like to propose our idea firstly and get feedbacks from community to do quick iteration.SQL Syntax (How to use disk quota)1 Specify the databases with disk quota enabled in GUC “diskquota_databases” in postgresql.conf and restart the database.2 DBA could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);3 Simulate a schema out of quota limit case: suppose table a1 and table a2 are both under schema s1.INSERT INTO a1 SELECT generate_series(1,1000);INSERT INTO a2 SELECT generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);ERROR: schema's disk space quota exceededDROP TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the following components.1. Quota Setting Store is where the disk quota setting to be stored and accessed. We plan to use catalog table pg_diskquota to store these information. pg_diskquota is like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /* diskquota name */ int16 quotatype; /* diskquota type name */ Oid quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /* diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size change of database objects. We plan to use stat collector to detect the ‘active’ table list at initial stage. But stat collector has some limitation on finding the active table which is in a running transaction. Details see TODO section.3. Quota Size Checker is where to calculate the size and compare with quota limit for database objects. According to Pavel’s comment, autovacuum launcher and worker process could be a good reference to disk quota. So we plan to use a disk quota launcher daemon process and several disk quota worker process to finish this work. Launcher process is responsible for starting worker process based on a user defined database list from GUC. Worker process will connect to its target database and monitor the disk usage for objects in this database. In init stage of worker process, it will call calculate_total_relation_size() to calculate the size for each user table. After init stage, worker process will refresh the disk model every N seconds. Refreshing will only recalculate the size of tables in ‘active’ table list, which is generated by Quata Change Detector to minimize the cost.4. Quota Enforcement Operator is where to check for the quota limitation at postgres backend side. We will firstly implement it in ExecCheckRTPerms() as pre-running enforcement. It will check the disk quota of tables being inserted or updated, and report error if table’s or table’s schema’s or table’s owner’s quota limit is exceeded. As a native feature, we plan to add more checkpoint to do running query enforcement. For example, if a disk quota lefts 10MB quota, a query could insert 1GB data. This query could be allowed in pre-running enforcement check, but will be cancelled in running query enforcement check. Therefore, it can improve the accurate of disk quota usage. To achieve this, we plan to add a checkpoint in lower API such as smgr_extened. Hence, the Quota Enforcement Operator will check the disk quota usage when smgr_extened is called. If the quota is over limited, current query will be cancelled. Highlight1. Native feature.Support native Create/Drop Disk Quota SQL statement.New catalog table pg_diskquota to store disk quota setting.2. Auto DML/DDL detection. Table create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change, Schema create/drop and Role create/drop will be detected by disk quota automatically. 3. Low cost disk quota checker.Worker process of disk quota need to refresh the disk usage model every N seconds. Since recalculate the file size using stat() system call is expensive for a large number of files, we use an ‘active’ table list to reduce the real work at each iteration. A basic experiment on our init stage implementation on database with 20K tables shows that the refresh cost is 1% cpu usage and will be finished within 50ms. Todo/LimitationBefore we propose our patch, we plan to enhance it with the following ideas:1. Setting database list with disk quota enabled dynamically without restart database. Since we have the disk quota launcher process, it could detect the new ‘diskquota_databases’ list and start/stop the corresponding disk quota worker process.2. Enforcement when query is running. Considering the case when there is 10MB quota left, but next query will insert 10GB data. Current enforcement design will allow this query to be executed. This is limited by the ‘active’ table detection is generated by stat collector. Postgres backend will only send table stat information to collector only when the transaction ends. We need a new way to detect the ‘active’ table even when this table is being modified inside a running transaction.3. Monitor unlimited number of databases. Current we set the max number of disk quota worker process to be 10 to reduce the affection normal workload. But how about if we want to monitor the disk quota of more than 10 databases? Our solution is to let disk quota launcher to manage a queue of database need to be monitored. And disk quota worker process consuming the queue and refresh the disk usage/quota for this database. After some periods, worker will return the database to the queue, and fetch the top database from queue to process. The period determine the delay of detecting disk quota change. To implement this feature, we need to support a subprocess of postmaster to rebind to another database instead of the database binded in InitPostgres().4. Support active table detection on vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman Flack for the former comments on disk quota feature. Any comments on how to improve disk quota feature are appreciated.On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \"implement some quotas on storage level?\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n\n> 2 Keep one worker process for each database. But using a parent/global\n> quota worker process to manage the lifecycle of database level worker\n> processes. It could handle the newly created database(avoid restart\n> database) and save resource when a database is not used. But this needs to\n> change worker process to be hierarchical. Postmaster becomes the grandfather\n> of database level worker processes in this case.\n\nI am using background workers this way in 9.5 at $work.\n\nIn my case, one worker lives forever, wakes up on a set period, and\nstarts a short-lived worker for every database, waiting for each\none before starting the next.\n\nIt was straightforward to implement. Looking back over the code,\nI see the global worker assigns its own PID to worker.bgw_notify_pid\nof each of its children, and also obtains a handle for each child\nfrom RegisterDynamicBackgroundWorker().\n\nI imagine the global quota worker would prefer to start workers\nfor every database and then just wait for notifications from any\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\n\n-Chap\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang","msg_date":"Fri, 21 Sep 2018 19:32:16 +0800","msg_from":"Hubert Zhang <hzhang@pivotal.io>","msg_from_op":true,"msg_subject":"Re: Proposal for disk quota feature"},{"msg_contents":"pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:\n\n>\n>\n>\n>\n> *Hi all,We redesign disk quota feature based on the comments from Pavel\n> Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk\n> quota feature is used to support multi-tenancy environment, different level\n> of database objects could be set a quota limit to avoid over use of disk\n> space. A common case could be as follows: DBA could enable disk quota on a\n> specified database list. DBA could set disk quota limit for\n> tables/schemas/roles in these databases. Separate disk quota worker process\n> will monitor the disk usage for these objects and detect the objects which\n> exceed their quota limit. Queries loading data into these “out of disk\n> quota” tables/schemas/roles will be cancelled.We are currently working at\n> init implementation stage. We would like to propose our idea firstly and\n> get feedbacks from community to do quick iteration.SQL Syntax (How to use\n> disk quota)1 Specify the databases with disk quota enabled in GUC\n> “diskquota_databases” in postgresql.conf and restart the database.2 DBA\n> could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\n> ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\n> (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\n> ‘3MB’);*\n>\n\njust fast reaction - why QUOTA object?\n\nIsn't ALTER SET enough?\n\nSome like\n\nALTER TABLE a1 SET quote = 1MB;\nALTER USER ...\nALTER SCHEMA ..\n\nNew DDL commans looks like too hard hammer .\n\n\n\n>\n>\n>\n>\n>\n>\n>\n>\n>\n>\n>\n>\n>\n>\n>\n>\n>\n> *3 Simulate a schema out of quota limit case: suppose table a1 and table\n> a2 are both under schema s1.INSERT INTO a1 SELECT\n> generate_series(1,1000);INSERT INTO a2 SELECT\n> generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\n> generate_series(1,1000);ERROR: schema's disk space quota exceededDROP\n> TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\n> generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\n> following components.1. Quota Setting Store is where the disk quota setting\n> to be stored and accessed. We plan to use catalog table pg_diskquota to\n> store these information. pg_diskquota is\n> like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\n> diskquota name */ int16 quotatype; /* diskquota type name */ Oid\n> quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\n> diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\n> MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\n> change of database objects. We plan to use stat collector to detect the\n> ‘active’ table list at initial stage. But stat collector has some\n> limitation on finding the active table which is in a running transaction.\n> Details see TODO section.3. Quota Size Checker is where to calculate the\n> size and compare with quota limit for database objects. According to\n> Pavel’s comment, autovacuum launcher and worker process could be a good\n> reference to disk quota. So we plan to use a disk quota launcher daemon\n> process and several disk quota worker process to finish this work. Launcher\n> process is responsible for starting worker process based on a user defined\n> database list from GUC. Worker process will connect to its target database\n> and monitor the disk usage for objects in this database. In init stage of\n> worker process, it will call calculate_total_relation_size() to calculate\n> the size for each user table. After init stage, worker process will refresh\n> the disk model every N seconds. Refreshing will only recalculate the size\n> of tables in ‘active’ table list, which is generated by Quata Change\n> Detector to minimize the cost.4. Quota Enforcement Operator is where to\n> check for the quota limitation at postgres backend side. We will firstly\n> implement it in ExecCheckRTPerms() as pre-running enforcement. It will\n> check the disk quota of tables being inserted or updated, and report error\n> if table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\n> As a native feature, we plan to add more checkpoint to do running query\n> enforcement. For example, if a disk quota lefts 10MB quota, a query could\n> insert 1GB data. This query could be allowed in pre-running enforcement\n> check, but will be cancelled in running query enforcement check. Therefore,\n> it can improve the accurate of disk quota usage. To achieve this, we plan\n> to add a checkpoint in lower API such as smgr_extened. Hence, the Quota\n> Enforcement Operator will check the disk quota usage when smgr_extened is\n> called. If the quota is over limited, current query will be cancelled.\n> Highlight1. Native feature.Support native Create/Drop Disk Quota SQL\n> statement.New catalog table pg_diskquota to store disk quota setting.2.\n> Auto DML/DDL detection. Table\n> create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\n> Schema create/drop and Role create/drop will be detected by disk quota\n> automatically. 3. Low cost disk quota checker.Worker process of disk quota\n> need to refresh the disk usage model every N seconds. Since recalculate the\n> file size using stat() system call is expensive for a large number of\n> files, we use an ‘active’ table list to reduce the real work at each\n> iteration. A basic experiment on our init stage implementation on database\n> with 20K tables shows that the refresh cost is 1% cpu usage and will be\n> finished within 50ms. Todo/LimitationBefore we propose our patch, we plan\n> to enhance it with the following ideas:1. Setting database list with disk\n> quota enabled dynamically without restart database. Since we have the disk\n> quota launcher process, it could detect the new ‘diskquota_databases’ list\n> and start/stop the corresponding disk quota worker process.2. Enforcement\n> when query is running. Considering the case when there is 10MB quota left,\n> but next query will insert 10GB data. Current enforcement design will allow\n> this query to be executed. This is limited by the ‘active’ table detection\n> is generated by stat collector. Postgres backend will only send table stat\n> information to collector only when the transaction ends. We need a new way\n> to detect the ‘active’ table even when this table is being modified inside\n> a running transaction.3. Monitor unlimited number of databases. Current we\n> set the max number of disk quota worker process to be 10 to reduce the\n> affection normal workload. But how about if we want to monitor the disk\n> quota of more than 10 databases? Our solution is to let disk quota launcher\n> to manage a queue of database need to be monitored. And disk quota worker\n> process consuming the queue and refresh the disk usage/quota for this\n> database. After some periods, worker will return the database to the queue,\n> and fetch the top database from queue to process. The period determine the\n> delay of detecting disk quota change. To implement this feature, we need to\n> support a subprocess of postmaster to rebind to another database instead of\n> the database binded in InitPostgres().4. Support active table detection on\n> vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\n> are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\n> Flack for the former comments on disk quota feature. Any comments on how to\n> improve disk quota feature are appreciated.*\n>\n>\n> On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com>\n> wrote:\n>\n>>\n>>\n>> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>\n>>> Thanks Pavel.\n>>> Your patch did enforcement on storage level(md.c or we could also use\n>>> smgr_extend). It's straight forward.\n>>> But I prefer to implement disk_quota as a feature with following\n>>> objectives:\n>>> 1 set/alter disk quota setting on different database objects, e.g. user,\n>>> database, schema etc. not only a general GUC, but we could set separate\n>>> quota limit for a specific objects.\n>>> 2 enforcement operator should work at two positions: before query is\n>>> running and when query is running. The latter one's implementation maybe\n>>> similar to your patch.\n>>>\n>>\n>> The patch was just example. The resource quotes should be more complex -\n>> per partition, table, schema, database, user - so GUC are possible, but not\n>> very user friendly.\n>>\n>> Our case is specific, but not too much. The servers are used for\n>> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\n>> SELECT). We need to solve limits immediately. The implementation is simple,\n>> so I did it. Same implementation on database level, or schema level needs\n>> some more locks, so it will not be too effective. The resource management\n>> can be complex very complex, and I expect so it will be hard work.\n>>\n>> Regards\n>>\n>> Pavel\n>>\n>>\n>>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com>\n>>> wrote:\n>>>\n>>>> Hi\n>>>>\n>>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>>>\n>>>>> Thanks Chapman.\n>>>>> @Pavel, could you please explain more about your second suggestion \"implement\n>>>>> some quotas on storage level?\"\n>>>>>\n>>>>\n>>>> See attached patch - it is very simple - and good enough for our\n>>>> purposes.\n>>>>\n>>>> Regards\n>>>>\n>>>> Pavel\n>>>>\n>>>>\n>>>>\n>>>>> We will not keep the long-lived processes attach to all databases(just\n>>>>> like you mentioned servers with thousands of databases)\n>>>>> And you are right, we could share ideas with autovacuum process, fork\n>>>>> worker processes in need.\n>>>>> \"autovacuum checks for tables that have had a large number of\n>>>>> inserted, updated or deleted tuples. These checks use the statistics\n>>>>> collection facility\"\n>>>>> diskquota process is similar to autovacuum at caring about insert, but\n>>>>> the difference is that it also care about vucuum full, truncate and drop.\n>>>>> While update and delete may not be interested since no file change happens.\n>>>>> So a separate diskquota process is preferred.\n>>>>>\n>>>>> So if we implemented disk quota as a full native feature, and in the\n>>>>> first initial version I prefer to implement the following features:\n>>>>> 1 Fork diskquota launcher process under Postmaster serverloop, which\n>>>>> is long-lived.\n>>>>> 2 Diskquota launcher process is responsible for creating diskquota\n>>>>> worker process for every database.\n>>>>> 3 DIskquota setting is stored in a separate catalog table for each\n>>>>> database.\n>>>>> 4 Initialization stage, Diskquota launcher process creates diskquota worker\n>>>>> process for all the databases(traverse like autovacuum). Worker process\n>>>>> calculates disk usage of db objects and their diskquota setting. If any\n>>>>> db object exceeds its quota limit, put them into the blacklist in the\n>>>>> shared memory, which will later be used by enforcement operator. Worker\n>>>>> process exits when works are done.\n>>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\n>>>>> process for the database with a large number of insert, copy, truncate,\n>>>>> drop etc. or create disk quota statement. Worker process updates the file\n>>>>> size for db objects containing the result relation, and compare with the\n>>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\n>>>>> remove from blacklist vice versa. Worker process exits when works are\n>>>>> done and a GUC could control the frequency of worker process restart to a\n>>>>> specific database. As you know, this GUC also controls the delay when we do\n>>>>> enforcement.\n>>>>> 6 Enforcement. When postgres backend executes queries, check the\n>>>>> blacklist in shared memory to determine whether the query is allowed(before\n>>>>> execute) or need rollback(is executing)?\n>>>>>\n>>>>> If we implemented disk quota as an extension, we could just use\n>>>>> background worker to start diskquota launcher process and use\n>>>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\n>>>>> processes by the launcher process as suggested by @Chapman.\n>>>>> Diskquota setting could be stored in user table in a separate schema for\n>>>>> each database(Schema and table created by create extension statement) just\n>>>>> like what Heikki has done in pg_quota project. But in this case, we need to\n>>>>> create extension for each database before diskquota worker process can be\n>>>>> set up for that database.\n>>>>>\n>>>>> Any comments on the above design and which is preferred, native\n>>>>> feature or extension as the POC?\n>>>>>\n>>>>>\n>>>>> -- Hubert\n>>>>>\n>>>>>\n>>>>>\n>>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <\n>>>>> pavel.stehule@gmail.com> wrote:\n>>>>>\n>>>>>>\n>>>>>>\n>>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\n>>>>>>\n>>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n>>>>>>>\n>>>>>>> > 2 Keep one worker process for each database. But using a\n>>>>>>> parent/global\n>>>>>>> > quota worker process to manage the lifecycle of database level\n>>>>>>> worker\n>>>>>>> > processes. It could handle the newly created database(avoid restart\n>>>>>>> > database) and save resource when a database is not used. But this\n>>>>>>> needs to\n>>>>>>> > change worker process to be hierarchical. Postmaster becomes the\n>>>>>>> grandfather\n>>>>>>> > of database level worker processes in this case.\n>>>>>>>\n>>>>>>> I am using background workers this way in 9.5 at $work.\n>>>>>>>\n>>>>>>> In my case, one worker lives forever, wakes up on a set period, and\n>>>>>>> starts a short-lived worker for every database, waiting for each\n>>>>>>> one before starting the next.\n>>>>>>>\n>>>>>>> It was straightforward to implement. Looking back over the code,\n>>>>>>> I see the global worker assigns its own PID to worker.bgw_notify_pid\n>>>>>>> of each of its children, and also obtains a handle for each child\n>>>>>>> from RegisterDynamicBackgroundWorker().\n>>>>>>>\n>>>>>>> I imagine the global quota worker would prefer to start workers\n>>>>>>> for every database and then just wait for notifications from any\n>>>>>>> of them, but that seems equally straightforward at first glance.\n>>>>>>>\n>>>>>>\n>>>>>> There are servers with thousands databases. Worker per database is\n>>>>>> not good idea.\n>>>>>>\n>>>>>> It should to share ideas, code with autovacuum process.\n>>>>>>\n>>>>>> Not sure, how to effective implementation based on bg workers can be.\n>>>>>> On servers with large set of databases, large set of tables it can identify\n>>>>>> too big table too late.\n>>>>>>\n>>>>>> Isn't better to implement some quotas on storage level?\n>>>>>>\n>>>>>> Regards\n>>>>>>\n>>>>>> Pavel\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>> -Chap\n>>>>>>>\n>>>>>>>\n>>>>>>\n>>>>>\n>>>>>\n>>>>> --\n>>>>> Thanks\n>>>>>\n>>>>> Hubert Zhang\n>>>>>\n>>>>\n>>>>\n>>>\n>>>\n>>> --\n>>> Thanks\n>>>\n>>> Hubert Zhang\n>>>\n>>\n>>\n>\n>\n> --\n> Thanks\n>\n> Hubert Zhang\n>\n\npá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:Hi all,We redesign disk quota feature based on the comments from Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk quota feature is used to support multi-tenancy environment, different level of database objects could be set a quota limit to avoid over use of disk space. A common case could be as follows: DBA could enable disk quota on a specified database list. DBA could set disk quota limit for tables/schemas/roles in these databases. Separate disk quota worker process will monitor the disk usage for these objects and detect the objects which exceed their quota limit. Queries loading data into these “out of disk quota” tables/schemas/roles will be cancelled.We are currently working at init implementation stage. We would like to propose our idea firstly and get feedbacks from community to do quick iteration.SQL Syntax (How to use disk quota)1 Specify the databases with disk quota enabled in GUC “diskquota_databases” in postgresql.conf and restart the database.2 DBA could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer . 3 Simulate a schema out of quota limit case: suppose table a1 and table a2 are both under schema s1.INSERT INTO a1 SELECT generate_series(1,1000);INSERT INTO a2 SELECT generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);ERROR: schema's disk space quota exceededDROP TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the following components.1. Quota Setting Store is where the disk quota setting to be stored and accessed. We plan to use catalog table pg_diskquota to store these information. pg_diskquota is like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /* diskquota name */ int16 quotatype; /* diskquota type name */ Oid quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /* diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size change of database objects. We plan to use stat collector to detect the ‘active’ table list at initial stage. But stat collector has some limitation on finding the active table which is in a running transaction. Details see TODO section.3. Quota Size Checker is where to calculate the size and compare with quota limit for database objects. According to Pavel’s comment, autovacuum launcher and worker process could be a good reference to disk quota. So we plan to use a disk quota launcher daemon process and several disk quota worker process to finish this work. Launcher process is responsible for starting worker process based on a user defined database list from GUC. Worker process will connect to its target database and monitor the disk usage for objects in this database. In init stage of worker process, it will call calculate_total_relation_size() to calculate the size for each user table. After init stage, worker process will refresh the disk model every N seconds. Refreshing will only recalculate the size of tables in ‘active’ table list, which is generated by Quata Change Detector to minimize the cost.4. Quota Enforcement Operator is where to check for the quota limitation at postgres backend side. We will firstly implement it in ExecCheckRTPerms() as pre-running enforcement. It will check the disk quota of tables being inserted or updated, and report error if table’s or table’s schema’s or table’s owner’s quota limit is exceeded. As a native feature, we plan to add more checkpoint to do running query enforcement. For example, if a disk quota lefts 10MB quota, a query could insert 1GB data. This query could be allowed in pre-running enforcement check, but will be cancelled in running query enforcement check. Therefore, it can improve the accurate of disk quota usage. To achieve this, we plan to add a checkpoint in lower API such as smgr_extened. Hence, the Quota Enforcement Operator will check the disk quota usage when smgr_extened is called. If the quota is over limited, current query will be cancelled. Highlight1. Native feature.Support native Create/Drop Disk Quota SQL statement.New catalog table pg_diskquota to store disk quota setting.2. Auto DML/DDL detection. Table create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change, Schema create/drop and Role create/drop will be detected by disk quota automatically. 3. Low cost disk quota checker.Worker process of disk quota need to refresh the disk usage model every N seconds. Since recalculate the file size using stat() system call is expensive for a large number of files, we use an ‘active’ table list to reduce the real work at each iteration. A basic experiment on our init stage implementation on database with 20K tables shows that the refresh cost is 1% cpu usage and will be finished within 50ms. Todo/LimitationBefore we propose our patch, we plan to enhance it with the following ideas:1. Setting database list with disk quota enabled dynamically without restart database. Since we have the disk quota launcher process, it could detect the new ‘diskquota_databases’ list and start/stop the corresponding disk quota worker process.2. Enforcement when query is running. Considering the case when there is 10MB quota left, but next query will insert 10GB data. Current enforcement design will allow this query to be executed. This is limited by the ‘active’ table detection is generated by stat collector. Postgres backend will only send table stat information to collector only when the transaction ends. We need a new way to detect the ‘active’ table even when this table is being modified inside a running transaction.3. Monitor unlimited number of databases. Current we set the max number of disk quota worker process to be 10 to reduce the affection normal workload. But how about if we want to monitor the disk quota of more than 10 databases? Our solution is to let disk quota launcher to manage a queue of database need to be monitored. And disk quota worker process consuming the queue and refresh the disk usage/quota for this database. After some periods, worker will return the database to the queue, and fetch the top database from queue to process. The period determine the delay of detecting disk quota change. To implement this feature, we need to support a subprocess of postmaster to rebind to another database instead of the database binded in InitPostgres().4. Support active table detection on vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman Flack for the former comments on disk quota feature. Any comments on how to improve disk quota feature are appreciated.On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \"implement some quotas on storage level?\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n\n> 2 Keep one worker process for each database. But using a parent/global\n> quota worker process to manage the lifecycle of database level worker\n> processes. It could handle the newly created database(avoid restart\n> database) and save resource when a database is not used. But this needs to\n> change worker process to be hierarchical. Postmaster becomes the grandfather\n> of database level worker processes in this case.\n\nI am using background workers this way in 9.5 at $work.\n\nIn my case, one worker lives forever, wakes up on a set period, and\nstarts a short-lived worker for every database, waiting for each\none before starting the next.\n\nIt was straightforward to implement. Looking back over the code,\nI see the global worker assigns its own PID to worker.bgw_notify_pid\nof each of its children, and also obtains a handle for each child\nfrom RegisterDynamicBackgroundWorker().\n\nI imagine the global quota worker would prefer to start workers\nfor every database and then just wait for notifications from any\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\n\n-Chap\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang","msg_date":"Fri, 21 Sep 2018 14:01:15 +0200","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":false,"msg_subject":"Re: Proposal for disk quota feature"},{"msg_contents":">\n> just fast reaction - why QUOTA object?\n> Isn't ALTER SET enough?\n> Some like\n> ALTER TABLE a1 SET quote = 1MB;\n> ALTER USER ...\n> ALTER SCHEMA ..\n> New DDL commans looks like too hard hammer .\n\n\nIt's an option. Prefer to consider quota setting store together:\nCREATE DISK QUOTA way is more nature to store quota setting in a separate\npg_diskquota catalog\nWhile ALTER SET way is more close to store quota setting in pg_class,\npg_role, pg_namespace. etc in an integrated way.\n(Note that here I mean nature/close is not must, ALTER SET could also store\nin pg_diskquota and vice versa.)\n\nHere are some differences I can think of:\n1 pg_role is a global catalog, not per database level. It's harder to\ntracker the user's disk usage in the whole clusters(considering 1000+\ndatabases). So the semantic of CREATE DISK QUOTA ON USER is limited: it\nonly tracks the user's disk usage inside the current database.\n2 using separate pg_diskquota could add more field except for quota limit\nwithout adding too many fields in pg_class, e.g. red zone to give the user\na warning or the current disk usage of the db objects.\n\nOn Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com>\nwrote:\n\n>\n>\n> pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:\n>\n>>\n>>\n>>\n>>\n>> *Hi all,We redesign disk quota feature based on the comments from Pavel\n>> Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk\n>> quota feature is used to support multi-tenancy environment, different level\n>> of database objects could be set a quota limit to avoid over use of disk\n>> space. A common case could be as follows: DBA could enable disk quota on a\n>> specified database list. DBA could set disk quota limit for\n>> tables/schemas/roles in these databases. Separate disk quota worker process\n>> will monitor the disk usage for these objects and detect the objects which\n>> exceed their quota limit. Queries loading data into these “out of disk\n>> quota” tables/schemas/roles will be cancelled.We are currently working at\n>> init implementation stage. We would like to propose our idea firstly and\n>> get feedbacks from community to do quick iteration.SQL Syntax (How to use\n>> disk quota)1 Specify the databases with disk quota enabled in GUC\n>> “diskquota_databases” in postgresql.conf and restart the database.2 DBA\n>> could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\n>> ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\n>> (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\n>> ‘3MB’);*\n>>\n>\n> just fast reaction - why QUOTA object?\n>\n> Isn't ALTER SET enough?\n>\n> Some like\n>\n> ALTER TABLE a1 SET quote = 1MB;\n> ALTER USER ...\n> ALTER SCHEMA ..\n>\n> New DDL commans looks like too hard hammer .\n>\n>\n>\n>>\n>>\n>>\n>>\n>>\n>>\n>>\n>>\n>>\n>>\n>>\n>>\n>>\n>>\n>>\n>>\n>>\n>> *3 Simulate a schema out of quota limit case: suppose table a1 and table\n>> a2 are both under schema s1.INSERT INTO a1 SELECT\n>> generate_series(1,1000);INSERT INTO a2 SELECT\n>> generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\n>> generate_series(1,1000);ERROR: schema's disk space quota exceededDROP\n>> TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\n>> generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\n>> following components.1. Quota Setting Store is where the disk quota setting\n>> to be stored and accessed. We plan to use catalog table pg_diskquota to\n>> store these information. pg_diskquota is\n>> like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\n>> diskquota name */ int16 quotatype; /* diskquota type name */ Oid\n>> quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\n>> diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\n>> MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\n>> change of database objects. We plan to use stat collector to detect the\n>> ‘active’ table list at initial stage. But stat collector has some\n>> limitation on finding the active table which is in a running transaction.\n>> Details see TODO section.3. Quota Size Checker is where to calculate the\n>> size and compare with quota limit for database objects. According to\n>> Pavel’s comment, autovacuum launcher and worker process could be a good\n>> reference to disk quota. So we plan to use a disk quota launcher daemon\n>> process and several disk quota worker process to finish this work. Launcher\n>> process is responsible for starting worker process based on a user defined\n>> database list from GUC. Worker process will connect to its target database\n>> and monitor the disk usage for objects in this database. In init stage of\n>> worker process, it will call calculate_total_relation_size() to calculate\n>> the size for each user table. After init stage, worker process will refresh\n>> the disk model every N seconds. Refreshing will only recalculate the size\n>> of tables in ‘active’ table list, which is generated by Quata Change\n>> Detector to minimize the cost.4. Quota Enforcement Operator is where to\n>> check for the quota limitation at postgres backend side. We will firstly\n>> implement it in ExecCheckRTPerms() as pre-running enforcement. It will\n>> check the disk quota of tables being inserted or updated, and report error\n>> if table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\n>> As a native feature, we plan to add more checkpoint to do running query\n>> enforcement. For example, if a disk quota lefts 10MB quota, a query could\n>> insert 1GB data. This query could be allowed in pre-running enforcement\n>> check, but will be cancelled in running query enforcement check. Therefore,\n>> it can improve the accurate of disk quota usage. To achieve this, we plan\n>> to add a checkpoint in lower API such as smgr_extened. Hence, the Quota\n>> Enforcement Operator will check the disk quota usage when smgr_extened is\n>> called. If the quota is over limited, current query will be cancelled.\n>> Highlight1. Native feature.Support native Create/Drop Disk Quota SQL\n>> statement.New catalog table pg_diskquota to store disk quota setting.2.\n>> Auto DML/DDL detection. Table\n>> create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\n>> Schema create/drop and Role create/drop will be detected by disk quota\n>> automatically. 3. Low cost disk quota checker.Worker process of disk quota\n>> need to refresh the disk usage model every N seconds. Since recalculate the\n>> file size using stat() system call is expensive for a large number of\n>> files, we use an ‘active’ table list to reduce the real work at each\n>> iteration. A basic experiment on our init stage implementation on database\n>> with 20K tables shows that the refresh cost is 1% cpu usage and will be\n>> finished within 50ms. Todo/LimitationBefore we propose our patch, we plan\n>> to enhance it with the following ideas:1. Setting database list with disk\n>> quota enabled dynamically without restart database. Since we have the disk\n>> quota launcher process, it could detect the new ‘diskquota_databases’ list\n>> and start/stop the corresponding disk quota worker process.2. Enforcement\n>> when query is running. Considering the case when there is 10MB quota left,\n>> but next query will insert 10GB data. Current enforcement design will allow\n>> this query to be executed. This is limited by the ‘active’ table detection\n>> is generated by stat collector. Postgres backend will only send table stat\n>> information to collector only when the transaction ends. We need a new way\n>> to detect the ‘active’ table even when this table is being modified inside\n>> a running transaction.3. Monitor unlimited number of databases. Current we\n>> set the max number of disk quota worker process to be 10 to reduce the\n>> affection normal workload. But how about if we want to monitor the disk\n>> quota of more than 10 databases? Our solution is to let disk quota launcher\n>> to manage a queue of database need to be monitored. And disk quota worker\n>> process consuming the queue and refresh the disk usage/quota for this\n>> database. After some periods, worker will return the database to the queue,\n>> and fetch the top database from queue to process. The period determine the\n>> delay of detecting disk quota change. To implement this feature, we need to\n>> support a subprocess of postmaster to rebind to another database instead of\n>> the database binded in InitPostgres().4. Support active table detection on\n>> vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\n>> are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\n>> Flack for the former comments on disk quota feature. Any comments on how to\n>> improve disk quota feature are appreciated.*\n>>\n>>\n>> On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com>\n>> wrote:\n>>\n>>>\n>>>\n>>> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>>\n>>>> Thanks Pavel.\n>>>> Your patch did enforcement on storage level(md.c or we could also use\n>>>> smgr_extend). It's straight forward.\n>>>> But I prefer to implement disk_quota as a feature with following\n>>>> objectives:\n>>>> 1 set/alter disk quota setting on different database objects, e.g.\n>>>> user, database, schema etc. not only a general GUC, but we could set\n>>>> separate quota limit for a specific objects.\n>>>> 2 enforcement operator should work at two positions: before query is\n>>>> running and when query is running. The latter one's implementation maybe\n>>>> similar to your patch.\n>>>>\n>>>\n>>> The patch was just example. The resource quotes should be more complex -\n>>> per partition, table, schema, database, user - so GUC are possible, but not\n>>> very user friendly.\n>>>\n>>> Our case is specific, but not too much. The servers are used for\n>>> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\n>>> SELECT). We need to solve limits immediately. The implementation is simple,\n>>> so I did it. Same implementation on database level, or schema level needs\n>>> some more locks, so it will not be too effective. The resource management\n>>> can be complex very complex, and I expect so it will be hard work.\n>>>\n>>> Regards\n>>>\n>>> Pavel\n>>>\n>>>\n>>>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com>\n>>>> wrote:\n>>>>\n>>>>> Hi\n>>>>>\n>>>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>>>>\n>>>>>> Thanks Chapman.\n>>>>>> @Pavel, could you please explain more about your second suggestion \"implement\n>>>>>> some quotas on storage level?\"\n>>>>>>\n>>>>>\n>>>>> See attached patch - it is very simple - and good enough for our\n>>>>> purposes.\n>>>>>\n>>>>> Regards\n>>>>>\n>>>>> Pavel\n>>>>>\n>>>>>\n>>>>>\n>>>>>> We will not keep the long-lived processes attach to all\n>>>>>> databases(just like you mentioned servers with thousands of databases\n>>>>>> )\n>>>>>> And you are right, we could share ideas with autovacuum process, fork\n>>>>>> worker processes in need.\n>>>>>> \"autovacuum checks for tables that have had a large number of\n>>>>>> inserted, updated or deleted tuples. These checks use the statistics\n>>>>>> collection facility\"\n>>>>>> diskquota process is similar to autovacuum at caring about insert,\n>>>>>> but the difference is that it also care about vucuum full, truncate and\n>>>>>> drop. While update and delete may not be interested since no file change\n>>>>>> happens. So a separate diskquota process is preferred.\n>>>>>>\n>>>>>> So if we implemented disk quota as a full native feature, and in the\n>>>>>> first initial version I prefer to implement the following features:\n>>>>>> 1 Fork diskquota launcher process under Postmaster serverloop, which\n>>>>>> is long-lived.\n>>>>>> 2 Diskquota launcher process is responsible for creating diskquota\n>>>>>> worker process for every database.\n>>>>>> 3 DIskquota setting is stored in a separate catalog table for each\n>>>>>> database.\n>>>>>> 4 Initialization stage, Diskquota launcher process creates diskquota worker\n>>>>>> process for all the databases(traverse like autovacuum). Worker process\n>>>>>> calculates disk usage of db objects and their diskquota setting. If any\n>>>>>> db object exceeds its quota limit, put them into the blacklist in the\n>>>>>> shared memory, which will later be used by enforcement operator. Worker\n>>>>>> process exits when works are done.\n>>>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\n>>>>>> process for the database with a large number of insert, copy, truncate,\n>>>>>> drop etc. or create disk quota statement. Worker process updates the file\n>>>>>> size for db objects containing the result relation, and compare with the\n>>>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\n>>>>>> remove from blacklist vice versa. Worker process exits when works\n>>>>>> are done and a GUC could control the frequency of worker process restart to\n>>>>>> a specific database. As you know, this GUC also controls the delay when we\n>>>>>> do enforcement.\n>>>>>> 6 Enforcement. When postgres backend executes queries, check the\n>>>>>> blacklist in shared memory to determine whether the query is allowed(before\n>>>>>> execute) or need rollback(is executing)?\n>>>>>>\n>>>>>> If we implemented disk quota as an extension, we could just use\n>>>>>> background worker to start diskquota launcher process and use\n>>>>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\n>>>>>> processes by the launcher process as suggested by @Chapman.\n>>>>>> Diskquota setting could be stored in user table in a separate schema for\n>>>>>> each database(Schema and table created by create extension statement) just\n>>>>>> like what Heikki has done in pg_quota project. But in this case, we need to\n>>>>>> create extension for each database before diskquota worker process can be\n>>>>>> set up for that database.\n>>>>>>\n>>>>>> Any comments on the above design and which is preferred, native\n>>>>>> feature or extension as the POC?\n>>>>>>\n>>>>>>\n>>>>>> -- Hubert\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <\n>>>>>> pavel.stehule@gmail.com> wrote:\n>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\n>>>>>>>\n>>>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n>>>>>>>>\n>>>>>>>> > 2 Keep one worker process for each database. But using a\n>>>>>>>> parent/global\n>>>>>>>> > quota worker process to manage the lifecycle of database level\n>>>>>>>> worker\n>>>>>>>> > processes. It could handle the newly created database(avoid\n>>>>>>>> restart\n>>>>>>>> > database) and save resource when a database is not used. But this\n>>>>>>>> needs to\n>>>>>>>> > change worker process to be hierarchical. Postmaster becomes the\n>>>>>>>> grandfather\n>>>>>>>> > of database level worker processes in this case.\n>>>>>>>>\n>>>>>>>> I am using background workers this way in 9.5 at $work.\n>>>>>>>>\n>>>>>>>> In my case, one worker lives forever, wakes up on a set period, and\n>>>>>>>> starts a short-lived worker for every database, waiting for each\n>>>>>>>> one before starting the next.\n>>>>>>>>\n>>>>>>>> It was straightforward to implement. Looking back over the code,\n>>>>>>>> I see the global worker assigns its own PID to worker.bgw_notify_pid\n>>>>>>>> of each of its children, and also obtains a handle for each child\n>>>>>>>> from RegisterDynamicBackgroundWorker().\n>>>>>>>>\n>>>>>>>> I imagine the global quota worker would prefer to start workers\n>>>>>>>> for every database and then just wait for notifications from any\n>>>>>>>> of them, but that seems equally straightforward at first glance.\n>>>>>>>>\n>>>>>>>\n>>>>>>> There are servers with thousands databases. Worker per database is\n>>>>>>> not good idea.\n>>>>>>>\n>>>>>>> It should to share ideas, code with autovacuum process.\n>>>>>>>\n>>>>>>> Not sure, how to effective implementation based on bg workers can\n>>>>>>> be. On servers with large set of databases, large set of tables it can\n>>>>>>> identify too big table too late.\n>>>>>>>\n>>>>>>> Isn't better to implement some quotas on storage level?\n>>>>>>>\n>>>>>>> Regards\n>>>>>>>\n>>>>>>> Pavel\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>> -Chap\n>>>>>>>>\n>>>>>>>>\n>>>>>>>\n>>>>>>\n>>>>>>\n>>>>>> --\n>>>>>> Thanks\n>>>>>>\n>>>>>> Hubert Zhang\n>>>>>>\n>>>>>\n>>>>>\n>>>>\n>>>>\n>>>> --\n>>>> Thanks\n>>>>\n>>>> Hubert Zhang\n>>>>\n>>>\n>>>\n>>\n>>\n>> --\n>> Thanks\n>>\n>> Hubert Zhang\n>>\n>\n\n-- \nThanks\n\nHubert Zhang\n\njust fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer .It's an option. Prefer to consider quota setting store together:CREATE DISK QUOTA way is more nature to store quota setting in a separate pg_diskquota catalogWhile ALTER SET way is more close to store quota setting in pg_class, pg_role, pg_namespace. etc in an integrated way.(Note that here I mean nature/close is not must, ALTER SET could also store in pg_diskquota and vice versa.)Here are some differences I can think of:1 pg_role is a global catalog, not per database level. It's harder to tracker the user's disk usage in the whole clusters(considering 1000+ databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it only tracks the user's disk usage inside the current database.2 using separate pg_diskquota could add more field except for quota limit without adding too many fields in pg_class, e.g. red zone to give the user a warning or the current disk usage of the db objects.On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:Hi all,We redesign disk quota feature based on the comments from Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk quota feature is used to support multi-tenancy environment, different level of database objects could be set a quota limit to avoid over use of disk space. A common case could be as follows: DBA could enable disk quota on a specified database list. DBA could set disk quota limit for tables/schemas/roles in these databases. Separate disk quota worker process will monitor the disk usage for these objects and detect the objects which exceed their quota limit. Queries loading data into these “out of disk quota” tables/schemas/roles will be cancelled.We are currently working at init implementation stage. We would like to propose our idea firstly and get feedbacks from community to do quick iteration.SQL Syntax (How to use disk quota)1 Specify the databases with disk quota enabled in GUC “diskquota_databases” in postgresql.conf and restart the database.2 DBA could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer . 3 Simulate a schema out of quota limit case: suppose table a1 and table a2 are both under schema s1.INSERT INTO a1 SELECT generate_series(1,1000);INSERT INTO a2 SELECT generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);ERROR: schema's disk space quota exceededDROP TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the following components.1. Quota Setting Store is where the disk quota setting to be stored and accessed. We plan to use catalog table pg_diskquota to store these information. pg_diskquota is like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /* diskquota name */ int16 quotatype; /* diskquota type name */ Oid quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /* diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size change of database objects. We plan to use stat collector to detect the ‘active’ table list at initial stage. But stat collector has some limitation on finding the active table which is in a running transaction. Details see TODO section.3. Quota Size Checker is where to calculate the size and compare with quota limit for database objects. According to Pavel’s comment, autovacuum launcher and worker process could be a good reference to disk quota. So we plan to use a disk quota launcher daemon process and several disk quota worker process to finish this work. Launcher process is responsible for starting worker process based on a user defined database list from GUC. Worker process will connect to its target database and monitor the disk usage for objects in this database. In init stage of worker process, it will call calculate_total_relation_size() to calculate the size for each user table. After init stage, worker process will refresh the disk model every N seconds. Refreshing will only recalculate the size of tables in ‘active’ table list, which is generated by Quata Change Detector to minimize the cost.4. Quota Enforcement Operator is where to check for the quota limitation at postgres backend side. We will firstly implement it in ExecCheckRTPerms() as pre-running enforcement. It will check the disk quota of tables being inserted or updated, and report error if table’s or table’s schema’s or table’s owner’s quota limit is exceeded. As a native feature, we plan to add more checkpoint to do running query enforcement. For example, if a disk quota lefts 10MB quota, a query could insert 1GB data. This query could be allowed in pre-running enforcement check, but will be cancelled in running query enforcement check. Therefore, it can improve the accurate of disk quota usage. To achieve this, we plan to add a checkpoint in lower API such as smgr_extened. Hence, the Quota Enforcement Operator will check the disk quota usage when smgr_extened is called. If the quota is over limited, current query will be cancelled. Highlight1. Native feature.Support native Create/Drop Disk Quota SQL statement.New catalog table pg_diskquota to store disk quota setting.2. Auto DML/DDL detection. Table create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change, Schema create/drop and Role create/drop will be detected by disk quota automatically. 3. Low cost disk quota checker.Worker process of disk quota need to refresh the disk usage model every N seconds. Since recalculate the file size using stat() system call is expensive for a large number of files, we use an ‘active’ table list to reduce the real work at each iteration. A basic experiment on our init stage implementation on database with 20K tables shows that the refresh cost is 1% cpu usage and will be finished within 50ms. Todo/LimitationBefore we propose our patch, we plan to enhance it with the following ideas:1. Setting database list with disk quota enabled dynamically without restart database. Since we have the disk quota launcher process, it could detect the new ‘diskquota_databases’ list and start/stop the corresponding disk quota worker process.2. Enforcement when query is running. Considering the case when there is 10MB quota left, but next query will insert 10GB data. Current enforcement design will allow this query to be executed. This is limited by the ‘active’ table detection is generated by stat collector. Postgres backend will only send table stat information to collector only when the transaction ends. We need a new way to detect the ‘active’ table even when this table is being modified inside a running transaction.3. Monitor unlimited number of databases. Current we set the max number of disk quota worker process to be 10 to reduce the affection normal workload. But how about if we want to monitor the disk quota of more than 10 databases? Our solution is to let disk quota launcher to manage a queue of database need to be monitored. And disk quota worker process consuming the queue and refresh the disk usage/quota for this database. After some periods, worker will return the database to the queue, and fetch the top database from queue to process. The period determine the delay of detecting disk quota change. To implement this feature, we need to support a subprocess of postmaster to rebind to another database instead of the database binded in InitPostgres().4. Support active table detection on vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman Flack for the former comments on disk quota feature. Any comments on how to improve disk quota feature are appreciated.On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \"implement some quotas on storage level?\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n\n> 2 Keep one worker process for each database. But using a parent/global\n> quota worker process to manage the lifecycle of database level worker\n> processes. It could handle the newly created database(avoid restart\n> database) and save resource when a database is not used. But this needs to\n> change worker process to be hierarchical. Postmaster becomes the grandfather\n> of database level worker processes in this case.\n\nI am using background workers this way in 9.5 at $work.\n\nIn my case, one worker lives forever, wakes up on a set period, and\nstarts a short-lived worker for every database, waiting for each\none before starting the next.\n\nIt was straightforward to implement. Looking back over the code,\nI see the global worker assigns its own PID to worker.bgw_notify_pid\nof each of its children, and also obtains a handle for each child\nfrom RegisterDynamicBackgroundWorker().\n\nI imagine the global quota worker would prefer to start workers\nfor every database and then just wait for notifications from any\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\n\n-Chap\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang","msg_date":"Fri, 21 Sep 2018 22:21:16 +0800","msg_from":"Hubert Zhang <hzhang@pivotal.io>","msg_from_op":true,"msg_subject":"Re: Proposal for disk quota feature"},{"msg_contents":"On Fri, Sep 21, 2018 at 9:21 AM Hubert Zhang <hzhang@pivotal.io> wrote:\n\n> just fast reaction - why QUOTA object?\n>> Isn't ALTER SET enough?\n>> Some like\n>> ALTER TABLE a1 SET quote = 1MB;\n>> ALTER USER ...\n>> ALTER SCHEMA ..\n>> New DDL commans looks like too hard hammer .\n>\n>\n> It's an option. Prefer to consider quota setting store together:\n> CREATE DISK QUOTA way is more nature to store quota setting in a separate\n> pg_diskquota catalog\n> While ALTER SET way is more close to store quota setting in pg_class,\n> pg_role, pg_namespace. etc in an integrated way.\n> (Note that here I mean nature/close is not must, ALTER SET could also\n> store in pg_diskquota and vice versa.)\n>\n> Here are some differences I can think of:\n> 1 pg_role is a global catalog, not per database level. It's harder to\n> tracker the user's disk usage in the whole clusters(considering 1000+\n> databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it\n> only tracks the user's disk usage inside the current database.\n> 2 using separate pg_diskquota could add more field except for quota limit\n> without adding too many fields in pg_class, e.g. red zone to give the user\n> a warning or the current disk usage of the db objects.\n>\n> On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com>\n> wrote:\n>\n>>\n>>\n>> pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io>\n>> napsal:\n>>\n>>>\n>>>\n>>>\n>>>\n>>> *Hi all,We redesign disk quota feature based on the comments from Pavel\n>>> Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk\n>>> quota feature is used to support multi-tenancy environment, different level\n>>> of database objects could be set a quota limit to avoid over use of disk\n>>> space. A common case could be as follows: DBA could enable disk quota on a\n>>> specified database list. DBA could set disk quota limit for\n>>> tables/schemas/roles in these databases. Separate disk quota worker process\n>>> will monitor the disk usage for these objects and detect the objects which\n>>> exceed their quota limit. Queries loading data into these “out of disk\n>>> quota” tables/schemas/roles will be cancelled.We are currently working at\n>>> init implementation stage. We would like to propose our idea firstly and\n>>> get feedbacks from community to do quick iteration.SQL Syntax (How to use\n>>> disk quota)1 Specify the databases with disk quota enabled in GUC\n>>> “diskquota_databases” in postgresql.conf and restart the database.2 DBA\n>>> could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\n>>> ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\n>>> (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\n>>> ‘3MB’);*\n>>>\n>>\n>> just fast reaction - why QUOTA object?\n>>\n>> Isn't ALTER SET enough?\n>>\n>> Some like\n>>\n>> ALTER TABLE a1 SET quote = 1MB;\n>> ALTER USER ...\n>> ALTER SCHEMA ..\n>>\n>> New DDL commans looks like too hard hammer .\n>>\n>>\n>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>> *3 Simulate a schema out of quota limit case: suppose table a1 and table\n>>> a2 are both under schema s1.INSERT INTO a1 SELECT\n>>> generate_series(1,1000);INSERT INTO a2 SELECT\n>>> generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\n>>> generate_series(1,1000);ERROR: schema's disk space quota exceededDROP\n>>> TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\n>>> generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\n>>> following components.1. Quota Setting Store is where the disk quota setting\n>>> to be stored and accessed. We plan to use catalog table pg_diskquota to\n>>> store these information. pg_diskquota is\n>>> like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\n>>> diskquota name */ int16 quotatype; /* diskquota type name */ Oid\n>>> quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\n>>> diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\n>>> MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\n>>> change of database objects. We plan to use stat collector to detect the\n>>> ‘active’ table list at initial stage. But stat collector has some\n>>> limitation on finding the active table which is in a running transaction.\n>>> Details see TODO section.3. Quota Size Checker is where to calculate the\n>>> size and compare with quota limit for database objects. According to\n>>> Pavel’s comment, autovacuum launcher and worker process could be a good\n>>> reference to disk quota. So we plan to use a disk quota launcher daemon\n>>> process and several disk quota worker process to finish this work. Launcher\n>>> process is responsible for starting worker process based on a user defined\n>>> database list from GUC. Worker process will connect to its target database\n>>> and monitor the disk usage for objects in this database. In init stage of\n>>> worker process, it will call calculate_total_relation_size() to calculate\n>>> the size for each user table. After init stage, worker process will refresh\n>>> the disk model every N seconds. Refreshing will only recalculate the size\n>>> of tables in ‘active’ table list, which is generated by Quata Change\n>>> Detector to minimize the cost.4. Quota Enforcement Operator is where to\n>>> check for the quota limitation at postgres backend side. We will firstly\n>>> implement it in ExecCheckRTPerms() as pre-running enforcement. It will\n>>> check the disk quota of tables being inserted or updated, and report error\n>>> if table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\n>>> As a native feature, we plan to add more checkpoint to do running query\n>>> enforcement. For example, if a disk quota lefts 10MB quota, a query could\n>>> insert 1GB data. This query could be allowed in pre-running enforcement\n>>> check, but will be cancelled in running query enforcement check. Therefore,\n>>> it can improve the accurate of disk quota usage. To achieve this, we plan\n>>> to add a checkpoint in lower API such as smgr_extened. Hence, the Quota\n>>> Enforcement Operator will check the disk quota usage when smgr_extened is\n>>> called. If the quota is over limited, current query will be cancelled.\n>>> Highlight1. Native feature.Support native Create/Drop Disk Quota SQL\n>>> statement.New catalog table pg_diskquota to store disk quota setting.2.\n>>> Auto DML/DDL detection. Table\n>>> create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\n>>> Schema create/drop and Role create/drop will be detected by disk quota\n>>> automatically. 3. Low cost disk quota checker.Worker process of disk quota\n>>> need to refresh the disk usage model every N seconds. Since recalculate the\n>>> file size using stat() system call is expensive for a large number of\n>>> files, we use an ‘active’ table list to reduce the real work at each\n>>> iteration. A basic experiment on our init stage implementation on database\n>>> with 20K tables shows that the refresh cost is 1% cpu usage and will be\n>>> finished within 50ms. Todo/LimitationBefore we propose our patch, we plan\n>>> to enhance it with the following ideas:1. Setting database list with disk\n>>> quota enabled dynamically without restart database. Since we have the disk\n>>> quota launcher process, it could detect the new ‘diskquota_databases’ list\n>>> and start/stop the corresponding disk quota worker process.2. Enforcement\n>>> when query is running. Considering the case when there is 10MB quota left,\n>>> but next query will insert 10GB data. Current enforcement design will allow\n>>> this query to be executed. This is limited by the ‘active’ table detection\n>>> is generated by stat collector. Postgres backend will only send table stat\n>>> information to collector only when the transaction ends. We need a new way\n>>> to detect the ‘active’ table even when this table is being modified inside\n>>> a running transaction.3. Monitor unlimited number of databases. Current we\n>>> set the max number of disk quota worker process to be 10 to reduce the\n>>> affection normal workload. But how about if we want to monitor the disk\n>>> quota of more than 10 databases? Our solution is to let disk quota launcher\n>>> to manage a queue of database need to be monitored. And disk quota worker\n>>> process consuming the queue and refresh the disk usage/quota for this\n>>> database. After some periods, worker will return the database to the queue,\n>>> and fetch the top database from queue to process. The period determine the\n>>> delay of detecting disk quota change. To implement this feature, we need to\n>>> support a subprocess of postmaster to rebind to another database instead of\n>>> the database binded in InitPostgres().4. Support active table detection on\n>>> vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\n>>> are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\n>>> Flack for the former comments on disk quota feature. Any comments on how to\n>>> improve disk quota feature are appreciated.*\n>>>\n>>>\n>>> On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com>\n>>> wrote:\n>>>\n>>>>\n>>>>\n>>>> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>>>\n>>>>> Thanks Pavel.\n>>>>> Your patch did enforcement on storage level(md.c or we could also use\n>>>>> smgr_extend). It's straight forward.\n>>>>> But I prefer to implement disk_quota as a feature with following\n>>>>> objectives:\n>>>>> 1 set/alter disk quota setting on different database objects, e.g.\n>>>>> user, database, schema etc. not only a general GUC, but we could set\n>>>>> separate quota limit for a specific objects.\n>>>>> 2 enforcement operator should work at two positions: before query is\n>>>>> running and when query is running. The latter one's implementation maybe\n>>>>> similar to your patch.\n>>>>>\n>>>>\n>>>> The patch was just example. The resource quotes should be more complex\n>>>> - per partition, table, schema, database, user - so GUC are possible, but\n>>>> not very user friendly.\n>>>>\n>>>> Our case is specific, but not too much. The servers are used for\n>>>> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\n>>>> SELECT). We need to solve limits immediately. The implementation is simple,\n>>>> so I did it. Same implementation on database level, or schema level needs\n>>>> some more locks, so it will not be too effective. The resource management\n>>>> can be complex very complex, and I expect so it will be hard work.\n>>>>\n>>>> Regards\n>>>>\n>>>> Pavel\n>>>>\n>>>>\n>>>>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com\n>>>>> > wrote:\n>>>>>\n>>>>>> Hi\n>>>>>>\n>>>>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>>>>>\n>>>>>>> Thanks Chapman.\n>>>>>>> @Pavel, could you please explain more about your second suggestion\n>>>>>>> \"implement some quotas on storage level?\"\n>>>>>>>\n>>>>>>\n>>>>>> See attached patch - it is very simple - and good enough for our\n>>>>>> purposes.\n>>>>>>\n>>>>>> Regards\n>>>>>>\n>>>>>> Pavel\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>> We will not keep the long-lived processes attach to all\n>>>>>>> databases(just like you mentioned servers with thousands of\n>>>>>>> databases)\n>>>>>>> And you are right, we could share ideas with autovacuum process,\n>>>>>>> fork worker processes in need.\n>>>>>>> \"autovacuum checks for tables that have had a large number of\n>>>>>>> inserted, updated or deleted tuples. These checks use the statistics\n>>>>>>> collection facility\"\n>>>>>>> diskquota process is similar to autovacuum at caring about insert,\n>>>>>>> but the difference is that it also care about vucuum full, truncate and\n>>>>>>> drop. While update and delete may not be interested since no file change\n>>>>>>> happens. So a separate diskquota process is preferred.\n>>>>>>>\n>>>>>>> So if we implemented disk quota as a full native feature, and in\n>>>>>>> the first initial version I prefer to implement the following features:\n>>>>>>> 1 Fork diskquota launcher process under Postmaster serverloop, which\n>>>>>>> is long-lived.\n>>>>>>> 2 Diskquota launcher process is responsible for creating diskquota\n>>>>>>> worker process for every database.\n>>>>>>> 3 DIskquota setting is stored in a separate catalog table for each\n>>>>>>> database.\n>>>>>>> 4 Initialization stage, Diskquota launcher process creates diskquota\n>>>>>>> worker process for all the databases(traverse like autovacuum).\n>>>>>>> Worker process calculates disk usage of db objects and\n>>>>>>> their diskquota setting. If any db object exceeds its quota limit, put them\n>>>>>>> into the blacklist in the shared memory, which will later be used by\n>>>>>>> enforcement operator. Worker process exits when works are done.\n>>>>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\n>>>>>>> process for the database with a large number of insert, copy, truncate,\n>>>>>>> drop etc. or create disk quota statement. Worker process updates the file\n>>>>>>> size for db objects containing the result relation, and compare with the\n>>>>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\n>>>>>>> remove from blacklist vice versa. Worker process exits when works\n>>>>>>> are done and a GUC could control the frequency of worker process restart to\n>>>>>>> a specific database. As you know, this GUC also controls the delay when we\n>>>>>>> do enforcement.\n>>>>>>> 6 Enforcement. When postgres backend executes queries, check the\n>>>>>>> blacklist in shared memory to determine whether the query is allowed(before\n>>>>>>> execute) or need rollback(is executing)?\n>>>>>>>\n>>>>>>> If we implemented disk quota as an extension, we could just use\n>>>>>>> background worker to start diskquota launcher process and use\n>>>>>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\n>>>>>>> processes by the launcher process as suggested by @Chapman.\n>>>>>>> Diskquota setting could be stored in user table in a separate schema for\n>>>>>>> each database(Schema and table created by create extension statement) just\n>>>>>>> like what Heikki has done in pg_quota project. But in this case, we need to\n>>>>>>> create extension for each database before diskquota worker process can be\n>>>>>>> set up for that database.\n>>>>>>>\n>>>>>>> Any comments on the above design and which is preferred, native\n>>>>>>> feature or extension as the POC?\n>>>>>>>\n>>>>>>>\n>>>>>>> -- Hubert\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <\n>>>>>>> pavel.stehule@gmail.com> wrote:\n>>>>>>>\n>>>>>>>>\n>>>>>>>>\n>>>>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\n>>>>>>>>\n>>>>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n>>>>>>>>>\n>>>>>>>>> > 2 Keep one worker process for each database. But using a\n>>>>>>>>> parent/global\n>>>>>>>>> > quota worker process to manage the lifecycle of database level\n>>>>>>>>> worker\n>>>>>>>>> > processes. It could handle the newly created database(avoid\n>>>>>>>>> restart\n>>>>>>>>> > database) and save resource when a database is not used. But\n>>>>>>>>> this needs to\n>>>>>>>>> > change worker process to be hierarchical. Postmaster becomes the\n>>>>>>>>> grandfather\n>>>>>>>>> > of database level worker processes in this case.\n>>>>>>>>>\n>>>>>>>>> I am using background workers this way in 9.5 at $work.\n>>>>>>>>>\n>>>>>>>>> In my case, one worker lives forever, wakes up on a set period, and\n>>>>>>>>> starts a short-lived worker for every database, waiting for each\n>>>>>>>>> one before starting the next.\n>>>>>>>>>\n>>>>>>>>> It was straightforward to implement. Looking back over the code,\n>>>>>>>>> I see the global worker assigns its own PID to\n>>>>>>>>> worker.bgw_notify_pid\n>>>>>>>>> of each of its children, and also obtains a handle for each child\n>>>>>>>>> from RegisterDynamicBackgroundWorker().\n>>>>>>>>>\n>>>>>>>>> I imagine the global quota worker would prefer to start workers\n>>>>>>>>> for every database and then just wait for notifications from any\n>>>>>>>>> of them, but that seems equally straightforward at first glance.\n>>>>>>>>>\n>>>>>>>>\n>>>>>>>> There are servers with thousands databases. Worker per database is\n>>>>>>>> not good idea.\n>>>>>>>>\n>>>>>>>> It should to share ideas, code with autovacuum process.\n>>>>>>>>\n>>>>>>>> Not sure, how to effective implementation based on bg workers can\n>>>>>>>> be. On servers with large set of databases, large set of tables it can\n>>>>>>>> identify too big table too late.\n>>>>>>>>\n>>>>>>>> Isn't better to implement some quotas on storage level?\n>>>>>>>>\n>>>>>>>\nFWIW, we are super interested in this feature at a role level especially in\nreporting systems where users can for example accidentally start a create\ntable statement overnight and sure enough after their Cartesian product\ngenerated 2TB of data they run the system out of disk.\n\nThat problem is really hard to manage currently and this feature would\nreally solve that.\n\nThanks,\nJeremy\n\n>\n\nOn Fri, Sep 21, 2018 at 9:21 AM Hubert Zhang <hzhang@pivotal.io> wrote:just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer .It's an option. Prefer to consider quota setting store together:CREATE DISK QUOTA way is more nature to store quota setting in a separate pg_diskquota catalogWhile ALTER SET way is more close to store quota setting in pg_class, pg_role, pg_namespace. etc in an integrated way.(Note that here I mean nature/close is not must, ALTER SET could also store in pg_diskquota and vice versa.)Here are some differences I can think of:1 pg_role is a global catalog, not per database level. It's harder to tracker the user's disk usage in the whole clusters(considering 1000+ databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it only tracks the user's disk usage inside the current database.2 using separate pg_diskquota could add more field except for quota limit without adding too many fields in pg_class, e.g. red zone to give the user a warning or the current disk usage of the db objects.On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:Hi all,We redesign disk quota feature based on the comments from Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk quota feature is used to support multi-tenancy environment, different level of database objects could be set a quota limit to avoid over use of disk space. A common case could be as follows: DBA could enable disk quota on a specified database list. DBA could set disk quota limit for tables/schemas/roles in these databases. Separate disk quota worker process will monitor the disk usage for these objects and detect the objects which exceed their quota limit. Queries loading data into these “out of disk quota” tables/schemas/roles will be cancelled.We are currently working at init implementation stage. We would like to propose our idea firstly and get feedbacks from community to do quick iteration.SQL Syntax (How to use disk quota)1 Specify the databases with disk quota enabled in GUC “diskquota_databases” in postgresql.conf and restart the database.2 DBA could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer . 3 Simulate a schema out of quota limit case: suppose table a1 and table a2 are both under schema s1.INSERT INTO a1 SELECT generate_series(1,1000);INSERT INTO a2 SELECT generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);ERROR: schema's disk space quota exceededDROP TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the following components.1. Quota Setting Store is where the disk quota setting to be stored and accessed. We plan to use catalog table pg_diskquota to store these information. pg_diskquota is like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /* diskquota name */ int16 quotatype; /* diskquota type name */ Oid quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /* diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size change of database objects. We plan to use stat collector to detect the ‘active’ table list at initial stage. But stat collector has some limitation on finding the active table which is in a running transaction. Details see TODO section.3. Quota Size Checker is where to calculate the size and compare with quota limit for database objects. According to Pavel’s comment, autovacuum launcher and worker process could be a good reference to disk quota. So we plan to use a disk quota launcher daemon process and several disk quota worker process to finish this work. Launcher process is responsible for starting worker process based on a user defined database list from GUC. Worker process will connect to its target database and monitor the disk usage for objects in this database. In init stage of worker process, it will call calculate_total_relation_size() to calculate the size for each user table. After init stage, worker process will refresh the disk model every N seconds. Refreshing will only recalculate the size of tables in ‘active’ table list, which is generated by Quata Change Detector to minimize the cost.4. Quota Enforcement Operator is where to check for the quota limitation at postgres backend side. We will firstly implement it in ExecCheckRTPerms() as pre-running enforcement. It will check the disk quota of tables being inserted or updated, and report error if table’s or table’s schema’s or table’s owner’s quota limit is exceeded. As a native feature, we plan to add more checkpoint to do running query enforcement. For example, if a disk quota lefts 10MB quota, a query could insert 1GB data. This query could be allowed in pre-running enforcement check, but will be cancelled in running query enforcement check. Therefore, it can improve the accurate of disk quota usage. To achieve this, we plan to add a checkpoint in lower API such as smgr_extened. Hence, the Quota Enforcement Operator will check the disk quota usage when smgr_extened is called. If the quota is over limited, current query will be cancelled. Highlight1. Native feature.Support native Create/Drop Disk Quota SQL statement.New catalog table pg_diskquota to store disk quota setting.2. Auto DML/DDL detection. Table create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change, Schema create/drop and Role create/drop will be detected by disk quota automatically. 3. Low cost disk quota checker.Worker process of disk quota need to refresh the disk usage model every N seconds. Since recalculate the file size using stat() system call is expensive for a large number of files, we use an ‘active’ table list to reduce the real work at each iteration. A basic experiment on our init stage implementation on database with 20K tables shows that the refresh cost is 1% cpu usage and will be finished within 50ms. Todo/LimitationBefore we propose our patch, we plan to enhance it with the following ideas:1. Setting database list with disk quota enabled dynamically without restart database. Since we have the disk quota launcher process, it could detect the new ‘diskquota_databases’ list and start/stop the corresponding disk quota worker process.2. Enforcement when query is running. Considering the case when there is 10MB quota left, but next query will insert 10GB data. Current enforcement design will allow this query to be executed. This is limited by the ‘active’ table detection is generated by stat collector. Postgres backend will only send table stat information to collector only when the transaction ends. We need a new way to detect the ‘active’ table even when this table is being modified inside a running transaction.3. Monitor unlimited number of databases. Current we set the max number of disk quota worker process to be 10 to reduce the affection normal workload. But how about if we want to monitor the disk quota of more than 10 databases? Our solution is to let disk quota launcher to manage a queue of database need to be monitored. And disk quota worker process consuming the queue and refresh the disk usage/quota for this database. After some periods, worker will return the database to the queue, and fetch the top database from queue to process. The period determine the delay of detecting disk quota change. To implement this feature, we need to support a subprocess of postmaster to rebind to another database instead of the database binded in InitPostgres().4. Support active table detection on vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman Flack for the former comments on disk quota feature. Any comments on how to improve disk quota feature are appreciated.On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \"implement some quotas on storage level?\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n\n> 2 Keep one worker process for each database. But using a parent/global\n> quota worker process to manage the lifecycle of database level worker\n> processes. It could handle the newly created database(avoid restart\n> database) and save resource when a database is not used. But this needs to\n> change worker process to be hierarchical. Postmaster becomes the grandfather\n> of database level worker processes in this case.\n\nI am using background workers this way in 9.5 at $work.\n\nIn my case, one worker lives forever, wakes up on a set period, and\nstarts a short-lived worker for every database, waiting for each\none before starting the next.\n\nIt was straightforward to implement. Looking back over the code,\nI see the global worker assigns its own PID to worker.bgw_notify_pid\nof each of its children, and also obtains a handle for each child\nfrom RegisterDynamicBackgroundWorker().\n\nI imagine the global quota worker would prefer to start workers\nfor every database and then just wait for notifications from any\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?FWIW, we are super interested in this feature at a role level especially in reporting systems where users can for example accidentally start a create table statement overnight and sure enough after their Cartesian product generated 2TB of data they run the system out of disk.That problem is really hard to manage currently and this feature would really solve that.Thanks,Jeremy","msg_date":"Fri, 21 Sep 2018 16:33:38 -0500","msg_from":"Jeremy Finzel <finzelj@gmail.com>","msg_from_op":false,"msg_subject":"Re: Proposal for disk quota feature"},{"msg_contents":"pá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:\n\n> just fast reaction - why QUOTA object?\n>> Isn't ALTER SET enough?\n>> Some like\n>> ALTER TABLE a1 SET quote = 1MB;\n>> ALTER USER ...\n>> ALTER SCHEMA ..\n>> New DDL commans looks like too hard hammer .\n>\n>\n> It's an option. Prefer to consider quota setting store together:\n> CREATE DISK QUOTA way is more nature to store quota setting in a separate\n> pg_diskquota catalog\n> While ALTER SET way is more close to store quota setting in pg_class,\n> pg_role, pg_namespace. etc in an integrated way.\n> (Note that here I mean nature/close is not must, ALTER SET could also\n> store in pg_diskquota and vice versa.)\n>\n\nI have not a problem with new special table for storing this information.\nBut it looks like redundant to current GUC configuration and limits. Can be\nmessy do some work with ALTER ROLE, and some work via CREATE QUOTE.\n\nRegards\n\nPavel\n\n\n> Here are some differences I can think of:\n> 1 pg_role is a global catalog, not per database level. It's harder to\n> tracker the user's disk usage in the whole clusters(considering 1000+\n> databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it\n> only tracks the user's disk usage inside the current database.\n> 2 using separate pg_diskquota could add more field except for quota limit\n> without adding too many fields in pg_class, e.g. red zone to give the user\n> a warning or the current disk usage of the db objects.\n>\n> On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com>\n> wrote:\n>\n>>\n>>\n>> pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io>\n>> napsal:\n>>\n>>>\n>>>\n>>>\n>>>\n>>> *Hi all,We redesign disk quota feature based on the comments from Pavel\n>>> Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk\n>>> quota feature is used to support multi-tenancy environment, different level\n>>> of database objects could be set a quota limit to avoid over use of disk\n>>> space. A common case could be as follows: DBA could enable disk quota on a\n>>> specified database list. DBA could set disk quota limit for\n>>> tables/schemas/roles in these databases. Separate disk quota worker process\n>>> will monitor the disk usage for these objects and detect the objects which\n>>> exceed their quota limit. Queries loading data into these “out of disk\n>>> quota” tables/schemas/roles will be cancelled.We are currently working at\n>>> init implementation stage. We would like to propose our idea firstly and\n>>> get feedbacks from community to do quick iteration.SQL Syntax (How to use\n>>> disk quota)1 Specify the databases with disk quota enabled in GUC\n>>> “diskquota_databases” in postgresql.conf and restart the database.2 DBA\n>>> could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\n>>> ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\n>>> (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\n>>> ‘3MB’);*\n>>>\n>>\n>> just fast reaction - why QUOTA object?\n>>\n>> Isn't ALTER SET enough?\n>>\n>> Some like\n>>\n>> ALTER TABLE a1 SET quote = 1MB;\n>> ALTER USER ...\n>> ALTER SCHEMA ..\n>>\n>> New DDL commans looks like too hard hammer .\n>>\n>>\n>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>>\n>>> *3 Simulate a schema out of quota limit case: suppose table a1 and table\n>>> a2 are both under schema s1.INSERT INTO a1 SELECT\n>>> generate_series(1,1000);INSERT INTO a2 SELECT\n>>> generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\n>>> generate_series(1,1000);ERROR: schema's disk space quota exceededDROP\n>>> TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\n>>> generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\n>>> following components.1. Quota Setting Store is where the disk quota setting\n>>> to be stored and accessed. We plan to use catalog table pg_diskquota to\n>>> store these information. pg_diskquota is\n>>> like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\n>>> diskquota name */ int16 quotatype; /* diskquota type name */ Oid\n>>> quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\n>>> diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\n>>> MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\n>>> change of database objects. We plan to use stat collector to detect the\n>>> ‘active’ table list at initial stage. But stat collector has some\n>>> limitation on finding the active table which is in a running transaction.\n>>> Details see TODO section.3. Quota Size Checker is where to calculate the\n>>> size and compare with quota limit for database objects. According to\n>>> Pavel’s comment, autovacuum launcher and worker process could be a good\n>>> reference to disk quota. So we plan to use a disk quota launcher daemon\n>>> process and several disk quota worker process to finish this work. Launcher\n>>> process is responsible for starting worker process based on a user defined\n>>> database list from GUC. Worker process will connect to its target database\n>>> and monitor the disk usage for objects in this database. In init stage of\n>>> worker process, it will call calculate_total_relation_size() to calculate\n>>> the size for each user table. After init stage, worker process will refresh\n>>> the disk model every N seconds. Refreshing will only recalculate the size\n>>> of tables in ‘active’ table list, which is generated by Quata Change\n>>> Detector to minimize the cost.4. Quota Enforcement Operator is where to\n>>> check for the quota limitation at postgres backend side. We will firstly\n>>> implement it in ExecCheckRTPerms() as pre-running enforcement. It will\n>>> check the disk quota of tables being inserted or updated, and report error\n>>> if table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\n>>> As a native feature, we plan to add more checkpoint to do running query\n>>> enforcement. For example, if a disk quota lefts 10MB quota, a query could\n>>> insert 1GB data. This query could be allowed in pre-running enforcement\n>>> check, but will be cancelled in running query enforcement check. Therefore,\n>>> it can improve the accurate of disk quota usage. To achieve this, we plan\n>>> to add a checkpoint in lower API such as smgr_extened. Hence, the Quota\n>>> Enforcement Operator will check the disk quota usage when smgr_extened is\n>>> called. If the quota is over limited, current query will be cancelled.\n>>> Highlight1. Native feature.Support native Create/Drop Disk Quota SQL\n>>> statement.New catalog table pg_diskquota to store disk quota setting.2.\n>>> Auto DML/DDL detection. Table\n>>> create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\n>>> Schema create/drop and Role create/drop will be detected by disk quota\n>>> automatically. 3. Low cost disk quota checker.Worker process of disk quota\n>>> need to refresh the disk usage model every N seconds. Since recalculate the\n>>> file size using stat() system call is expensive for a large number of\n>>> files, we use an ‘active’ table list to reduce the real work at each\n>>> iteration. A basic experiment on our init stage implementation on database\n>>> with 20K tables shows that the refresh cost is 1% cpu usage and will be\n>>> finished within 50ms. Todo/LimitationBefore we propose our patch, we plan\n>>> to enhance it with the following ideas:1. Setting database list with disk\n>>> quota enabled dynamically without restart database. Since we have the disk\n>>> quota launcher process, it could detect the new ‘diskquota_databases’ list\n>>> and start/stop the corresponding disk quota worker process.2. Enforcement\n>>> when query is running. Considering the case when there is 10MB quota left,\n>>> but next query will insert 10GB data. Current enforcement design will allow\n>>> this query to be executed. This is limited by the ‘active’ table detection\n>>> is generated by stat collector. Postgres backend will only send table stat\n>>> information to collector only when the transaction ends. We need a new way\n>>> to detect the ‘active’ table even when this table is being modified inside\n>>> a running transaction.3. Monitor unlimited number of databases. Current we\n>>> set the max number of disk quota worker process to be 10 to reduce the\n>>> affection normal workload. But how about if we want to monitor the disk\n>>> quota of more than 10 databases? Our solution is to let disk quota launcher\n>>> to manage a queue of database need to be monitored. And disk quota worker\n>>> process consuming the queue and refresh the disk usage/quota for this\n>>> database. After some periods, worker will return the database to the queue,\n>>> and fetch the top database from queue to process. The period determine the\n>>> delay of detecting disk quota change. To implement this feature, we need to\n>>> support a subprocess of postmaster to rebind to another database instead of\n>>> the database binded in InitPostgres().4. Support active table detection on\n>>> vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\n>>> are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\n>>> Flack for the former comments on disk quota feature. Any comments on how to\n>>> improve disk quota feature are appreciated.*\n>>>\n>>>\n>>> On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com>\n>>> wrote:\n>>>\n>>>>\n>>>>\n>>>> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>>>\n>>>>> Thanks Pavel.\n>>>>> Your patch did enforcement on storage level(md.c or we could also use\n>>>>> smgr_extend). It's straight forward.\n>>>>> But I prefer to implement disk_quota as a feature with following\n>>>>> objectives:\n>>>>> 1 set/alter disk quota setting on different database objects, e.g.\n>>>>> user, database, schema etc. not only a general GUC, but we could set\n>>>>> separate quota limit for a specific objects.\n>>>>> 2 enforcement operator should work at two positions: before query is\n>>>>> running and when query is running. The latter one's implementation maybe\n>>>>> similar to your patch.\n>>>>>\n>>>>\n>>>> The patch was just example. The resource quotes should be more complex\n>>>> - per partition, table, schema, database, user - so GUC are possible, but\n>>>> not very user friendly.\n>>>>\n>>>> Our case is specific, but not too much. The servers are used for\n>>>> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\n>>>> SELECT). We need to solve limits immediately. The implementation is simple,\n>>>> so I did it. Same implementation on database level, or schema level needs\n>>>> some more locks, so it will not be too effective. The resource management\n>>>> can be complex very complex, and I expect so it will be hard work.\n>>>>\n>>>> Regards\n>>>>\n>>>> Pavel\n>>>>\n>>>>\n>>>>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com\n>>>>> > wrote:\n>>>>>\n>>>>>> Hi\n>>>>>>\n>>>>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>>>>>\n>>>>>>> Thanks Chapman.\n>>>>>>> @Pavel, could you please explain more about your second suggestion\n>>>>>>> \"implement some quotas on storage level?\"\n>>>>>>>\n>>>>>>\n>>>>>> See attached patch - it is very simple - and good enough for our\n>>>>>> purposes.\n>>>>>>\n>>>>>> Regards\n>>>>>>\n>>>>>> Pavel\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>> We will not keep the long-lived processes attach to all\n>>>>>>> databases(just like you mentioned servers with thousands of\n>>>>>>> databases)\n>>>>>>> And you are right, we could share ideas with autovacuum process,\n>>>>>>> fork worker processes in need.\n>>>>>>> \"autovacuum checks for tables that have had a large number of\n>>>>>>> inserted, updated or deleted tuples. These checks use the statistics\n>>>>>>> collection facility\"\n>>>>>>> diskquota process is similar to autovacuum at caring about insert,\n>>>>>>> but the difference is that it also care about vucuum full, truncate and\n>>>>>>> drop. While update and delete may not be interested since no file change\n>>>>>>> happens. So a separate diskquota process is preferred.\n>>>>>>>\n>>>>>>> So if we implemented disk quota as a full native feature, and in\n>>>>>>> the first initial version I prefer to implement the following features:\n>>>>>>> 1 Fork diskquota launcher process under Postmaster serverloop, which\n>>>>>>> is long-lived.\n>>>>>>> 2 Diskquota launcher process is responsible for creating diskquota\n>>>>>>> worker process for every database.\n>>>>>>> 3 DIskquota setting is stored in a separate catalog table for each\n>>>>>>> database.\n>>>>>>> 4 Initialization stage, Diskquota launcher process creates diskquota\n>>>>>>> worker process for all the databases(traverse like autovacuum).\n>>>>>>> Worker process calculates disk usage of db objects and\n>>>>>>> their diskquota setting. If any db object exceeds its quota limit, put them\n>>>>>>> into the blacklist in the shared memory, which will later be used by\n>>>>>>> enforcement operator. Worker process exits when works are done.\n>>>>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\n>>>>>>> process for the database with a large number of insert, copy, truncate,\n>>>>>>> drop etc. or create disk quota statement. Worker process updates the file\n>>>>>>> size for db objects containing the result relation, and compare with the\n>>>>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\n>>>>>>> remove from blacklist vice versa. Worker process exits when works\n>>>>>>> are done and a GUC could control the frequency of worker process restart to\n>>>>>>> a specific database. As you know, this GUC also controls the delay when we\n>>>>>>> do enforcement.\n>>>>>>> 6 Enforcement. When postgres backend executes queries, check the\n>>>>>>> blacklist in shared memory to determine whether the query is allowed(before\n>>>>>>> execute) or need rollback(is executing)?\n>>>>>>>\n>>>>>>> If we implemented disk quota as an extension, we could just use\n>>>>>>> background worker to start diskquota launcher process and use\n>>>>>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\n>>>>>>> processes by the launcher process as suggested by @Chapman.\n>>>>>>> Diskquota setting could be stored in user table in a separate schema for\n>>>>>>> each database(Schema and table created by create extension statement) just\n>>>>>>> like what Heikki has done in pg_quota project. But in this case, we need to\n>>>>>>> create extension for each database before diskquota worker process can be\n>>>>>>> set up for that database.\n>>>>>>>\n>>>>>>> Any comments on the above design and which is preferred, native\n>>>>>>> feature or extension as the POC?\n>>>>>>>\n>>>>>>>\n>>>>>>> -- Hubert\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <\n>>>>>>> pavel.stehule@gmail.com> wrote:\n>>>>>>>\n>>>>>>>>\n>>>>>>>>\n>>>>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\n>>>>>>>>\n>>>>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n>>>>>>>>>\n>>>>>>>>> > 2 Keep one worker process for each database. But using a\n>>>>>>>>> parent/global\n>>>>>>>>> > quota worker process to manage the lifecycle of database level\n>>>>>>>>> worker\n>>>>>>>>> > processes. It could handle the newly created database(avoid\n>>>>>>>>> restart\n>>>>>>>>> > database) and save resource when a database is not used. But\n>>>>>>>>> this needs to\n>>>>>>>>> > change worker process to be hierarchical. Postmaster becomes the\n>>>>>>>>> grandfather\n>>>>>>>>> > of database level worker processes in this case.\n>>>>>>>>>\n>>>>>>>>> I am using background workers this way in 9.5 at $work.\n>>>>>>>>>\n>>>>>>>>> In my case, one worker lives forever, wakes up on a set period, and\n>>>>>>>>> starts a short-lived worker for every database, waiting for each\n>>>>>>>>> one before starting the next.\n>>>>>>>>>\n>>>>>>>>> It was straightforward to implement. Looking back over the code,\n>>>>>>>>> I see the global worker assigns its own PID to\n>>>>>>>>> worker.bgw_notify_pid\n>>>>>>>>> of each of its children, and also obtains a handle for each child\n>>>>>>>>> from RegisterDynamicBackgroundWorker().\n>>>>>>>>>\n>>>>>>>>> I imagine the global quota worker would prefer to start workers\n>>>>>>>>> for every database and then just wait for notifications from any\n>>>>>>>>> of them, but that seems equally straightforward at first glance.\n>>>>>>>>>\n>>>>>>>>\n>>>>>>>> There are servers with thousands databases. Worker per database is\n>>>>>>>> not good idea.\n>>>>>>>>\n>>>>>>>> It should to share ideas, code with autovacuum process.\n>>>>>>>>\n>>>>>>>> Not sure, how to effective implementation based on bg workers can\n>>>>>>>> be. On servers with large set of databases, large set of tables it can\n>>>>>>>> identify too big table too late.\n>>>>>>>>\n>>>>>>>> Isn't better to implement some quotas on storage level?\n>>>>>>>>\n>>>>>>>> Regards\n>>>>>>>>\n>>>>>>>> Pavel\n>>>>>>>>\n>>>>>>>>\n>>>>>>>>\n>>>>>>>>> -Chap\n>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>> --\n>>>>>>> Thanks\n>>>>>>>\n>>>>>>> Hubert Zhang\n>>>>>>>\n>>>>>>\n>>>>>>\n>>>>>\n>>>>>\n>>>>> --\n>>>>> Thanks\n>>>>>\n>>>>> Hubert Zhang\n>>>>>\n>>>>\n>>>>\n>>>\n>>>\n>>> --\n>>> Thanks\n>>>\n>>> Hubert Zhang\n>>>\n>>\n>\n> --\n> Thanks\n>\n> Hubert Zhang\n>\n\npá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer .It's an option. Prefer to consider quota setting store together:CREATE DISK QUOTA way is more nature to store quota setting in a separate pg_diskquota catalogWhile ALTER SET way is more close to store quota setting in pg_class, pg_role, pg_namespace. etc in an integrated way.(Note that here I mean nature/close is not must, ALTER SET could also store in pg_diskquota and vice versa.)I have not a problem with new special table for storing this information. But it looks like redundant to current GUC configuration and limits. Can be messy do some work with ALTER ROLE, and some work via CREATE QUOTE.RegardsPavel Here are some differences I can think of:1 pg_role is a global catalog, not per database level. It's harder to tracker the user's disk usage in the whole clusters(considering 1000+ databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it only tracks the user's disk usage inside the current database.2 using separate pg_diskquota could add more field except for quota limit without adding too many fields in pg_class, e.g. red zone to give the user a warning or the current disk usage of the db objects.On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:Hi all,We redesign disk quota feature based on the comments from Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk quota feature is used to support multi-tenancy environment, different level of database objects could be set a quota limit to avoid over use of disk space. A common case could be as follows: DBA could enable disk quota on a specified database list. DBA could set disk quota limit for tables/schemas/roles in these databases. Separate disk quota worker process will monitor the disk usage for these objects and detect the objects which exceed their quota limit. Queries loading data into these “out of disk quota” tables/schemas/roles will be cancelled.We are currently working at init implementation stage. We would like to propose our idea firstly and get feedbacks from community to do quick iteration.SQL Syntax (How to use disk quota)1 Specify the databases with disk quota enabled in GUC “diskquota_databases” in postgresql.conf and restart the database.2 DBA could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer . 3 Simulate a schema out of quota limit case: suppose table a1 and table a2 are both under schema s1.INSERT INTO a1 SELECT generate_series(1,1000);INSERT INTO a2 SELECT generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);ERROR: schema's disk space quota exceededDROP TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the following components.1. Quota Setting Store is where the disk quota setting to be stored and accessed. We plan to use catalog table pg_diskquota to store these information. pg_diskquota is like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /* diskquota name */ int16 quotatype; /* diskquota type name */ Oid quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /* diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size change of database objects. We plan to use stat collector to detect the ‘active’ table list at initial stage. But stat collector has some limitation on finding the active table which is in a running transaction. Details see TODO section.3. Quota Size Checker is where to calculate the size and compare with quota limit for database objects. According to Pavel’s comment, autovacuum launcher and worker process could be a good reference to disk quota. So we plan to use a disk quota launcher daemon process and several disk quota worker process to finish this work. Launcher process is responsible for starting worker process based on a user defined database list from GUC. Worker process will connect to its target database and monitor the disk usage for objects in this database. In init stage of worker process, it will call calculate_total_relation_size() to calculate the size for each user table. After init stage, worker process will refresh the disk model every N seconds. Refreshing will only recalculate the size of tables in ‘active’ table list, which is generated by Quata Change Detector to minimize the cost.4. Quota Enforcement Operator is where to check for the quota limitation at postgres backend side. We will firstly implement it in ExecCheckRTPerms() as pre-running enforcement. It will check the disk quota of tables being inserted or updated, and report error if table’s or table’s schema’s or table’s owner’s quota limit is exceeded. As a native feature, we plan to add more checkpoint to do running query enforcement. For example, if a disk quota lefts 10MB quota, a query could insert 1GB data. This query could be allowed in pre-running enforcement check, but will be cancelled in running query enforcement check. Therefore, it can improve the accurate of disk quota usage. To achieve this, we plan to add a checkpoint in lower API such as smgr_extened. Hence, the Quota Enforcement Operator will check the disk quota usage when smgr_extened is called. If the quota is over limited, current query will be cancelled. Highlight1. Native feature.Support native Create/Drop Disk Quota SQL statement.New catalog table pg_diskquota to store disk quota setting.2. Auto DML/DDL detection. Table create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change, Schema create/drop and Role create/drop will be detected by disk quota automatically. 3. Low cost disk quota checker.Worker process of disk quota need to refresh the disk usage model every N seconds. Since recalculate the file size using stat() system call is expensive for a large number of files, we use an ‘active’ table list to reduce the real work at each iteration. A basic experiment on our init stage implementation on database with 20K tables shows that the refresh cost is 1% cpu usage and will be finished within 50ms. Todo/LimitationBefore we propose our patch, we plan to enhance it with the following ideas:1. Setting database list with disk quota enabled dynamically without restart database. Since we have the disk quota launcher process, it could detect the new ‘diskquota_databases’ list and start/stop the corresponding disk quota worker process.2. Enforcement when query is running. Considering the case when there is 10MB quota left, but next query will insert 10GB data. Current enforcement design will allow this query to be executed. This is limited by the ‘active’ table detection is generated by stat collector. Postgres backend will only send table stat information to collector only when the transaction ends. We need a new way to detect the ‘active’ table even when this table is being modified inside a running transaction.3. Monitor unlimited number of databases. Current we set the max number of disk quota worker process to be 10 to reduce the affection normal workload. But how about if we want to monitor the disk quota of more than 10 databases? Our solution is to let disk quota launcher to manage a queue of database need to be monitored. And disk quota worker process consuming the queue and refresh the disk usage/quota for this database. After some periods, worker will return the database to the queue, and fetch the top database from queue to process. The period determine the delay of detecting disk quota change. To implement this feature, we need to support a subprocess of postmaster to rebind to another database instead of the database binded in InitPostgres().4. Support active table detection on vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman Flack for the former comments on disk quota feature. Any comments on how to improve disk quota feature are appreciated.On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \"implement some quotas on storage level?\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n\n> 2 Keep one worker process for each database. But using a parent/global\n> quota worker process to manage the lifecycle of database level worker\n> processes. It could handle the newly created database(avoid restart\n> database) and save resource when a database is not used. But this needs to\n> change worker process to be hierarchical. Postmaster becomes the grandfather\n> of database level worker processes in this case.\n\nI am using background workers this way in 9.5 at $work.\n\nIn my case, one worker lives forever, wakes up on a set period, and\nstarts a short-lived worker for every database, waiting for each\none before starting the next.\n\nIt was straightforward to implement. Looking back over the code,\nI see the global worker assigns its own PID to worker.bgw_notify_pid\nof each of its children, and also obtains a handle for each child\nfrom RegisterDynamicBackgroundWorker().\n\nI imagine the global quota worker would prefer to start workers\nfor every database and then just wait for notifications from any\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\n\n-Chap\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang","msg_date":"Sat, 22 Sep 2018 05:17:14 +0200","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":false,"msg_subject":"Re: Proposal for disk quota feature"},{"msg_contents":">\n> But it looks like redundant to current GUC configuration and limits\n\nwhat do you mean by current GUC configuration? Is that the general block\nnumber limit in your patch? If yes, the difference between GUC and\npg_diskquota catalog is that pg_diskquota will store different quota limit\nfor the different role, schema or table instead of a single GUC value.\n\nOn Sat, Sep 22, 2018 at 11:17 AM Pavel Stehule <pavel.stehule@gmail.com>\nwrote:\n\n>\n>\n> pá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:\n>\n>> just fast reaction - why QUOTA object?\n>>> Isn't ALTER SET enough?\n>>> Some like\n>>> ALTER TABLE a1 SET quote = 1MB;\n>>> ALTER USER ...\n>>> ALTER SCHEMA ..\n>>> New DDL commans looks like too hard hammer .\n>>\n>>\n>> It's an option. Prefer to consider quota setting store together:\n>> CREATE DISK QUOTA way is more nature to store quota setting in a separate\n>> pg_diskquota catalog\n>> While ALTER SET way is more close to store quota setting in pg_class,\n>> pg_role, pg_namespace. etc in an integrated way.\n>> (Note that here I mean nature/close is not must, ALTER SET could also\n>> store in pg_diskquota and vice versa.)\n>>\n>\n> I have not a problem with new special table for storing this information.\n> But it looks like redundant to current GUC configuration and limits. Can be\n> messy do some work with ALTER ROLE, and some work via CREATE QUOTE.\n>\n> Regards\n>\n> Pavel\n>\n>\n>> Here are some differences I can think of:\n>> 1 pg_role is a global catalog, not per database level. It's harder to\n>> tracker the user's disk usage in the whole clusters(considering 1000+\n>> databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it\n>> only tracks the user's disk usage inside the current database.\n>> 2 using separate pg_diskquota could add more field except for quota limit\n>> without adding too many fields in pg_class, e.g. red zone to give the user\n>> a warning or the current disk usage of the db objects.\n>>\n>> On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com>\n>> wrote:\n>>\n>>>\n>>>\n>>> pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io>\n>>> napsal:\n>>>\n>>>>\n>>>>\n>>>>\n>>>>\n>>>> *Hi all,We redesign disk quota feature based on the comments from Pavel\n>>>> Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk\n>>>> quota feature is used to support multi-tenancy environment, different level\n>>>> of database objects could be set a quota limit to avoid over use of disk\n>>>> space. A common case could be as follows: DBA could enable disk quota on a\n>>>> specified database list. DBA could set disk quota limit for\n>>>> tables/schemas/roles in these databases. Separate disk quota worker process\n>>>> will monitor the disk usage for these objects and detect the objects which\n>>>> exceed their quota limit. Queries loading data into these “out of disk\n>>>> quota” tables/schemas/roles will be cancelled.We are currently working at\n>>>> init implementation stage. We would like to propose our idea firstly and\n>>>> get feedbacks from community to do quick iteration.SQL Syntax (How to use\n>>>> disk quota)1 Specify the databases with disk quota enabled in GUC\n>>>> “diskquota_databases” in postgresql.conf and restart the database.2 DBA\n>>>> could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\n>>>> ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\n>>>> (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\n>>>> ‘3MB’);*\n>>>>\n>>>\n>>> just fast reaction - why QUOTA object?\n>>>\n>>> Isn't ALTER SET enough?\n>>>\n>>> Some like\n>>>\n>>> ALTER TABLE a1 SET quote = 1MB;\n>>> ALTER USER ...\n>>> ALTER SCHEMA ..\n>>>\n>>> New DDL commans looks like too hard hammer .\n>>>\n>>>\n>>>\n>>>>\n>>>>\n>>>>\n>>>>\n>>>>\n>>>>\n>>>>\n>>>>\n>>>>\n>>>>\n>>>>\n>>>>\n>>>>\n>>>>\n>>>>\n>>>>\n>>>>\n>>>> *3 Simulate a schema out of quota limit case: suppose table a1 and\n>>>> table a2 are both under schema s1.INSERT INTO a1 SELECT\n>>>> generate_series(1,1000);INSERT INTO a2 SELECT\n>>>> generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\n>>>> generate_series(1,1000);ERROR: schema's disk space quota exceededDROP\n>>>> TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\n>>>> generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\n>>>> following components.1. Quota Setting Store is where the disk quota setting\n>>>> to be stored and accessed. We plan to use catalog table pg_diskquota to\n>>>> store these information. pg_diskquota is\n>>>> like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\n>>>> diskquota name */ int16 quotatype; /* diskquota type name */ Oid\n>>>> quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\n>>>> diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\n>>>> MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\n>>>> change of database objects. We plan to use stat collector to detect the\n>>>> ‘active’ table list at initial stage. But stat collector has some\n>>>> limitation on finding the active table which is in a running transaction.\n>>>> Details see TODO section.3. Quota Size Checker is where to calculate the\n>>>> size and compare with quota limit for database objects. According to\n>>>> Pavel’s comment, autovacuum launcher and worker process could be a good\n>>>> reference to disk quota. So we plan to use a disk quota launcher daemon\n>>>> process and several disk quota worker process to finish this work. Launcher\n>>>> process is responsible for starting worker process based on a user defined\n>>>> database list from GUC. Worker process will connect to its target database\n>>>> and monitor the disk usage for objects in this database. In init stage of\n>>>> worker process, it will call calculate_total_relation_size() to calculate\n>>>> the size for each user table. After init stage, worker process will refresh\n>>>> the disk model every N seconds. Refreshing will only recalculate the size\n>>>> of tables in ‘active’ table list, which is generated by Quata Change\n>>>> Detector to minimize the cost.4. Quota Enforcement Operator is where to\n>>>> check for the quota limitation at postgres backend side. We will firstly\n>>>> implement it in ExecCheckRTPerms() as pre-running enforcement. It will\n>>>> check the disk quota of tables being inserted or updated, and report error\n>>>> if table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\n>>>> As a native feature, we plan to add more checkpoint to do running query\n>>>> enforcement. For example, if a disk quota lefts 10MB quota, a query could\n>>>> insert 1GB data. This query could be allowed in pre-running enforcement\n>>>> check, but will be cancelled in running query enforcement check. Therefore,\n>>>> it can improve the accurate of disk quota usage. To achieve this, we plan\n>>>> to add a checkpoint in lower API such as smgr_extened. Hence, the Quota\n>>>> Enforcement Operator will check the disk quota usage when smgr_extened is\n>>>> called. If the quota is over limited, current query will be cancelled.\n>>>> Highlight1. Native feature.Support native Create/Drop Disk Quota SQL\n>>>> statement.New catalog table pg_diskquota to store disk quota setting.2.\n>>>> Auto DML/DDL detection. Table\n>>>> create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\n>>>> Schema create/drop and Role create/drop will be detected by disk quota\n>>>> automatically. 3. Low cost disk quota checker.Worker process of disk quota\n>>>> need to refresh the disk usage model every N seconds. Since recalculate the\n>>>> file size using stat() system call is expensive for a large number of\n>>>> files, we use an ‘active’ table list to reduce the real work at each\n>>>> iteration. A basic experiment on our init stage implementation on database\n>>>> with 20K tables shows that the refresh cost is 1% cpu usage and will be\n>>>> finished within 50ms. Todo/LimitationBefore we propose our patch, we plan\n>>>> to enhance it with the following ideas:1. Setting database list with disk\n>>>> quota enabled dynamically without restart database. Since we have the disk\n>>>> quota launcher process, it could detect the new ‘diskquota_databases’ list\n>>>> and start/stop the corresponding disk quota worker process.2. Enforcement\n>>>> when query is running. Considering the case when there is 10MB quota left,\n>>>> but next query will insert 10GB data. Current enforcement design will allow\n>>>> this query to be executed. This is limited by the ‘active’ table detection\n>>>> is generated by stat collector. Postgres backend will only send table stat\n>>>> information to collector only when the transaction ends. We need a new way\n>>>> to detect the ‘active’ table even when this table is being modified inside\n>>>> a running transaction.3. Monitor unlimited number of databases. Current we\n>>>> set the max number of disk quota worker process to be 10 to reduce the\n>>>> affection normal workload. But how about if we want to monitor the disk\n>>>> quota of more than 10 databases? Our solution is to let disk quota launcher\n>>>> to manage a queue of database need to be monitored. And disk quota worker\n>>>> process consuming the queue and refresh the disk usage/quota for this\n>>>> database. After some periods, worker will return the database to the queue,\n>>>> and fetch the top database from queue to process. The period determine the\n>>>> delay of detecting disk quota change. To implement this feature, we need to\n>>>> support a subprocess of postmaster to rebind to another database instead of\n>>>> the database binded in InitPostgres().4. Support active table detection on\n>>>> vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\n>>>> are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\n>>>> Flack for the former comments on disk quota feature. Any comments on how to\n>>>> improve disk quota feature are appreciated.*\n>>>>\n>>>>\n>>>> On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com\n>>>> > wrote:\n>>>>\n>>>>>\n>>>>>\n>>>>> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>>>>\n>>>>>> Thanks Pavel.\n>>>>>> Your patch did enforcement on storage level(md.c or we could also use\n>>>>>> smgr_extend). It's straight forward.\n>>>>>> But I prefer to implement disk_quota as a feature with following\n>>>>>> objectives:\n>>>>>> 1 set/alter disk quota setting on different database objects, e.g.\n>>>>>> user, database, schema etc. not only a general GUC, but we could set\n>>>>>> separate quota limit for a specific objects.\n>>>>>> 2 enforcement operator should work at two positions: before query is\n>>>>>> running and when query is running. The latter one's implementation maybe\n>>>>>> similar to your patch.\n>>>>>>\n>>>>>\n>>>>> The patch was just example. The resource quotes should be more complex\n>>>>> - per partition, table, schema, database, user - so GUC are possible, but\n>>>>> not very user friendly.\n>>>>>\n>>>>> Our case is specific, but not too much. The servers are used for\n>>>>> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\n>>>>> SELECT). We need to solve limits immediately. The implementation is simple,\n>>>>> so I did it. Same implementation on database level, or schema level needs\n>>>>> some more locks, so it will not be too effective. The resource management\n>>>>> can be complex very complex, and I expect so it will be hard work.\n>>>>>\n>>>>> Regards\n>>>>>\n>>>>> Pavel\n>>>>>\n>>>>>\n>>>>>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <\n>>>>>> pavel.stehule@gmail.com> wrote:\n>>>>>>\n>>>>>>> Hi\n>>>>>>>\n>>>>>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>>>>>>\n>>>>>>>> Thanks Chapman.\n>>>>>>>> @Pavel, could you please explain more about your second\n>>>>>>>> suggestion \"implement some quotas on storage level?\"\n>>>>>>>>\n>>>>>>>\n>>>>>>> See attached patch - it is very simple - and good enough for our\n>>>>>>> purposes.\n>>>>>>>\n>>>>>>> Regards\n>>>>>>>\n>>>>>>> Pavel\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>> We will not keep the long-lived processes attach to all\n>>>>>>>> databases(just like you mentioned servers with thousands of\n>>>>>>>> databases)\n>>>>>>>> And you are right, we could share ideas with autovacuum process,\n>>>>>>>> fork worker processes in need.\n>>>>>>>> \"autovacuum checks for tables that have had a large number of\n>>>>>>>> inserted, updated or deleted tuples. These checks use the statistics\n>>>>>>>> collection facility\"\n>>>>>>>> diskquota process is similar to autovacuum at caring about insert,\n>>>>>>>> but the difference is that it also care about vucuum full, truncate and\n>>>>>>>> drop. While update and delete may not be interested since no file change\n>>>>>>>> happens. So a separate diskquota process is preferred.\n>>>>>>>>\n>>>>>>>> So if we implemented disk quota as a full native feature, and in\n>>>>>>>> the first initial version I prefer to implement the following features:\n>>>>>>>> 1 Fork diskquota launcher process under Postmaster serverloop,\n>>>>>>>> which is long-lived.\n>>>>>>>> 2 Diskquota launcher process is responsible for creating diskquota\n>>>>>>>> worker process for every database.\n>>>>>>>> 3 DIskquota setting is stored in a separate catalog table for each\n>>>>>>>> database.\n>>>>>>>> 4 Initialization stage, Diskquota launcher process creates\n>>>>>>>> diskquota worker process for all the databases(traverse like\n>>>>>>>> autovacuum). Worker process calculates disk usage of db objects and\n>>>>>>>> their diskquota setting. If any db object exceeds its quota limit, put them\n>>>>>>>> into the blacklist in the shared memory, which will later be used by\n>>>>>>>> enforcement operator. Worker process exits when works are done.\n>>>>>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\n>>>>>>>> process for the database with a large number of insert, copy, truncate,\n>>>>>>>> drop etc. or create disk quota statement. Worker process updates the file\n>>>>>>>> size for db objects containing the result relation, and compare with the\n>>>>>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\n>>>>>>>> remove from blacklist vice versa. Worker process exits when works\n>>>>>>>> are done and a GUC could control the frequency of worker process restart to\n>>>>>>>> a specific database. As you know, this GUC also controls the delay when we\n>>>>>>>> do enforcement.\n>>>>>>>> 6 Enforcement. When postgres backend executes queries, check the\n>>>>>>>> blacklist in shared memory to determine whether the query is allowed(before\n>>>>>>>> execute) or need rollback(is executing)?\n>>>>>>>>\n>>>>>>>> If we implemented disk quota as an extension, we could just use\n>>>>>>>> background worker to start diskquota launcher process and use\n>>>>>>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\n>>>>>>>> processes by the launcher process as suggested by @Chapman.\n>>>>>>>> Diskquota setting could be stored in user table in a separate schema for\n>>>>>>>> each database(Schema and table created by create extension statement) just\n>>>>>>>> like what Heikki has done in pg_quota project. But in this case, we need to\n>>>>>>>> create extension for each database before diskquota worker process can be\n>>>>>>>> set up for that database.\n>>>>>>>>\n>>>>>>>> Any comments on the above design and which is preferred, native\n>>>>>>>> feature or extension as the POC?\n>>>>>>>>\n>>>>>>>>\n>>>>>>>> -- Hubert\n>>>>>>>>\n>>>>>>>>\n>>>>>>>>\n>>>>>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <\n>>>>>>>> pavel.stehule@gmail.com> wrote:\n>>>>>>>>\n>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\n>>>>>>>>>\n>>>>>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n>>>>>>>>>>\n>>>>>>>>>> > 2 Keep one worker process for each database. But using a\n>>>>>>>>>> parent/global\n>>>>>>>>>> > quota worker process to manage the lifecycle of database level\n>>>>>>>>>> worker\n>>>>>>>>>> > processes. It could handle the newly created database(avoid\n>>>>>>>>>> restart\n>>>>>>>>>> > database) and save resource when a database is not used. But\n>>>>>>>>>> this needs to\n>>>>>>>>>> > change worker process to be hierarchical. Postmaster becomes\n>>>>>>>>>> the grandfather\n>>>>>>>>>> > of database level worker processes in this case.\n>>>>>>>>>>\n>>>>>>>>>> I am using background workers this way in 9.5 at $work.\n>>>>>>>>>>\n>>>>>>>>>> In my case, one worker lives forever, wakes up on a set period,\n>>>>>>>>>> and\n>>>>>>>>>> starts a short-lived worker for every database, waiting for each\n>>>>>>>>>> one before starting the next.\n>>>>>>>>>>\n>>>>>>>>>> It was straightforward to implement. Looking back over the code,\n>>>>>>>>>> I see the global worker assigns its own PID to\n>>>>>>>>>> worker.bgw_notify_pid\n>>>>>>>>>> of each of its children, and also obtains a handle for each child\n>>>>>>>>>> from RegisterDynamicBackgroundWorker().\n>>>>>>>>>>\n>>>>>>>>>> I imagine the global quota worker would prefer to start workers\n>>>>>>>>>> for every database and then just wait for notifications from any\n>>>>>>>>>> of them, but that seems equally straightforward at first glance.\n>>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>> There are servers with thousands databases. Worker per database is\n>>>>>>>>> not good idea.\n>>>>>>>>>\n>>>>>>>>> It should to share ideas, code with autovacuum process.\n>>>>>>>>>\n>>>>>>>>> Not sure, how to effective implementation based on bg workers can\n>>>>>>>>> be. On servers with large set of databases, large set of tables it can\n>>>>>>>>> identify too big table too late.\n>>>>>>>>>\n>>>>>>>>> Isn't better to implement some quotas on storage level?\n>>>>>>>>>\n>>>>>>>>> Regards\n>>>>>>>>>\n>>>>>>>>> Pavel\n>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>>> -Chap\n>>>>>>>>>>\n>>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>\n>>>>>>>>\n>>>>>>>> --\n>>>>>>>> Thanks\n>>>>>>>>\n>>>>>>>> Hubert Zhang\n>>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>\n>>>>>>\n>>>>>> --\n>>>>>> Thanks\n>>>>>>\n>>>>>> Hubert Zhang\n>>>>>>\n>>>>>\n>>>>>\n>>>>\n>>>>\n>>>> --\n>>>> Thanks\n>>>>\n>>>> Hubert Zhang\n>>>>\n>>>\n>>\n>> --\n>> Thanks\n>>\n>> Hubert Zhang\n>>\n>\n\n-- \nThanks\n\nHubert Zhang\n\nBut it looks like redundant to current GUC configuration and limitswhat do you mean by current GUC configuration? Is that the general block number limit in your patch? If yes, the difference between GUC and pg_diskquota catalog is that pg_diskquota will store different quota limit for the different role, schema or table instead of a single GUC value.On Sat, Sep 22, 2018 at 11:17 AM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer .It's an option. Prefer to consider quota setting store together:CREATE DISK QUOTA way is more nature to store quota setting in a separate pg_diskquota catalogWhile ALTER SET way is more close to store quota setting in pg_class, pg_role, pg_namespace. etc in an integrated way.(Note that here I mean nature/close is not must, ALTER SET could also store in pg_diskquota and vice versa.)I have not a problem with new special table for storing this information. But it looks like redundant to current GUC configuration and limits. Can be messy do some work with ALTER ROLE, and some work via CREATE QUOTE.RegardsPavel Here are some differences I can think of:1 pg_role is a global catalog, not per database level. It's harder to tracker the user's disk usage in the whole clusters(considering 1000+ databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it only tracks the user's disk usage inside the current database.2 using separate pg_diskquota could add more field except for quota limit without adding too many fields in pg_class, e.g. red zone to give the user a warning or the current disk usage of the db objects.On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:Hi all,We redesign disk quota feature based on the comments from Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk quota feature is used to support multi-tenancy environment, different level of database objects could be set a quota limit to avoid over use of disk space. A common case could be as follows: DBA could enable disk quota on a specified database list. DBA could set disk quota limit for tables/schemas/roles in these databases. Separate disk quota worker process will monitor the disk usage for these objects and detect the objects which exceed their quota limit. Queries loading data into these “out of disk quota” tables/schemas/roles will be cancelled.We are currently working at init implementation stage. We would like to propose our idea firstly and get feedbacks from community to do quick iteration.SQL Syntax (How to use disk quota)1 Specify the databases with disk quota enabled in GUC “diskquota_databases” in postgresql.conf and restart the database.2 DBA could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer . 3 Simulate a schema out of quota limit case: suppose table a1 and table a2 are both under schema s1.INSERT INTO a1 SELECT generate_series(1,1000);INSERT INTO a2 SELECT generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);ERROR: schema's disk space quota exceededDROP TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the following components.1. Quota Setting Store is where the disk quota setting to be stored and accessed. We plan to use catalog table pg_diskquota to store these information. pg_diskquota is like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /* diskquota name */ int16 quotatype; /* diskquota type name */ Oid quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /* diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size change of database objects. We plan to use stat collector to detect the ‘active’ table list at initial stage. But stat collector has some limitation on finding the active table which is in a running transaction. Details see TODO section.3. Quota Size Checker is where to calculate the size and compare with quota limit for database objects. According to Pavel’s comment, autovacuum launcher and worker process could be a good reference to disk quota. So we plan to use a disk quota launcher daemon process and several disk quota worker process to finish this work. Launcher process is responsible for starting worker process based on a user defined database list from GUC. Worker process will connect to its target database and monitor the disk usage for objects in this database. In init stage of worker process, it will call calculate_total_relation_size() to calculate the size for each user table. After init stage, worker process will refresh the disk model every N seconds. Refreshing will only recalculate the size of tables in ‘active’ table list, which is generated by Quata Change Detector to minimize the cost.4. Quota Enforcement Operator is where to check for the quota limitation at postgres backend side. We will firstly implement it in ExecCheckRTPerms() as pre-running enforcement. It will check the disk quota of tables being inserted or updated, and report error if table’s or table’s schema’s or table’s owner’s quota limit is exceeded. As a native feature, we plan to add more checkpoint to do running query enforcement. For example, if a disk quota lefts 10MB quota, a query could insert 1GB data. This query could be allowed in pre-running enforcement check, but will be cancelled in running query enforcement check. Therefore, it can improve the accurate of disk quota usage. To achieve this, we plan to add a checkpoint in lower API such as smgr_extened. Hence, the Quota Enforcement Operator will check the disk quota usage when smgr_extened is called. If the quota is over limited, current query will be cancelled. Highlight1. Native feature.Support native Create/Drop Disk Quota SQL statement.New catalog table pg_diskquota to store disk quota setting.2. Auto DML/DDL detection. Table create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change, Schema create/drop and Role create/drop will be detected by disk quota automatically. 3. Low cost disk quota checker.Worker process of disk quota need to refresh the disk usage model every N seconds. Since recalculate the file size using stat() system call is expensive for a large number of files, we use an ‘active’ table list to reduce the real work at each iteration. A basic experiment on our init stage implementation on database with 20K tables shows that the refresh cost is 1% cpu usage and will be finished within 50ms. Todo/LimitationBefore we propose our patch, we plan to enhance it with the following ideas:1. Setting database list with disk quota enabled dynamically without restart database. Since we have the disk quota launcher process, it could detect the new ‘diskquota_databases’ list and start/stop the corresponding disk quota worker process.2. Enforcement when query is running. Considering the case when there is 10MB quota left, but next query will insert 10GB data. Current enforcement design will allow this query to be executed. This is limited by the ‘active’ table detection is generated by stat collector. Postgres backend will only send table stat information to collector only when the transaction ends. We need a new way to detect the ‘active’ table even when this table is being modified inside a running transaction.3. Monitor unlimited number of databases. Current we set the max number of disk quota worker process to be 10 to reduce the affection normal workload. But how about if we want to monitor the disk quota of more than 10 databases? Our solution is to let disk quota launcher to manage a queue of database need to be monitored. And disk quota worker process consuming the queue and refresh the disk usage/quota for this database. After some periods, worker will return the database to the queue, and fetch the top database from queue to process. The period determine the delay of detecting disk quota change. To implement this feature, we need to support a subprocess of postmaster to rebind to another database instead of the database binded in InitPostgres().4. Support active table detection on vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman Flack for the former comments on disk quota feature. Any comments on how to improve disk quota feature are appreciated.On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \"implement some quotas on storage level?\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n\n> 2 Keep one worker process for each database. But using a parent/global\n> quota worker process to manage the lifecycle of database level worker\n> processes. It could handle the newly created database(avoid restart\n> database) and save resource when a database is not used. But this needs to\n> change worker process to be hierarchical. Postmaster becomes the grandfather\n> of database level worker processes in this case.\n\nI am using background workers this way in 9.5 at $work.\n\nIn my case, one worker lives forever, wakes up on a set period, and\nstarts a short-lived worker for every database, waiting for each\none before starting the next.\n\nIt was straightforward to implement. Looking back over the code,\nI see the global worker assigns its own PID to worker.bgw_notify_pid\nof each of its children, and also obtains a handle for each child\nfrom RegisterDynamicBackgroundWorker().\n\nI imagine the global quota worker would prefer to start workers\nfor every database and then just wait for notifications from any\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\n\n-Chap\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang\n\n-- ThanksHubert Zhang","msg_date":"Sat, 22 Sep 2018 14:48:03 +0800","msg_from":"Hubert Zhang <hzhang@pivotal.io>","msg_from_op":true,"msg_subject":"Re: Proposal for disk quota feature"},{"msg_contents":"so 22. 9. 2018 v 8:48 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:\n\n> But it looks like redundant to current GUC configuration and limits\n>\n> what do you mean by current GUC configuration? Is that the general block\n> number limit in your patch? If yes, the difference between GUC and\n> pg_diskquota catalog is that pg_diskquota will store different quota limit\n> for the different role, schema or table instead of a single GUC value.\n>\n\nstorage is not relevant in this moment.\n\nI don't see to consistent to sets some limits via SET command, or ALTER X\nSET, and some other with CREATE QUOTA ON.\n\nThe quotas or object limits, resource limits are pretty useful and\nnecessary, but I don't see these like new type of objects, it is much more\nsome property of current objects. Because we have one syntax for this\npurpose I prefer it. Because is not good to have two syntaxes for similar\npurpose.\n\nSo instead CREATE DISC QUATA ON SCHEMA xxx some value I prefer\n\nALTER SCHEMA xxx SET disc_quota = xxx;\n\nThe functionality is +/- same. But ALTER XX SET was introduce first, and I\ndon't feel comfortable to have any new syntax for similar purpose\n\nRegards\n\nPavel\n\n\n\n\n\n>\n> On Sat, Sep 22, 2018 at 11:17 AM Pavel Stehule <pavel.stehule@gmail.com>\n> wrote:\n>\n>>\n>>\n>> pá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io>\n>> napsal:\n>>\n>>> just fast reaction - why QUOTA object?\n>>>> Isn't ALTER SET enough?\n>>>> Some like\n>>>> ALTER TABLE a1 SET quote = 1MB;\n>>>> ALTER USER ...\n>>>> ALTER SCHEMA ..\n>>>> New DDL commans looks like too hard hammer .\n>>>\n>>>\n>>> It's an option. Prefer to consider quota setting store together:\n>>> CREATE DISK QUOTA way is more nature to store quota setting in a\n>>> separate pg_diskquota catalog\n>>> While ALTER SET way is more close to store quota setting in pg_class,\n>>> pg_role, pg_namespace. etc in an integrated way.\n>>> (Note that here I mean nature/close is not must, ALTER SET could also\n>>> store in pg_diskquota and vice versa.)\n>>>\n>>\n>> I have not a problem with new special table for storing this information.\n>> But it looks like redundant to current GUC configuration and limits. Can be\n>> messy do some work with ALTER ROLE, and some work via CREATE QUOTE.\n>>\n>> Regards\n>>\n>> Pavel\n>>\n>>\n>>> Here are some differences I can think of:\n>>> 1 pg_role is a global catalog, not per database level. It's harder to\n>>> tracker the user's disk usage in the whole clusters(considering 1000+\n>>> databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it\n>>> only tracks the user's disk usage inside the current database.\n>>> 2 using separate pg_diskquota could add more field except for quota\n>>> limit without adding too many fields in pg_class, e.g. red zone to give the\n>>> user a warning or the current disk usage of the db objects.\n>>>\n>>> On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com>\n>>> wrote:\n>>>\n>>>>\n>>>>\n>>>> pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io>\n>>>> napsal:\n>>>>\n>>>>>\n>>>>>\n>>>>>\n>>>>>\n>>>>> *Hi all,We redesign disk quota feature based on the comments from\n>>>>> Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically,\n>>>>> disk quota feature is used to support multi-tenancy environment, different\n>>>>> level of database objects could be set a quota limit to avoid over use of\n>>>>> disk space. A common case could be as follows: DBA could enable disk quota\n>>>>> on a specified database list. DBA could set disk quota limit for\n>>>>> tables/schemas/roles in these databases. Separate disk quota worker process\n>>>>> will monitor the disk usage for these objects and detect the objects which\n>>>>> exceed their quota limit. Queries loading data into these “out of disk\n>>>>> quota” tables/schemas/roles will be cancelled.We are currently working at\n>>>>> init implementation stage. We would like to propose our idea firstly and\n>>>>> get feedbacks from community to do quick iteration.SQL Syntax (How to use\n>>>>> disk quota)1 Specify the databases with disk quota enabled in GUC\n>>>>> “diskquota_databases” in postgresql.conf and restart the database.2 DBA\n>>>>> could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\n>>>>> ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\n>>>>> (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\n>>>>> ‘3MB’);*\n>>>>>\n>>>>\n>>>> just fast reaction - why QUOTA object?\n>>>>\n>>>> Isn't ALTER SET enough?\n>>>>\n>>>> Some like\n>>>>\n>>>> ALTER TABLE a1 SET quote = 1MB;\n>>>> ALTER USER ...\n>>>> ALTER SCHEMA ..\n>>>>\n>>>> New DDL commans looks like too hard hammer .\n>>>>\n>>>>\n>>>>\n>>>>>\n>>>>>\n>>>>>\n>>>>>\n>>>>>\n>>>>>\n>>>>>\n>>>>>\n>>>>>\n>>>>>\n>>>>>\n>>>>>\n>>>>>\n>>>>>\n>>>>>\n>>>>>\n>>>>>\n>>>>> *3 Simulate a schema out of quota limit case: suppose table a1 and\n>>>>> table a2 are both under schema s1.INSERT INTO a1 SELECT\n>>>>> generate_series(1,1000);INSERT INTO a2 SELECT\n>>>>> generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\n>>>>> generate_series(1,1000);ERROR: schema's disk space quota exceededDROP\n>>>>> TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\n>>>>> generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\n>>>>> following components.1. Quota Setting Store is where the disk quota setting\n>>>>> to be stored and accessed. We plan to use catalog table pg_diskquota to\n>>>>> store these information. pg_diskquota is\n>>>>> like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\n>>>>> diskquota name */ int16 quotatype; /* diskquota type name */ Oid\n>>>>> quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\n>>>>> diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\n>>>>> MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\n>>>>> change of database objects. We plan to use stat collector to detect the\n>>>>> ‘active’ table list at initial stage. But stat collector has some\n>>>>> limitation on finding the active table which is in a running transaction.\n>>>>> Details see TODO section.3. Quota Size Checker is where to calculate the\n>>>>> size and compare with quota limit for database objects. According to\n>>>>> Pavel’s comment, autovacuum launcher and worker process could be a good\n>>>>> reference to disk quota. So we plan to use a disk quota launcher daemon\n>>>>> process and several disk quota worker process to finish this work. Launcher\n>>>>> process is responsible for starting worker process based on a user defined\n>>>>> database list from GUC. Worker process will connect to its target database\n>>>>> and monitor the disk usage for objects in this database. In init stage of\n>>>>> worker process, it will call calculate_total_relation_size() to calculate\n>>>>> the size for each user table. After init stage, worker process will refresh\n>>>>> the disk model every N seconds. Refreshing will only recalculate the size\n>>>>> of tables in ‘active’ table list, which is generated by Quata Change\n>>>>> Detector to minimize the cost.4. Quota Enforcement Operator is where to\n>>>>> check for the quota limitation at postgres backend side. We will firstly\n>>>>> implement it in ExecCheckRTPerms() as pre-running enforcement. It will\n>>>>> check the disk quota of tables being inserted or updated, and report error\n>>>>> if table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\n>>>>> As a native feature, we plan to add more checkpoint to do running query\n>>>>> enforcement. For example, if a disk quota lefts 10MB quota, a query could\n>>>>> insert 1GB data. This query could be allowed in pre-running enforcement\n>>>>> check, but will be cancelled in running query enforcement check. Therefore,\n>>>>> it can improve the accurate of disk quota usage. To achieve this, we plan\n>>>>> to add a checkpoint in lower API such as smgr_extened. Hence, the Quota\n>>>>> Enforcement Operator will check the disk quota usage when smgr_extened is\n>>>>> called. If the quota is over limited, current query will be cancelled.\n>>>>> Highlight1. Native feature.Support native Create/Drop Disk Quota SQL\n>>>>> statement.New catalog table pg_diskquota to store disk quota setting.2.\n>>>>> Auto DML/DDL detection. Table\n>>>>> create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\n>>>>> Schema create/drop and Role create/drop will be detected by disk quota\n>>>>> automatically. 3. Low cost disk quota checker.Worker process of disk quota\n>>>>> need to refresh the disk usage model every N seconds. Since recalculate the\n>>>>> file size using stat() system call is expensive for a large number of\n>>>>> files, we use an ‘active’ table list to reduce the real work at each\n>>>>> iteration. A basic experiment on our init stage implementation on database\n>>>>> with 20K tables shows that the refresh cost is 1% cpu usage and will be\n>>>>> finished within 50ms. Todo/LimitationBefore we propose our patch, we plan\n>>>>> to enhance it with the following ideas:1. Setting database list with disk\n>>>>> quota enabled dynamically without restart database. Since we have the disk\n>>>>> quota launcher process, it could detect the new ‘diskquota_databases’ list\n>>>>> and start/stop the corresponding disk quota worker process.2. Enforcement\n>>>>> when query is running. Considering the case when there is 10MB quota left,\n>>>>> but next query will insert 10GB data. Current enforcement design will allow\n>>>>> this query to be executed. This is limited by the ‘active’ table detection\n>>>>> is generated by stat collector. Postgres backend will only send table stat\n>>>>> information to collector only when the transaction ends. We need a new way\n>>>>> to detect the ‘active’ table even when this table is being modified inside\n>>>>> a running transaction.3. Monitor unlimited number of databases. Current we\n>>>>> set the max number of disk quota worker process to be 10 to reduce the\n>>>>> affection normal workload. But how about if we want to monitor the disk\n>>>>> quota of more than 10 databases? Our solution is to let disk quota launcher\n>>>>> to manage a queue of database need to be monitored. And disk quota worker\n>>>>> process consuming the queue and refresh the disk usage/quota for this\n>>>>> database. After some periods, worker will return the database to the queue,\n>>>>> and fetch the top database from queue to process. The period determine the\n>>>>> delay of detecting disk quota change. To implement this feature, we need to\n>>>>> support a subprocess of postmaster to rebind to another database instead of\n>>>>> the database binded in InitPostgres().4. Support active table detection on\n>>>>> vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\n>>>>> are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\n>>>>> Flack for the former comments on disk quota feature. Any comments on how to\n>>>>> improve disk quota feature are appreciated.*\n>>>>>\n>>>>>\n>>>>> On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <\n>>>>> pavel.stehule@gmail.com> wrote:\n>>>>>\n>>>>>>\n>>>>>>\n>>>>>> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>>>>>\n>>>>>>> Thanks Pavel.\n>>>>>>> Your patch did enforcement on storage level(md.c or we could also\n>>>>>>> use smgr_extend). It's straight forward.\n>>>>>>> But I prefer to implement disk_quota as a feature with following\n>>>>>>> objectives:\n>>>>>>> 1 set/alter disk quota setting on different database objects, e.g.\n>>>>>>> user, database, schema etc. not only a general GUC, but we could set\n>>>>>>> separate quota limit for a specific objects.\n>>>>>>> 2 enforcement operator should work at two positions: before query is\n>>>>>>> running and when query is running. The latter one's implementation maybe\n>>>>>>> similar to your patch.\n>>>>>>>\n>>>>>>\n>>>>>> The patch was just example. The resource quotes should be more\n>>>>>> complex - per partition, table, schema, database, user - so GUC are\n>>>>>> possible, but not very user friendly.\n>>>>>>\n>>>>>> Our case is specific, but not too much. The servers are used for\n>>>>>> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\n>>>>>> SELECT). We need to solve limits immediately. The implementation is simple,\n>>>>>> so I did it. Same implementation on database level, or schema level needs\n>>>>>> some more locks, so it will not be too effective. The resource management\n>>>>>> can be complex very complex, and I expect so it will be hard work.\n>>>>>>\n>>>>>> Regards\n>>>>>>\n>>>>>> Pavel\n>>>>>>\n>>>>>>\n>>>>>>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <\n>>>>>>> pavel.stehule@gmail.com> wrote:\n>>>>>>>\n>>>>>>>> Hi\n>>>>>>>>\n>>>>>>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>>>>>>>\n>>>>>>>>> Thanks Chapman.\n>>>>>>>>> @Pavel, could you please explain more about your second\n>>>>>>>>> suggestion \"implement some quotas on storage level?\"\n>>>>>>>>>\n>>>>>>>>\n>>>>>>>> See attached patch - it is very simple - and good enough for our\n>>>>>>>> purposes.\n>>>>>>>>\n>>>>>>>> Regards\n>>>>>>>>\n>>>>>>>> Pavel\n>>>>>>>>\n>>>>>>>>\n>>>>>>>>\n>>>>>>>>> We will not keep the long-lived processes attach to all\n>>>>>>>>> databases(just like you mentioned servers with thousands of\n>>>>>>>>> databases)\n>>>>>>>>> And you are right, we could share ideas with autovacuum process,\n>>>>>>>>> fork worker processes in need.\n>>>>>>>>> \"autovacuum checks for tables that have had a large number of\n>>>>>>>>> inserted, updated or deleted tuples. These checks use the statistics\n>>>>>>>>> collection facility\"\n>>>>>>>>> diskquota process is similar to autovacuum at caring about insert,\n>>>>>>>>> but the difference is that it also care about vucuum full, truncate and\n>>>>>>>>> drop. While update and delete may not be interested since no file change\n>>>>>>>>> happens. So a separate diskquota process is preferred.\n>>>>>>>>>\n>>>>>>>>> So if we implemented disk quota as a full native feature, and in\n>>>>>>>>> the first initial version I prefer to implement the following features:\n>>>>>>>>> 1 Fork diskquota launcher process under Postmaster serverloop,\n>>>>>>>>> which is long-lived.\n>>>>>>>>> 2 Diskquota launcher process is responsible for\n>>>>>>>>> creating diskquota worker process for every database.\n>>>>>>>>> 3 DIskquota setting is stored in a separate catalog table for each\n>>>>>>>>> database.\n>>>>>>>>> 4 Initialization stage, Diskquota launcher process creates\n>>>>>>>>> diskquota worker process for all the databases(traverse like\n>>>>>>>>> autovacuum). Worker process calculates disk usage of db objects and\n>>>>>>>>> their diskquota setting. If any db object exceeds its quota limit, put them\n>>>>>>>>> into the blacklist in the shared memory, which will later be used by\n>>>>>>>>> enforcement operator. Worker process exits when works are done.\n>>>>>>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\n>>>>>>>>> process for the database with a large number of insert, copy, truncate,\n>>>>>>>>> drop etc. or create disk quota statement. Worker process updates the file\n>>>>>>>>> size for db objects containing the result relation, and compare with the\n>>>>>>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\n>>>>>>>>> remove from blacklist vice versa. Worker process exits when works\n>>>>>>>>> are done and a GUC could control the frequency of worker process restart to\n>>>>>>>>> a specific database. As you know, this GUC also controls the delay when we\n>>>>>>>>> do enforcement.\n>>>>>>>>> 6 Enforcement. When postgres backend executes queries, check the\n>>>>>>>>> blacklist in shared memory to determine whether the query is allowed(before\n>>>>>>>>> execute) or need rollback(is executing)?\n>>>>>>>>>\n>>>>>>>>> If we implemented disk quota as an extension, we could just use\n>>>>>>>>> background worker to start diskquota launcher process and use\n>>>>>>>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\n>>>>>>>>> processes by the launcher process as suggested by @Chapman.\n>>>>>>>>> Diskquota setting could be stored in user table in a separate schema for\n>>>>>>>>> each database(Schema and table created by create extension statement) just\n>>>>>>>>> like what Heikki has done in pg_quota project. But in this case, we need to\n>>>>>>>>> create extension for each database before diskquota worker process can be\n>>>>>>>>> set up for that database.\n>>>>>>>>>\n>>>>>>>>> Any comments on the above design and which is preferred, native\n>>>>>>>>> feature or extension as the POC?\n>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>> -- Hubert\n>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <\n>>>>>>>>> pavel.stehule@gmail.com> wrote:\n>>>>>>>>>\n>>>>>>>>>>\n>>>>>>>>>>\n>>>>>>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\n>>>>>>>>>>\n>>>>>>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n>>>>>>>>>>>\n>>>>>>>>>>> > 2 Keep one worker process for each database. But using a\n>>>>>>>>>>> parent/global\n>>>>>>>>>>> > quota worker process to manage the lifecycle of database level\n>>>>>>>>>>> worker\n>>>>>>>>>>> > processes. It could handle the newly created database(avoid\n>>>>>>>>>>> restart\n>>>>>>>>>>> > database) and save resource when a database is not used. But\n>>>>>>>>>>> this needs to\n>>>>>>>>>>> > change worker process to be hierarchical. Postmaster becomes\n>>>>>>>>>>> the grandfather\n>>>>>>>>>>> > of database level worker processes in this case.\n>>>>>>>>>>>\n>>>>>>>>>>> I am using background workers this way in 9.5 at $work.\n>>>>>>>>>>>\n>>>>>>>>>>> In my case, one worker lives forever, wakes up on a set period,\n>>>>>>>>>>> and\n>>>>>>>>>>> starts a short-lived worker for every database, waiting for each\n>>>>>>>>>>> one before starting the next.\n>>>>>>>>>>>\n>>>>>>>>>>> It was straightforward to implement. Looking back over the code,\n>>>>>>>>>>> I see the global worker assigns its own PID to\n>>>>>>>>>>> worker.bgw_notify_pid\n>>>>>>>>>>> of each of its children, and also obtains a handle for each child\n>>>>>>>>>>> from RegisterDynamicBackgroundWorker().\n>>>>>>>>>>>\n>>>>>>>>>>> I imagine the global quota worker would prefer to start workers\n>>>>>>>>>>> for every database and then just wait for notifications from any\n>>>>>>>>>>> of them, but that seems equally straightforward at first glance.\n>>>>>>>>>>>\n>>>>>>>>>>\n>>>>>>>>>> There are servers with thousands databases. Worker per database\n>>>>>>>>>> is not good idea.\n>>>>>>>>>>\n>>>>>>>>>> It should to share ideas, code with autovacuum process.\n>>>>>>>>>>\n>>>>>>>>>> Not sure, how to effective implementation based on bg workers can\n>>>>>>>>>> be. On servers with large set of databases, large set of tables it can\n>>>>>>>>>> identify too big table too late.\n>>>>>>>>>>\n>>>>>>>>>> Isn't better to implement some quotas on storage level?\n>>>>>>>>>>\n>>>>>>>>>> Regards\n>>>>>>>>>>\n>>>>>>>>>> Pavel\n>>>>>>>>>>\n>>>>>>>>>>\n>>>>>>>>>>\n>>>>>>>>>>> -Chap\n>>>>>>>>>>>\n>>>>>>>>>>>\n>>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>> --\n>>>>>>>>> Thanks\n>>>>>>>>>\n>>>>>>>>> Hubert Zhang\n>>>>>>>>>\n>>>>>>>>\n>>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>> --\n>>>>>>> Thanks\n>>>>>>>\n>>>>>>> Hubert Zhang\n>>>>>>>\n>>>>>>\n>>>>>>\n>>>>>\n>>>>>\n>>>>> --\n>>>>> Thanks\n>>>>>\n>>>>> Hubert Zhang\n>>>>>\n>>>>\n>>>\n>>> --\n>>> Thanks\n>>>\n>>> Hubert Zhang\n>>>\n>>\n>\n> --\n> Thanks\n>\n> Hubert Zhang\n>\n\nso 22. 9. 2018 v 8:48 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:But it looks like redundant to current GUC configuration and limitswhat do you mean by current GUC configuration? Is that the general block number limit in your patch? If yes, the difference between GUC and pg_diskquota catalog is that pg_diskquota will store different quota limit for the different role, schema or table instead of a single GUC value.storage is not relevant in this moment.I don't see to consistent to sets some limits via SET command, or ALTER X SET, and some other with CREATE QUOTA ON.The quotas or object limits, resource limits are pretty useful and necessary, but I don't see these like new type of objects, it is much more some property of current objects. Because we have one syntax for this purpose I prefer it. Because is not good to have two syntaxes for similar purpose.So instead CREATE DISC QUATA ON SCHEMA xxx some value I preferALTER SCHEMA xxx SET disc_quota = xxx;The functionality is +/- same. But ALTER XX SET was introduce first, and I don't feel comfortable to have any new syntax for similar purposeRegardsPavel On Sat, Sep 22, 2018 at 11:17 AM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer .It's an option. Prefer to consider quota setting store together:CREATE DISK QUOTA way is more nature to store quota setting in a separate pg_diskquota catalogWhile ALTER SET way is more close to store quota setting in pg_class, pg_role, pg_namespace. etc in an integrated way.(Note that here I mean nature/close is not must, ALTER SET could also store in pg_diskquota and vice versa.)I have not a problem with new special table for storing this information. But it looks like redundant to current GUC configuration and limits. Can be messy do some work with ALTER ROLE, and some work via CREATE QUOTE.RegardsPavel Here are some differences I can think of:1 pg_role is a global catalog, not per database level. It's harder to tracker the user's disk usage in the whole clusters(considering 1000+ databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it only tracks the user's disk usage inside the current database.2 using separate pg_diskquota could add more field except for quota limit without adding too many fields in pg_class, e.g. red zone to give the user a warning or the current disk usage of the db objects.On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:Hi all,We redesign disk quota feature based on the comments from Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk quota feature is used to support multi-tenancy environment, different level of database objects could be set a quota limit to avoid over use of disk space. A common case could be as follows: DBA could enable disk quota on a specified database list. DBA could set disk quota limit for tables/schemas/roles in these databases. Separate disk quota worker process will monitor the disk usage for these objects and detect the objects which exceed their quota limit. Queries loading data into these “out of disk quota” tables/schemas/roles will be cancelled.We are currently working at init implementation stage. We would like to propose our idea firstly and get feedbacks from community to do quick iteration.SQL Syntax (How to use disk quota)1 Specify the databases with disk quota enabled in GUC “diskquota_databases” in postgresql.conf and restart the database.2 DBA could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer . 3 Simulate a schema out of quota limit case: suppose table a1 and table a2 are both under schema s1.INSERT INTO a1 SELECT generate_series(1,1000);INSERT INTO a2 SELECT generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);ERROR: schema's disk space quota exceededDROP TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the following components.1. Quota Setting Store is where the disk quota setting to be stored and accessed. We plan to use catalog table pg_diskquota to store these information. pg_diskquota is like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /* diskquota name */ int16 quotatype; /* diskquota type name */ Oid quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /* diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size change of database objects. We plan to use stat collector to detect the ‘active’ table list at initial stage. But stat collector has some limitation on finding the active table which is in a running transaction. Details see TODO section.3. Quota Size Checker is where to calculate the size and compare with quota limit for database objects. According to Pavel’s comment, autovacuum launcher and worker process could be a good reference to disk quota. So we plan to use a disk quota launcher daemon process and several disk quota worker process to finish this work. Launcher process is responsible for starting worker process based on a user defined database list from GUC. Worker process will connect to its target database and monitor the disk usage for objects in this database. In init stage of worker process, it will call calculate_total_relation_size() to calculate the size for each user table. After init stage, worker process will refresh the disk model every N seconds. Refreshing will only recalculate the size of tables in ‘active’ table list, which is generated by Quata Change Detector to minimize the cost.4. Quota Enforcement Operator is where to check for the quota limitation at postgres backend side. We will firstly implement it in ExecCheckRTPerms() as pre-running enforcement. It will check the disk quota of tables being inserted or updated, and report error if table’s or table’s schema’s or table’s owner’s quota limit is exceeded. As a native feature, we plan to add more checkpoint to do running query enforcement. For example, if a disk quota lefts 10MB quota, a query could insert 1GB data. This query could be allowed in pre-running enforcement check, but will be cancelled in running query enforcement check. Therefore, it can improve the accurate of disk quota usage. To achieve this, we plan to add a checkpoint in lower API such as smgr_extened. Hence, the Quota Enforcement Operator will check the disk quota usage when smgr_extened is called. If the quota is over limited, current query will be cancelled. Highlight1. Native feature.Support native Create/Drop Disk Quota SQL statement.New catalog table pg_diskquota to store disk quota setting.2. Auto DML/DDL detection. Table create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change, Schema create/drop and Role create/drop will be detected by disk quota automatically. 3. Low cost disk quota checker.Worker process of disk quota need to refresh the disk usage model every N seconds. Since recalculate the file size using stat() system call is expensive for a large number of files, we use an ‘active’ table list to reduce the real work at each iteration. A basic experiment on our init stage implementation on database with 20K tables shows that the refresh cost is 1% cpu usage and will be finished within 50ms. Todo/LimitationBefore we propose our patch, we plan to enhance it with the following ideas:1. Setting database list with disk quota enabled dynamically without restart database. Since we have the disk quota launcher process, it could detect the new ‘diskquota_databases’ list and start/stop the corresponding disk quota worker process.2. Enforcement when query is running. Considering the case when there is 10MB quota left, but next query will insert 10GB data. Current enforcement design will allow this query to be executed. This is limited by the ‘active’ table detection is generated by stat collector. Postgres backend will only send table stat information to collector only when the transaction ends. We need a new way to detect the ‘active’ table even when this table is being modified inside a running transaction.3. Monitor unlimited number of databases. Current we set the max number of disk quota worker process to be 10 to reduce the affection normal workload. But how about if we want to monitor the disk quota of more than 10 databases? Our solution is to let disk quota launcher to manage a queue of database need to be monitored. And disk quota worker process consuming the queue and refresh the disk usage/quota for this database. After some periods, worker will return the database to the queue, and fetch the top database from queue to process. The period determine the delay of detecting disk quota change. To implement this feature, we need to support a subprocess of postmaster to rebind to another database instead of the database binded in InitPostgres().4. Support active table detection on vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman Flack for the former comments on disk quota feature. Any comments on how to improve disk quota feature are appreciated.On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \"implement some quotas on storage level?\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n\n> 2 Keep one worker process for each database. But using a parent/global\n> quota worker process to manage the lifecycle of database level worker\n> processes. It could handle the newly created database(avoid restart\n> database) and save resource when a database is not used. But this needs to\n> change worker process to be hierarchical. Postmaster becomes the grandfather\n> of database level worker processes in this case.\n\nI am using background workers this way in 9.5 at $work.\n\nIn my case, one worker lives forever, wakes up on a set period, and\nstarts a short-lived worker for every database, waiting for each\none before starting the next.\n\nIt was straightforward to implement. Looking back over the code,\nI see the global worker assigns its own PID to worker.bgw_notify_pid\nof each of its children, and also obtains a handle for each child\nfrom RegisterDynamicBackgroundWorker().\n\nI imagine the global quota worker would prefer to start workers\nfor every database and then just wait for notifications from any\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\n\n-Chap\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang\n\n-- ThanksHubert Zhang","msg_date":"Sat, 22 Sep 2018 09:02:25 +0200","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":false,"msg_subject":"Re: Proposal for disk quota feature"},{"msg_contents":">\n> The quotas or object limits, resource limits are pretty useful and\n> necessary, but I don't see these like new type of objects, it is much more\n> some property of current objects. Because we have one syntax for this\n> purpose I prefer it. Because is not good to have two syntaxes for similar\n> purpose.\n\nSCHEMA and TABLE are OK for me, But as I mentioned before, ROLE is a\nspecial case when using ALTER SET at this moment.\nTABLE and SCHEMA are both database level, e.g. pg_class and pg_namespace\nboth residents in one database. But ROLE is cluster-level. They don't\nbelong to a database. ALTER ROLE XXX SET disk_quota = xxx means to set the\nquota for the user on all the databases in the first glance. But in our\nfirst stage design, ROLE's quota is bind to a specific database. E.g. Role\nJack could have 10GB quota on database A and 2GB quota on database B.\n\nSQL syntax is not hard to modify, I don't think this should block the main\ndesign of disk quota feature. Is there any comment on the design and\narchitecture? If no, we'll firstly submit our patch and involve more\ndiscussion?\n\nOn Sat, Sep 22, 2018 at 3:03 PM Pavel Stehule <pavel.stehule@gmail.com>\nwrote:\n\n>\n>\n> so 22. 9. 2018 v 8:48 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:\n>\n>> But it looks like redundant to current GUC configuration and limits\n>>\n>> what do you mean by current GUC configuration? Is that the general block\n>> number limit in your patch? If yes, the difference between GUC and\n>> pg_diskquota catalog is that pg_diskquota will store different quota limit\n>> for the different role, schema or table instead of a single GUC value.\n>>\n>\n> storage is not relevant in this moment.\n>\n> I don't see to consistent to sets some limits via SET command, or ALTER X\n> SET, and some other with CREATE QUOTA ON.\n>\n> The quotas or object limits, resource limits are pretty useful and\n> necessary, but I don't see these like new type of objects, it is much more\n> some property of current objects. Because we have one syntax for this\n> purpose I prefer it. Because is not good to have two syntaxes for similar\n> purpose.\n>\n> So instead CREATE DISC QUATA ON SCHEMA xxx some value I prefer\n>\n> ALTER SCHEMA xxx SET disc_quota = xxx;\n>\n> The functionality is +/- same. But ALTER XX SET was introduce first, and I\n> don't feel comfortable to have any new syntax for similar purpose\n>\n> Regards\n>\n> Pavel\n>\n>\n>\n>\n>\n>>\n>> On Sat, Sep 22, 2018 at 11:17 AM Pavel Stehule <pavel.stehule@gmail.com>\n>> wrote:\n>>\n>>>\n>>>\n>>> pá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io>\n>>> napsal:\n>>>\n>>>> just fast reaction - why QUOTA object?\n>>>>> Isn't ALTER SET enough?\n>>>>> Some like\n>>>>> ALTER TABLE a1 SET quote = 1MB;\n>>>>> ALTER USER ...\n>>>>> ALTER SCHEMA ..\n>>>>> New DDL commans looks like too hard hammer .\n>>>>\n>>>>\n>>>> It's an option. Prefer to consider quota setting store together:\n>>>> CREATE DISK QUOTA way is more nature to store quota setting in a\n>>>> separate pg_diskquota catalog\n>>>> While ALTER SET way is more close to store quota setting in pg_class,\n>>>> pg_role, pg_namespace. etc in an integrated way.\n>>>> (Note that here I mean nature/close is not must, ALTER SET could also\n>>>> store in pg_diskquota and vice versa.)\n>>>>\n>>>\n>>> I have not a problem with new special table for storing this\n>>> information. But it looks like redundant to current GUC configuration and\n>>> limits. Can be messy do some work with ALTER ROLE, and some work via CREATE\n>>> QUOTE.\n>>>\n>>> Regards\n>>>\n>>> Pavel\n>>>\n>>>\n>>>> Here are some differences I can think of:\n>>>> 1 pg_role is a global catalog, not per database level. It's harder to\n>>>> tracker the user's disk usage in the whole clusters(considering 1000+\n>>>> databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it\n>>>> only tracks the user's disk usage inside the current database.\n>>>> 2 using separate pg_diskquota could add more field except for quota\n>>>> limit without adding too many fields in pg_class, e.g. red zone to give the\n>>>> user a warning or the current disk usage of the db objects.\n>>>>\n>>>> On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com>\n>>>> wrote:\n>>>>\n>>>>>\n>>>>>\n>>>>> pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io>\n>>>>> napsal:\n>>>>>\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>> *Hi all,We redesign disk quota feature based on the comments from\n>>>>>> Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically,\n>>>>>> disk quota feature is used to support multi-tenancy environment, different\n>>>>>> level of database objects could be set a quota limit to avoid over use of\n>>>>>> disk space. A common case could be as follows: DBA could enable disk quota\n>>>>>> on a specified database list. DBA could set disk quota limit for\n>>>>>> tables/schemas/roles in these databases. Separate disk quota worker process\n>>>>>> will monitor the disk usage for these objects and detect the objects which\n>>>>>> exceed their quota limit. Queries loading data into these “out of disk\n>>>>>> quota” tables/schemas/roles will be cancelled.We are currently working at\n>>>>>> init implementation stage. We would like to propose our idea firstly and\n>>>>>> get feedbacks from community to do quick iteration.SQL Syntax (How to use\n>>>>>> disk quota)1 Specify the databases with disk quota enabled in GUC\n>>>>>> “diskquota_databases” in postgresql.conf and restart the database.2 DBA\n>>>>>> could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\n>>>>>> ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\n>>>>>> (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\n>>>>>> ‘3MB’);*\n>>>>>>\n>>>>>\n>>>>> just fast reaction - why QUOTA object?\n>>>>>\n>>>>> Isn't ALTER SET enough?\n>>>>>\n>>>>> Some like\n>>>>>\n>>>>> ALTER TABLE a1 SET quote = 1MB;\n>>>>> ALTER USER ...\n>>>>> ALTER SCHEMA ..\n>>>>>\n>>>>> New DDL commans looks like too hard hammer .\n>>>>>\n>>>>>\n>>>>>\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>> *3 Simulate a schema out of quota limit case: suppose table a1 and\n>>>>>> table a2 are both under schema s1.INSERT INTO a1 SELECT\n>>>>>> generate_series(1,1000);INSERT INTO a2 SELECT\n>>>>>> generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\n>>>>>> generate_series(1,1000);ERROR: schema's disk space quota exceededDROP\n>>>>>> TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\n>>>>>> generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\n>>>>>> following components.1. Quota Setting Store is where the disk quota setting\n>>>>>> to be stored and accessed. We plan to use catalog table pg_diskquota to\n>>>>>> store these information. pg_diskquota is\n>>>>>> like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\n>>>>>> diskquota name */ int16 quotatype; /* diskquota type name */ Oid\n>>>>>> quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\n>>>>>> diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\n>>>>>> MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\n>>>>>> change of database objects. We plan to use stat collector to detect the\n>>>>>> ‘active’ table list at initial stage. But stat collector has some\n>>>>>> limitation on finding the active table which is in a running transaction.\n>>>>>> Details see TODO section.3. Quota Size Checker is where to calculate the\n>>>>>> size and compare with quota limit for database objects. According to\n>>>>>> Pavel’s comment, autovacuum launcher and worker process could be a good\n>>>>>> reference to disk quota. So we plan to use a disk quota launcher daemon\n>>>>>> process and several disk quota worker process to finish this work. Launcher\n>>>>>> process is responsible for starting worker process based on a user defined\n>>>>>> database list from GUC. Worker process will connect to its target database\n>>>>>> and monitor the disk usage for objects in this database. In init stage of\n>>>>>> worker process, it will call calculate_total_relation_size() to calculate\n>>>>>> the size for each user table. After init stage, worker process will refresh\n>>>>>> the disk model every N seconds. Refreshing will only recalculate the size\n>>>>>> of tables in ‘active’ table list, which is generated by Quata Change\n>>>>>> Detector to minimize the cost.4. Quota Enforcement Operator is where to\n>>>>>> check for the quota limitation at postgres backend side. We will firstly\n>>>>>> implement it in ExecCheckRTPerms() as pre-running enforcement. It will\n>>>>>> check the disk quota of tables being inserted or updated, and report error\n>>>>>> if table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\n>>>>>> As a native feature, we plan to add more checkpoint to do running query\n>>>>>> enforcement. For example, if a disk quota lefts 10MB quota, a query could\n>>>>>> insert 1GB data. This query could be allowed in pre-running enforcement\n>>>>>> check, but will be cancelled in running query enforcement check. Therefore,\n>>>>>> it can improve the accurate of disk quota usage. To achieve this, we plan\n>>>>>> to add a checkpoint in lower API such as smgr_extened. Hence, the Quota\n>>>>>> Enforcement Operator will check the disk quota usage when smgr_extened is\n>>>>>> called. If the quota is over limited, current query will be cancelled.\n>>>>>> Highlight1. Native feature.Support native Create/Drop Disk Quota SQL\n>>>>>> statement.New catalog table pg_diskquota to store disk quota setting.2.\n>>>>>> Auto DML/DDL detection. Table\n>>>>>> create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\n>>>>>> Schema create/drop and Role create/drop will be detected by disk quota\n>>>>>> automatically. 3. Low cost disk quota checker.Worker process of disk quota\n>>>>>> need to refresh the disk usage model every N seconds. Since recalculate the\n>>>>>> file size using stat() system call is expensive for a large number of\n>>>>>> files, we use an ‘active’ table list to reduce the real work at each\n>>>>>> iteration. A basic experiment on our init stage implementation on database\n>>>>>> with 20K tables shows that the refresh cost is 1% cpu usage and will be\n>>>>>> finished within 50ms. Todo/LimitationBefore we propose our patch, we plan\n>>>>>> to enhance it with the following ideas:1. Setting database list with disk\n>>>>>> quota enabled dynamically without restart database. Since we have the disk\n>>>>>> quota launcher process, it could detect the new ‘diskquota_databases’ list\n>>>>>> and start/stop the corresponding disk quota worker process.2. Enforcement\n>>>>>> when query is running. Considering the case when there is 10MB quota left,\n>>>>>> but next query will insert 10GB data. Current enforcement design will allow\n>>>>>> this query to be executed. This is limited by the ‘active’ table detection\n>>>>>> is generated by stat collector. Postgres backend will only send table stat\n>>>>>> information to collector only when the transaction ends. We need a new way\n>>>>>> to detect the ‘active’ table even when this table is being modified inside\n>>>>>> a running transaction.3. Monitor unlimited number of databases. Current we\n>>>>>> set the max number of disk quota worker process to be 10 to reduce the\n>>>>>> affection normal workload. But how about if we want to monitor the disk\n>>>>>> quota of more than 10 databases? Our solution is to let disk quota launcher\n>>>>>> to manage a queue of database need to be monitored. And disk quota worker\n>>>>>> process consuming the queue and refresh the disk usage/quota for this\n>>>>>> database. After some periods, worker will return the database to the queue,\n>>>>>> and fetch the top database from queue to process. The period determine the\n>>>>>> delay of detecting disk quota change. To implement this feature, we need to\n>>>>>> support a subprocess of postmaster to rebind to another database instead of\n>>>>>> the database binded in InitPostgres().4. Support active table detection on\n>>>>>> vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\n>>>>>> are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\n>>>>>> Flack for the former comments on disk quota feature. Any comments on how to\n>>>>>> improve disk quota feature are appreciated.*\n>>>>>>\n>>>>>>\n>>>>>> On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <\n>>>>>> pavel.stehule@gmail.com> wrote:\n>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>>>>>>\n>>>>>>>> Thanks Pavel.\n>>>>>>>> Your patch did enforcement on storage level(md.c or we could also\n>>>>>>>> use smgr_extend). It's straight forward.\n>>>>>>>> But I prefer to implement disk_quota as a feature with following\n>>>>>>>> objectives:\n>>>>>>>> 1 set/alter disk quota setting on different database objects, e.g.\n>>>>>>>> user, database, schema etc. not only a general GUC, but we could set\n>>>>>>>> separate quota limit for a specific objects.\n>>>>>>>> 2 enforcement operator should work at two positions: before query\n>>>>>>>> is running and when query is running. The latter one's implementation maybe\n>>>>>>>> similar to your patch.\n>>>>>>>>\n>>>>>>>\n>>>>>>> The patch was just example. The resource quotes should be more\n>>>>>>> complex - per partition, table, schema, database, user - so GUC are\n>>>>>>> possible, but not very user friendly.\n>>>>>>>\n>>>>>>> Our case is specific, but not too much. The servers are used for\n>>>>>>> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\n>>>>>>> SELECT). We need to solve limits immediately. The implementation is simple,\n>>>>>>> so I did it. Same implementation on database level, or schema level needs\n>>>>>>> some more locks, so it will not be too effective. The resource management\n>>>>>>> can be complex very complex, and I expect so it will be hard work.\n>>>>>>>\n>>>>>>> Regards\n>>>>>>>\n>>>>>>> Pavel\n>>>>>>>\n>>>>>>>\n>>>>>>>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <\n>>>>>>>> pavel.stehule@gmail.com> wrote:\n>>>>>>>>\n>>>>>>>>> Hi\n>>>>>>>>>\n>>>>>>>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>>>>>>>>\n>>>>>>>>>> Thanks Chapman.\n>>>>>>>>>> @Pavel, could you please explain more about your second\n>>>>>>>>>> suggestion \"implement some quotas on storage level?\"\n>>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>> See attached patch - it is very simple - and good enough for our\n>>>>>>>>> purposes.\n>>>>>>>>>\n>>>>>>>>> Regards\n>>>>>>>>>\n>>>>>>>>> Pavel\n>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>>> We will not keep the long-lived processes attach to all\n>>>>>>>>>> databases(just like you mentioned servers with thousands of\n>>>>>>>>>> databases)\n>>>>>>>>>> And you are right, we could share ideas with autovacuum process,\n>>>>>>>>>> fork worker processes in need.\n>>>>>>>>>> \"autovacuum checks for tables that have had a large number of\n>>>>>>>>>> inserted, updated or deleted tuples. These checks use the statistics\n>>>>>>>>>> collection facility\"\n>>>>>>>>>> diskquota process is similar to autovacuum at caring about\n>>>>>>>>>> insert, but the difference is that it also care about vucuum full, truncate\n>>>>>>>>>> and drop. While update and delete may not be interested since no file\n>>>>>>>>>> change happens. So a separate diskquota process is preferred.\n>>>>>>>>>>\n>>>>>>>>>> So if we implemented disk quota as a full native feature, and in\n>>>>>>>>>> the first initial version I prefer to implement the following features:\n>>>>>>>>>> 1 Fork diskquota launcher process under Postmaster serverloop,\n>>>>>>>>>> which is long-lived.\n>>>>>>>>>> 2 Diskquota launcher process is responsible for\n>>>>>>>>>> creating diskquota worker process for every database.\n>>>>>>>>>> 3 DIskquota setting is stored in a separate catalog table for\n>>>>>>>>>> each database.\n>>>>>>>>>> 4 Initialization stage, Diskquota launcher process creates\n>>>>>>>>>> diskquota worker process for all the databases(traverse like\n>>>>>>>>>> autovacuum). Worker process calculates disk usage of db objects and\n>>>>>>>>>> their diskquota setting. If any db object exceeds its quota limit, put them\n>>>>>>>>>> into the blacklist in the shared memory, which will later be used by\n>>>>>>>>>> enforcement operator. Worker process exits when works are done.\n>>>>>>>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\n>>>>>>>>>> process for the database with a large number of insert, copy, truncate,\n>>>>>>>>>> drop etc. or create disk quota statement. Worker process updates the file\n>>>>>>>>>> size for db objects containing the result relation, and compare with the\n>>>>>>>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\n>>>>>>>>>> remove from blacklist vice versa. Worker process exits when\n>>>>>>>>>> works are done and a GUC could control the frequency of worker process\n>>>>>>>>>> restart to a specific database. As you know, this GUC also controls the\n>>>>>>>>>> delay when we do enforcement.\n>>>>>>>>>> 6 Enforcement. When postgres backend executes queries, check the\n>>>>>>>>>> blacklist in shared memory to determine whether the query is allowed(before\n>>>>>>>>>> execute) or need rollback(is executing)?\n>>>>>>>>>>\n>>>>>>>>>> If we implemented disk quota as an extension, we could just use\n>>>>>>>>>> background worker to start diskquota launcher process and use\n>>>>>>>>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\n>>>>>>>>>> processes by the launcher process as suggested by @Chapman.\n>>>>>>>>>> Diskquota setting could be stored in user table in a separate schema for\n>>>>>>>>>> each database(Schema and table created by create extension statement) just\n>>>>>>>>>> like what Heikki has done in pg_quota project. But in this case, we need to\n>>>>>>>>>> create extension for each database before diskquota worker process can be\n>>>>>>>>>> set up for that database.\n>>>>>>>>>>\n>>>>>>>>>> Any comments on the above design and which is preferred, native\n>>>>>>>>>> feature or extension as the POC?\n>>>>>>>>>>\n>>>>>>>>>>\n>>>>>>>>>> -- Hubert\n>>>>>>>>>>\n>>>>>>>>>>\n>>>>>>>>>>\n>>>>>>>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <\n>>>>>>>>>> pavel.stehule@gmail.com> wrote:\n>>>>>>>>>>\n>>>>>>>>>>>\n>>>>>>>>>>>\n>>>>>>>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>\n>>>>>>>>>>> :\n>>>>>>>>>>>\n>>>>>>>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n>>>>>>>>>>>>\n>>>>>>>>>>>> > 2 Keep one worker process for each database. But using a\n>>>>>>>>>>>> parent/global\n>>>>>>>>>>>> > quota worker process to manage the lifecycle of database\n>>>>>>>>>>>> level worker\n>>>>>>>>>>>> > processes. It could handle the newly created database(avoid\n>>>>>>>>>>>> restart\n>>>>>>>>>>>> > database) and save resource when a database is not used. But\n>>>>>>>>>>>> this needs to\n>>>>>>>>>>>> > change worker process to be hierarchical. Postmaster becomes\n>>>>>>>>>>>> the grandfather\n>>>>>>>>>>>> > of database level worker processes in this case.\n>>>>>>>>>>>>\n>>>>>>>>>>>> I am using background workers this way in 9.5 at $work.\n>>>>>>>>>>>>\n>>>>>>>>>>>> In my case, one worker lives forever, wakes up on a set period,\n>>>>>>>>>>>> and\n>>>>>>>>>>>> starts a short-lived worker for every database, waiting for each\n>>>>>>>>>>>> one before starting the next.\n>>>>>>>>>>>>\n>>>>>>>>>>>> It was straightforward to implement. Looking back over the code,\n>>>>>>>>>>>> I see the global worker assigns its own PID to\n>>>>>>>>>>>> worker.bgw_notify_pid\n>>>>>>>>>>>> of each of its children, and also obtains a handle for each\n>>>>>>>>>>>> child\n>>>>>>>>>>>> from RegisterDynamicBackgroundWorker().\n>>>>>>>>>>>>\n>>>>>>>>>>>> I imagine the global quota worker would prefer to start workers\n>>>>>>>>>>>> for every database and then just wait for notifications from any\n>>>>>>>>>>>> of them, but that seems equally straightforward at first glance.\n>>>>>>>>>>>>\n>>>>>>>>>>>\n>>>>>>>>>>> There are servers with thousands databases. Worker per database\n>>>>>>>>>>> is not good idea.\n>>>>>>>>>>>\n>>>>>>>>>>> It should to share ideas, code with autovacuum process.\n>>>>>>>>>>>\n>>>>>>>>>>> Not sure, how to effective implementation based on bg workers\n>>>>>>>>>>> can be. On servers with large set of databases, large set of tables it can\n>>>>>>>>>>> identify too big table too late.\n>>>>>>>>>>>\n>>>>>>>>>>> Isn't better to implement some quotas on storage level?\n>>>>>>>>>>>\n>>>>>>>>>>> Regards\n>>>>>>>>>>>\n>>>>>>>>>>> Pavel\n>>>>>>>>>>>\n>>>>>>>>>>>\n>>>>>>>>>>>\n>>>>>>>>>>>> -Chap\n>>>>>>>>>>>>\n>>>>>>>>>>>>\n>>>>>>>>>>>\n>>>>>>>>>>\n>>>>>>>>>>\n>>>>>>>>>> --\n>>>>>>>>>> Thanks\n>>>>>>>>>>\n>>>>>>>>>> Hubert Zhang\n>>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>\n>>>>>>>>\n>>>>>>>> --\n>>>>>>>> Thanks\n>>>>>>>>\n>>>>>>>> Hubert Zhang\n>>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>\n>>>>>>\n>>>>>> --\n>>>>>> Thanks\n>>>>>>\n>>>>>> Hubert Zhang\n>>>>>>\n>>>>>\n>>>>\n>>>> --\n>>>> Thanks\n>>>>\n>>>> Hubert Zhang\n>>>>\n>>>\n>>\n>> --\n>> Thanks\n>>\n>> Hubert Zhang\n>>\n>\n\n-- \nThanks\n\nHubert Zhang\n\nThe quotas or object limits, resource limits are pretty useful and necessary, but I don't see these like new type of objects, it is much more some property of current objects. Because we have one syntax for this purpose I prefer it. Because is not good to have two syntaxes for similar purpose.SCHEMA and TABLE are OK for me, But as I mentioned before, ROLE is a special case when using ALTER SET at this moment.TABLE and SCHEMA are both database level, e.g. pg_class and pg_namespace both residents in one database. But ROLE is cluster-level. They don't belong to a database. ALTER ROLE XXX SET disk_quota = xxx means to set the quota for the user on all the databases in the first glance. But in our first stage design, ROLE's quota is bind to a specific database. E.g. Role Jack could have 10GB quota on database A and 2GB quota on database B. SQL syntax is not hard to modify, I don't think this should block the main design of disk quota feature. Is there any comment on the design and architecture? If no, we'll firstly submit our patch and involve more discussion?On Sat, Sep 22, 2018 at 3:03 PM Pavel Stehule <pavel.stehule@gmail.com> wrote:so 22. 9. 2018 v 8:48 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:But it looks like redundant to current GUC configuration and limitswhat do you mean by current GUC configuration? Is that the general block number limit in your patch? If yes, the difference between GUC and pg_diskquota catalog is that pg_diskquota will store different quota limit for the different role, schema or table instead of a single GUC value.storage is not relevant in this moment.I don't see to consistent to sets some limits via SET command, or ALTER X SET, and some other with CREATE QUOTA ON.The quotas or object limits, resource limits are pretty useful and necessary, but I don't see these like new type of objects, it is much more some property of current objects. Because we have one syntax for this purpose I prefer it. Because is not good to have two syntaxes for similar purpose.So instead CREATE DISC QUATA ON SCHEMA xxx some value I preferALTER SCHEMA xxx SET disc_quota = xxx;The functionality is +/- same. But ALTER XX SET was introduce first, and I don't feel comfortable to have any new syntax for similar purposeRegardsPavel On Sat, Sep 22, 2018 at 11:17 AM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer .It's an option. Prefer to consider quota setting store together:CREATE DISK QUOTA way is more nature to store quota setting in a separate pg_diskquota catalogWhile ALTER SET way is more close to store quota setting in pg_class, pg_role, pg_namespace. etc in an integrated way.(Note that here I mean nature/close is not must, ALTER SET could also store in pg_diskquota and vice versa.)I have not a problem with new special table for storing this information. But it looks like redundant to current GUC configuration and limits. Can be messy do some work with ALTER ROLE, and some work via CREATE QUOTE.RegardsPavel Here are some differences I can think of:1 pg_role is a global catalog, not per database level. It's harder to tracker the user's disk usage in the whole clusters(considering 1000+ databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it only tracks the user's disk usage inside the current database.2 using separate pg_diskquota could add more field except for quota limit without adding too many fields in pg_class, e.g. red zone to give the user a warning or the current disk usage of the db objects.On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:Hi all,We redesign disk quota feature based on the comments from Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk quota feature is used to support multi-tenancy environment, different level of database objects could be set a quota limit to avoid over use of disk space. A common case could be as follows: DBA could enable disk quota on a specified database list. DBA could set disk quota limit for tables/schemas/roles in these databases. Separate disk quota worker process will monitor the disk usage for these objects and detect the objects which exceed their quota limit. Queries loading data into these “out of disk quota” tables/schemas/roles will be cancelled.We are currently working at init implementation stage. We would like to propose our idea firstly and get feedbacks from community to do quick iteration.SQL Syntax (How to use disk quota)1 Specify the databases with disk quota enabled in GUC “diskquota_databases” in postgresql.conf and restart the database.2 DBA could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer . 3 Simulate a schema out of quota limit case: suppose table a1 and table a2 are both under schema s1.INSERT INTO a1 SELECT generate_series(1,1000);INSERT INTO a2 SELECT generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);ERROR: schema's disk space quota exceededDROP TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the following components.1. Quota Setting Store is where the disk quota setting to be stored and accessed. We plan to use catalog table pg_diskquota to store these information. pg_diskquota is like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /* diskquota name */ int16 quotatype; /* diskquota type name */ Oid quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /* diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size change of database objects. We plan to use stat collector to detect the ‘active’ table list at initial stage. But stat collector has some limitation on finding the active table which is in a running transaction. Details see TODO section.3. Quota Size Checker is where to calculate the size and compare with quota limit for database objects. According to Pavel’s comment, autovacuum launcher and worker process could be a good reference to disk quota. So we plan to use a disk quota launcher daemon process and several disk quota worker process to finish this work. Launcher process is responsible for starting worker process based on a user defined database list from GUC. Worker process will connect to its target database and monitor the disk usage for objects in this database. In init stage of worker process, it will call calculate_total_relation_size() to calculate the size for each user table. After init stage, worker process will refresh the disk model every N seconds. Refreshing will only recalculate the size of tables in ‘active’ table list, which is generated by Quata Change Detector to minimize the cost.4. Quota Enforcement Operator is where to check for the quota limitation at postgres backend side. We will firstly implement it in ExecCheckRTPerms() as pre-running enforcement. It will check the disk quota of tables being inserted or updated, and report error if table’s or table’s schema’s or table’s owner’s quota limit is exceeded. As a native feature, we plan to add more checkpoint to do running query enforcement. For example, if a disk quota lefts 10MB quota, a query could insert 1GB data. This query could be allowed in pre-running enforcement check, but will be cancelled in running query enforcement check. Therefore, it can improve the accurate of disk quota usage. To achieve this, we plan to add a checkpoint in lower API such as smgr_extened. Hence, the Quota Enforcement Operator will check the disk quota usage when smgr_extened is called. If the quota is over limited, current query will be cancelled. Highlight1. Native feature.Support native Create/Drop Disk Quota SQL statement.New catalog table pg_diskquota to store disk quota setting.2. Auto DML/DDL detection. Table create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change, Schema create/drop and Role create/drop will be detected by disk quota automatically. 3. Low cost disk quota checker.Worker process of disk quota need to refresh the disk usage model every N seconds. Since recalculate the file size using stat() system call is expensive for a large number of files, we use an ‘active’ table list to reduce the real work at each iteration. A basic experiment on our init stage implementation on database with 20K tables shows that the refresh cost is 1% cpu usage and will be finished within 50ms. Todo/LimitationBefore we propose our patch, we plan to enhance it with the following ideas:1. Setting database list with disk quota enabled dynamically without restart database. Since we have the disk quota launcher process, it could detect the new ‘diskquota_databases’ list and start/stop the corresponding disk quota worker process.2. Enforcement when query is running. Considering the case when there is 10MB quota left, but next query will insert 10GB data. Current enforcement design will allow this query to be executed. This is limited by the ‘active’ table detection is generated by stat collector. Postgres backend will only send table stat information to collector only when the transaction ends. We need a new way to detect the ‘active’ table even when this table is being modified inside a running transaction.3. Monitor unlimited number of databases. Current we set the max number of disk quota worker process to be 10 to reduce the affection normal workload. But how about if we want to monitor the disk quota of more than 10 databases? Our solution is to let disk quota launcher to manage a queue of database need to be monitored. And disk quota worker process consuming the queue and refresh the disk usage/quota for this database. After some periods, worker will return the database to the queue, and fetch the top database from queue to process. The period determine the delay of detecting disk quota change. To implement this feature, we need to support a subprocess of postmaster to rebind to another database instead of the database binded in InitPostgres().4. Support active table detection on vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman Flack for the former comments on disk quota feature. Any comments on how to improve disk quota feature are appreciated.On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \"implement some quotas on storage level?\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n\n> 2 Keep one worker process for each database. But using a parent/global\n> quota worker process to manage the lifecycle of database level worker\n> processes. It could handle the newly created database(avoid restart\n> database) and save resource when a database is not used. But this needs to\n> change worker process to be hierarchical. Postmaster becomes the grandfather\n> of database level worker processes in this case.\n\nI am using background workers this way in 9.5 at $work.\n\nIn my case, one worker lives forever, wakes up on a set period, and\nstarts a short-lived worker for every database, waiting for each\none before starting the next.\n\nIt was straightforward to implement. Looking back over the code,\nI see the global worker assigns its own PID to worker.bgw_notify_pid\nof each of its children, and also obtains a handle for each child\nfrom RegisterDynamicBackgroundWorker().\n\nI imagine the global quota worker would prefer to start workers\nfor every database and then just wait for notifications from any\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\n\n-Chap\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang\n\n\n-- ThanksHubert Zhang\n\n-- ThanksHubert Zhang\n\n-- ThanksHubert Zhang","msg_date":"Mon, 24 Sep 2018 21:56:08 +0800","msg_from":"Hubert Zhang <hzhang@pivotal.io>","msg_from_op":true,"msg_subject":"Re: Proposal for disk quota feature"},{"msg_contents":"Hi All,\n\nWe prepare a WIP patch which is implemented based on our redesigned\nproposal. This patch is for demonstration our ideas and currently can\nsupport create/delete disk quota for table/schema/user.\n\nHere is the usage,\n--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------\n1 Specify the databases with disk quota enabled in GUC\n“diskquota_databases” in postgresql.conf and restart the database.\n\n2 DBA could set disk quota limit for table/schema/role.\n\nCREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);\nCREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);\nCREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);\n\n3 Simulate a schema out of quota limit case: suppose table a1 and table a2\nare both under schema s1.\n\nINSERT INTO a1 SELECT generate_series(1,1000);\nINSERT INTO a2 SELECT generate_series(1,3000000);\nSELECT pg_sleep(5)\n\nINSERT INTO a1 SELECT generate_series(1,1000);\nERROR: schema's disk space quota exceeded\n\nDROP TABLE a2;\nSELECT pg_sleep(5)\n\nINSERT INTO a1 SELECT generate_series(1,1000);\nINSERT 0 1000\n------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------\n\nThis patch works well on small size of databases (tens of databases) as\neach database in GUC diskquota_databases will have a background worker\nprocess for monitoring purpose.\n\nThanks very much for Heikki's suggestion and comments.\nAll comments are very welcome!\n\nOn Mon, Sep 24, 2018 at 9:56 PM Hubert Zhang <hzhang@pivotal.io> wrote:\n\n> The quotas or object limits, resource limits are pretty useful and\n>> necessary, but I don't see these like new type of objects, it is much more\n>> some property of current objects. Because we have one syntax for this\n>> purpose I prefer it. Because is not good to have two syntaxes for similar\n>> purpose.\n>\n> SCHEMA and TABLE are OK for me, But as I mentioned before, ROLE is a\n> special case when using ALTER SET at this moment.\n> TABLE and SCHEMA are both database level, e.g. pg_class and pg_namespace\n> both residents in one database. But ROLE is cluster-level. They don't\n> belong to a database. ALTER ROLE XXX SET disk_quota = xxx means to set the\n> quota for the user on all the databases in the first glance. But in our\n> first stage design, ROLE's quota is bind to a specific database. E.g. Role\n> Jack could have 10GB quota on database A and 2GB quota on database B.\n>\n> SQL syntax is not hard to modify, I don't think this should block the\n> main design of disk quota feature. Is there any comment on the design and\n> architecture? If no, we'll firstly submit our patch and involve more\n> discussion?\n>\n> On Sat, Sep 22, 2018 at 3:03 PM Pavel Stehule <pavel.stehule@gmail.com>\n> wrote:\n>\n>>\n>>\n>> so 22. 9. 2018 v 8:48 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:\n>>\n>>> But it looks like redundant to current GUC configuration and limits\n>>>\n>>> what do you mean by current GUC configuration? Is that the general block\n>>> number limit in your patch? If yes, the difference between GUC and\n>>> pg_diskquota catalog is that pg_diskquota will store different quota limit\n>>> for the different role, schema or table instead of a single GUC value.\n>>>\n>>\n>> storage is not relevant in this moment.\n>>\n>> I don't see to consistent to sets some limits via SET command, or ALTER X\n>> SET, and some other with CREATE QUOTA ON.\n>>\n>> The quotas or object limits, resource limits are pretty useful and\n>> necessary, but I don't see these like new type of objects, it is much more\n>> some property of current objects. Because we have one syntax for this\n>> purpose I prefer it. Because is not good to have two syntaxes for similar\n>> purpose.\n>>\n>> So instead CREATE DISC QUATA ON SCHEMA xxx some value I prefer\n>>\n>> ALTER SCHEMA xxx SET disc_quota = xxx;\n>>\n>> The functionality is +/- same. But ALTER XX SET was introduce first, and\n>> I don't feel comfortable to have any new syntax for similar purpose\n>>\n>> Regards\n>>\n>> Pavel\n>>\n>>\n>>\n>>\n>>\n>>>\n>>> On Sat, Sep 22, 2018 at 11:17 AM Pavel Stehule <pavel.stehule@gmail.com>\n>>> wrote:\n>>>\n>>>>\n>>>>\n>>>> pá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io>\n>>>> napsal:\n>>>>\n>>>>> just fast reaction - why QUOTA object?\n>>>>>> Isn't ALTER SET enough?\n>>>>>> Some like\n>>>>>> ALTER TABLE a1 SET quote = 1MB;\n>>>>>> ALTER USER ...\n>>>>>> ALTER SCHEMA ..\n>>>>>> New DDL commans looks like too hard hammer .\n>>>>>\n>>>>>\n>>>>> It's an option. Prefer to consider quota setting store together:\n>>>>> CREATE DISK QUOTA way is more nature to store quota setting in a\n>>>>> separate pg_diskquota catalog\n>>>>> While ALTER SET way is more close to store quota setting in pg_class,\n>>>>> pg_role, pg_namespace. etc in an integrated way.\n>>>>> (Note that here I mean nature/close is not must, ALTER SET could also\n>>>>> store in pg_diskquota and vice versa.)\n>>>>>\n>>>>\n>>>> I have not a problem with new special table for storing this\n>>>> information. But it looks like redundant to current GUC configuration and\n>>>> limits. Can be messy do some work with ALTER ROLE, and some work via CREATE\n>>>> QUOTE.\n>>>>\n>>>> Regards\n>>>>\n>>>> Pavel\n>>>>\n>>>>\n>>>>> Here are some differences I can think of:\n>>>>> 1 pg_role is a global catalog, not per database level. It's harder to\n>>>>> tracker the user's disk usage in the whole clusters(considering 1000+\n>>>>> databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it\n>>>>> only tracks the user's disk usage inside the current database.\n>>>>> 2 using separate pg_diskquota could add more field except for quota\n>>>>> limit without adding too many fields in pg_class, e.g. red zone to give the\n>>>>> user a warning or the current disk usage of the db objects.\n>>>>>\n>>>>> On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com>\n>>>>> wrote:\n>>>>>\n>>>>>>\n>>>>>>\n>>>>>> pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io>\n>>>>>> napsal:\n>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>> *Hi all,We redesign disk quota feature based on the comments from\n>>>>>>> Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically,\n>>>>>>> disk quota feature is used to support multi-tenancy environment, different\n>>>>>>> level of database objects could be set a quota limit to avoid over use of\n>>>>>>> disk space. A common case could be as follows: DBA could enable disk quota\n>>>>>>> on a specified database list. DBA could set disk quota limit for\n>>>>>>> tables/schemas/roles in these databases. Separate disk quota worker process\n>>>>>>> will monitor the disk usage for these objects and detect the objects which\n>>>>>>> exceed their quota limit. Queries loading data into these “out of disk\n>>>>>>> quota” tables/schemas/roles will be cancelled.We are currently working at\n>>>>>>> init implementation stage. We would like to propose our idea firstly and\n>>>>>>> get feedbacks from community to do quick iteration.SQL Syntax (How to use\n>>>>>>> disk quota)1 Specify the databases with disk quota enabled in GUC\n>>>>>>> “diskquota_databases” in postgresql.conf and restart the database.2 DBA\n>>>>>>> could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\n>>>>>>> ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\n>>>>>>> (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\n>>>>>>> ‘3MB’);*\n>>>>>>>\n>>>>>>\n>>>>>> just fast reaction - why QUOTA object?\n>>>>>>\n>>>>>> Isn't ALTER SET enough?\n>>>>>>\n>>>>>> Some like\n>>>>>>\n>>>>>> ALTER TABLE a1 SET quote = 1MB;\n>>>>>> ALTER USER ...\n>>>>>> ALTER SCHEMA ..\n>>>>>>\n>>>>>> New DDL commans looks like too hard hammer .\n>>>>>>\n>>>>>>\n>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>> *3 Simulate a schema out of quota limit case: suppose table a1 and\n>>>>>>> table a2 are both under schema s1.INSERT INTO a1 SELECT\n>>>>>>> generate_series(1,1000);INSERT INTO a2 SELECT\n>>>>>>> generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\n>>>>>>> generate_series(1,1000);ERROR: schema's disk space quota exceededDROP\n>>>>>>> TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\n>>>>>>> generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\n>>>>>>> following components.1. Quota Setting Store is where the disk quota setting\n>>>>>>> to be stored and accessed. We plan to use catalog table pg_diskquota to\n>>>>>>> store these information. pg_diskquota is\n>>>>>>> like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\n>>>>>>> diskquota name */ int16 quotatype; /* diskquota type name */ Oid\n>>>>>>> quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\n>>>>>>> diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\n>>>>>>> MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\n>>>>>>> change of database objects. We plan to use stat collector to detect the\n>>>>>>> ‘active’ table list at initial stage. But stat collector has some\n>>>>>>> limitation on finding the active table which is in a running transaction.\n>>>>>>> Details see TODO section.3. Quota Size Checker is where to calculate the\n>>>>>>> size and compare with quota limit for database objects. According to\n>>>>>>> Pavel’s comment, autovacuum launcher and worker process could be a good\n>>>>>>> reference to disk quota. So we plan to use a disk quota launcher daemon\n>>>>>>> process and several disk quota worker process to finish this work. Launcher\n>>>>>>> process is responsible for starting worker process based on a user defined\n>>>>>>> database list from GUC. Worker process will connect to its target database\n>>>>>>> and monitor the disk usage for objects in this database. In init stage of\n>>>>>>> worker process, it will call calculate_total_relation_size() to calculate\n>>>>>>> the size for each user table. After init stage, worker process will refresh\n>>>>>>> the disk model every N seconds. Refreshing will only recalculate the size\n>>>>>>> of tables in ‘active’ table list, which is generated by Quata Change\n>>>>>>> Detector to minimize the cost.4. Quota Enforcement Operator is where to\n>>>>>>> check for the quota limitation at postgres backend side. We will firstly\n>>>>>>> implement it in ExecCheckRTPerms() as pre-running enforcement. It will\n>>>>>>> check the disk quota of tables being inserted or updated, and report error\n>>>>>>> if table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\n>>>>>>> As a native feature, we plan to add more checkpoint to do running query\n>>>>>>> enforcement. For example, if a disk quota lefts 10MB quota, a query could\n>>>>>>> insert 1GB data. This query could be allowed in pre-running enforcement\n>>>>>>> check, but will be cancelled in running query enforcement check. Therefore,\n>>>>>>> it can improve the accurate of disk quota usage. To achieve this, we plan\n>>>>>>> to add a checkpoint in lower API such as smgr_extened. Hence, the Quota\n>>>>>>> Enforcement Operator will check the disk quota usage when smgr_extened is\n>>>>>>> called. If the quota is over limited, current query will be cancelled.\n>>>>>>> Highlight1. Native feature.Support native Create/Drop Disk Quota SQL\n>>>>>>> statement.New catalog table pg_diskquota to store disk quota setting.2.\n>>>>>>> Auto DML/DDL detection. Table\n>>>>>>> create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\n>>>>>>> Schema create/drop and Role create/drop will be detected by disk quota\n>>>>>>> automatically. 3. Low cost disk quota checker.Worker process of disk quota\n>>>>>>> need to refresh the disk usage model every N seconds. Since recalculate the\n>>>>>>> file size using stat() system call is expensive for a large number of\n>>>>>>> files, we use an ‘active’ table list to reduce the real work at each\n>>>>>>> iteration. A basic experiment on our init stage implementation on database\n>>>>>>> with 20K tables shows that the refresh cost is 1% cpu usage and will be\n>>>>>>> finished within 50ms. Todo/LimitationBefore we propose our patch, we plan\n>>>>>>> to enhance it with the following ideas:1. Setting database list with disk\n>>>>>>> quota enabled dynamically without restart database. Since we have the disk\n>>>>>>> quota launcher process, it could detect the new ‘diskquota_databases’ list\n>>>>>>> and start/stop the corresponding disk quota worker process.2. Enforcement\n>>>>>>> when query is running. Considering the case when there is 10MB quota left,\n>>>>>>> but next query will insert 10GB data. Current enforcement design will allow\n>>>>>>> this query to be executed. This is limited by the ‘active’ table detection\n>>>>>>> is generated by stat collector. Postgres backend will only send table stat\n>>>>>>> information to collector only when the transaction ends. We need a new way\n>>>>>>> to detect the ‘active’ table even when this table is being modified inside\n>>>>>>> a running transaction.3. Monitor unlimited number of databases. Current we\n>>>>>>> set the max number of disk quota worker process to be 10 to reduce the\n>>>>>>> affection normal workload. But how about if we want to monitor the disk\n>>>>>>> quota of more than 10 databases? Our solution is to let disk quota launcher\n>>>>>>> to manage a queue of database need to be monitored. And disk quota worker\n>>>>>>> process consuming the queue and refresh the disk usage/quota for this\n>>>>>>> database. After some periods, worker will return the database to the queue,\n>>>>>>> and fetch the top database from queue to process. The period determine the\n>>>>>>> delay of detecting disk quota change. To implement this feature, we need to\n>>>>>>> support a subprocess of postmaster to rebind to another database instead of\n>>>>>>> the database binded in InitPostgres().4. Support active table detection on\n>>>>>>> vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\n>>>>>>> are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\n>>>>>>> Flack for the former comments on disk quota feature. Any comments on how to\n>>>>>>> improve disk quota feature are appreciated.*\n>>>>>>>\n>>>>>>>\n>>>>>>> On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <\n>>>>>>> pavel.stehule@gmail.com> wrote:\n>>>>>>>\n>>>>>>>>\n>>>>>>>>\n>>>>>>>> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>>>>>>>\n>>>>>>>>> Thanks Pavel.\n>>>>>>>>> Your patch did enforcement on storage level(md.c or we could also\n>>>>>>>>> use smgr_extend). It's straight forward.\n>>>>>>>>> But I prefer to implement disk_quota as a feature with following\n>>>>>>>>> objectives:\n>>>>>>>>> 1 set/alter disk quota setting on different database objects, e.g.\n>>>>>>>>> user, database, schema etc. not only a general GUC, but we could set\n>>>>>>>>> separate quota limit for a specific objects.\n>>>>>>>>> 2 enforcement operator should work at two positions: before query\n>>>>>>>>> is running and when query is running. The latter one's implementation maybe\n>>>>>>>>> similar to your patch.\n>>>>>>>>>\n>>>>>>>>\n>>>>>>>> The patch was just example. The resource quotes should be more\n>>>>>>>> complex - per partition, table, schema, database, user - so GUC are\n>>>>>>>> possible, but not very user friendly.\n>>>>>>>>\n>>>>>>>> Our case is specific, but not too much. The servers are used for\n>>>>>>>> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\n>>>>>>>> SELECT). We need to solve limits immediately. The implementation is simple,\n>>>>>>>> so I did it. Same implementation on database level, or schema level needs\n>>>>>>>> some more locks, so it will not be too effective. The resource management\n>>>>>>>> can be complex very complex, and I expect so it will be hard work.\n>>>>>>>>\n>>>>>>>> Regards\n>>>>>>>>\n>>>>>>>> Pavel\n>>>>>>>>\n>>>>>>>>\n>>>>>>>>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <\n>>>>>>>>> pavel.stehule@gmail.com> wrote:\n>>>>>>>>>\n>>>>>>>>>> Hi\n>>>>>>>>>>\n>>>>>>>>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\n>>>>>>>>>>\n>>>>>>>>>>> Thanks Chapman.\n>>>>>>>>>>> @Pavel, could you please explain more about your second\n>>>>>>>>>>> suggestion \"implement some quotas on storage level?\"\n>>>>>>>>>>>\n>>>>>>>>>>\n>>>>>>>>>> See attached patch - it is very simple - and good enough for our\n>>>>>>>>>> purposes.\n>>>>>>>>>>\n>>>>>>>>>> Regards\n>>>>>>>>>>\n>>>>>>>>>> Pavel\n>>>>>>>>>>\n>>>>>>>>>>\n>>>>>>>>>>\n>>>>>>>>>>> We will not keep the long-lived processes attach to all\n>>>>>>>>>>> databases(just like you mentioned servers with thousands of\n>>>>>>>>>>> databases)\n>>>>>>>>>>> And you are right, we could share ideas with autovacuum process,\n>>>>>>>>>>> fork worker processes in need.\n>>>>>>>>>>> \"autovacuum checks for tables that have had a large number of\n>>>>>>>>>>> inserted, updated or deleted tuples. These checks use the statistics\n>>>>>>>>>>> collection facility\"\n>>>>>>>>>>> diskquota process is similar to autovacuum at caring about\n>>>>>>>>>>> insert, but the difference is that it also care about vucuum full, truncate\n>>>>>>>>>>> and drop. While update and delete may not be interested since no file\n>>>>>>>>>>> change happens. So a separate diskquota process is preferred.\n>>>>>>>>>>>\n>>>>>>>>>>> So if we implemented disk quota as a full native feature, and\n>>>>>>>>>>> in the first initial version I prefer to implement the following features:\n>>>>>>>>>>> 1 Fork diskquota launcher process under Postmaster serverloop,\n>>>>>>>>>>> which is long-lived.\n>>>>>>>>>>> 2 Diskquota launcher process is responsible for\n>>>>>>>>>>> creating diskquota worker process for every database.\n>>>>>>>>>>> 3 DIskquota setting is stored in a separate catalog table for\n>>>>>>>>>>> each database.\n>>>>>>>>>>> 4 Initialization stage, Diskquota launcher process creates\n>>>>>>>>>>> diskquota worker process for all the databases(traverse like\n>>>>>>>>>>> autovacuum). Worker process calculates disk usage of db objects and\n>>>>>>>>>>> their diskquota setting. If any db object exceeds its quota limit, put them\n>>>>>>>>>>> into the blacklist in the shared memory, which will later be used by\n>>>>>>>>>>> enforcement operator. Worker process exits when works are done.\n>>>>>>>>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\n>>>>>>>>>>> process for the database with a large number of insert, copy, truncate,\n>>>>>>>>>>> drop etc. or create disk quota statement. Worker process updates the file\n>>>>>>>>>>> size for db objects containing the result relation, and compare with the\n>>>>>>>>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\n>>>>>>>>>>> remove from blacklist vice versa. Worker process exits when\n>>>>>>>>>>> works are done and a GUC could control the frequency of worker process\n>>>>>>>>>>> restart to a specific database. As you know, this GUC also controls the\n>>>>>>>>>>> delay when we do enforcement.\n>>>>>>>>>>> 6 Enforcement. When postgres backend executes queries, check the\n>>>>>>>>>>> blacklist in shared memory to determine whether the query is allowed(before\n>>>>>>>>>>> execute) or need rollback(is executing)?\n>>>>>>>>>>>\n>>>>>>>>>>> If we implemented disk quota as an extension, we could just use\n>>>>>>>>>>> background worker to start diskquota launcher process and use\n>>>>>>>>>>> RegisterDynamicBackgroundWorker() to fork child diskquota\n>>>>>>>>>>> worker processes by the launcher process as suggested by @Chapman.\n>>>>>>>>>>> Diskquota setting could be stored in user table in a separate schema for\n>>>>>>>>>>> each database(Schema and table created by create extension statement) just\n>>>>>>>>>>> like what Heikki has done in pg_quota project. But in this case, we need to\n>>>>>>>>>>> create extension for each database before diskquota worker process can be\n>>>>>>>>>>> set up for that database.\n>>>>>>>>>>>\n>>>>>>>>>>> Any comments on the above design and which is preferred, native\n>>>>>>>>>>> feature or extension as the POC?\n>>>>>>>>>>>\n>>>>>>>>>>>\n>>>>>>>>>>> -- Hubert\n>>>>>>>>>>>\n>>>>>>>>>>>\n>>>>>>>>>>>\n>>>>>>>>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <\n>>>>>>>>>>> pavel.stehule@gmail.com> wrote:\n>>>>>>>>>>>\n>>>>>>>>>>>>\n>>>>>>>>>>>>\n>>>>>>>>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net\n>>>>>>>>>>>> >:\n>>>>>>>>>>>>\n>>>>>>>>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\n>>>>>>>>>>>>>\n>>>>>>>>>>>>> > 2 Keep one worker process for each database. But using a\n>>>>>>>>>>>>> parent/global\n>>>>>>>>>>>>> > quota worker process to manage the lifecycle of database\n>>>>>>>>>>>>> level worker\n>>>>>>>>>>>>> > processes. It could handle the newly created database(avoid\n>>>>>>>>>>>>> restart\n>>>>>>>>>>>>> > database) and save resource when a database is not used. But\n>>>>>>>>>>>>> this needs to\n>>>>>>>>>>>>> > change worker process to be hierarchical. Postmaster becomes\n>>>>>>>>>>>>> the grandfather\n>>>>>>>>>>>>> > of database level worker processes in this case.\n>>>>>>>>>>>>>\n>>>>>>>>>>>>> I am using background workers this way in 9.5 at $work.\n>>>>>>>>>>>>>\n>>>>>>>>>>>>> In my case, one worker lives forever, wakes up on a set\n>>>>>>>>>>>>> period, and\n>>>>>>>>>>>>> starts a short-lived worker for every database, waiting for\n>>>>>>>>>>>>> each\n>>>>>>>>>>>>> one before starting the next.\n>>>>>>>>>>>>>\n>>>>>>>>>>>>> It was straightforward to implement. Looking back over the\n>>>>>>>>>>>>> code,\n>>>>>>>>>>>>> I see the global worker assigns its own PID to\n>>>>>>>>>>>>> worker.bgw_notify_pid\n>>>>>>>>>>>>> of each of its children, and also obtains a handle for each\n>>>>>>>>>>>>> child\n>>>>>>>>>>>>> from RegisterDynamicBackgroundWorker().\n>>>>>>>>>>>>>\n>>>>>>>>>>>>> I imagine the global quota worker would prefer to start workers\n>>>>>>>>>>>>> for every database and then just wait for notifications from\n>>>>>>>>>>>>> any\n>>>>>>>>>>>>> of them, but that seems equally straightforward at first\n>>>>>>>>>>>>> glance.\n>>>>>>>>>>>>>\n>>>>>>>>>>>>\n>>>>>>>>>>>> There are servers with thousands databases. Worker per database\n>>>>>>>>>>>> is not good idea.\n>>>>>>>>>>>>\n>>>>>>>>>>>> It should to share ideas, code with autovacuum process.\n>>>>>>>>>>>>\n>>>>>>>>>>>> Not sure, how to effective implementation based on bg workers\n>>>>>>>>>>>> can be. On servers with large set of databases, large set of tables it can\n>>>>>>>>>>>> identify too big table too late.\n>>>>>>>>>>>>\n>>>>>>>>>>>> Isn't better to implement some quotas on storage level?\n>>>>>>>>>>>>\n>>>>>>>>>>>> Regards\n>>>>>>>>>>>>\n>>>>>>>>>>>> Pavel\n>>>>>>>>>>>>\n>>>>>>>>>>>>\n>>>>>>>>>>>>\n>>>>>>>>>>>>> -Chap\n>>>>>>>>>>>>>\n>>>>>>>>>>>>>\n>>>>>>>>>>>>\n>>>>>>>>>>>\n>>>>>>>>>>>\n>>>>>>>>>>> --\n>>>>>>>>>>> Thanks\n>>>>>>>>>>>\n>>>>>>>>>>> Hubert Zhang\n>>>>>>>>>>>\n>>>>>>>>>>\n>>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>>\n>>>>>>>>> --\n>>>>>>>>> Thanks\n>>>>>>>>>\n>>>>>>>>> Hubert Zhang\n>>>>>>>>>\n>>>>>>>>\n>>>>>>>>\n>>>>>>>\n>>>>>>>\n>>>>>>> --\n>>>>>>> Thanks\n>>>>>>>\n>>>>>>> Hubert Zhang\n>>>>>>>\n>>>>>>\n>>>>>\n>>>>> --\n>>>>> Thanks\n>>>>>\n>>>>> Hubert Zhang\n>>>>>\n>>>>\n>>>\n>>> --\n>>> Thanks\n>>>\n>>> Hubert Zhang\n>>>\n>>\n>\n> --\n> Thanks\n>\n> Hubert Zhang\n>\n\n\n-- \nRegards,\nHaozhou","msg_date":"Wed, 26 Sep 2018 18:10:13 +0800","msg_from":"Haozhou Wang <hawang@pivotal.io>","msg_from_op":false,"msg_subject":"Re: Proposal for disk quota feature"}],"string":"[\n {\n \"msg_contents\": \"Hi all,\\nWe want to introduce disk quota feature into Postgres.\\n\\n*Why disk quota*\\n*In a multi-tenant environment, there is a requirement to limit the disk\\nquota that database/schema/table can be written or a user can consume for\\ndifferent organizations.*\\n*Meanwhile, other databases such as Oracle, Teradata, DB2 have already\\nsupported disk quota feature.*\\n\\n*Heikki has already implemented disk quota feature in Postgres as an\\nextension pg_quota <https://github.com/hlinnaka/pg_quota>. We plan to\\nenhance disk quota feature based on Heikki's implementation.*\\n\\n*Scope*\\n\\nThe scope of disk quota feature is to limit the disk usage of\\ndatabase/schema/table objects and users.\\n\\nHere table means heap table, ao table, index table, partition table and\\nassociated table( toast table, visible table, large object etc.). Schema\\ndisk quota is the disk quota of all the tables in this schema. Database\\ndisk quota is the disk quota of all the tables in this database.\\n\\nUser's quota is the size of all the tables whose owner are this user.\\nOut of Scope: Note that spill files, xlogs, clogs and logs are not\\nconsidered for database object level disk quota at this stage.\\n\\n\\n*Design*\\nWe propose disk quota with the following components:\\n\\n1. Quota Setting Store is where the disk quota setting to be stored and\\naccessed. DBA or object owner uses SQL queries to configure the disk quota\\nfor each database objects.\\n\\n\\n\\n*2. Quota Change Detector is the monitor of size change of database objects\\nin Postgres. It will write change information to shared memory to notify\\nQuota Size Checker. The implementation of Change Detector could be hooks in\\nsmgr_extend/smgr_unlink/smgr_truncate when modifying the size of a heap\\ntable. The hooks will write to shared memory in a batched way to reduce the\\nimpact on OLTP performance.3. Quota Size Checker is implemented as a worker\\nprocess. It maintains the current disk usage for each database objects and\\nusers, and compare them with settings in Quota Setting Store. If it detects\\nthe disk usage hit the quota redzone(either upon or below), it will notify\\nQuota Enforcement Operator. 4. Quota Enforcement Operator has two roles:\\none is to check the disk quota permission before queries are\\nexecuted(QueryBeforeRun Check), the other is to cancel the running queries\\nwhen it reaches the disk quota limit dynamically(QueryRunning Check). Quota\\nEnforcement Operator uses shared memory to store the enforcement\\ninformation to guarantee a quick check. *\\n\\n\\n*To implement the right proof of concept, we want to receive feedback from\\nthe community from the following aspects: *\\nQ1. User Interface: when setting a disk quota,\\n Option 1 is to use *insert into quota.config(user1, 10G)*\\n Option 2 is to use UDF *select set_quota(\\\"role\\\",\\\"user1\\\",10G)*\\n Option 3 is to use native SQL syntax *create disk quota on ROLE user1\\n10G, or create disk quota on SCHEMA s1 25G;*\\nQ2. Quota Setting Store using user table or catalog?\\n Option 1 is to create a schema called quota for each database and write\\nquota settings into quota.config table, only DBA could modify it. This\\ncorresponds to Q1.option1\\n Option 2 is to store quota settings into the catalog. For Schema and\\nTable write them to database level catalog. For database or user, write\\nthem to either database level or global catalog.\\n\\nI personally prefer Q1's option3 and Q2's option2, since it makes disk\\nquota more like a native feature. We could support the quota worker process\\nimplementation as an extension for now, but in the long term, disk quota is\\nlike a fundamental feature of a database and should be a native feature\\njust like other databases. So store quota conf into catalog and supply\\nnative syntax is better.\\n\\nOpen Problem\\nWe prepare to implement Quota Size Checker as a worker process. Worker\\nprocess needs to connect to a database to build the disk usage map and\\nquota map(e.g. a user/shcema's disk usage on a given database) But one\\nworker process can only bind to one database(InitPostgres(dbname)) at the\\ninitialization stage. It results in we need a separate worker process for\\neach database. The solution to this problem is not straightforward, here\\nare some ideas:\\n1 To make worker process could retrieve and cache information from all the\\ndatabases. As Tom Lane pointed out that it needs to flush all the database\\nspecific thing, like relcache syscache etc.\\n2 Keep one worker process for each database. But using a parent/global\\nquota worker process to manage the lifecycle of database level worker\\nprocesses. It could handle the newly created database(avoid restart\\ndatabase) and save resource when a database is not used. But this needs to\\nchange worker process to be hierarchical. Postmaster becomes the grandfather\\n of database level worker processes in this case.\\n\\nAny better ideas on it?\\n\\n\\n\\n-- \\nThanks\\n\\nHubert Zhang\\n\\nHi all,We want to introduce disk quota feature into Postgres.Why disk quotaIn a multi-tenant environment, there is a requirement to limit the disk quota that database/schema/table can be written or a user can consume for different organizations.Meanwhile, other databases such as Oracle, Teradata, DB2 have already supported disk quota feature.Heikki has already implemented disk quota feature in Postgres as an extension pg_quota. We plan to enhance disk quota feature based on Heikki's implementation.ScopeThe scope of disk quota feature is to limit the disk usage of database/schema/table objects and users.Here table means heap table, ao table, index table, partition table and associated table( toast table, visible table, large object etc.). Schema disk quota is the disk quota of all the tables in this schema. Database disk quota is the disk quota of all the tables in this database.User's quota is the size of all the tables whose owner are this user.Out of Scope: Note that spill files, xlogs, clogs and logs are not considered for database object level disk quota at this stage.DesignWe propose disk quota with the following components:1. Quota Setting Store is where the disk quota setting to be stored and accessed. DBA or object owner uses SQL queries to configure the disk quota for each database objects. 2. Quota Change Detector is the monitor of size change of database objects in Postgres. It will write change information to shared memory to notify Quota Size Checker. The implementation of Change Detector could be hooks in smgr_extend/smgr_unlink/smgr_truncate when modifying the size of a heap table. The hooks will write to shared memory in a batched way to reduce the impact on OLTP performance.3. Quota Size Checker is implemented as a worker process. It maintains the current disk usage for each database objects and users, and compare them with settings in Quota Setting Store. If it detects the disk usage hit the quota redzone(either upon or below), it will notify Quota Enforcement Operator. 4. Quota Enforcement Operator has two roles: one is to check the disk quota permission before queries are executed(QueryBeforeRun Check), the other is to cancel the running queries when it reaches the disk quota limit dynamically(QueryRunning Check). Quota Enforcement Operator uses shared memory to store the enforcement information to guarantee a quick check. To implement the right proof of concept, we want to receive feedback from the community from the following aspects: Q1. User Interface: when setting a disk quota, Option 1 is to use insert into quota.config(user1, 10G) Option 2 is to use UDF select set_quota(\\\"role\\\",\\\"user1\\\",10G) Option 3 is to use native SQL syntax create disk quota on ROLE user1 10G, or create disk quota on SCHEMA s1 25G;Q2. Quota Setting Store using user table or catalog? Option 1 is to create a schema called quota for each database and write quota settings into quota.config table, only DBA could modify it. This corresponds to Q1.option1 Option 2 is to store quota settings into the catalog. For Schema and Table write them to database level catalog. For database or user, write them to either database level or global catalog. I personally prefer Q1's option3 and Q2's option2, since it makes disk quota more like a native feature. We could support the quota worker process implementation as an extension for now, but in the long term, disk quota is like a fundamental feature of a database and should be a native feature just like other databases. So store quota conf into catalog and supply native syntax is better.Open ProblemWe prepare to implement Quota Size Checker as a worker process. Worker process needs to connect to a database to build the disk usage map and quota map(e.g. a user/shcema's disk usage on a given database) But one worker process can only bind to one database(InitPostgres(dbname)) at the initialization stage. It results in we need a separate worker process for each database. The solution to this problem is not straightforward, here are some ideas:1 To make worker process could retrieve and cache information from all the databases. As Tom Lane pointed out that it needs to flush all the database specific thing, like relcache syscache etc.2 Keep one worker process for each database. But using a parent/global quota worker process to manage the lifecycle of database level worker processes. It could handle the newly created database(avoid restart database) and save resource when a database is not used. But this needs to change worker process to be hierarchical. Postmaster becomes the grandfather of database level worker processes in this case.Any better ideas on it?-- ThanksHubert Zhang\",\n \"msg_date\": \"Thu, 30 Aug 2018 21:57:41 +0800\",\n \"msg_from\": \"Hubert Zhang <hzhang@pivotal.io>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Proposal for disk quota feature\"\n },\n {\n \"msg_contents\": \"On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n\\n> 2 Keep one worker process for each database. But using a parent/global\\n> quota worker process to manage the lifecycle of database level worker\\n> processes. It could handle the newly created database(avoid restart\\n> database) and save resource when a database is not used. But this needs to\\n> change worker process to be hierarchical. Postmaster becomes the grandfather\\n> of database level worker processes in this case.\\n\\nI am using background workers this way in 9.5 at $work.\\n\\nIn my case, one worker lives forever, wakes up on a set period, and\\nstarts a short-lived worker for every database, waiting for each\\none before starting the next.\\n\\nIt was straightforward to implement. Looking back over the code,\\nI see the global worker assigns its own PID to worker.bgw_notify_pid\\nof each of its children, and also obtains a handle for each child\\nfrom RegisterDynamicBackgroundWorker().\\n\\nI imagine the global quota worker would prefer to start workers\\nfor every database and then just wait for notifications from any\\nof them, but that seems equally straightforward at first glance.\\n\\n-Chap\\n\\n\",\n \"msg_date\": \"Thu, 30 Aug 2018 10:22:37 -0400\",\n \"msg_from\": \"Chapman Flack <chap@anastigmatix.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal for disk quota feature\"\n },\n {\n \"msg_contents\": \"2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\\n\\n> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n>\\n> > 2 Keep one worker process for each database. But using a parent/global\\n> > quota worker process to manage the lifecycle of database level worker\\n> > processes. It could handle the newly created database(avoid restart\\n> > database) and save resource when a database is not used. But this needs\\n> to\\n> > change worker process to be hierarchical. Postmaster becomes the\\n> grandfather\\n> > of database level worker processes in this case.\\n>\\n> I am using background workers this way in 9.5 at $work.\\n>\\n> In my case, one worker lives forever, wakes up on a set period, and\\n> starts a short-lived worker for every database, waiting for each\\n> one before starting the next.\\n>\\n> It was straightforward to implement. Looking back over the code,\\n> I see the global worker assigns its own PID to worker.bgw_notify_pid\\n> of each of its children, and also obtains a handle for each child\\n> from RegisterDynamicBackgroundWorker().\\n>\\n> I imagine the global quota worker would prefer to start workers\\n> for every database and then just wait for notifications from any\\n> of them, but that seems equally straightforward at first glance.\\n>\\n\\nThere are servers with thousands databases. Worker per database is not good\\nidea.\\n\\nIt should to share ideas, code with autovacuum process.\\n\\nNot sure, how to effective implementation based on bg workers can be. On\\nservers with large set of databases, large set of tables it can identify\\ntoo big table too late.\\n\\nIsn't better to implement some quotas on storage level?\\n\\nRegards\\n\\nPavel\\n\\n\\n\\n> -Chap\\n>\\n>\\n\\n2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n\\n> 2 Keep one worker process for each database. But using a parent/global\\n> quota worker process to manage the lifecycle of database level worker\\n> processes. It could handle the newly created database(avoid restart\\n> database) and save resource when a database is not used. But this needs to\\n> change worker process to be hierarchical. Postmaster becomes the grandfather\\n> of database level worker processes in this case.\\n\\nI am using background workers this way in 9.5 at $work.\\n\\nIn my case, one worker lives forever, wakes up on a set period, and\\nstarts a short-lived worker for every database, waiting for each\\none before starting the next.\\n\\nIt was straightforward to implement. Looking back over the code,\\nI see the global worker assigns its own PID to worker.bgw_notify_pid\\nof each of its children, and also obtains a handle for each child\\nfrom RegisterDynamicBackgroundWorker().\\n\\nI imagine the global quota worker would prefer to start workers\\nfor every database and then just wait for notifications from any\\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\\n\\n-Chap\",\n \"msg_date\": \"Thu, 30 Aug 2018 21:32:58 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal for disk quota feature\"\n },\n {\n \"msg_contents\": \"Thanks Chapman.\\n@Pavel, could you please explain more about your second suggestion \\\"implement\\nsome quotas on storage level?\\\"\\nWe will not keep the long-lived processes attach to all databases(just like\\nyou mentioned servers with thousands of databases)\\nAnd you are right, we could share ideas with autovacuum process, fork\\nworker processes in need.\\n\\\"autovacuum checks for tables that have had a large number of inserted,\\nupdated or deleted tuples. These checks use the statistics collection\\nfacility\\\"\\ndiskquota process is similar to autovacuum at caring about insert, but the\\ndifference is that it also care about vucuum full, truncate and drop. While\\nupdate and delete may not be interested since no file change happens. So a\\nseparate diskquota process is preferred.\\n\\nSo if we implemented disk quota as a full native feature, and in the first\\ninitial version I prefer to implement the following features:\\n1 Fork diskquota launcher process under Postmaster serverloop, which is\\nlong-lived.\\n2 Diskquota launcher process is responsible for creating diskquota worker\\nprocess for every database.\\n3 DIskquota setting is stored in a separate catalog table for each database.\\n4 Initialization stage, Diskquota launcher process creates diskquota worker\\nprocess for all the databases(traverse like autovacuum). Worker process\\ncalculates disk usage of db objects and their diskquota setting. If any\\ndb object exceeds its quota limit, put them into the blacklist in the\\nshared memory, which will later be used by enforcement operator. Worker\\nprocess exits when works are done.\\n5 Running stage, Diskquota launcher process creates diskquota worker\\nprocess for the database with a large number of insert, copy, truncate,\\ndrop etc. or create disk quota statement. Worker process updates the file\\nsize for db objects containing the result relation, and compare with the\\ndiskquota setting. Again, if exceeds quota limit, put them into blacklist,\\nremove from blacklist vice versa. Worker process exits when works are done\\nand a GUC could control the frequency of worker process restart to a\\nspecific database. As you know, this GUC also controls the delay when we do\\nenforcement.\\n6 Enforcement. When postgres backend executes queries, check the blacklist\\nin shared memory to determine whether the query is allowed(before execute)\\nor need rollback(is executing)?\\n\\nIf we implemented disk quota as an extension, we could just use background\\nworker to start diskquota launcher process and use\\nRegisterDynamicBackgroundWorker() to fork child diskquota worker processes\\nby the launcher process as suggested by @Chapman. Diskquota setting could\\nbe stored in user table in a separate schema for each database(Schema and\\ntable created by create extension statement) just like what Heikki has done\\nin pg_quota project. But in this case, we need to create extension for each\\ndatabase before diskquota worker process can be set up for that database.\\n\\nAny comments on the above design and which is preferred, native feature or\\nextension as the POC?\\n\\n\\n-- Hubert\\n\\n\\n\\nOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com>\\nwrote:\\n\\n>\\n>\\n> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\\n>\\n>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n>>\\n>> > 2 Keep one worker process for each database. But using a parent/global\\n>> > quota worker process to manage the lifecycle of database level worker\\n>> > processes. It could handle the newly created database(avoid restart\\n>> > database) and save resource when a database is not used. But this needs\\n>> to\\n>> > change worker process to be hierarchical. Postmaster becomes the\\n>> grandfather\\n>> > of database level worker processes in this case.\\n>>\\n>> I am using background workers this way in 9.5 at $work.\\n>>\\n>> In my case, one worker lives forever, wakes up on a set period, and\\n>> starts a short-lived worker for every database, waiting for each\\n>> one before starting the next.\\n>>\\n>> It was straightforward to implement. Looking back over the code,\\n>> I see the global worker assigns its own PID to worker.bgw_notify_pid\\n>> of each of its children, and also obtains a handle for each child\\n>> from RegisterDynamicBackgroundWorker().\\n>>\\n>> I imagine the global quota worker would prefer to start workers\\n>> for every database and then just wait for notifications from any\\n>> of them, but that seems equally straightforward at first glance.\\n>>\\n>\\n> There are servers with thousands databases. Worker per database is not\\n> good idea.\\n>\\n> It should to share ideas, code with autovacuum process.\\n>\\n> Not sure, how to effective implementation based on bg workers can be. On\\n> servers with large set of databases, large set of tables it can identify\\n> too big table too late.\\n>\\n> Isn't better to implement some quotas on storage level?\\n>\\n> Regards\\n>\\n> Pavel\\n>\\n>\\n>\\n>> -Chap\\n>>\\n>>\\n>\\n\\n\\n-- \\nThanks\\n\\nHubert Zhang\\n\\nThanks Chapman.@Pavel, could you please explain more about your second suggestion \\\"implement some quotas on storage level?\\\"We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \\\"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\\\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n\\n> 2 Keep one worker process for each database. But using a parent/global\\n> quota worker process to manage the lifecycle of database level worker\\n> processes. It could handle the newly created database(avoid restart\\n> database) and save resource when a database is not used. But this needs to\\n> change worker process to be hierarchical. Postmaster becomes the grandfather\\n> of database level worker processes in this case.\\n\\nI am using background workers this way in 9.5 at $work.\\n\\nIn my case, one worker lives forever, wakes up on a set period, and\\nstarts a short-lived worker for every database, waiting for each\\none before starting the next.\\n\\nIt was straightforward to implement. Looking back over the code,\\nI see the global worker assigns its own PID to worker.bgw_notify_pid\\nof each of its children, and also obtains a handle for each child\\nfrom RegisterDynamicBackgroundWorker().\\n\\nI imagine the global quota worker would prefer to start workers\\nfor every database and then just wait for notifications from any\\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\\n\\n-Chap\\n\\n\\n-- ThanksHubert Zhang\",\n \"msg_date\": \"Sun, 2 Sep 2018 20:18:56 +0800\",\n \"msg_from\": \"Hubert Zhang <hzhang@pivotal.io>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal for disk quota feature\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n\\n> Thanks Chapman.\\n> @Pavel, could you please explain more about your second suggestion \\\"implement\\n> some quotas on storage level?\\\"\\n>\\n\\nSee attached patch - it is very simple - and good enough for our purposes.\\n\\nRegards\\n\\nPavel\\n\\n\\n\\n> We will not keep the long-lived processes attach to all databases(just\\n> like you mentioned servers with thousands of databases)\\n> And you are right, we could share ideas with autovacuum process, fork\\n> worker processes in need.\\n> \\\"autovacuum checks for tables that have had a large number of inserted,\\n> updated or deleted tuples. These checks use the statistics collection\\n> facility\\\"\\n> diskquota process is similar to autovacuum at caring about insert, but the\\n> difference is that it also care about vucuum full, truncate and drop. While\\n> update and delete may not be interested since no file change happens. So a\\n> separate diskquota process is preferred.\\n>\\n> So if we implemented disk quota as a full native feature, and in the\\n> first initial version I prefer to implement the following features:\\n> 1 Fork diskquota launcher process under Postmaster serverloop, which is\\n> long-lived.\\n> 2 Diskquota launcher process is responsible for creating diskquota worker\\n> process for every database.\\n> 3 DIskquota setting is stored in a separate catalog table for each\\n> database.\\n> 4 Initialization stage, Diskquota launcher process creates diskquota worker\\n> process for all the databases(traverse like autovacuum). Worker process\\n> calculates disk usage of db objects and their diskquota setting. If any\\n> db object exceeds its quota limit, put them into the blacklist in the\\n> shared memory, which will later be used by enforcement operator. Worker\\n> process exits when works are done.\\n> 5 Running stage, Diskquota launcher process creates diskquota worker\\n> process for the database with a large number of insert, copy, truncate,\\n> drop etc. or create disk quota statement. Worker process updates the file\\n> size for db objects containing the result relation, and compare with the\\n> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\\n> remove from blacklist vice versa. Worker process exits when works are\\n> done and a GUC could control the frequency of worker process restart to a\\n> specific database. As you know, this GUC also controls the delay when we do\\n> enforcement.\\n> 6 Enforcement. When postgres backend executes queries, check the blacklist\\n> in shared memory to determine whether the query is allowed(before execute)\\n> or need rollback(is executing)?\\n>\\n> If we implemented disk quota as an extension, we could just use background\\n> worker to start diskquota launcher process and use\\n> RegisterDynamicBackgroundWorker() to fork child diskquota worker\\n> processes by the launcher process as suggested by @Chapman.\\n> Diskquota setting could be stored in user table in a separate schema for\\n> each database(Schema and table created by create extension statement) just\\n> like what Heikki has done in pg_quota project. But in this case, we need to\\n> create extension for each database before diskquota worker process can be\\n> set up for that database.\\n>\\n> Any comments on the above design and which is preferred, native feature or\\n> extension as the POC?\\n>\\n>\\n> -- Hubert\\n>\\n>\\n>\\n> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com>\\n> wrote:\\n>\\n>>\\n>>\\n>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\\n>>\\n>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n>>>\\n>>> > 2 Keep one worker process for each database. But using a parent/global\\n>>> > quota worker process to manage the lifecycle of database level worker\\n>>> > processes. It could handle the newly created database(avoid restart\\n>>> > database) and save resource when a database is not used. But this\\n>>> needs to\\n>>> > change worker process to be hierarchical. Postmaster becomes the\\n>>> grandfather\\n>>> > of database level worker processes in this case.\\n>>>\\n>>> I am using background workers this way in 9.5 at $work.\\n>>>\\n>>> In my case, one worker lives forever, wakes up on a set period, and\\n>>> starts a short-lived worker for every database, waiting for each\\n>>> one before starting the next.\\n>>>\\n>>> It was straightforward to implement. Looking back over the code,\\n>>> I see the global worker assigns its own PID to worker.bgw_notify_pid\\n>>> of each of its children, and also obtains a handle for each child\\n>>> from RegisterDynamicBackgroundWorker().\\n>>>\\n>>> I imagine the global quota worker would prefer to start workers\\n>>> for every database and then just wait for notifications from any\\n>>> of them, but that seems equally straightforward at first glance.\\n>>>\\n>>\\n>> There are servers with thousands databases. Worker per database is not\\n>> good idea.\\n>>\\n>> It should to share ideas, code with autovacuum process.\\n>>\\n>> Not sure, how to effective implementation based on bg workers can be. On\\n>> servers with large set of databases, large set of tables it can identify\\n>> too big table too late.\\n>>\\n>> Isn't better to implement some quotas on storage level?\\n>>\\n>> Regards\\n>>\\n>> Pavel\\n>>\\n>>\\n>>\\n>>> -Chap\\n>>>\\n>>>\\n>>\\n>\\n>\\n> --\\n> Thanks\\n>\\n> Hubert Zhang\\n>\",\n \"msg_date\": \"Sun, 2 Sep 2018 14:44:40 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal for disk quota feature\"\n },\n {\n \"msg_contents\": \"Thanks Pavel.\\nYour patch did enforcement on storage level(md.c or we could also use\\nsmgr_extend). It's straight forward.\\nBut I prefer to implement disk_quota as a feature with following objectives:\\n1 set/alter disk quota setting on different database objects, e.g. user,\\ndatabase, schema etc. not only a general GUC, but we could set separate\\nquota limit for a specific objects.\\n2 enforcement operator should work at two positions: before query is\\nrunning and when query is running. The latter one's implementation maybe\\nsimilar to your patch.\\n\\nOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com>\\nwrote:\\n\\n> Hi\\n>\\n> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>\\n>> Thanks Chapman.\\n>> @Pavel, could you please explain more about your second suggestion \\\"implement\\n>> some quotas on storage level?\\\"\\n>>\\n>\\n> See attached patch - it is very simple - and good enough for our purposes.\\n>\\n> Regards\\n>\\n> Pavel\\n>\\n>\\n>\\n>> We will not keep the long-lived processes attach to all databases(just\\n>> like you mentioned servers with thousands of databases)\\n>> And you are right, we could share ideas with autovacuum process, fork\\n>> worker processes in need.\\n>> \\\"autovacuum checks for tables that have had a large number of inserted,\\n>> updated or deleted tuples. These checks use the statistics collection\\n>> facility\\\"\\n>> diskquota process is similar to autovacuum at caring about insert, but\\n>> the difference is that it also care about vucuum full, truncate and drop.\\n>> While update and delete may not be interested since no file change happens.\\n>> So a separate diskquota process is preferred.\\n>>\\n>> So if we implemented disk quota as a full native feature, and in the\\n>> first initial version I prefer to implement the following features:\\n>> 1 Fork diskquota launcher process under Postmaster serverloop, which is\\n>> long-lived.\\n>> 2 Diskquota launcher process is responsible for creating diskquota\\n>> worker process for every database.\\n>> 3 DIskquota setting is stored in a separate catalog table for each\\n>> database.\\n>> 4 Initialization stage, Diskquota launcher process creates diskquota worker\\n>> process for all the databases(traverse like autovacuum). Worker process\\n>> calculates disk usage of db objects and their diskquota setting. If any\\n>> db object exceeds its quota limit, put them into the blacklist in the\\n>> shared memory, which will later be used by enforcement operator. Worker\\n>> process exits when works are done.\\n>> 5 Running stage, Diskquota launcher process creates diskquota worker\\n>> process for the database with a large number of insert, copy, truncate,\\n>> drop etc. or create disk quota statement. Worker process updates the file\\n>> size for db objects containing the result relation, and compare with the\\n>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\\n>> remove from blacklist vice versa. Worker process exits when works are\\n>> done and a GUC could control the frequency of worker process restart to a\\n>> specific database. As you know, this GUC also controls the delay when we do\\n>> enforcement.\\n>> 6 Enforcement. When postgres backend executes queries, check the\\n>> blacklist in shared memory to determine whether the query is allowed(before\\n>> execute) or need rollback(is executing)?\\n>>\\n>> If we implemented disk quota as an extension, we could just use\\n>> background worker to start diskquota launcher process and use\\n>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\\n>> processes by the launcher process as suggested by @Chapman.\\n>> Diskquota setting could be stored in user table in a separate schema for\\n>> each database(Schema and table created by create extension statement) just\\n>> like what Heikki has done in pg_quota project. But in this case, we need to\\n>> create extension for each database before diskquota worker process can be\\n>> set up for that database.\\n>>\\n>> Any comments on the above design and which is preferred, native feature\\n>> or extension as the POC?\\n>>\\n>>\\n>> -- Hubert\\n>>\\n>>\\n>>\\n>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com>\\n>> wrote:\\n>>\\n>>>\\n>>>\\n>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\\n>>>\\n>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n>>>>\\n>>>> > 2 Keep one worker process for each database. But using a parent/global\\n>>>> > quota worker process to manage the lifecycle of database level worker\\n>>>> > processes. It could handle the newly created database(avoid restart\\n>>>> > database) and save resource when a database is not used. But this\\n>>>> needs to\\n>>>> > change worker process to be hierarchical. Postmaster becomes the\\n>>>> grandfather\\n>>>> > of database level worker processes in this case.\\n>>>>\\n>>>> I am using background workers this way in 9.5 at $work.\\n>>>>\\n>>>> In my case, one worker lives forever, wakes up on a set period, and\\n>>>> starts a short-lived worker for every database, waiting for each\\n>>>> one before starting the next.\\n>>>>\\n>>>> It was straightforward to implement. Looking back over the code,\\n>>>> I see the global worker assigns its own PID to worker.bgw_notify_pid\\n>>>> of each of its children, and also obtains a handle for each child\\n>>>> from RegisterDynamicBackgroundWorker().\\n>>>>\\n>>>> I imagine the global quota worker would prefer to start workers\\n>>>> for every database and then just wait for notifications from any\\n>>>> of them, but that seems equally straightforward at first glance.\\n>>>>\\n>>>\\n>>> There are servers with thousands databases. Worker per database is not\\n>>> good idea.\\n>>>\\n>>> It should to share ideas, code with autovacuum process.\\n>>>\\n>>> Not sure, how to effective implementation based on bg workers can be. On\\n>>> servers with large set of databases, large set of tables it can identify\\n>>> too big table too late.\\n>>>\\n>>> Isn't better to implement some quotas on storage level?\\n>>>\\n>>> Regards\\n>>>\\n>>> Pavel\\n>>>\\n>>>\\n>>>\\n>>>> -Chap\\n>>>>\\n>>>>\\n>>>\\n>>\\n>>\\n>> --\\n>> Thanks\\n>>\\n>> Hubert Zhang\\n>>\\n>\\n>\\n\\n\\n-- \\nThanks\\n\\nHubert Zhang\\n\\nThanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \\\"implement some quotas on storage level?\\\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \\\"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\\\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n\\n> 2 Keep one worker process for each database. But using a parent/global\\n> quota worker process to manage the lifecycle of database level worker\\n> processes. It could handle the newly created database(avoid restart\\n> database) and save resource when a database is not used. But this needs to\\n> change worker process to be hierarchical. Postmaster becomes the grandfather\\n> of database level worker processes in this case.\\n\\nI am using background workers this way in 9.5 at $work.\\n\\nIn my case, one worker lives forever, wakes up on a set period, and\\nstarts a short-lived worker for every database, waiting for each\\none before starting the next.\\n\\nIt was straightforward to implement. Looking back over the code,\\nI see the global worker assigns its own PID to worker.bgw_notify_pid\\nof each of its children, and also obtains a handle for each child\\nfrom RegisterDynamicBackgroundWorker().\\n\\nI imagine the global quota worker would prefer to start workers\\nfor every database and then just wait for notifications from any\\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\\n\\n-Chap\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\",\n \"msg_date\": \"Mon, 3 Sep 2018 09:49:22 +0800\",\n \"msg_from\": \"Hubert Zhang <hzhang@pivotal.io>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal for disk quota feature\"\n },\n {\n \"msg_contents\": \"2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n\\n> Thanks Pavel.\\n> Your patch did enforcement on storage level(md.c or we could also use\\n> smgr_extend). It's straight forward.\\n> But I prefer to implement disk_quota as a feature with following\\n> objectives:\\n> 1 set/alter disk quota setting on different database objects, e.g. user,\\n> database, schema etc. not only a general GUC, but we could set separate\\n> quota limit for a specific objects.\\n> 2 enforcement operator should work at two positions: before query is\\n> running and when query is running. The latter one's implementation maybe\\n> similar to your patch.\\n>\\n\\nThe patch was just example. The resource quotes should be more complex -\\nper partition, table, schema, database, user - so GUC are possible, but not\\nvery user friendly.\\n\\nOur case is specific, but not too much. The servers are used for\\nmultidimensional analyses - and some tables can grow too fast (COPY, INSERT\\nSELECT). We need to solve limits immediately. The implementation is simple,\\nso I did it. Same implementation on database level, or schema level needs\\nsome more locks, so it will not be too effective. The resource management\\ncan be complex very complex, and I expect so it will be hard work.\\n\\nRegards\\n\\nPavel\\n\\n\\n> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com>\\n> wrote:\\n>\\n>> Hi\\n>>\\n>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>\\n>>> Thanks Chapman.\\n>>> @Pavel, could you please explain more about your second suggestion \\\"implement\\n>>> some quotas on storage level?\\\"\\n>>>\\n>>\\n>> See attached patch - it is very simple - and good enough for our\\n>> purposes.\\n>>\\n>> Regards\\n>>\\n>> Pavel\\n>>\\n>>\\n>>\\n>>> We will not keep the long-lived processes attach to all databases(just\\n>>> like you mentioned servers with thousands of databases)\\n>>> And you are right, we could share ideas with autovacuum process, fork\\n>>> worker processes in need.\\n>>> \\\"autovacuum checks for tables that have had a large number of inserted,\\n>>> updated or deleted tuples. These checks use the statistics collection\\n>>> facility\\\"\\n>>> diskquota process is similar to autovacuum at caring about insert, but\\n>>> the difference is that it also care about vucuum full, truncate and drop.\\n>>> While update and delete may not be interested since no file change happens.\\n>>> So a separate diskquota process is preferred.\\n>>>\\n>>> So if we implemented disk quota as a full native feature, and in the\\n>>> first initial version I prefer to implement the following features:\\n>>> 1 Fork diskquota launcher process under Postmaster serverloop, which is\\n>>> long-lived.\\n>>> 2 Diskquota launcher process is responsible for creating diskquota\\n>>> worker process for every database.\\n>>> 3 DIskquota setting is stored in a separate catalog table for each\\n>>> database.\\n>>> 4 Initialization stage, Diskquota launcher process creates diskquota worker\\n>>> process for all the databases(traverse like autovacuum). Worker process\\n>>> calculates disk usage of db objects and their diskquota setting. If any\\n>>> db object exceeds its quota limit, put them into the blacklist in the\\n>>> shared memory, which will later be used by enforcement operator. Worker\\n>>> process exits when works are done.\\n>>> 5 Running stage, Diskquota launcher process creates diskquota worker\\n>>> process for the database with a large number of insert, copy, truncate,\\n>>> drop etc. or create disk quota statement. Worker process updates the file\\n>>> size for db objects containing the result relation, and compare with the\\n>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\\n>>> remove from blacklist vice versa. Worker process exits when works are\\n>>> done and a GUC could control the frequency of worker process restart to a\\n>>> specific database. As you know, this GUC also controls the delay when we do\\n>>> enforcement.\\n>>> 6 Enforcement. When postgres backend executes queries, check the\\n>>> blacklist in shared memory to determine whether the query is allowed(before\\n>>> execute) or need rollback(is executing)?\\n>>>\\n>>> If we implemented disk quota as an extension, we could just use\\n>>> background worker to start diskquota launcher process and use\\n>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\\n>>> processes by the launcher process as suggested by @Chapman.\\n>>> Diskquota setting could be stored in user table in a separate schema for\\n>>> each database(Schema and table created by create extension statement) just\\n>>> like what Heikki has done in pg_quota project. But in this case, we need to\\n>>> create extension for each database before diskquota worker process can be\\n>>> set up for that database.\\n>>>\\n>>> Any comments on the above design and which is preferred, native feature\\n>>> or extension as the POC?\\n>>>\\n>>>\\n>>> -- Hubert\\n>>>\\n>>>\\n>>>\\n>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com>\\n>>> wrote:\\n>>>\\n>>>>\\n>>>>\\n>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\\n>>>>\\n>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n>>>>>\\n>>>>> > 2 Keep one worker process for each database. But using a\\n>>>>> parent/global\\n>>>>> > quota worker process to manage the lifecycle of database level worker\\n>>>>> > processes. It could handle the newly created database(avoid restart\\n>>>>> > database) and save resource when a database is not used. But this\\n>>>>> needs to\\n>>>>> > change worker process to be hierarchical. Postmaster becomes the\\n>>>>> grandfather\\n>>>>> > of database level worker processes in this case.\\n>>>>>\\n>>>>> I am using background workers this way in 9.5 at $work.\\n>>>>>\\n>>>>> In my case, one worker lives forever, wakes up on a set period, and\\n>>>>> starts a short-lived worker for every database, waiting for each\\n>>>>> one before starting the next.\\n>>>>>\\n>>>>> It was straightforward to implement. Looking back over the code,\\n>>>>> I see the global worker assigns its own PID to worker.bgw_notify_pid\\n>>>>> of each of its children, and also obtains a handle for each child\\n>>>>> from RegisterDynamicBackgroundWorker().\\n>>>>>\\n>>>>> I imagine the global quota worker would prefer to start workers\\n>>>>> for every database and then just wait for notifications from any\\n>>>>> of them, but that seems equally straightforward at first glance.\\n>>>>>\\n>>>>\\n>>>> There are servers with thousands databases. Worker per database is not\\n>>>> good idea.\\n>>>>\\n>>>> It should to share ideas, code with autovacuum process.\\n>>>>\\n>>>> Not sure, how to effective implementation based on bg workers can be.\\n>>>> On servers with large set of databases, large set of tables it can identify\\n>>>> too big table too late.\\n>>>>\\n>>>> Isn't better to implement some quotas on storage level?\\n>>>>\\n>>>> Regards\\n>>>>\\n>>>> Pavel\\n>>>>\\n>>>>\\n>>>>\\n>>>>> -Chap\\n>>>>>\\n>>>>>\\n>>>>\\n>>>\\n>>>\\n>>> --\\n>>> Thanks\\n>>>\\n>>> Hubert Zhang\\n>>>\\n>>\\n>>\\n>\\n>\\n> --\\n> Thanks\\n>\\n> Hubert Zhang\\n>\\n\\n2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \\\"implement some quotas on storage level?\\\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \\\"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\\\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n\\n> 2 Keep one worker process for each database. But using a parent/global\\n> quota worker process to manage the lifecycle of database level worker\\n> processes. It could handle the newly created database(avoid restart\\n> database) and save resource when a database is not used. But this needs to\\n> change worker process to be hierarchical. Postmaster becomes the grandfather\\n> of database level worker processes in this case.\\n\\nI am using background workers this way in 9.5 at $work.\\n\\nIn my case, one worker lives forever, wakes up on a set period, and\\nstarts a short-lived worker for every database, waiting for each\\none before starting the next.\\n\\nIt was straightforward to implement. Looking back over the code,\\nI see the global worker assigns its own PID to worker.bgw_notify_pid\\nof each of its children, and also obtains a handle for each child\\nfrom RegisterDynamicBackgroundWorker().\\n\\nI imagine the global quota worker would prefer to start workers\\nfor every database and then just wait for notifications from any\\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\\n\\n-Chap\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\",\n \"msg_date\": \"Mon, 3 Sep 2018 06:05:40 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal for disk quota feature\"\n },\n {\n \"msg_contents\": \"*Hi all,We redesign disk quota feature based on the comments from Pavel\\nStehule and Chapman Flack. Here are the new design.OverviewBasically, disk\\nquota feature is used to support multi-tenancy environment, different level\\nof database objects could be set a quota limit to avoid over use of disk\\nspace. A common case could be as follows: DBA could enable disk quota on a\\nspecified database list. DBA could set disk quota limit for\\ntables/schemas/roles in these databases. Separate disk quota worker process\\nwill monitor the disk usage for these objects and detect the objects which\\nexceed their quota limit. Queries loading data into these “out of disk\\nquota” tables/schemas/roles will be cancelled.We are currently working at\\ninit implementation stage. We would like to propose our idea firstly and\\nget feedbacks from community to do quick iteration.SQL Syntax (How to use\\ndisk quota)1 Specify the databases with disk quota enabled in GUC\\n“diskquota_databases” in postgresql.conf and restart the database.2 DBA\\ncould set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\\nON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\\n(quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\\n‘3MB’);3 Simulate a schema out of quota limit case: suppose table a1 and\\ntable a2 are both under schema s1.INSERT INTO a1 SELECT\\ngenerate_series(1,1000);INSERT INTO a2 SELECT\\ngenerate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\\ngenerate_series(1,1000);ERROR: schema's disk space quota exceededDROP\\nTABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\\ngenerate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\\nfollowing components.1. Quota Setting Store is where the disk quota setting\\nto be stored and accessed. We plan to use catalog table pg_diskquota to\\nstore these information. pg_diskquota is\\nlike:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\\ndiskquota name */ int16 quotatype; /* diskquota type name */ Oid\\nquotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\\ndiskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\\nMB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\\nchange of database objects. We plan to use stat collector to detect the\\n‘active’ table list at initial stage. But stat collector has some\\nlimitation on finding the active table which is in a running transaction.\\nDetails see TODO section.3. Quota Size Checker is where to calculate the\\nsize and compare with quota limit for database objects. According to\\nPavel’s comment, autovacuum launcher and worker process could be a good\\nreference to disk quota. So we plan to use a disk quota launcher daemon\\nprocess and several disk quota worker process to finish this work. Launcher\\nprocess is responsible for starting worker process based on a user defined\\ndatabase list from GUC. Worker process will connect to its target database\\nand monitor the disk usage for objects in this database. In init stage of\\nworker process, it will call calculate_total_relation_size() to calculate\\nthe size for each user table. After init stage, worker process will refresh\\nthe disk model every N seconds. Refreshing will only recalculate the size\\nof tables in ‘active’ table list, which is generated by Quata Change\\nDetector to minimize the cost.4. Quota Enforcement Operator is where to\\ncheck for the quota limitation at postgres backend side. We will firstly\\nimplement it in ExecCheckRTPerms() as pre-running enforcement. It will\\ncheck the disk quota of tables being inserted or updated, and report error\\nif table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\\nAs a native feature, we plan to add more checkpoint to do running query\\nenforcement. For example, if a disk quota lefts 10MB quota, a query could\\ninsert 1GB data. This query could be allowed in pre-running enforcement\\ncheck, but will be cancelled in running query enforcement check. Therefore,\\nit can improve the accurate of disk quota usage. To achieve this, we plan\\nto add a checkpoint in lower API such as smgr_extened. Hence, the Quota\\nEnforcement Operator will check the disk quota usage when smgr_extened is\\ncalled. If the quota is over limited, current query will be cancelled.\\nHighlight1. Native feature.Support native Create/Drop Disk Quota SQL\\nstatement.New catalog table pg_diskquota to store disk quota setting.2.\\nAuto DML/DDL detection. Table\\ncreate/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\\n Schema create/drop and Role create/drop will be detected by disk quota\\nautomatically. 3. Low cost disk quota checker.Worker process of disk quota\\nneed to refresh the disk usage model every N seconds. Since recalculate the\\nfile size using stat() system call is expensive for a large number of\\nfiles, we use an ‘active’ table list to reduce the real work at each\\niteration. A basic experiment on our init stage implementation on database\\nwith 20K tables shows that the refresh cost is 1% cpu usage and will be\\nfinished within 50ms. Todo/LimitationBefore we propose our patch, we plan\\nto enhance it with the following ideas:1. Setting database list with disk\\nquota enabled dynamically without restart database. Since we have the disk\\nquota launcher process, it could detect the new ‘diskquota_databases’ list\\nand start/stop the corresponding disk quota worker process.2. Enforcement\\nwhen query is running. Considering the case when there is 10MB quota left,\\nbut next query will insert 10GB data. Current enforcement design will allow\\nthis query to be executed. This is limited by the ‘active’ table detection\\nis generated by stat collector. Postgres backend will only send table stat\\ninformation to collector only when the transaction ends. We need a new way\\nto detect the ‘active’ table even when this table is being modified inside\\na running transaction.3. Monitor unlimited number of databases. Current we\\nset the max number of disk quota worker process to be 10 to reduce the\\naffection normal workload. But how about if we want to monitor the disk\\nquota of more than 10 databases? Our solution is to let disk quota launcher\\nto manage a queue of database need to be monitored. And disk quota worker\\nprocess consuming the queue and refresh the disk usage/quota for this\\ndatabase. After some periods, worker will return the database to the queue,\\nand fetch the top database from queue to process. The period determine the\\ndelay of detecting disk quota change. To implement this feature, we need to\\nsupport a subprocess of postmaster to rebind to another database instead of\\nthe database binded in InitPostgres().4. Support active table detection on\\nvacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\\nare not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\\nFlack for the former comments on disk quota feature. Any comments on how to\\nimprove disk quota feature are appreciated.*\\n\\n\\nOn Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com>\\nwrote:\\n\\n>\\n>\\n> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>\\n>> Thanks Pavel.\\n>> Your patch did enforcement on storage level(md.c or we could also use\\n>> smgr_extend). It's straight forward.\\n>> But I prefer to implement disk_quota as a feature with following\\n>> objectives:\\n>> 1 set/alter disk quota setting on different database objects, e.g. user,\\n>> database, schema etc. not only a general GUC, but we could set separate\\n>> quota limit for a specific objects.\\n>> 2 enforcement operator should work at two positions: before query is\\n>> running and when query is running. The latter one's implementation maybe\\n>> similar to your patch.\\n>>\\n>\\n> The patch was just example. The resource quotes should be more complex -\\n> per partition, table, schema, database, user - so GUC are possible, but not\\n> very user friendly.\\n>\\n> Our case is specific, but not too much. The servers are used for\\n> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\\n> SELECT). We need to solve limits immediately. The implementation is simple,\\n> so I did it. Same implementation on database level, or schema level needs\\n> some more locks, so it will not be too effective. The resource management\\n> can be complex very complex, and I expect so it will be hard work.\\n>\\n> Regards\\n>\\n> Pavel\\n>\\n>\\n>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com>\\n>> wrote:\\n>>\\n>>> Hi\\n>>>\\n>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>>\\n>>>> Thanks Chapman.\\n>>>> @Pavel, could you please explain more about your second suggestion \\\"implement\\n>>>> some quotas on storage level?\\\"\\n>>>>\\n>>>\\n>>> See attached patch - it is very simple - and good enough for our\\n>>> purposes.\\n>>>\\n>>> Regards\\n>>>\\n>>> Pavel\\n>>>\\n>>>\\n>>>\\n>>>> We will not keep the long-lived processes attach to all databases(just\\n>>>> like you mentioned servers with thousands of databases)\\n>>>> And you are right, we could share ideas with autovacuum process, fork\\n>>>> worker processes in need.\\n>>>> \\\"autovacuum checks for tables that have had a large number of\\n>>>> inserted, updated or deleted tuples. These checks use the statistics\\n>>>> collection facility\\\"\\n>>>> diskquota process is similar to autovacuum at caring about insert, but\\n>>>> the difference is that it also care about vucuum full, truncate and drop.\\n>>>> While update and delete may not be interested since no file change happens.\\n>>>> So a separate diskquota process is preferred.\\n>>>>\\n>>>> So if we implemented disk quota as a full native feature, and in the\\n>>>> first initial version I prefer to implement the following features:\\n>>>> 1 Fork diskquota launcher process under Postmaster serverloop, which is\\n>>>> long-lived.\\n>>>> 2 Diskquota launcher process is responsible for creating diskquota\\n>>>> worker process for every database.\\n>>>> 3 DIskquota setting is stored in a separate catalog table for each\\n>>>> database.\\n>>>> 4 Initialization stage, Diskquota launcher process creates diskquota worker\\n>>>> process for all the databases(traverse like autovacuum). Worker process\\n>>>> calculates disk usage of db objects and their diskquota setting. If any\\n>>>> db object exceeds its quota limit, put them into the blacklist in the\\n>>>> shared memory, which will later be used by enforcement operator. Worker\\n>>>> process exits when works are done.\\n>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\\n>>>> process for the database with a large number of insert, copy, truncate,\\n>>>> drop etc. or create disk quota statement. Worker process updates the file\\n>>>> size for db objects containing the result relation, and compare with the\\n>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\\n>>>> remove from blacklist vice versa. Worker process exits when works are\\n>>>> done and a GUC could control the frequency of worker process restart to a\\n>>>> specific database. As you know, this GUC also controls the delay when we do\\n>>>> enforcement.\\n>>>> 6 Enforcement. When postgres backend executes queries, check the\\n>>>> blacklist in shared memory to determine whether the query is allowed(before\\n>>>> execute) or need rollback(is executing)?\\n>>>>\\n>>>> If we implemented disk quota as an extension, we could just use\\n>>>> background worker to start diskquota launcher process and use\\n>>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\\n>>>> processes by the launcher process as suggested by @Chapman.\\n>>>> Diskquota setting could be stored in user table in a separate schema for\\n>>>> each database(Schema and table created by create extension statement) just\\n>>>> like what Heikki has done in pg_quota project. But in this case, we need to\\n>>>> create extension for each database before diskquota worker process can be\\n>>>> set up for that database.\\n>>>>\\n>>>> Any comments on the above design and which is preferred, native feature\\n>>>> or extension as the POC?\\n>>>>\\n>>>>\\n>>>> -- Hubert\\n>>>>\\n>>>>\\n>>>>\\n>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com\\n>>>> > wrote:\\n>>>>\\n>>>>>\\n>>>>>\\n>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\\n>>>>>\\n>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n>>>>>>\\n>>>>>> > 2 Keep one worker process for each database. But using a\\n>>>>>> parent/global\\n>>>>>> > quota worker process to manage the lifecycle of database level\\n>>>>>> worker\\n>>>>>> > processes. It could handle the newly created database(avoid restart\\n>>>>>> > database) and save resource when a database is not used. But this\\n>>>>>> needs to\\n>>>>>> > change worker process to be hierarchical. Postmaster becomes the\\n>>>>>> grandfather\\n>>>>>> > of database level worker processes in this case.\\n>>>>>>\\n>>>>>> I am using background workers this way in 9.5 at $work.\\n>>>>>>\\n>>>>>> In my case, one worker lives forever, wakes up on a set period, and\\n>>>>>> starts a short-lived worker for every database, waiting for each\\n>>>>>> one before starting the next.\\n>>>>>>\\n>>>>>> It was straightforward to implement. Looking back over the code,\\n>>>>>> I see the global worker assigns its own PID to worker.bgw_notify_pid\\n>>>>>> of each of its children, and also obtains a handle for each child\\n>>>>>> from RegisterDynamicBackgroundWorker().\\n>>>>>>\\n>>>>>> I imagine the global quota worker would prefer to start workers\\n>>>>>> for every database and then just wait for notifications from any\\n>>>>>> of them, but that seems equally straightforward at first glance.\\n>>>>>>\\n>>>>>\\n>>>>> There are servers with thousands databases. Worker per database is not\\n>>>>> good idea.\\n>>>>>\\n>>>>> It should to share ideas, code with autovacuum process.\\n>>>>>\\n>>>>> Not sure, how to effective implementation based on bg workers can be.\\n>>>>> On servers with large set of databases, large set of tables it can identify\\n>>>>> too big table too late.\\n>>>>>\\n>>>>> Isn't better to implement some quotas on storage level?\\n>>>>>\\n>>>>> Regards\\n>>>>>\\n>>>>> Pavel\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>> -Chap\\n>>>>>>\\n>>>>>>\\n>>>>>\\n>>>>\\n>>>>\\n>>>> --\\n>>>> Thanks\\n>>>>\\n>>>> Hubert Zhang\\n>>>>\\n>>>\\n>>>\\n>>\\n>>\\n>> --\\n>> Thanks\\n>>\\n>> Hubert Zhang\\n>>\\n>\\n>\\n\\n\\n-- \\nThanks\\n\\nHubert Zhang\\n\\nHi all,We redesign disk quota feature based on the comments from Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk quota feature is used to support multi-tenancy environment, different level of database objects could be set a quota limit to avoid over use of disk space. A common case could be as follows: DBA could enable disk quota on a specified database list. DBA could set disk quota limit for tables/schemas/roles in these databases. Separate disk quota worker process will monitor the disk usage for these objects and detect the objects which exceed their quota limit. Queries loading data into these “out of disk quota” tables/schemas/roles will be cancelled.We are currently working at init implementation stage. We would like to propose our idea firstly and get feedbacks from community to do quick iteration.SQL Syntax (How to use disk quota)1 Specify the databases with disk quota enabled in GUC “diskquota_databases” in postgresql.conf and restart the database.2 DBA could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);3 Simulate a schema out of quota limit case: suppose table a1 and table a2 are both under schema s1.INSERT INTO a1 SELECT generate_series(1,1000);INSERT INTO a2 SELECT generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);ERROR: schema's disk space quota exceededDROP TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the following components.1. Quota Setting Store is where the disk quota setting to be stored and accessed. We plan to use catalog table pg_diskquota to store these information. pg_diskquota is like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /* diskquota name */ int16 quotatype; /* diskquota type name */ Oid quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /* diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size change of database objects. We plan to use stat collector to detect the ‘active’ table list at initial stage. But stat collector has some limitation on finding the active table which is in a running transaction. Details see TODO section.3. Quota Size Checker is where to calculate the size and compare with quota limit for database objects. According to Pavel’s comment, autovacuum launcher and worker process could be a good reference to disk quota. So we plan to use a disk quota launcher daemon process and several disk quota worker process to finish this work. Launcher process is responsible for starting worker process based on a user defined database list from GUC. Worker process will connect to its target database and monitor the disk usage for objects in this database. In init stage of worker process, it will call calculate_total_relation_size() to calculate the size for each user table. After init stage, worker process will refresh the disk model every N seconds. Refreshing will only recalculate the size of tables in ‘active’ table list, which is generated by Quata Change Detector to minimize the cost.4. Quota Enforcement Operator is where to check for the quota limitation at postgres backend side. We will firstly implement it in ExecCheckRTPerms() as pre-running enforcement. It will check the disk quota of tables being inserted or updated, and report error if table’s or table’s schema’s or table’s owner’s quota limit is exceeded. As a native feature, we plan to add more checkpoint to do running query enforcement. For example, if a disk quota lefts 10MB quota, a query could insert 1GB data. This query could be allowed in pre-running enforcement check, but will be cancelled in running query enforcement check. Therefore, it can improve the accurate of disk quota usage. To achieve this, we plan to add a checkpoint in lower API such as smgr_extened. Hence, the Quota Enforcement Operator will check the disk quota usage when smgr_extened is called. If the quota is over limited, current query will be cancelled. Highlight1. Native feature.Support native Create/Drop Disk Quota SQL statement.New catalog table pg_diskquota to store disk quota setting.2. Auto DML/DDL detection. Table create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change, Schema create/drop and Role create/drop will be detected by disk quota automatically. 3. Low cost disk quota checker.Worker process of disk quota need to refresh the disk usage model every N seconds. Since recalculate the file size using stat() system call is expensive for a large number of files, we use an ‘active’ table list to reduce the real work at each iteration. A basic experiment on our init stage implementation on database with 20K tables shows that the refresh cost is 1% cpu usage and will be finished within 50ms. Todo/LimitationBefore we propose our patch, we plan to enhance it with the following ideas:1. Setting database list with disk quota enabled dynamically without restart database. Since we have the disk quota launcher process, it could detect the new ‘diskquota_databases’ list and start/stop the corresponding disk quota worker process.2. Enforcement when query is running. Considering the case when there is 10MB quota left, but next query will insert 10GB data. Current enforcement design will allow this query to be executed. This is limited by the ‘active’ table detection is generated by stat collector. Postgres backend will only send table stat information to collector only when the transaction ends. We need a new way to detect the ‘active’ table even when this table is being modified inside a running transaction.3. Monitor unlimited number of databases. Current we set the max number of disk quota worker process to be 10 to reduce the affection normal workload. But how about if we want to monitor the disk quota of more than 10 databases? Our solution is to let disk quota launcher to manage a queue of database need to be monitored. And disk quota worker process consuming the queue and refresh the disk usage/quota for this database. After some periods, worker will return the database to the queue, and fetch the top database from queue to process. The period determine the delay of detecting disk quota change. To implement this feature, we need to support a subprocess of postmaster to rebind to another database instead of the database binded in InitPostgres().4. Support active table detection on vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman Flack for the former comments on disk quota feature. Any comments on how to improve disk quota feature are appreciated.On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \\\"implement some quotas on storage level?\\\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \\\"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\\\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n\\n> 2 Keep one worker process for each database. But using a parent/global\\n> quota worker process to manage the lifecycle of database level worker\\n> processes. It could handle the newly created database(avoid restart\\n> database) and save resource when a database is not used. But this needs to\\n> change worker process to be hierarchical. Postmaster becomes the grandfather\\n> of database level worker processes in this case.\\n\\nI am using background workers this way in 9.5 at $work.\\n\\nIn my case, one worker lives forever, wakes up on a set period, and\\nstarts a short-lived worker for every database, waiting for each\\none before starting the next.\\n\\nIt was straightforward to implement. Looking back over the code,\\nI see the global worker assigns its own PID to worker.bgw_notify_pid\\nof each of its children, and also obtains a handle for each child\\nfrom RegisterDynamicBackgroundWorker().\\n\\nI imagine the global quota worker would prefer to start workers\\nfor every database and then just wait for notifications from any\\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\\n\\n-Chap\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\",\n \"msg_date\": \"Fri, 21 Sep 2018 19:32:16 +0800\",\n \"msg_from\": \"Hubert Zhang <hzhang@pivotal.io>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal for disk quota feature\"\n },\n {\n \"msg_contents\": \"pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:\\n\\n>\\n>\\n>\\n>\\n> *Hi all,We redesign disk quota feature based on the comments from Pavel\\n> Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk\\n> quota feature is used to support multi-tenancy environment, different level\\n> of database objects could be set a quota limit to avoid over use of disk\\n> space. A common case could be as follows: DBA could enable disk quota on a\\n> specified database list. DBA could set disk quota limit for\\n> tables/schemas/roles in these databases. Separate disk quota worker process\\n> will monitor the disk usage for these objects and detect the objects which\\n> exceed their quota limit. Queries loading data into these “out of disk\\n> quota” tables/schemas/roles will be cancelled.We are currently working at\\n> init implementation stage. We would like to propose our idea firstly and\\n> get feedbacks from community to do quick iteration.SQL Syntax (How to use\\n> disk quota)1 Specify the databases with disk quota enabled in GUC\\n> “diskquota_databases” in postgresql.conf and restart the database.2 DBA\\n> could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\\n> ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\\n> (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\\n> ‘3MB’);*\\n>\\n\\njust fast reaction - why QUOTA object?\\n\\nIsn't ALTER SET enough?\\n\\nSome like\\n\\nALTER TABLE a1 SET quote = 1MB;\\nALTER USER ...\\nALTER SCHEMA ..\\n\\nNew DDL commans looks like too hard hammer .\\n\\n\\n\\n>\\n>\\n>\\n>\\n>\\n>\\n>\\n>\\n>\\n>\\n>\\n>\\n>\\n>\\n>\\n>\\n>\\n> *3 Simulate a schema out of quota limit case: suppose table a1 and table\\n> a2 are both under schema s1.INSERT INTO a1 SELECT\\n> generate_series(1,1000);INSERT INTO a2 SELECT\\n> generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\\n> generate_series(1,1000);ERROR: schema's disk space quota exceededDROP\\n> TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\\n> generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\\n> following components.1. Quota Setting Store is where the disk quota setting\\n> to be stored and accessed. We plan to use catalog table pg_diskquota to\\n> store these information. pg_diskquota is\\n> like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\\n> diskquota name */ int16 quotatype; /* diskquota type name */ Oid\\n> quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\\n> diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\\n> MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\\n> change of database objects. We plan to use stat collector to detect the\\n> ‘active’ table list at initial stage. But stat collector has some\\n> limitation on finding the active table which is in a running transaction.\\n> Details see TODO section.3. Quota Size Checker is where to calculate the\\n> size and compare with quota limit for database objects. According to\\n> Pavel’s comment, autovacuum launcher and worker process could be a good\\n> reference to disk quota. So we plan to use a disk quota launcher daemon\\n> process and several disk quota worker process to finish this work. Launcher\\n> process is responsible for starting worker process based on a user defined\\n> database list from GUC. Worker process will connect to its target database\\n> and monitor the disk usage for objects in this database. In init stage of\\n> worker process, it will call calculate_total_relation_size() to calculate\\n> the size for each user table. After init stage, worker process will refresh\\n> the disk model every N seconds. Refreshing will only recalculate the size\\n> of tables in ‘active’ table list, which is generated by Quata Change\\n> Detector to minimize the cost.4. Quota Enforcement Operator is where to\\n> check for the quota limitation at postgres backend side. We will firstly\\n> implement it in ExecCheckRTPerms() as pre-running enforcement. It will\\n> check the disk quota of tables being inserted or updated, and report error\\n> if table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\\n> As a native feature, we plan to add more checkpoint to do running query\\n> enforcement. For example, if a disk quota lefts 10MB quota, a query could\\n> insert 1GB data. This query could be allowed in pre-running enforcement\\n> check, but will be cancelled in running query enforcement check. Therefore,\\n> it can improve the accurate of disk quota usage. To achieve this, we plan\\n> to add a checkpoint in lower API such as smgr_extened. Hence, the Quota\\n> Enforcement Operator will check the disk quota usage when smgr_extened is\\n> called. If the quota is over limited, current query will be cancelled.\\n> Highlight1. Native feature.Support native Create/Drop Disk Quota SQL\\n> statement.New catalog table pg_diskquota to store disk quota setting.2.\\n> Auto DML/DDL detection. Table\\n> create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\\n> Schema create/drop and Role create/drop will be detected by disk quota\\n> automatically. 3. Low cost disk quota checker.Worker process of disk quota\\n> need to refresh the disk usage model every N seconds. Since recalculate the\\n> file size using stat() system call is expensive for a large number of\\n> files, we use an ‘active’ table list to reduce the real work at each\\n> iteration. A basic experiment on our init stage implementation on database\\n> with 20K tables shows that the refresh cost is 1% cpu usage and will be\\n> finished within 50ms. Todo/LimitationBefore we propose our patch, we plan\\n> to enhance it with the following ideas:1. Setting database list with disk\\n> quota enabled dynamically without restart database. Since we have the disk\\n> quota launcher process, it could detect the new ‘diskquota_databases’ list\\n> and start/stop the corresponding disk quota worker process.2. Enforcement\\n> when query is running. Considering the case when there is 10MB quota left,\\n> but next query will insert 10GB data. Current enforcement design will allow\\n> this query to be executed. This is limited by the ‘active’ table detection\\n> is generated by stat collector. Postgres backend will only send table stat\\n> information to collector only when the transaction ends. We need a new way\\n> to detect the ‘active’ table even when this table is being modified inside\\n> a running transaction.3. Monitor unlimited number of databases. Current we\\n> set the max number of disk quota worker process to be 10 to reduce the\\n> affection normal workload. But how about if we want to monitor the disk\\n> quota of more than 10 databases? Our solution is to let disk quota launcher\\n> to manage a queue of database need to be monitored. And disk quota worker\\n> process consuming the queue and refresh the disk usage/quota for this\\n> database. After some periods, worker will return the database to the queue,\\n> and fetch the top database from queue to process. The period determine the\\n> delay of detecting disk quota change. To implement this feature, we need to\\n> support a subprocess of postmaster to rebind to another database instead of\\n> the database binded in InitPostgres().4. Support active table detection on\\n> vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\\n> are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\\n> Flack for the former comments on disk quota feature. Any comments on how to\\n> improve disk quota feature are appreciated.*\\n>\\n>\\n> On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com>\\n> wrote:\\n>\\n>>\\n>>\\n>> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>\\n>>> Thanks Pavel.\\n>>> Your patch did enforcement on storage level(md.c or we could also use\\n>>> smgr_extend). It's straight forward.\\n>>> But I prefer to implement disk_quota as a feature with following\\n>>> objectives:\\n>>> 1 set/alter disk quota setting on different database objects, e.g. user,\\n>>> database, schema etc. not only a general GUC, but we could set separate\\n>>> quota limit for a specific objects.\\n>>> 2 enforcement operator should work at two positions: before query is\\n>>> running and when query is running. The latter one's implementation maybe\\n>>> similar to your patch.\\n>>>\\n>>\\n>> The patch was just example. The resource quotes should be more complex -\\n>> per partition, table, schema, database, user - so GUC are possible, but not\\n>> very user friendly.\\n>>\\n>> Our case is specific, but not too much. The servers are used for\\n>> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\\n>> SELECT). We need to solve limits immediately. The implementation is simple,\\n>> so I did it. Same implementation on database level, or schema level needs\\n>> some more locks, so it will not be too effective. The resource management\\n>> can be complex very complex, and I expect so it will be hard work.\\n>>\\n>> Regards\\n>>\\n>> Pavel\\n>>\\n>>\\n>>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com>\\n>>> wrote:\\n>>>\\n>>>> Hi\\n>>>>\\n>>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>>>\\n>>>>> Thanks Chapman.\\n>>>>> @Pavel, could you please explain more about your second suggestion \\\"implement\\n>>>>> some quotas on storage level?\\\"\\n>>>>>\\n>>>>\\n>>>> See attached patch - it is very simple - and good enough for our\\n>>>> purposes.\\n>>>>\\n>>>> Regards\\n>>>>\\n>>>> Pavel\\n>>>>\\n>>>>\\n>>>>\\n>>>>> We will not keep the long-lived processes attach to all databases(just\\n>>>>> like you mentioned servers with thousands of databases)\\n>>>>> And you are right, we could share ideas with autovacuum process, fork\\n>>>>> worker processes in need.\\n>>>>> \\\"autovacuum checks for tables that have had a large number of\\n>>>>> inserted, updated or deleted tuples. These checks use the statistics\\n>>>>> collection facility\\\"\\n>>>>> diskquota process is similar to autovacuum at caring about insert, but\\n>>>>> the difference is that it also care about vucuum full, truncate and drop.\\n>>>>> While update and delete may not be interested since no file change happens.\\n>>>>> So a separate diskquota process is preferred.\\n>>>>>\\n>>>>> So if we implemented disk quota as a full native feature, and in the\\n>>>>> first initial version I prefer to implement the following features:\\n>>>>> 1 Fork diskquota launcher process under Postmaster serverloop, which\\n>>>>> is long-lived.\\n>>>>> 2 Diskquota launcher process is responsible for creating diskquota\\n>>>>> worker process for every database.\\n>>>>> 3 DIskquota setting is stored in a separate catalog table for each\\n>>>>> database.\\n>>>>> 4 Initialization stage, Diskquota launcher process creates diskquota worker\\n>>>>> process for all the databases(traverse like autovacuum). Worker process\\n>>>>> calculates disk usage of db objects and their diskquota setting. If any\\n>>>>> db object exceeds its quota limit, put them into the blacklist in the\\n>>>>> shared memory, which will later be used by enforcement operator. Worker\\n>>>>> process exits when works are done.\\n>>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\\n>>>>> process for the database with a large number of insert, copy, truncate,\\n>>>>> drop etc. or create disk quota statement. Worker process updates the file\\n>>>>> size for db objects containing the result relation, and compare with the\\n>>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\\n>>>>> remove from blacklist vice versa. Worker process exits when works are\\n>>>>> done and a GUC could control the frequency of worker process restart to a\\n>>>>> specific database. As you know, this GUC also controls the delay when we do\\n>>>>> enforcement.\\n>>>>> 6 Enforcement. When postgres backend executes queries, check the\\n>>>>> blacklist in shared memory to determine whether the query is allowed(before\\n>>>>> execute) or need rollback(is executing)?\\n>>>>>\\n>>>>> If we implemented disk quota as an extension, we could just use\\n>>>>> background worker to start diskquota launcher process and use\\n>>>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\\n>>>>> processes by the launcher process as suggested by @Chapman.\\n>>>>> Diskquota setting could be stored in user table in a separate schema for\\n>>>>> each database(Schema and table created by create extension statement) just\\n>>>>> like what Heikki has done in pg_quota project. But in this case, we need to\\n>>>>> create extension for each database before diskquota worker process can be\\n>>>>> set up for that database.\\n>>>>>\\n>>>>> Any comments on the above design and which is preferred, native\\n>>>>> feature or extension as the POC?\\n>>>>>\\n>>>>>\\n>>>>> -- Hubert\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <\\n>>>>> pavel.stehule@gmail.com> wrote:\\n>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\\n>>>>>>\\n>>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n>>>>>>>\\n>>>>>>> > 2 Keep one worker process for each database. But using a\\n>>>>>>> parent/global\\n>>>>>>> > quota worker process to manage the lifecycle of database level\\n>>>>>>> worker\\n>>>>>>> > processes. It could handle the newly created database(avoid restart\\n>>>>>>> > database) and save resource when a database is not used. But this\\n>>>>>>> needs to\\n>>>>>>> > change worker process to be hierarchical. Postmaster becomes the\\n>>>>>>> grandfather\\n>>>>>>> > of database level worker processes in this case.\\n>>>>>>>\\n>>>>>>> I am using background workers this way in 9.5 at $work.\\n>>>>>>>\\n>>>>>>> In my case, one worker lives forever, wakes up on a set period, and\\n>>>>>>> starts a short-lived worker for every database, waiting for each\\n>>>>>>> one before starting the next.\\n>>>>>>>\\n>>>>>>> It was straightforward to implement. Looking back over the code,\\n>>>>>>> I see the global worker assigns its own PID to worker.bgw_notify_pid\\n>>>>>>> of each of its children, and also obtains a handle for each child\\n>>>>>>> from RegisterDynamicBackgroundWorker().\\n>>>>>>>\\n>>>>>>> I imagine the global quota worker would prefer to start workers\\n>>>>>>> for every database and then just wait for notifications from any\\n>>>>>>> of them, but that seems equally straightforward at first glance.\\n>>>>>>>\\n>>>>>>\\n>>>>>> There are servers with thousands databases. Worker per database is\\n>>>>>> not good idea.\\n>>>>>>\\n>>>>>> It should to share ideas, code with autovacuum process.\\n>>>>>>\\n>>>>>> Not sure, how to effective implementation based on bg workers can be.\\n>>>>>> On servers with large set of databases, large set of tables it can identify\\n>>>>>> too big table too late.\\n>>>>>>\\n>>>>>> Isn't better to implement some quotas on storage level?\\n>>>>>>\\n>>>>>> Regards\\n>>>>>>\\n>>>>>> Pavel\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>> -Chap\\n>>>>>>>\\n>>>>>>>\\n>>>>>>\\n>>>>>\\n>>>>>\\n>>>>> --\\n>>>>> Thanks\\n>>>>>\\n>>>>> Hubert Zhang\\n>>>>>\\n>>>>\\n>>>>\\n>>>\\n>>>\\n>>> --\\n>>> Thanks\\n>>>\\n>>> Hubert Zhang\\n>>>\\n>>\\n>>\\n>\\n>\\n> --\\n> Thanks\\n>\\n> Hubert Zhang\\n>\\n\\npá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:Hi all,We redesign disk quota feature based on the comments from Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk quota feature is used to support multi-tenancy environment, different level of database objects could be set a quota limit to avoid over use of disk space. A common case could be as follows: DBA could enable disk quota on a specified database list. DBA could set disk quota limit for tables/schemas/roles in these databases. Separate disk quota worker process will monitor the disk usage for these objects and detect the objects which exceed their quota limit. Queries loading data into these “out of disk quota” tables/schemas/roles will be cancelled.We are currently working at init implementation stage. We would like to propose our idea firstly and get feedbacks from community to do quick iteration.SQL Syntax (How to use disk quota)1 Specify the databases with disk quota enabled in GUC “diskquota_databases” in postgresql.conf and restart the database.2 DBA could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer . 3 Simulate a schema out of quota limit case: suppose table a1 and table a2 are both under schema s1.INSERT INTO a1 SELECT generate_series(1,1000);INSERT INTO a2 SELECT generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);ERROR: schema's disk space quota exceededDROP TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the following components.1. Quota Setting Store is where the disk quota setting to be stored and accessed. We plan to use catalog table pg_diskquota to store these information. pg_diskquota is like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /* diskquota name */ int16 quotatype; /* diskquota type name */ Oid quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /* diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size change of database objects. We plan to use stat collector to detect the ‘active’ table list at initial stage. But stat collector has some limitation on finding the active table which is in a running transaction. Details see TODO section.3. Quota Size Checker is where to calculate the size and compare with quota limit for database objects. According to Pavel’s comment, autovacuum launcher and worker process could be a good reference to disk quota. So we plan to use a disk quota launcher daemon process and several disk quota worker process to finish this work. Launcher process is responsible for starting worker process based on a user defined database list from GUC. Worker process will connect to its target database and monitor the disk usage for objects in this database. In init stage of worker process, it will call calculate_total_relation_size() to calculate the size for each user table. After init stage, worker process will refresh the disk model every N seconds. Refreshing will only recalculate the size of tables in ‘active’ table list, which is generated by Quata Change Detector to minimize the cost.4. Quota Enforcement Operator is where to check for the quota limitation at postgres backend side. We will firstly implement it in ExecCheckRTPerms() as pre-running enforcement. It will check the disk quota of tables being inserted or updated, and report error if table’s or table’s schema’s or table’s owner’s quota limit is exceeded. As a native feature, we plan to add more checkpoint to do running query enforcement. For example, if a disk quota lefts 10MB quota, a query could insert 1GB data. This query could be allowed in pre-running enforcement check, but will be cancelled in running query enforcement check. Therefore, it can improve the accurate of disk quota usage. To achieve this, we plan to add a checkpoint in lower API such as smgr_extened. Hence, the Quota Enforcement Operator will check the disk quota usage when smgr_extened is called. If the quota is over limited, current query will be cancelled. Highlight1. Native feature.Support native Create/Drop Disk Quota SQL statement.New catalog table pg_diskquota to store disk quota setting.2. Auto DML/DDL detection. Table create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change, Schema create/drop and Role create/drop will be detected by disk quota automatically. 3. Low cost disk quota checker.Worker process of disk quota need to refresh the disk usage model every N seconds. Since recalculate the file size using stat() system call is expensive for a large number of files, we use an ‘active’ table list to reduce the real work at each iteration. A basic experiment on our init stage implementation on database with 20K tables shows that the refresh cost is 1% cpu usage and will be finished within 50ms. Todo/LimitationBefore we propose our patch, we plan to enhance it with the following ideas:1. Setting database list with disk quota enabled dynamically without restart database. Since we have the disk quota launcher process, it could detect the new ‘diskquota_databases’ list and start/stop the corresponding disk quota worker process.2. Enforcement when query is running. Considering the case when there is 10MB quota left, but next query will insert 10GB data. Current enforcement design will allow this query to be executed. This is limited by the ‘active’ table detection is generated by stat collector. Postgres backend will only send table stat information to collector only when the transaction ends. We need a new way to detect the ‘active’ table even when this table is being modified inside a running transaction.3. Monitor unlimited number of databases. Current we set the max number of disk quota worker process to be 10 to reduce the affection normal workload. But how about if we want to monitor the disk quota of more than 10 databases? Our solution is to let disk quota launcher to manage a queue of database need to be monitored. And disk quota worker process consuming the queue and refresh the disk usage/quota for this database. After some periods, worker will return the database to the queue, and fetch the top database from queue to process. The period determine the delay of detecting disk quota change. To implement this feature, we need to support a subprocess of postmaster to rebind to another database instead of the database binded in InitPostgres().4. Support active table detection on vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman Flack for the former comments on disk quota feature. Any comments on how to improve disk quota feature are appreciated.On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \\\"implement some quotas on storage level?\\\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \\\"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\\\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n\\n> 2 Keep one worker process for each database. But using a parent/global\\n> quota worker process to manage the lifecycle of database level worker\\n> processes. It could handle the newly created database(avoid restart\\n> database) and save resource when a database is not used. But this needs to\\n> change worker process to be hierarchical. Postmaster becomes the grandfather\\n> of database level worker processes in this case.\\n\\nI am using background workers this way in 9.5 at $work.\\n\\nIn my case, one worker lives forever, wakes up on a set period, and\\nstarts a short-lived worker for every database, waiting for each\\none before starting the next.\\n\\nIt was straightforward to implement. Looking back over the code,\\nI see the global worker assigns its own PID to worker.bgw_notify_pid\\nof each of its children, and also obtains a handle for each child\\nfrom RegisterDynamicBackgroundWorker().\\n\\nI imagine the global quota worker would prefer to start workers\\nfor every database and then just wait for notifications from any\\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\\n\\n-Chap\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\",\n \"msg_date\": \"Fri, 21 Sep 2018 14:01:15 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal for disk quota feature\"\n },\n {\n \"msg_contents\": \">\\n> just fast reaction - why QUOTA object?\\n> Isn't ALTER SET enough?\\n> Some like\\n> ALTER TABLE a1 SET quote = 1MB;\\n> ALTER USER ...\\n> ALTER SCHEMA ..\\n> New DDL commans looks like too hard hammer .\\n\\n\\nIt's an option. Prefer to consider quota setting store together:\\nCREATE DISK QUOTA way is more nature to store quota setting in a separate\\npg_diskquota catalog\\nWhile ALTER SET way is more close to store quota setting in pg_class,\\npg_role, pg_namespace. etc in an integrated way.\\n(Note that here I mean nature/close is not must, ALTER SET could also store\\nin pg_diskquota and vice versa.)\\n\\nHere are some differences I can think of:\\n1 pg_role is a global catalog, not per database level. It's harder to\\ntracker the user's disk usage in the whole clusters(considering 1000+\\ndatabases). So the semantic of CREATE DISK QUOTA ON USER is limited: it\\nonly tracks the user's disk usage inside the current database.\\n2 using separate pg_diskquota could add more field except for quota limit\\nwithout adding too many fields in pg_class, e.g. red zone to give the user\\na warning or the current disk usage of the db objects.\\n\\nOn Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com>\\nwrote:\\n\\n>\\n>\\n> pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:\\n>\\n>>\\n>>\\n>>\\n>>\\n>> *Hi all,We redesign disk quota feature based on the comments from Pavel\\n>> Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk\\n>> quota feature is used to support multi-tenancy environment, different level\\n>> of database objects could be set a quota limit to avoid over use of disk\\n>> space. A common case could be as follows: DBA could enable disk quota on a\\n>> specified database list. DBA could set disk quota limit for\\n>> tables/schemas/roles in these databases. Separate disk quota worker process\\n>> will monitor the disk usage for these objects and detect the objects which\\n>> exceed their quota limit. Queries loading data into these “out of disk\\n>> quota” tables/schemas/roles will be cancelled.We are currently working at\\n>> init implementation stage. We would like to propose our idea firstly and\\n>> get feedbacks from community to do quick iteration.SQL Syntax (How to use\\n>> disk quota)1 Specify the databases with disk quota enabled in GUC\\n>> “diskquota_databases” in postgresql.conf and restart the database.2 DBA\\n>> could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\\n>> ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\\n>> (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\\n>> ‘3MB’);*\\n>>\\n>\\n> just fast reaction - why QUOTA object?\\n>\\n> Isn't ALTER SET enough?\\n>\\n> Some like\\n>\\n> ALTER TABLE a1 SET quote = 1MB;\\n> ALTER USER ...\\n> ALTER SCHEMA ..\\n>\\n> New DDL commans looks like too hard hammer .\\n>\\n>\\n>\\n>>\\n>>\\n>>\\n>>\\n>>\\n>>\\n>>\\n>>\\n>>\\n>>\\n>>\\n>>\\n>>\\n>>\\n>>\\n>>\\n>>\\n>> *3 Simulate a schema out of quota limit case: suppose table a1 and table\\n>> a2 are both under schema s1.INSERT INTO a1 SELECT\\n>> generate_series(1,1000);INSERT INTO a2 SELECT\\n>> generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\\n>> generate_series(1,1000);ERROR: schema's disk space quota exceededDROP\\n>> TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\\n>> generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\\n>> following components.1. Quota Setting Store is where the disk quota setting\\n>> to be stored and accessed. We plan to use catalog table pg_diskquota to\\n>> store these information. pg_diskquota is\\n>> like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\\n>> diskquota name */ int16 quotatype; /* diskquota type name */ Oid\\n>> quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\\n>> diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\\n>> MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\\n>> change of database objects. We plan to use stat collector to detect the\\n>> ‘active’ table list at initial stage. But stat collector has some\\n>> limitation on finding the active table which is in a running transaction.\\n>> Details see TODO section.3. Quota Size Checker is where to calculate the\\n>> size and compare with quota limit for database objects. According to\\n>> Pavel’s comment, autovacuum launcher and worker process could be a good\\n>> reference to disk quota. So we plan to use a disk quota launcher daemon\\n>> process and several disk quota worker process to finish this work. Launcher\\n>> process is responsible for starting worker process based on a user defined\\n>> database list from GUC. Worker process will connect to its target database\\n>> and monitor the disk usage for objects in this database. In init stage of\\n>> worker process, it will call calculate_total_relation_size() to calculate\\n>> the size for each user table. After init stage, worker process will refresh\\n>> the disk model every N seconds. Refreshing will only recalculate the size\\n>> of tables in ‘active’ table list, which is generated by Quata Change\\n>> Detector to minimize the cost.4. Quota Enforcement Operator is where to\\n>> check for the quota limitation at postgres backend side. We will firstly\\n>> implement it in ExecCheckRTPerms() as pre-running enforcement. It will\\n>> check the disk quota of tables being inserted or updated, and report error\\n>> if table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\\n>> As a native feature, we plan to add more checkpoint to do running query\\n>> enforcement. For example, if a disk quota lefts 10MB quota, a query could\\n>> insert 1GB data. This query could be allowed in pre-running enforcement\\n>> check, but will be cancelled in running query enforcement check. Therefore,\\n>> it can improve the accurate of disk quota usage. To achieve this, we plan\\n>> to add a checkpoint in lower API such as smgr_extened. Hence, the Quota\\n>> Enforcement Operator will check the disk quota usage when smgr_extened is\\n>> called. If the quota is over limited, current query will be cancelled.\\n>> Highlight1. Native feature.Support native Create/Drop Disk Quota SQL\\n>> statement.New catalog table pg_diskquota to store disk quota setting.2.\\n>> Auto DML/DDL detection. Table\\n>> create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\\n>> Schema create/drop and Role create/drop will be detected by disk quota\\n>> automatically. 3. Low cost disk quota checker.Worker process of disk quota\\n>> need to refresh the disk usage model every N seconds. Since recalculate the\\n>> file size using stat() system call is expensive for a large number of\\n>> files, we use an ‘active’ table list to reduce the real work at each\\n>> iteration. A basic experiment on our init stage implementation on database\\n>> with 20K tables shows that the refresh cost is 1% cpu usage and will be\\n>> finished within 50ms. Todo/LimitationBefore we propose our patch, we plan\\n>> to enhance it with the following ideas:1. Setting database list with disk\\n>> quota enabled dynamically without restart database. Since we have the disk\\n>> quota launcher process, it could detect the new ‘diskquota_databases’ list\\n>> and start/stop the corresponding disk quota worker process.2. Enforcement\\n>> when query is running. Considering the case when there is 10MB quota left,\\n>> but next query will insert 10GB data. Current enforcement design will allow\\n>> this query to be executed. This is limited by the ‘active’ table detection\\n>> is generated by stat collector. Postgres backend will only send table stat\\n>> information to collector only when the transaction ends. We need a new way\\n>> to detect the ‘active’ table even when this table is being modified inside\\n>> a running transaction.3. Monitor unlimited number of databases. Current we\\n>> set the max number of disk quota worker process to be 10 to reduce the\\n>> affection normal workload. But how about if we want to monitor the disk\\n>> quota of more than 10 databases? Our solution is to let disk quota launcher\\n>> to manage a queue of database need to be monitored. And disk quota worker\\n>> process consuming the queue and refresh the disk usage/quota for this\\n>> database. After some periods, worker will return the database to the queue,\\n>> and fetch the top database from queue to process. The period determine the\\n>> delay of detecting disk quota change. To implement this feature, we need to\\n>> support a subprocess of postmaster to rebind to another database instead of\\n>> the database binded in InitPostgres().4. Support active table detection on\\n>> vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\\n>> are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\\n>> Flack for the former comments on disk quota feature. Any comments on how to\\n>> improve disk quota feature are appreciated.*\\n>>\\n>>\\n>> On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com>\\n>> wrote:\\n>>\\n>>>\\n>>>\\n>>> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>>\\n>>>> Thanks Pavel.\\n>>>> Your patch did enforcement on storage level(md.c or we could also use\\n>>>> smgr_extend). It's straight forward.\\n>>>> But I prefer to implement disk_quota as a feature with following\\n>>>> objectives:\\n>>>> 1 set/alter disk quota setting on different database objects, e.g.\\n>>>> user, database, schema etc. not only a general GUC, but we could set\\n>>>> separate quota limit for a specific objects.\\n>>>> 2 enforcement operator should work at two positions: before query is\\n>>>> running and when query is running. The latter one's implementation maybe\\n>>>> similar to your patch.\\n>>>>\\n>>>\\n>>> The patch was just example. The resource quotes should be more complex -\\n>>> per partition, table, schema, database, user - so GUC are possible, but not\\n>>> very user friendly.\\n>>>\\n>>> Our case is specific, but not too much. The servers are used for\\n>>> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\\n>>> SELECT). We need to solve limits immediately. The implementation is simple,\\n>>> so I did it. Same implementation on database level, or schema level needs\\n>>> some more locks, so it will not be too effective. The resource management\\n>>> can be complex very complex, and I expect so it will be hard work.\\n>>>\\n>>> Regards\\n>>>\\n>>> Pavel\\n>>>\\n>>>\\n>>>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com>\\n>>>> wrote:\\n>>>>\\n>>>>> Hi\\n>>>>>\\n>>>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>>>>\\n>>>>>> Thanks Chapman.\\n>>>>>> @Pavel, could you please explain more about your second suggestion \\\"implement\\n>>>>>> some quotas on storage level?\\\"\\n>>>>>>\\n>>>>>\\n>>>>> See attached patch - it is very simple - and good enough for our\\n>>>>> purposes.\\n>>>>>\\n>>>>> Regards\\n>>>>>\\n>>>>> Pavel\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>> We will not keep the long-lived processes attach to all\\n>>>>>> databases(just like you mentioned servers with thousands of databases\\n>>>>>> )\\n>>>>>> And you are right, we could share ideas with autovacuum process, fork\\n>>>>>> worker processes in need.\\n>>>>>> \\\"autovacuum checks for tables that have had a large number of\\n>>>>>> inserted, updated or deleted tuples. These checks use the statistics\\n>>>>>> collection facility\\\"\\n>>>>>> diskquota process is similar to autovacuum at caring about insert,\\n>>>>>> but the difference is that it also care about vucuum full, truncate and\\n>>>>>> drop. While update and delete may not be interested since no file change\\n>>>>>> happens. So a separate diskquota process is preferred.\\n>>>>>>\\n>>>>>> So if we implemented disk quota as a full native feature, and in the\\n>>>>>> first initial version I prefer to implement the following features:\\n>>>>>> 1 Fork diskquota launcher process under Postmaster serverloop, which\\n>>>>>> is long-lived.\\n>>>>>> 2 Diskquota launcher process is responsible for creating diskquota\\n>>>>>> worker process for every database.\\n>>>>>> 3 DIskquota setting is stored in a separate catalog table for each\\n>>>>>> database.\\n>>>>>> 4 Initialization stage, Diskquota launcher process creates diskquota worker\\n>>>>>> process for all the databases(traverse like autovacuum). Worker process\\n>>>>>> calculates disk usage of db objects and their diskquota setting. If any\\n>>>>>> db object exceeds its quota limit, put them into the blacklist in the\\n>>>>>> shared memory, which will later be used by enforcement operator. Worker\\n>>>>>> process exits when works are done.\\n>>>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\\n>>>>>> process for the database with a large number of insert, copy, truncate,\\n>>>>>> drop etc. or create disk quota statement. Worker process updates the file\\n>>>>>> size for db objects containing the result relation, and compare with the\\n>>>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\\n>>>>>> remove from blacklist vice versa. Worker process exits when works\\n>>>>>> are done and a GUC could control the frequency of worker process restart to\\n>>>>>> a specific database. As you know, this GUC also controls the delay when we\\n>>>>>> do enforcement.\\n>>>>>> 6 Enforcement. When postgres backend executes queries, check the\\n>>>>>> blacklist in shared memory to determine whether the query is allowed(before\\n>>>>>> execute) or need rollback(is executing)?\\n>>>>>>\\n>>>>>> If we implemented disk quota as an extension, we could just use\\n>>>>>> background worker to start diskquota launcher process and use\\n>>>>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\\n>>>>>> processes by the launcher process as suggested by @Chapman.\\n>>>>>> Diskquota setting could be stored in user table in a separate schema for\\n>>>>>> each database(Schema and table created by create extension statement) just\\n>>>>>> like what Heikki has done in pg_quota project. But in this case, we need to\\n>>>>>> create extension for each database before diskquota worker process can be\\n>>>>>> set up for that database.\\n>>>>>>\\n>>>>>> Any comments on the above design and which is preferred, native\\n>>>>>> feature or extension as the POC?\\n>>>>>>\\n>>>>>>\\n>>>>>> -- Hubert\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <\\n>>>>>> pavel.stehule@gmail.com> wrote:\\n>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\\n>>>>>>>\\n>>>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n>>>>>>>>\\n>>>>>>>> > 2 Keep one worker process for each database. But using a\\n>>>>>>>> parent/global\\n>>>>>>>> > quota worker process to manage the lifecycle of database level\\n>>>>>>>> worker\\n>>>>>>>> > processes. It could handle the newly created database(avoid\\n>>>>>>>> restart\\n>>>>>>>> > database) and save resource when a database is not used. But this\\n>>>>>>>> needs to\\n>>>>>>>> > change worker process to be hierarchical. Postmaster becomes the\\n>>>>>>>> grandfather\\n>>>>>>>> > of database level worker processes in this case.\\n>>>>>>>>\\n>>>>>>>> I am using background workers this way in 9.5 at $work.\\n>>>>>>>>\\n>>>>>>>> In my case, one worker lives forever, wakes up on a set period, and\\n>>>>>>>> starts a short-lived worker for every database, waiting for each\\n>>>>>>>> one before starting the next.\\n>>>>>>>>\\n>>>>>>>> It was straightforward to implement. Looking back over the code,\\n>>>>>>>> I see the global worker assigns its own PID to worker.bgw_notify_pid\\n>>>>>>>> of each of its children, and also obtains a handle for each child\\n>>>>>>>> from RegisterDynamicBackgroundWorker().\\n>>>>>>>>\\n>>>>>>>> I imagine the global quota worker would prefer to start workers\\n>>>>>>>> for every database and then just wait for notifications from any\\n>>>>>>>> of them, but that seems equally straightforward at first glance.\\n>>>>>>>>\\n>>>>>>>\\n>>>>>>> There are servers with thousands databases. Worker per database is\\n>>>>>>> not good idea.\\n>>>>>>>\\n>>>>>>> It should to share ideas, code with autovacuum process.\\n>>>>>>>\\n>>>>>>> Not sure, how to effective implementation based on bg workers can\\n>>>>>>> be. On servers with large set of databases, large set of tables it can\\n>>>>>>> identify too big table too late.\\n>>>>>>>\\n>>>>>>> Isn't better to implement some quotas on storage level?\\n>>>>>>>\\n>>>>>>> Regards\\n>>>>>>>\\n>>>>>>> Pavel\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>> -Chap\\n>>>>>>>>\\n>>>>>>>>\\n>>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>> --\\n>>>>>> Thanks\\n>>>>>>\\n>>>>>> Hubert Zhang\\n>>>>>>\\n>>>>>\\n>>>>>\\n>>>>\\n>>>>\\n>>>> --\\n>>>> Thanks\\n>>>>\\n>>>> Hubert Zhang\\n>>>>\\n>>>\\n>>>\\n>>\\n>>\\n>> --\\n>> Thanks\\n>>\\n>> Hubert Zhang\\n>>\\n>\\n\\n-- \\nThanks\\n\\nHubert Zhang\\n\\njust fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer .It's an option. Prefer to consider quota setting store together:CREATE DISK QUOTA way is more nature to store quota setting in a separate pg_diskquota catalogWhile ALTER SET way is more close to store quota setting in pg_class, pg_role, pg_namespace. etc in an integrated way.(Note that here I mean nature/close is not must, ALTER SET could also store in pg_diskquota and vice versa.)Here are some differences I can think of:1 pg_role is a global catalog, not per database level. It's harder to tracker the user's disk usage in the whole clusters(considering 1000+ databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it only tracks the user's disk usage inside the current database.2 using separate pg_diskquota could add more field except for quota limit without adding too many fields in pg_class, e.g. red zone to give the user a warning or the current disk usage of the db objects.On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:Hi all,We redesign disk quota feature based on the comments from Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk quota feature is used to support multi-tenancy environment, different level of database objects could be set a quota limit to avoid over use of disk space. A common case could be as follows: DBA could enable disk quota on a specified database list. DBA could set disk quota limit for tables/schemas/roles in these databases. Separate disk quota worker process will monitor the disk usage for these objects and detect the objects which exceed their quota limit. Queries loading data into these “out of disk quota” tables/schemas/roles will be cancelled.We are currently working at init implementation stage. We would like to propose our idea firstly and get feedbacks from community to do quick iteration.SQL Syntax (How to use disk quota)1 Specify the databases with disk quota enabled in GUC “diskquota_databases” in postgresql.conf and restart the database.2 DBA could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer . 3 Simulate a schema out of quota limit case: suppose table a1 and table a2 are both under schema s1.INSERT INTO a1 SELECT generate_series(1,1000);INSERT INTO a2 SELECT generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);ERROR: schema's disk space quota exceededDROP TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the following components.1. Quota Setting Store is where the disk quota setting to be stored and accessed. We plan to use catalog table pg_diskquota to store these information. pg_diskquota is like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /* diskquota name */ int16 quotatype; /* diskquota type name */ Oid quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /* diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size change of database objects. We plan to use stat collector to detect the ‘active’ table list at initial stage. But stat collector has some limitation on finding the active table which is in a running transaction. Details see TODO section.3. Quota Size Checker is where to calculate the size and compare with quota limit for database objects. According to Pavel’s comment, autovacuum launcher and worker process could be a good reference to disk quota. So we plan to use a disk quota launcher daemon process and several disk quota worker process to finish this work. Launcher process is responsible for starting worker process based on a user defined database list from GUC. Worker process will connect to its target database and monitor the disk usage for objects in this database. In init stage of worker process, it will call calculate_total_relation_size() to calculate the size for each user table. After init stage, worker process will refresh the disk model every N seconds. Refreshing will only recalculate the size of tables in ‘active’ table list, which is generated by Quata Change Detector to minimize the cost.4. Quota Enforcement Operator is where to check for the quota limitation at postgres backend side. We will firstly implement it in ExecCheckRTPerms() as pre-running enforcement. It will check the disk quota of tables being inserted or updated, and report error if table’s or table’s schema’s or table’s owner’s quota limit is exceeded. As a native feature, we plan to add more checkpoint to do running query enforcement. For example, if a disk quota lefts 10MB quota, a query could insert 1GB data. This query could be allowed in pre-running enforcement check, but will be cancelled in running query enforcement check. Therefore, it can improve the accurate of disk quota usage. To achieve this, we plan to add a checkpoint in lower API such as smgr_extened. Hence, the Quota Enforcement Operator will check the disk quota usage when smgr_extened is called. If the quota is over limited, current query will be cancelled. Highlight1. Native feature.Support native Create/Drop Disk Quota SQL statement.New catalog table pg_diskquota to store disk quota setting.2. Auto DML/DDL detection. Table create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change, Schema create/drop and Role create/drop will be detected by disk quota automatically. 3. Low cost disk quota checker.Worker process of disk quota need to refresh the disk usage model every N seconds. Since recalculate the file size using stat() system call is expensive for a large number of files, we use an ‘active’ table list to reduce the real work at each iteration. A basic experiment on our init stage implementation on database with 20K tables shows that the refresh cost is 1% cpu usage and will be finished within 50ms. Todo/LimitationBefore we propose our patch, we plan to enhance it with the following ideas:1. Setting database list with disk quota enabled dynamically without restart database. Since we have the disk quota launcher process, it could detect the new ‘diskquota_databases’ list and start/stop the corresponding disk quota worker process.2. Enforcement when query is running. Considering the case when there is 10MB quota left, but next query will insert 10GB data. Current enforcement design will allow this query to be executed. This is limited by the ‘active’ table detection is generated by stat collector. Postgres backend will only send table stat information to collector only when the transaction ends. We need a new way to detect the ‘active’ table even when this table is being modified inside a running transaction.3. Monitor unlimited number of databases. Current we set the max number of disk quota worker process to be 10 to reduce the affection normal workload. But how about if we want to monitor the disk quota of more than 10 databases? Our solution is to let disk quota launcher to manage a queue of database need to be monitored. And disk quota worker process consuming the queue and refresh the disk usage/quota for this database. After some periods, worker will return the database to the queue, and fetch the top database from queue to process. The period determine the delay of detecting disk quota change. To implement this feature, we need to support a subprocess of postmaster to rebind to another database instead of the database binded in InitPostgres().4. Support active table detection on vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman Flack for the former comments on disk quota feature. Any comments on how to improve disk quota feature are appreciated.On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \\\"implement some quotas on storage level?\\\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \\\"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\\\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n\\n> 2 Keep one worker process for each database. But using a parent/global\\n> quota worker process to manage the lifecycle of database level worker\\n> processes. It could handle the newly created database(avoid restart\\n> database) and save resource when a database is not used. But this needs to\\n> change worker process to be hierarchical. Postmaster becomes the grandfather\\n> of database level worker processes in this case.\\n\\nI am using background workers this way in 9.5 at $work.\\n\\nIn my case, one worker lives forever, wakes up on a set period, and\\nstarts a short-lived worker for every database, waiting for each\\none before starting the next.\\n\\nIt was straightforward to implement. Looking back over the code,\\nI see the global worker assigns its own PID to worker.bgw_notify_pid\\nof each of its children, and also obtains a handle for each child\\nfrom RegisterDynamicBackgroundWorker().\\n\\nI imagine the global quota worker would prefer to start workers\\nfor every database and then just wait for notifications from any\\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\\n\\n-Chap\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\",\n \"msg_date\": \"Fri, 21 Sep 2018 22:21:16 +0800\",\n \"msg_from\": \"Hubert Zhang <hzhang@pivotal.io>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal for disk quota feature\"\n },\n {\n \"msg_contents\": \"On Fri, Sep 21, 2018 at 9:21 AM Hubert Zhang <hzhang@pivotal.io> wrote:\\n\\n> just fast reaction - why QUOTA object?\\n>> Isn't ALTER SET enough?\\n>> Some like\\n>> ALTER TABLE a1 SET quote = 1MB;\\n>> ALTER USER ...\\n>> ALTER SCHEMA ..\\n>> New DDL commans looks like too hard hammer .\\n>\\n>\\n> It's an option. Prefer to consider quota setting store together:\\n> CREATE DISK QUOTA way is more nature to store quota setting in a separate\\n> pg_diskquota catalog\\n> While ALTER SET way is more close to store quota setting in pg_class,\\n> pg_role, pg_namespace. etc in an integrated way.\\n> (Note that here I mean nature/close is not must, ALTER SET could also\\n> store in pg_diskquota and vice versa.)\\n>\\n> Here are some differences I can think of:\\n> 1 pg_role is a global catalog, not per database level. It's harder to\\n> tracker the user's disk usage in the whole clusters(considering 1000+\\n> databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it\\n> only tracks the user's disk usage inside the current database.\\n> 2 using separate pg_diskquota could add more field except for quota limit\\n> without adding too many fields in pg_class, e.g. red zone to give the user\\n> a warning or the current disk usage of the db objects.\\n>\\n> On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com>\\n> wrote:\\n>\\n>>\\n>>\\n>> pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io>\\n>> napsal:\\n>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>> *Hi all,We redesign disk quota feature based on the comments from Pavel\\n>>> Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk\\n>>> quota feature is used to support multi-tenancy environment, different level\\n>>> of database objects could be set a quota limit to avoid over use of disk\\n>>> space. A common case could be as follows: DBA could enable disk quota on a\\n>>> specified database list. DBA could set disk quota limit for\\n>>> tables/schemas/roles in these databases. Separate disk quota worker process\\n>>> will monitor the disk usage for these objects and detect the objects which\\n>>> exceed their quota limit. Queries loading data into these “out of disk\\n>>> quota” tables/schemas/roles will be cancelled.We are currently working at\\n>>> init implementation stage. We would like to propose our idea firstly and\\n>>> get feedbacks from community to do quick iteration.SQL Syntax (How to use\\n>>> disk quota)1 Specify the databases with disk quota enabled in GUC\\n>>> “diskquota_databases” in postgresql.conf and restart the database.2 DBA\\n>>> could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\\n>>> ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\\n>>> (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\\n>>> ‘3MB’);*\\n>>>\\n>>\\n>> just fast reaction - why QUOTA object?\\n>>\\n>> Isn't ALTER SET enough?\\n>>\\n>> Some like\\n>>\\n>> ALTER TABLE a1 SET quote = 1MB;\\n>> ALTER USER ...\\n>> ALTER SCHEMA ..\\n>>\\n>> New DDL commans looks like too hard hammer .\\n>>\\n>>\\n>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>> *3 Simulate a schema out of quota limit case: suppose table a1 and table\\n>>> a2 are both under schema s1.INSERT INTO a1 SELECT\\n>>> generate_series(1,1000);INSERT INTO a2 SELECT\\n>>> generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\\n>>> generate_series(1,1000);ERROR: schema's disk space quota exceededDROP\\n>>> TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\\n>>> generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\\n>>> following components.1. Quota Setting Store is where the disk quota setting\\n>>> to be stored and accessed. We plan to use catalog table pg_diskquota to\\n>>> store these information. pg_diskquota is\\n>>> like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\\n>>> diskquota name */ int16 quotatype; /* diskquota type name */ Oid\\n>>> quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\\n>>> diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\\n>>> MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\\n>>> change of database objects. We plan to use stat collector to detect the\\n>>> ‘active’ table list at initial stage. But stat collector has some\\n>>> limitation on finding the active table which is in a running transaction.\\n>>> Details see TODO section.3. Quota Size Checker is where to calculate the\\n>>> size and compare with quota limit for database objects. According to\\n>>> Pavel’s comment, autovacuum launcher and worker process could be a good\\n>>> reference to disk quota. So we plan to use a disk quota launcher daemon\\n>>> process and several disk quota worker process to finish this work. Launcher\\n>>> process is responsible for starting worker process based on a user defined\\n>>> database list from GUC. Worker process will connect to its target database\\n>>> and monitor the disk usage for objects in this database. In init stage of\\n>>> worker process, it will call calculate_total_relation_size() to calculate\\n>>> the size for each user table. After init stage, worker process will refresh\\n>>> the disk model every N seconds. Refreshing will only recalculate the size\\n>>> of tables in ‘active’ table list, which is generated by Quata Change\\n>>> Detector to minimize the cost.4. Quota Enforcement Operator is where to\\n>>> check for the quota limitation at postgres backend side. We will firstly\\n>>> implement it in ExecCheckRTPerms() as pre-running enforcement. It will\\n>>> check the disk quota of tables being inserted or updated, and report error\\n>>> if table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\\n>>> As a native feature, we plan to add more checkpoint to do running query\\n>>> enforcement. For example, if a disk quota lefts 10MB quota, a query could\\n>>> insert 1GB data. This query could be allowed in pre-running enforcement\\n>>> check, but will be cancelled in running query enforcement check. Therefore,\\n>>> it can improve the accurate of disk quota usage. To achieve this, we plan\\n>>> to add a checkpoint in lower API such as smgr_extened. Hence, the Quota\\n>>> Enforcement Operator will check the disk quota usage when smgr_extened is\\n>>> called. If the quota is over limited, current query will be cancelled.\\n>>> Highlight1. Native feature.Support native Create/Drop Disk Quota SQL\\n>>> statement.New catalog table pg_diskquota to store disk quota setting.2.\\n>>> Auto DML/DDL detection. Table\\n>>> create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\\n>>> Schema create/drop and Role create/drop will be detected by disk quota\\n>>> automatically. 3. Low cost disk quota checker.Worker process of disk quota\\n>>> need to refresh the disk usage model every N seconds. Since recalculate the\\n>>> file size using stat() system call is expensive for a large number of\\n>>> files, we use an ‘active’ table list to reduce the real work at each\\n>>> iteration. A basic experiment on our init stage implementation on database\\n>>> with 20K tables shows that the refresh cost is 1% cpu usage and will be\\n>>> finished within 50ms. Todo/LimitationBefore we propose our patch, we plan\\n>>> to enhance it with the following ideas:1. Setting database list with disk\\n>>> quota enabled dynamically without restart database. Since we have the disk\\n>>> quota launcher process, it could detect the new ‘diskquota_databases’ list\\n>>> and start/stop the corresponding disk quota worker process.2. Enforcement\\n>>> when query is running. Considering the case when there is 10MB quota left,\\n>>> but next query will insert 10GB data. Current enforcement design will allow\\n>>> this query to be executed. This is limited by the ‘active’ table detection\\n>>> is generated by stat collector. Postgres backend will only send table stat\\n>>> information to collector only when the transaction ends. We need a new way\\n>>> to detect the ‘active’ table even when this table is being modified inside\\n>>> a running transaction.3. Monitor unlimited number of databases. Current we\\n>>> set the max number of disk quota worker process to be 10 to reduce the\\n>>> affection normal workload. But how about if we want to monitor the disk\\n>>> quota of more than 10 databases? Our solution is to let disk quota launcher\\n>>> to manage a queue of database need to be monitored. And disk quota worker\\n>>> process consuming the queue and refresh the disk usage/quota for this\\n>>> database. After some periods, worker will return the database to the queue,\\n>>> and fetch the top database from queue to process. The period determine the\\n>>> delay of detecting disk quota change. To implement this feature, we need to\\n>>> support a subprocess of postmaster to rebind to another database instead of\\n>>> the database binded in InitPostgres().4. Support active table detection on\\n>>> vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\\n>>> are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\\n>>> Flack for the former comments on disk quota feature. Any comments on how to\\n>>> improve disk quota feature are appreciated.*\\n>>>\\n>>>\\n>>> On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com>\\n>>> wrote:\\n>>>\\n>>>>\\n>>>>\\n>>>> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>>>\\n>>>>> Thanks Pavel.\\n>>>>> Your patch did enforcement on storage level(md.c or we could also use\\n>>>>> smgr_extend). It's straight forward.\\n>>>>> But I prefer to implement disk_quota as a feature with following\\n>>>>> objectives:\\n>>>>> 1 set/alter disk quota setting on different database objects, e.g.\\n>>>>> user, database, schema etc. not only a general GUC, but we could set\\n>>>>> separate quota limit for a specific objects.\\n>>>>> 2 enforcement operator should work at two positions: before query is\\n>>>>> running and when query is running. The latter one's implementation maybe\\n>>>>> similar to your patch.\\n>>>>>\\n>>>>\\n>>>> The patch was just example. The resource quotes should be more complex\\n>>>> - per partition, table, schema, database, user - so GUC are possible, but\\n>>>> not very user friendly.\\n>>>>\\n>>>> Our case is specific, but not too much. The servers are used for\\n>>>> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\\n>>>> SELECT). We need to solve limits immediately. The implementation is simple,\\n>>>> so I did it. Same implementation on database level, or schema level needs\\n>>>> some more locks, so it will not be too effective. The resource management\\n>>>> can be complex very complex, and I expect so it will be hard work.\\n>>>>\\n>>>> Regards\\n>>>>\\n>>>> Pavel\\n>>>>\\n>>>>\\n>>>>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com\\n>>>>> > wrote:\\n>>>>>\\n>>>>>> Hi\\n>>>>>>\\n>>>>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>>>>>\\n>>>>>>> Thanks Chapman.\\n>>>>>>> @Pavel, could you please explain more about your second suggestion\\n>>>>>>> \\\"implement some quotas on storage level?\\\"\\n>>>>>>>\\n>>>>>>\\n>>>>>> See attached patch - it is very simple - and good enough for our\\n>>>>>> purposes.\\n>>>>>>\\n>>>>>> Regards\\n>>>>>>\\n>>>>>> Pavel\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>> We will not keep the long-lived processes attach to all\\n>>>>>>> databases(just like you mentioned servers with thousands of\\n>>>>>>> databases)\\n>>>>>>> And you are right, we could share ideas with autovacuum process,\\n>>>>>>> fork worker processes in need.\\n>>>>>>> \\\"autovacuum checks for tables that have had a large number of\\n>>>>>>> inserted, updated or deleted tuples. These checks use the statistics\\n>>>>>>> collection facility\\\"\\n>>>>>>> diskquota process is similar to autovacuum at caring about insert,\\n>>>>>>> but the difference is that it also care about vucuum full, truncate and\\n>>>>>>> drop. While update and delete may not be interested since no file change\\n>>>>>>> happens. So a separate diskquota process is preferred.\\n>>>>>>>\\n>>>>>>> So if we implemented disk quota as a full native feature, and in\\n>>>>>>> the first initial version I prefer to implement the following features:\\n>>>>>>> 1 Fork diskquota launcher process under Postmaster serverloop, which\\n>>>>>>> is long-lived.\\n>>>>>>> 2 Diskquota launcher process is responsible for creating diskquota\\n>>>>>>> worker process for every database.\\n>>>>>>> 3 DIskquota setting is stored in a separate catalog table for each\\n>>>>>>> database.\\n>>>>>>> 4 Initialization stage, Diskquota launcher process creates diskquota\\n>>>>>>> worker process for all the databases(traverse like autovacuum).\\n>>>>>>> Worker process calculates disk usage of db objects and\\n>>>>>>> their diskquota setting. If any db object exceeds its quota limit, put them\\n>>>>>>> into the blacklist in the shared memory, which will later be used by\\n>>>>>>> enforcement operator. Worker process exits when works are done.\\n>>>>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\\n>>>>>>> process for the database with a large number of insert, copy, truncate,\\n>>>>>>> drop etc. or create disk quota statement. Worker process updates the file\\n>>>>>>> size for db objects containing the result relation, and compare with the\\n>>>>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\\n>>>>>>> remove from blacklist vice versa. Worker process exits when works\\n>>>>>>> are done and a GUC could control the frequency of worker process restart to\\n>>>>>>> a specific database. As you know, this GUC also controls the delay when we\\n>>>>>>> do enforcement.\\n>>>>>>> 6 Enforcement. When postgres backend executes queries, check the\\n>>>>>>> blacklist in shared memory to determine whether the query is allowed(before\\n>>>>>>> execute) or need rollback(is executing)?\\n>>>>>>>\\n>>>>>>> If we implemented disk quota as an extension, we could just use\\n>>>>>>> background worker to start diskquota launcher process and use\\n>>>>>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\\n>>>>>>> processes by the launcher process as suggested by @Chapman.\\n>>>>>>> Diskquota setting could be stored in user table in a separate schema for\\n>>>>>>> each database(Schema and table created by create extension statement) just\\n>>>>>>> like what Heikki has done in pg_quota project. But in this case, we need to\\n>>>>>>> create extension for each database before diskquota worker process can be\\n>>>>>>> set up for that database.\\n>>>>>>>\\n>>>>>>> Any comments on the above design and which is preferred, native\\n>>>>>>> feature or extension as the POC?\\n>>>>>>>\\n>>>>>>>\\n>>>>>>> -- Hubert\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <\\n>>>>>>> pavel.stehule@gmail.com> wrote:\\n>>>>>>>\\n>>>>>>>>\\n>>>>>>>>\\n>>>>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\\n>>>>>>>>\\n>>>>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n>>>>>>>>>\\n>>>>>>>>> > 2 Keep one worker process for each database. But using a\\n>>>>>>>>> parent/global\\n>>>>>>>>> > quota worker process to manage the lifecycle of database level\\n>>>>>>>>> worker\\n>>>>>>>>> > processes. It could handle the newly created database(avoid\\n>>>>>>>>> restart\\n>>>>>>>>> > database) and save resource when a database is not used. But\\n>>>>>>>>> this needs to\\n>>>>>>>>> > change worker process to be hierarchical. Postmaster becomes the\\n>>>>>>>>> grandfather\\n>>>>>>>>> > of database level worker processes in this case.\\n>>>>>>>>>\\n>>>>>>>>> I am using background workers this way in 9.5 at $work.\\n>>>>>>>>>\\n>>>>>>>>> In my case, one worker lives forever, wakes up on a set period, and\\n>>>>>>>>> starts a short-lived worker for every database, waiting for each\\n>>>>>>>>> one before starting the next.\\n>>>>>>>>>\\n>>>>>>>>> It was straightforward to implement. Looking back over the code,\\n>>>>>>>>> I see the global worker assigns its own PID to\\n>>>>>>>>> worker.bgw_notify_pid\\n>>>>>>>>> of each of its children, and also obtains a handle for each child\\n>>>>>>>>> from RegisterDynamicBackgroundWorker().\\n>>>>>>>>>\\n>>>>>>>>> I imagine the global quota worker would prefer to start workers\\n>>>>>>>>> for every database and then just wait for notifications from any\\n>>>>>>>>> of them, but that seems equally straightforward at first glance.\\n>>>>>>>>>\\n>>>>>>>>\\n>>>>>>>> There are servers with thousands databases. Worker per database is\\n>>>>>>>> not good idea.\\n>>>>>>>>\\n>>>>>>>> It should to share ideas, code with autovacuum process.\\n>>>>>>>>\\n>>>>>>>> Not sure, how to effective implementation based on bg workers can\\n>>>>>>>> be. On servers with large set of databases, large set of tables it can\\n>>>>>>>> identify too big table too late.\\n>>>>>>>>\\n>>>>>>>> Isn't better to implement some quotas on storage level?\\n>>>>>>>>\\n>>>>>>>\\nFWIW, we are super interested in this feature at a role level especially in\\nreporting systems where users can for example accidentally start a create\\ntable statement overnight and sure enough after their Cartesian product\\ngenerated 2TB of data they run the system out of disk.\\n\\nThat problem is really hard to manage currently and this feature would\\nreally solve that.\\n\\nThanks,\\nJeremy\\n\\n>\\n\\nOn Fri, Sep 21, 2018 at 9:21 AM Hubert Zhang <hzhang@pivotal.io> wrote:just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer .It's an option. Prefer to consider quota setting store together:CREATE DISK QUOTA way is more nature to store quota setting in a separate pg_diskquota catalogWhile ALTER SET way is more close to store quota setting in pg_class, pg_role, pg_namespace. etc in an integrated way.(Note that here I mean nature/close is not must, ALTER SET could also store in pg_diskquota and vice versa.)Here are some differences I can think of:1 pg_role is a global catalog, not per database level. It's harder to tracker the user's disk usage in the whole clusters(considering 1000+ databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it only tracks the user's disk usage inside the current database.2 using separate pg_diskquota could add more field except for quota limit without adding too many fields in pg_class, e.g. red zone to give the user a warning or the current disk usage of the db objects.On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:Hi all,We redesign disk quota feature based on the comments from Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk quota feature is used to support multi-tenancy environment, different level of database objects could be set a quota limit to avoid over use of disk space. A common case could be as follows: DBA could enable disk quota on a specified database list. DBA could set disk quota limit for tables/schemas/roles in these databases. Separate disk quota worker process will monitor the disk usage for these objects and detect the objects which exceed their quota limit. Queries loading data into these “out of disk quota” tables/schemas/roles will be cancelled.We are currently working at init implementation stage. We would like to propose our idea firstly and get feedbacks from community to do quick iteration.SQL Syntax (How to use disk quota)1 Specify the databases with disk quota enabled in GUC “diskquota_databases” in postgresql.conf and restart the database.2 DBA could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer . 3 Simulate a schema out of quota limit case: suppose table a1 and table a2 are both under schema s1.INSERT INTO a1 SELECT generate_series(1,1000);INSERT INTO a2 SELECT generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);ERROR: schema's disk space quota exceededDROP TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the following components.1. Quota Setting Store is where the disk quota setting to be stored and accessed. We plan to use catalog table pg_diskquota to store these information. pg_diskquota is like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /* diskquota name */ int16 quotatype; /* diskquota type name */ Oid quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /* diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size change of database objects. We plan to use stat collector to detect the ‘active’ table list at initial stage. But stat collector has some limitation on finding the active table which is in a running transaction. Details see TODO section.3. Quota Size Checker is where to calculate the size and compare with quota limit for database objects. According to Pavel’s comment, autovacuum launcher and worker process could be a good reference to disk quota. So we plan to use a disk quota launcher daemon process and several disk quota worker process to finish this work. Launcher process is responsible for starting worker process based on a user defined database list from GUC. Worker process will connect to its target database and monitor the disk usage for objects in this database. In init stage of worker process, it will call calculate_total_relation_size() to calculate the size for each user table. After init stage, worker process will refresh the disk model every N seconds. Refreshing will only recalculate the size of tables in ‘active’ table list, which is generated by Quata Change Detector to minimize the cost.4. Quota Enforcement Operator is where to check for the quota limitation at postgres backend side. We will firstly implement it in ExecCheckRTPerms() as pre-running enforcement. It will check the disk quota of tables being inserted or updated, and report error if table’s or table’s schema’s or table’s owner’s quota limit is exceeded. As a native feature, we plan to add more checkpoint to do running query enforcement. For example, if a disk quota lefts 10MB quota, a query could insert 1GB data. This query could be allowed in pre-running enforcement check, but will be cancelled in running query enforcement check. Therefore, it can improve the accurate of disk quota usage. To achieve this, we plan to add a checkpoint in lower API such as smgr_extened. Hence, the Quota Enforcement Operator will check the disk quota usage when smgr_extened is called. If the quota is over limited, current query will be cancelled. Highlight1. Native feature.Support native Create/Drop Disk Quota SQL statement.New catalog table pg_diskquota to store disk quota setting.2. Auto DML/DDL detection. Table create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change, Schema create/drop and Role create/drop will be detected by disk quota automatically. 3. Low cost disk quota checker.Worker process of disk quota need to refresh the disk usage model every N seconds. Since recalculate the file size using stat() system call is expensive for a large number of files, we use an ‘active’ table list to reduce the real work at each iteration. A basic experiment on our init stage implementation on database with 20K tables shows that the refresh cost is 1% cpu usage and will be finished within 50ms. Todo/LimitationBefore we propose our patch, we plan to enhance it with the following ideas:1. Setting database list with disk quota enabled dynamically without restart database. Since we have the disk quota launcher process, it could detect the new ‘diskquota_databases’ list and start/stop the corresponding disk quota worker process.2. Enforcement when query is running. Considering the case when there is 10MB quota left, but next query will insert 10GB data. Current enforcement design will allow this query to be executed. This is limited by the ‘active’ table detection is generated by stat collector. Postgres backend will only send table stat information to collector only when the transaction ends. We need a new way to detect the ‘active’ table even when this table is being modified inside a running transaction.3. Monitor unlimited number of databases. Current we set the max number of disk quota worker process to be 10 to reduce the affection normal workload. But how about if we want to monitor the disk quota of more than 10 databases? Our solution is to let disk quota launcher to manage a queue of database need to be monitored. And disk quota worker process consuming the queue and refresh the disk usage/quota for this database. After some periods, worker will return the database to the queue, and fetch the top database from queue to process. The period determine the delay of detecting disk quota change. To implement this feature, we need to support a subprocess of postmaster to rebind to another database instead of the database binded in InitPostgres().4. Support active table detection on vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman Flack for the former comments on disk quota feature. Any comments on how to improve disk quota feature are appreciated.On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \\\"implement some quotas on storage level?\\\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \\\"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\\\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n\\n> 2 Keep one worker process for each database. But using a parent/global\\n> quota worker process to manage the lifecycle of database level worker\\n> processes. It could handle the newly created database(avoid restart\\n> database) and save resource when a database is not used. But this needs to\\n> change worker process to be hierarchical. Postmaster becomes the grandfather\\n> of database level worker processes in this case.\\n\\nI am using background workers this way in 9.5 at $work.\\n\\nIn my case, one worker lives forever, wakes up on a set period, and\\nstarts a short-lived worker for every database, waiting for each\\none before starting the next.\\n\\nIt was straightforward to implement. Looking back over the code,\\nI see the global worker assigns its own PID to worker.bgw_notify_pid\\nof each of its children, and also obtains a handle for each child\\nfrom RegisterDynamicBackgroundWorker().\\n\\nI imagine the global quota worker would prefer to start workers\\nfor every database and then just wait for notifications from any\\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?FWIW, we are super interested in this feature at a role level especially in reporting systems where users can for example accidentally start a create table statement overnight and sure enough after their Cartesian product generated 2TB of data they run the system out of disk.That problem is really hard to manage currently and this feature would really solve that.Thanks,Jeremy\",\n \"msg_date\": \"Fri, 21 Sep 2018 16:33:38 -0500\",\n \"msg_from\": \"Jeremy Finzel <finzelj@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal for disk quota feature\"\n },\n {\n \"msg_contents\": \"pá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:\\n\\n> just fast reaction - why QUOTA object?\\n>> Isn't ALTER SET enough?\\n>> Some like\\n>> ALTER TABLE a1 SET quote = 1MB;\\n>> ALTER USER ...\\n>> ALTER SCHEMA ..\\n>> New DDL commans looks like too hard hammer .\\n>\\n>\\n> It's an option. Prefer to consider quota setting store together:\\n> CREATE DISK QUOTA way is more nature to store quota setting in a separate\\n> pg_diskquota catalog\\n> While ALTER SET way is more close to store quota setting in pg_class,\\n> pg_role, pg_namespace. etc in an integrated way.\\n> (Note that here I mean nature/close is not must, ALTER SET could also\\n> store in pg_diskquota and vice versa.)\\n>\\n\\nI have not a problem with new special table for storing this information.\\nBut it looks like redundant to current GUC configuration and limits. Can be\\nmessy do some work with ALTER ROLE, and some work via CREATE QUOTE.\\n\\nRegards\\n\\nPavel\\n\\n\\n> Here are some differences I can think of:\\n> 1 pg_role is a global catalog, not per database level. It's harder to\\n> tracker the user's disk usage in the whole clusters(considering 1000+\\n> databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it\\n> only tracks the user's disk usage inside the current database.\\n> 2 using separate pg_diskquota could add more field except for quota limit\\n> without adding too many fields in pg_class, e.g. red zone to give the user\\n> a warning or the current disk usage of the db objects.\\n>\\n> On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com>\\n> wrote:\\n>\\n>>\\n>>\\n>> pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io>\\n>> napsal:\\n>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>> *Hi all,We redesign disk quota feature based on the comments from Pavel\\n>>> Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk\\n>>> quota feature is used to support multi-tenancy environment, different level\\n>>> of database objects could be set a quota limit to avoid over use of disk\\n>>> space. A common case could be as follows: DBA could enable disk quota on a\\n>>> specified database list. DBA could set disk quota limit for\\n>>> tables/schemas/roles in these databases. Separate disk quota worker process\\n>>> will monitor the disk usage for these objects and detect the objects which\\n>>> exceed their quota limit. Queries loading data into these “out of disk\\n>>> quota” tables/schemas/roles will be cancelled.We are currently working at\\n>>> init implementation stage. We would like to propose our idea firstly and\\n>>> get feedbacks from community to do quick iteration.SQL Syntax (How to use\\n>>> disk quota)1 Specify the databases with disk quota enabled in GUC\\n>>> “diskquota_databases” in postgresql.conf and restart the database.2 DBA\\n>>> could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\\n>>> ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\\n>>> (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\\n>>> ‘3MB’);*\\n>>>\\n>>\\n>> just fast reaction - why QUOTA object?\\n>>\\n>> Isn't ALTER SET enough?\\n>>\\n>> Some like\\n>>\\n>> ALTER TABLE a1 SET quote = 1MB;\\n>> ALTER USER ...\\n>> ALTER SCHEMA ..\\n>>\\n>> New DDL commans looks like too hard hammer .\\n>>\\n>>\\n>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>> *3 Simulate a schema out of quota limit case: suppose table a1 and table\\n>>> a2 are both under schema s1.INSERT INTO a1 SELECT\\n>>> generate_series(1,1000);INSERT INTO a2 SELECT\\n>>> generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\\n>>> generate_series(1,1000);ERROR: schema's disk space quota exceededDROP\\n>>> TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\\n>>> generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\\n>>> following components.1. Quota Setting Store is where the disk quota setting\\n>>> to be stored and accessed. We plan to use catalog table pg_diskquota to\\n>>> store these information. pg_diskquota is\\n>>> like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\\n>>> diskquota name */ int16 quotatype; /* diskquota type name */ Oid\\n>>> quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\\n>>> diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\\n>>> MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\\n>>> change of database objects. We plan to use stat collector to detect the\\n>>> ‘active’ table list at initial stage. But stat collector has some\\n>>> limitation on finding the active table which is in a running transaction.\\n>>> Details see TODO section.3. Quota Size Checker is where to calculate the\\n>>> size and compare with quota limit for database objects. According to\\n>>> Pavel’s comment, autovacuum launcher and worker process could be a good\\n>>> reference to disk quota. So we plan to use a disk quota launcher daemon\\n>>> process and several disk quota worker process to finish this work. Launcher\\n>>> process is responsible for starting worker process based on a user defined\\n>>> database list from GUC. Worker process will connect to its target database\\n>>> and monitor the disk usage for objects in this database. In init stage of\\n>>> worker process, it will call calculate_total_relation_size() to calculate\\n>>> the size for each user table. After init stage, worker process will refresh\\n>>> the disk model every N seconds. Refreshing will only recalculate the size\\n>>> of tables in ‘active’ table list, which is generated by Quata Change\\n>>> Detector to minimize the cost.4. Quota Enforcement Operator is where to\\n>>> check for the quota limitation at postgres backend side. We will firstly\\n>>> implement it in ExecCheckRTPerms() as pre-running enforcement. It will\\n>>> check the disk quota of tables being inserted or updated, and report error\\n>>> if table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\\n>>> As a native feature, we plan to add more checkpoint to do running query\\n>>> enforcement. For example, if a disk quota lefts 10MB quota, a query could\\n>>> insert 1GB data. This query could be allowed in pre-running enforcement\\n>>> check, but will be cancelled in running query enforcement check. Therefore,\\n>>> it can improve the accurate of disk quota usage. To achieve this, we plan\\n>>> to add a checkpoint in lower API such as smgr_extened. Hence, the Quota\\n>>> Enforcement Operator will check the disk quota usage when smgr_extened is\\n>>> called. If the quota is over limited, current query will be cancelled.\\n>>> Highlight1. Native feature.Support native Create/Drop Disk Quota SQL\\n>>> statement.New catalog table pg_diskquota to store disk quota setting.2.\\n>>> Auto DML/DDL detection. Table\\n>>> create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\\n>>> Schema create/drop and Role create/drop will be detected by disk quota\\n>>> automatically. 3. Low cost disk quota checker.Worker process of disk quota\\n>>> need to refresh the disk usage model every N seconds. Since recalculate the\\n>>> file size using stat() system call is expensive for a large number of\\n>>> files, we use an ‘active’ table list to reduce the real work at each\\n>>> iteration. A basic experiment on our init stage implementation on database\\n>>> with 20K tables shows that the refresh cost is 1% cpu usage and will be\\n>>> finished within 50ms. Todo/LimitationBefore we propose our patch, we plan\\n>>> to enhance it with the following ideas:1. Setting database list with disk\\n>>> quota enabled dynamically without restart database. Since we have the disk\\n>>> quota launcher process, it could detect the new ‘diskquota_databases’ list\\n>>> and start/stop the corresponding disk quota worker process.2. Enforcement\\n>>> when query is running. Considering the case when there is 10MB quota left,\\n>>> but next query will insert 10GB data. Current enforcement design will allow\\n>>> this query to be executed. This is limited by the ‘active’ table detection\\n>>> is generated by stat collector. Postgres backend will only send table stat\\n>>> information to collector only when the transaction ends. We need a new way\\n>>> to detect the ‘active’ table even when this table is being modified inside\\n>>> a running transaction.3. Monitor unlimited number of databases. Current we\\n>>> set the max number of disk quota worker process to be 10 to reduce the\\n>>> affection normal workload. But how about if we want to monitor the disk\\n>>> quota of more than 10 databases? Our solution is to let disk quota launcher\\n>>> to manage a queue of database need to be monitored. And disk quota worker\\n>>> process consuming the queue and refresh the disk usage/quota for this\\n>>> database. After some periods, worker will return the database to the queue,\\n>>> and fetch the top database from queue to process. The period determine the\\n>>> delay of detecting disk quota change. To implement this feature, we need to\\n>>> support a subprocess of postmaster to rebind to another database instead of\\n>>> the database binded in InitPostgres().4. Support active table detection on\\n>>> vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\\n>>> are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\\n>>> Flack for the former comments on disk quota feature. Any comments on how to\\n>>> improve disk quota feature are appreciated.*\\n>>>\\n>>>\\n>>> On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com>\\n>>> wrote:\\n>>>\\n>>>>\\n>>>>\\n>>>> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>>>\\n>>>>> Thanks Pavel.\\n>>>>> Your patch did enforcement on storage level(md.c or we could also use\\n>>>>> smgr_extend). It's straight forward.\\n>>>>> But I prefer to implement disk_quota as a feature with following\\n>>>>> objectives:\\n>>>>> 1 set/alter disk quota setting on different database objects, e.g.\\n>>>>> user, database, schema etc. not only a general GUC, but we could set\\n>>>>> separate quota limit for a specific objects.\\n>>>>> 2 enforcement operator should work at two positions: before query is\\n>>>>> running and when query is running. The latter one's implementation maybe\\n>>>>> similar to your patch.\\n>>>>>\\n>>>>\\n>>>> The patch was just example. The resource quotes should be more complex\\n>>>> - per partition, table, schema, database, user - so GUC are possible, but\\n>>>> not very user friendly.\\n>>>>\\n>>>> Our case is specific, but not too much. The servers are used for\\n>>>> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\\n>>>> SELECT). We need to solve limits immediately. The implementation is simple,\\n>>>> so I did it. Same implementation on database level, or schema level needs\\n>>>> some more locks, so it will not be too effective. The resource management\\n>>>> can be complex very complex, and I expect so it will be hard work.\\n>>>>\\n>>>> Regards\\n>>>>\\n>>>> Pavel\\n>>>>\\n>>>>\\n>>>>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com\\n>>>>> > wrote:\\n>>>>>\\n>>>>>> Hi\\n>>>>>>\\n>>>>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>>>>>\\n>>>>>>> Thanks Chapman.\\n>>>>>>> @Pavel, could you please explain more about your second suggestion\\n>>>>>>> \\\"implement some quotas on storage level?\\\"\\n>>>>>>>\\n>>>>>>\\n>>>>>> See attached patch - it is very simple - and good enough for our\\n>>>>>> purposes.\\n>>>>>>\\n>>>>>> Regards\\n>>>>>>\\n>>>>>> Pavel\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>> We will not keep the long-lived processes attach to all\\n>>>>>>> databases(just like you mentioned servers with thousands of\\n>>>>>>> databases)\\n>>>>>>> And you are right, we could share ideas with autovacuum process,\\n>>>>>>> fork worker processes in need.\\n>>>>>>> \\\"autovacuum checks for tables that have had a large number of\\n>>>>>>> inserted, updated or deleted tuples. These checks use the statistics\\n>>>>>>> collection facility\\\"\\n>>>>>>> diskquota process is similar to autovacuum at caring about insert,\\n>>>>>>> but the difference is that it also care about vucuum full, truncate and\\n>>>>>>> drop. While update and delete may not be interested since no file change\\n>>>>>>> happens. So a separate diskquota process is preferred.\\n>>>>>>>\\n>>>>>>> So if we implemented disk quota as a full native feature, and in\\n>>>>>>> the first initial version I prefer to implement the following features:\\n>>>>>>> 1 Fork diskquota launcher process under Postmaster serverloop, which\\n>>>>>>> is long-lived.\\n>>>>>>> 2 Diskquota launcher process is responsible for creating diskquota\\n>>>>>>> worker process for every database.\\n>>>>>>> 3 DIskquota setting is stored in a separate catalog table for each\\n>>>>>>> database.\\n>>>>>>> 4 Initialization stage, Diskquota launcher process creates diskquota\\n>>>>>>> worker process for all the databases(traverse like autovacuum).\\n>>>>>>> Worker process calculates disk usage of db objects and\\n>>>>>>> their diskquota setting. If any db object exceeds its quota limit, put them\\n>>>>>>> into the blacklist in the shared memory, which will later be used by\\n>>>>>>> enforcement operator. Worker process exits when works are done.\\n>>>>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\\n>>>>>>> process for the database with a large number of insert, copy, truncate,\\n>>>>>>> drop etc. or create disk quota statement. Worker process updates the file\\n>>>>>>> size for db objects containing the result relation, and compare with the\\n>>>>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\\n>>>>>>> remove from blacklist vice versa. Worker process exits when works\\n>>>>>>> are done and a GUC could control the frequency of worker process restart to\\n>>>>>>> a specific database. As you know, this GUC also controls the delay when we\\n>>>>>>> do enforcement.\\n>>>>>>> 6 Enforcement. When postgres backend executes queries, check the\\n>>>>>>> blacklist in shared memory to determine whether the query is allowed(before\\n>>>>>>> execute) or need rollback(is executing)?\\n>>>>>>>\\n>>>>>>> If we implemented disk quota as an extension, we could just use\\n>>>>>>> background worker to start diskquota launcher process and use\\n>>>>>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\\n>>>>>>> processes by the launcher process as suggested by @Chapman.\\n>>>>>>> Diskquota setting could be stored in user table in a separate schema for\\n>>>>>>> each database(Schema and table created by create extension statement) just\\n>>>>>>> like what Heikki has done in pg_quota project. But in this case, we need to\\n>>>>>>> create extension for each database before diskquota worker process can be\\n>>>>>>> set up for that database.\\n>>>>>>>\\n>>>>>>> Any comments on the above design and which is preferred, native\\n>>>>>>> feature or extension as the POC?\\n>>>>>>>\\n>>>>>>>\\n>>>>>>> -- Hubert\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <\\n>>>>>>> pavel.stehule@gmail.com> wrote:\\n>>>>>>>\\n>>>>>>>>\\n>>>>>>>>\\n>>>>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\\n>>>>>>>>\\n>>>>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n>>>>>>>>>\\n>>>>>>>>> > 2 Keep one worker process for each database. But using a\\n>>>>>>>>> parent/global\\n>>>>>>>>> > quota worker process to manage the lifecycle of database level\\n>>>>>>>>> worker\\n>>>>>>>>> > processes. It could handle the newly created database(avoid\\n>>>>>>>>> restart\\n>>>>>>>>> > database) and save resource when a database is not used. But\\n>>>>>>>>> this needs to\\n>>>>>>>>> > change worker process to be hierarchical. Postmaster becomes the\\n>>>>>>>>> grandfather\\n>>>>>>>>> > of database level worker processes in this case.\\n>>>>>>>>>\\n>>>>>>>>> I am using background workers this way in 9.5 at $work.\\n>>>>>>>>>\\n>>>>>>>>> In my case, one worker lives forever, wakes up on a set period, and\\n>>>>>>>>> starts a short-lived worker for every database, waiting for each\\n>>>>>>>>> one before starting the next.\\n>>>>>>>>>\\n>>>>>>>>> It was straightforward to implement. Looking back over the code,\\n>>>>>>>>> I see the global worker assigns its own PID to\\n>>>>>>>>> worker.bgw_notify_pid\\n>>>>>>>>> of each of its children, and also obtains a handle for each child\\n>>>>>>>>> from RegisterDynamicBackgroundWorker().\\n>>>>>>>>>\\n>>>>>>>>> I imagine the global quota worker would prefer to start workers\\n>>>>>>>>> for every database and then just wait for notifications from any\\n>>>>>>>>> of them, but that seems equally straightforward at first glance.\\n>>>>>>>>>\\n>>>>>>>>\\n>>>>>>>> There are servers with thousands databases. Worker per database is\\n>>>>>>>> not good idea.\\n>>>>>>>>\\n>>>>>>>> It should to share ideas, code with autovacuum process.\\n>>>>>>>>\\n>>>>>>>> Not sure, how to effective implementation based on bg workers can\\n>>>>>>>> be. On servers with large set of databases, large set of tables it can\\n>>>>>>>> identify too big table too late.\\n>>>>>>>>\\n>>>>>>>> Isn't better to implement some quotas on storage level?\\n>>>>>>>>\\n>>>>>>>> Regards\\n>>>>>>>>\\n>>>>>>>> Pavel\\n>>>>>>>>\\n>>>>>>>>\\n>>>>>>>>\\n>>>>>>>>> -Chap\\n>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>> --\\n>>>>>>> Thanks\\n>>>>>>>\\n>>>>>>> Hubert Zhang\\n>>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>\\n>>>>>\\n>>>>> --\\n>>>>> Thanks\\n>>>>>\\n>>>>> Hubert Zhang\\n>>>>>\\n>>>>\\n>>>>\\n>>>\\n>>>\\n>>> --\\n>>> Thanks\\n>>>\\n>>> Hubert Zhang\\n>>>\\n>>\\n>\\n> --\\n> Thanks\\n>\\n> Hubert Zhang\\n>\\n\\npá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer .It's an option. Prefer to consider quota setting store together:CREATE DISK QUOTA way is more nature to store quota setting in a separate pg_diskquota catalogWhile ALTER SET way is more close to store quota setting in pg_class, pg_role, pg_namespace. etc in an integrated way.(Note that here I mean nature/close is not must, ALTER SET could also store in pg_diskquota and vice versa.)I have not a problem with new special table for storing this information. But it looks like redundant to current GUC configuration and limits. Can be messy do some work with ALTER ROLE, and some work via CREATE QUOTE.RegardsPavel Here are some differences I can think of:1 pg_role is a global catalog, not per database level. It's harder to tracker the user's disk usage in the whole clusters(considering 1000+ databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it only tracks the user's disk usage inside the current database.2 using separate pg_diskquota could add more field except for quota limit without adding too many fields in pg_class, e.g. red zone to give the user a warning or the current disk usage of the db objects.On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:Hi all,We redesign disk quota feature based on the comments from Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk quota feature is used to support multi-tenancy environment, different level of database objects could be set a quota limit to avoid over use of disk space. A common case could be as follows: DBA could enable disk quota on a specified database list. DBA could set disk quota limit for tables/schemas/roles in these databases. Separate disk quota worker process will monitor the disk usage for these objects and detect the objects which exceed their quota limit. Queries loading data into these “out of disk quota” tables/schemas/roles will be cancelled.We are currently working at init implementation stage. We would like to propose our idea firstly and get feedbacks from community to do quick iteration.SQL Syntax (How to use disk quota)1 Specify the databases with disk quota enabled in GUC “diskquota_databases” in postgresql.conf and restart the database.2 DBA could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer . 3 Simulate a schema out of quota limit case: suppose table a1 and table a2 are both under schema s1.INSERT INTO a1 SELECT generate_series(1,1000);INSERT INTO a2 SELECT generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);ERROR: schema's disk space quota exceededDROP TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the following components.1. Quota Setting Store is where the disk quota setting to be stored and accessed. We plan to use catalog table pg_diskquota to store these information. pg_diskquota is like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /* diskquota name */ int16 quotatype; /* diskquota type name */ Oid quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /* diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size change of database objects. We plan to use stat collector to detect the ‘active’ table list at initial stage. But stat collector has some limitation on finding the active table which is in a running transaction. Details see TODO section.3. Quota Size Checker is where to calculate the size and compare with quota limit for database objects. According to Pavel’s comment, autovacuum launcher and worker process could be a good reference to disk quota. So we plan to use a disk quota launcher daemon process and several disk quota worker process to finish this work. Launcher process is responsible for starting worker process based on a user defined database list from GUC. Worker process will connect to its target database and monitor the disk usage for objects in this database. In init stage of worker process, it will call calculate_total_relation_size() to calculate the size for each user table. After init stage, worker process will refresh the disk model every N seconds. Refreshing will only recalculate the size of tables in ‘active’ table list, which is generated by Quata Change Detector to minimize the cost.4. Quota Enforcement Operator is where to check for the quota limitation at postgres backend side. We will firstly implement it in ExecCheckRTPerms() as pre-running enforcement. It will check the disk quota of tables being inserted or updated, and report error if table’s or table’s schema’s or table’s owner’s quota limit is exceeded. As a native feature, we plan to add more checkpoint to do running query enforcement. For example, if a disk quota lefts 10MB quota, a query could insert 1GB data. This query could be allowed in pre-running enforcement check, but will be cancelled in running query enforcement check. Therefore, it can improve the accurate of disk quota usage. To achieve this, we plan to add a checkpoint in lower API such as smgr_extened. Hence, the Quota Enforcement Operator will check the disk quota usage when smgr_extened is called. If the quota is over limited, current query will be cancelled. Highlight1. Native feature.Support native Create/Drop Disk Quota SQL statement.New catalog table pg_diskquota to store disk quota setting.2. Auto DML/DDL detection. Table create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change, Schema create/drop and Role create/drop will be detected by disk quota automatically. 3. Low cost disk quota checker.Worker process of disk quota need to refresh the disk usage model every N seconds. Since recalculate the file size using stat() system call is expensive for a large number of files, we use an ‘active’ table list to reduce the real work at each iteration. A basic experiment on our init stage implementation on database with 20K tables shows that the refresh cost is 1% cpu usage and will be finished within 50ms. Todo/LimitationBefore we propose our patch, we plan to enhance it with the following ideas:1. Setting database list with disk quota enabled dynamically without restart database. Since we have the disk quota launcher process, it could detect the new ‘diskquota_databases’ list and start/stop the corresponding disk quota worker process.2. Enforcement when query is running. Considering the case when there is 10MB quota left, but next query will insert 10GB data. Current enforcement design will allow this query to be executed. This is limited by the ‘active’ table detection is generated by stat collector. Postgres backend will only send table stat information to collector only when the transaction ends. We need a new way to detect the ‘active’ table even when this table is being modified inside a running transaction.3. Monitor unlimited number of databases. Current we set the max number of disk quota worker process to be 10 to reduce the affection normal workload. But how about if we want to monitor the disk quota of more than 10 databases? Our solution is to let disk quota launcher to manage a queue of database need to be monitored. And disk quota worker process consuming the queue and refresh the disk usage/quota for this database. After some periods, worker will return the database to the queue, and fetch the top database from queue to process. The period determine the delay of detecting disk quota change. To implement this feature, we need to support a subprocess of postmaster to rebind to another database instead of the database binded in InitPostgres().4. Support active table detection on vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman Flack for the former comments on disk quota feature. Any comments on how to improve disk quota feature are appreciated.On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \\\"implement some quotas on storage level?\\\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \\\"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\\\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n\\n> 2 Keep one worker process for each database. But using a parent/global\\n> quota worker process to manage the lifecycle of database level worker\\n> processes. It could handle the newly created database(avoid restart\\n> database) and save resource when a database is not used. But this needs to\\n> change worker process to be hierarchical. Postmaster becomes the grandfather\\n> of database level worker processes in this case.\\n\\nI am using background workers this way in 9.5 at $work.\\n\\nIn my case, one worker lives forever, wakes up on a set period, and\\nstarts a short-lived worker for every database, waiting for each\\none before starting the next.\\n\\nIt was straightforward to implement. Looking back over the code,\\nI see the global worker assigns its own PID to worker.bgw_notify_pid\\nof each of its children, and also obtains a handle for each child\\nfrom RegisterDynamicBackgroundWorker().\\n\\nI imagine the global quota worker would prefer to start workers\\nfor every database and then just wait for notifications from any\\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\\n\\n-Chap\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\",\n \"msg_date\": \"Sat, 22 Sep 2018 05:17:14 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal for disk quota feature\"\n },\n {\n \"msg_contents\": \">\\n> But it looks like redundant to current GUC configuration and limits\\n\\nwhat do you mean by current GUC configuration? Is that the general block\\nnumber limit in your patch? If yes, the difference between GUC and\\npg_diskquota catalog is that pg_diskquota will store different quota limit\\nfor the different role, schema or table instead of a single GUC value.\\n\\nOn Sat, Sep 22, 2018 at 11:17 AM Pavel Stehule <pavel.stehule@gmail.com>\\nwrote:\\n\\n>\\n>\\n> pá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:\\n>\\n>> just fast reaction - why QUOTA object?\\n>>> Isn't ALTER SET enough?\\n>>> Some like\\n>>> ALTER TABLE a1 SET quote = 1MB;\\n>>> ALTER USER ...\\n>>> ALTER SCHEMA ..\\n>>> New DDL commans looks like too hard hammer .\\n>>\\n>>\\n>> It's an option. Prefer to consider quota setting store together:\\n>> CREATE DISK QUOTA way is more nature to store quota setting in a separate\\n>> pg_diskquota catalog\\n>> While ALTER SET way is more close to store quota setting in pg_class,\\n>> pg_role, pg_namespace. etc in an integrated way.\\n>> (Note that here I mean nature/close is not must, ALTER SET could also\\n>> store in pg_diskquota and vice versa.)\\n>>\\n>\\n> I have not a problem with new special table for storing this information.\\n> But it looks like redundant to current GUC configuration and limits. Can be\\n> messy do some work with ALTER ROLE, and some work via CREATE QUOTE.\\n>\\n> Regards\\n>\\n> Pavel\\n>\\n>\\n>> Here are some differences I can think of:\\n>> 1 pg_role is a global catalog, not per database level. It's harder to\\n>> tracker the user's disk usage in the whole clusters(considering 1000+\\n>> databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it\\n>> only tracks the user's disk usage inside the current database.\\n>> 2 using separate pg_diskquota could add more field except for quota limit\\n>> without adding too many fields in pg_class, e.g. red zone to give the user\\n>> a warning or the current disk usage of the db objects.\\n>>\\n>> On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com>\\n>> wrote:\\n>>\\n>>>\\n>>>\\n>>> pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io>\\n>>> napsal:\\n>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>> *Hi all,We redesign disk quota feature based on the comments from Pavel\\n>>>> Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk\\n>>>> quota feature is used to support multi-tenancy environment, different level\\n>>>> of database objects could be set a quota limit to avoid over use of disk\\n>>>> space. A common case could be as follows: DBA could enable disk quota on a\\n>>>> specified database list. DBA could set disk quota limit for\\n>>>> tables/schemas/roles in these databases. Separate disk quota worker process\\n>>>> will monitor the disk usage for these objects and detect the objects which\\n>>>> exceed their quota limit. Queries loading data into these “out of disk\\n>>>> quota” tables/schemas/roles will be cancelled.We are currently working at\\n>>>> init implementation stage. We would like to propose our idea firstly and\\n>>>> get feedbacks from community to do quick iteration.SQL Syntax (How to use\\n>>>> disk quota)1 Specify the databases with disk quota enabled in GUC\\n>>>> “diskquota_databases” in postgresql.conf and restart the database.2 DBA\\n>>>> could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\\n>>>> ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\\n>>>> (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\\n>>>> ‘3MB’);*\\n>>>>\\n>>>\\n>>> just fast reaction - why QUOTA object?\\n>>>\\n>>> Isn't ALTER SET enough?\\n>>>\\n>>> Some like\\n>>>\\n>>> ALTER TABLE a1 SET quote = 1MB;\\n>>> ALTER USER ...\\n>>> ALTER SCHEMA ..\\n>>>\\n>>> New DDL commans looks like too hard hammer .\\n>>>\\n>>>\\n>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>> *3 Simulate a schema out of quota limit case: suppose table a1 and\\n>>>> table a2 are both under schema s1.INSERT INTO a1 SELECT\\n>>>> generate_series(1,1000);INSERT INTO a2 SELECT\\n>>>> generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\\n>>>> generate_series(1,1000);ERROR: schema's disk space quota exceededDROP\\n>>>> TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\\n>>>> generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\\n>>>> following components.1. Quota Setting Store is where the disk quota setting\\n>>>> to be stored and accessed. We plan to use catalog table pg_diskquota to\\n>>>> store these information. pg_diskquota is\\n>>>> like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\\n>>>> diskquota name */ int16 quotatype; /* diskquota type name */ Oid\\n>>>> quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\\n>>>> diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\\n>>>> MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\\n>>>> change of database objects. We plan to use stat collector to detect the\\n>>>> ‘active’ table list at initial stage. But stat collector has some\\n>>>> limitation on finding the active table which is in a running transaction.\\n>>>> Details see TODO section.3. Quota Size Checker is where to calculate the\\n>>>> size and compare with quota limit for database objects. According to\\n>>>> Pavel’s comment, autovacuum launcher and worker process could be a good\\n>>>> reference to disk quota. So we plan to use a disk quota launcher daemon\\n>>>> process and several disk quota worker process to finish this work. Launcher\\n>>>> process is responsible for starting worker process based on a user defined\\n>>>> database list from GUC. Worker process will connect to its target database\\n>>>> and monitor the disk usage for objects in this database. In init stage of\\n>>>> worker process, it will call calculate_total_relation_size() to calculate\\n>>>> the size for each user table. After init stage, worker process will refresh\\n>>>> the disk model every N seconds. Refreshing will only recalculate the size\\n>>>> of tables in ‘active’ table list, which is generated by Quata Change\\n>>>> Detector to minimize the cost.4. Quota Enforcement Operator is where to\\n>>>> check for the quota limitation at postgres backend side. We will firstly\\n>>>> implement it in ExecCheckRTPerms() as pre-running enforcement. It will\\n>>>> check the disk quota of tables being inserted or updated, and report error\\n>>>> if table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\\n>>>> As a native feature, we plan to add more checkpoint to do running query\\n>>>> enforcement. For example, if a disk quota lefts 10MB quota, a query could\\n>>>> insert 1GB data. This query could be allowed in pre-running enforcement\\n>>>> check, but will be cancelled in running query enforcement check. Therefore,\\n>>>> it can improve the accurate of disk quota usage. To achieve this, we plan\\n>>>> to add a checkpoint in lower API such as smgr_extened. Hence, the Quota\\n>>>> Enforcement Operator will check the disk quota usage when smgr_extened is\\n>>>> called. If the quota is over limited, current query will be cancelled.\\n>>>> Highlight1. Native feature.Support native Create/Drop Disk Quota SQL\\n>>>> statement.New catalog table pg_diskquota to store disk quota setting.2.\\n>>>> Auto DML/DDL detection. Table\\n>>>> create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\\n>>>> Schema create/drop and Role create/drop will be detected by disk quota\\n>>>> automatically. 3. Low cost disk quota checker.Worker process of disk quota\\n>>>> need to refresh the disk usage model every N seconds. Since recalculate the\\n>>>> file size using stat() system call is expensive for a large number of\\n>>>> files, we use an ‘active’ table list to reduce the real work at each\\n>>>> iteration. A basic experiment on our init stage implementation on database\\n>>>> with 20K tables shows that the refresh cost is 1% cpu usage and will be\\n>>>> finished within 50ms. Todo/LimitationBefore we propose our patch, we plan\\n>>>> to enhance it with the following ideas:1. Setting database list with disk\\n>>>> quota enabled dynamically without restart database. Since we have the disk\\n>>>> quota launcher process, it could detect the new ‘diskquota_databases’ list\\n>>>> and start/stop the corresponding disk quota worker process.2. Enforcement\\n>>>> when query is running. Considering the case when there is 10MB quota left,\\n>>>> but next query will insert 10GB data. Current enforcement design will allow\\n>>>> this query to be executed. This is limited by the ‘active’ table detection\\n>>>> is generated by stat collector. Postgres backend will only send table stat\\n>>>> information to collector only when the transaction ends. We need a new way\\n>>>> to detect the ‘active’ table even when this table is being modified inside\\n>>>> a running transaction.3. Monitor unlimited number of databases. Current we\\n>>>> set the max number of disk quota worker process to be 10 to reduce the\\n>>>> affection normal workload. But how about if we want to monitor the disk\\n>>>> quota of more than 10 databases? Our solution is to let disk quota launcher\\n>>>> to manage a queue of database need to be monitored. And disk quota worker\\n>>>> process consuming the queue and refresh the disk usage/quota for this\\n>>>> database. After some periods, worker will return the database to the queue,\\n>>>> and fetch the top database from queue to process. The period determine the\\n>>>> delay of detecting disk quota change. To implement this feature, we need to\\n>>>> support a subprocess of postmaster to rebind to another database instead of\\n>>>> the database binded in InitPostgres().4. Support active table detection on\\n>>>> vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\\n>>>> are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\\n>>>> Flack for the former comments on disk quota feature. Any comments on how to\\n>>>> improve disk quota feature are appreciated.*\\n>>>>\\n>>>>\\n>>>> On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com\\n>>>> > wrote:\\n>>>>\\n>>>>>\\n>>>>>\\n>>>>> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>>>>\\n>>>>>> Thanks Pavel.\\n>>>>>> Your patch did enforcement on storage level(md.c or we could also use\\n>>>>>> smgr_extend). It's straight forward.\\n>>>>>> But I prefer to implement disk_quota as a feature with following\\n>>>>>> objectives:\\n>>>>>> 1 set/alter disk quota setting on different database objects, e.g.\\n>>>>>> user, database, schema etc. not only a general GUC, but we could set\\n>>>>>> separate quota limit for a specific objects.\\n>>>>>> 2 enforcement operator should work at two positions: before query is\\n>>>>>> running and when query is running. The latter one's implementation maybe\\n>>>>>> similar to your patch.\\n>>>>>>\\n>>>>>\\n>>>>> The patch was just example. The resource quotes should be more complex\\n>>>>> - per partition, table, schema, database, user - so GUC are possible, but\\n>>>>> not very user friendly.\\n>>>>>\\n>>>>> Our case is specific, but not too much. The servers are used for\\n>>>>> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\\n>>>>> SELECT). We need to solve limits immediately. The implementation is simple,\\n>>>>> so I did it. Same implementation on database level, or schema level needs\\n>>>>> some more locks, so it will not be too effective. The resource management\\n>>>>> can be complex very complex, and I expect so it will be hard work.\\n>>>>>\\n>>>>> Regards\\n>>>>>\\n>>>>> Pavel\\n>>>>>\\n>>>>>\\n>>>>>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <\\n>>>>>> pavel.stehule@gmail.com> wrote:\\n>>>>>>\\n>>>>>>> Hi\\n>>>>>>>\\n>>>>>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>>>>>>\\n>>>>>>>> Thanks Chapman.\\n>>>>>>>> @Pavel, could you please explain more about your second\\n>>>>>>>> suggestion \\\"implement some quotas on storage level?\\\"\\n>>>>>>>>\\n>>>>>>>\\n>>>>>>> See attached patch - it is very simple - and good enough for our\\n>>>>>>> purposes.\\n>>>>>>>\\n>>>>>>> Regards\\n>>>>>>>\\n>>>>>>> Pavel\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>> We will not keep the long-lived processes attach to all\\n>>>>>>>> databases(just like you mentioned servers with thousands of\\n>>>>>>>> databases)\\n>>>>>>>> And you are right, we could share ideas with autovacuum process,\\n>>>>>>>> fork worker processes in need.\\n>>>>>>>> \\\"autovacuum checks for tables that have had a large number of\\n>>>>>>>> inserted, updated or deleted tuples. These checks use the statistics\\n>>>>>>>> collection facility\\\"\\n>>>>>>>> diskquota process is similar to autovacuum at caring about insert,\\n>>>>>>>> but the difference is that it also care about vucuum full, truncate and\\n>>>>>>>> drop. While update and delete may not be interested since no file change\\n>>>>>>>> happens. So a separate diskquota process is preferred.\\n>>>>>>>>\\n>>>>>>>> So if we implemented disk quota as a full native feature, and in\\n>>>>>>>> the first initial version I prefer to implement the following features:\\n>>>>>>>> 1 Fork diskquota launcher process under Postmaster serverloop,\\n>>>>>>>> which is long-lived.\\n>>>>>>>> 2 Diskquota launcher process is responsible for creating diskquota\\n>>>>>>>> worker process for every database.\\n>>>>>>>> 3 DIskquota setting is stored in a separate catalog table for each\\n>>>>>>>> database.\\n>>>>>>>> 4 Initialization stage, Diskquota launcher process creates\\n>>>>>>>> diskquota worker process for all the databases(traverse like\\n>>>>>>>> autovacuum). Worker process calculates disk usage of db objects and\\n>>>>>>>> their diskquota setting. If any db object exceeds its quota limit, put them\\n>>>>>>>> into the blacklist in the shared memory, which will later be used by\\n>>>>>>>> enforcement operator. Worker process exits when works are done.\\n>>>>>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\\n>>>>>>>> process for the database with a large number of insert, copy, truncate,\\n>>>>>>>> drop etc. or create disk quota statement. Worker process updates the file\\n>>>>>>>> size for db objects containing the result relation, and compare with the\\n>>>>>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\\n>>>>>>>> remove from blacklist vice versa. Worker process exits when works\\n>>>>>>>> are done and a GUC could control the frequency of worker process restart to\\n>>>>>>>> a specific database. As you know, this GUC also controls the delay when we\\n>>>>>>>> do enforcement.\\n>>>>>>>> 6 Enforcement. When postgres backend executes queries, check the\\n>>>>>>>> blacklist in shared memory to determine whether the query is allowed(before\\n>>>>>>>> execute) or need rollback(is executing)?\\n>>>>>>>>\\n>>>>>>>> If we implemented disk quota as an extension, we could just use\\n>>>>>>>> background worker to start diskquota launcher process and use\\n>>>>>>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\\n>>>>>>>> processes by the launcher process as suggested by @Chapman.\\n>>>>>>>> Diskquota setting could be stored in user table in a separate schema for\\n>>>>>>>> each database(Schema and table created by create extension statement) just\\n>>>>>>>> like what Heikki has done in pg_quota project. But in this case, we need to\\n>>>>>>>> create extension for each database before diskquota worker process can be\\n>>>>>>>> set up for that database.\\n>>>>>>>>\\n>>>>>>>> Any comments on the above design and which is preferred, native\\n>>>>>>>> feature or extension as the POC?\\n>>>>>>>>\\n>>>>>>>>\\n>>>>>>>> -- Hubert\\n>>>>>>>>\\n>>>>>>>>\\n>>>>>>>>\\n>>>>>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <\\n>>>>>>>> pavel.stehule@gmail.com> wrote:\\n>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\\n>>>>>>>>>\\n>>>>>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n>>>>>>>>>>\\n>>>>>>>>>> > 2 Keep one worker process for each database. But using a\\n>>>>>>>>>> parent/global\\n>>>>>>>>>> > quota worker process to manage the lifecycle of database level\\n>>>>>>>>>> worker\\n>>>>>>>>>> > processes. It could handle the newly created database(avoid\\n>>>>>>>>>> restart\\n>>>>>>>>>> > database) and save resource when a database is not used. But\\n>>>>>>>>>> this needs to\\n>>>>>>>>>> > change worker process to be hierarchical. Postmaster becomes\\n>>>>>>>>>> the grandfather\\n>>>>>>>>>> > of database level worker processes in this case.\\n>>>>>>>>>>\\n>>>>>>>>>> I am using background workers this way in 9.5 at $work.\\n>>>>>>>>>>\\n>>>>>>>>>> In my case, one worker lives forever, wakes up on a set period,\\n>>>>>>>>>> and\\n>>>>>>>>>> starts a short-lived worker for every database, waiting for each\\n>>>>>>>>>> one before starting the next.\\n>>>>>>>>>>\\n>>>>>>>>>> It was straightforward to implement. Looking back over the code,\\n>>>>>>>>>> I see the global worker assigns its own PID to\\n>>>>>>>>>> worker.bgw_notify_pid\\n>>>>>>>>>> of each of its children, and also obtains a handle for each child\\n>>>>>>>>>> from RegisterDynamicBackgroundWorker().\\n>>>>>>>>>>\\n>>>>>>>>>> I imagine the global quota worker would prefer to start workers\\n>>>>>>>>>> for every database and then just wait for notifications from any\\n>>>>>>>>>> of them, but that seems equally straightforward at first glance.\\n>>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>> There are servers with thousands databases. Worker per database is\\n>>>>>>>>> not good idea.\\n>>>>>>>>>\\n>>>>>>>>> It should to share ideas, code with autovacuum process.\\n>>>>>>>>>\\n>>>>>>>>> Not sure, how to effective implementation based on bg workers can\\n>>>>>>>>> be. On servers with large set of databases, large set of tables it can\\n>>>>>>>>> identify too big table too late.\\n>>>>>>>>>\\n>>>>>>>>> Isn't better to implement some quotas on storage level?\\n>>>>>>>>>\\n>>>>>>>>> Regards\\n>>>>>>>>>\\n>>>>>>>>> Pavel\\n>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>>> -Chap\\n>>>>>>>>>>\\n>>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>\\n>>>>>>>>\\n>>>>>>>> --\\n>>>>>>>> Thanks\\n>>>>>>>>\\n>>>>>>>> Hubert Zhang\\n>>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>> --\\n>>>>>> Thanks\\n>>>>>>\\n>>>>>> Hubert Zhang\\n>>>>>>\\n>>>>>\\n>>>>>\\n>>>>\\n>>>>\\n>>>> --\\n>>>> Thanks\\n>>>>\\n>>>> Hubert Zhang\\n>>>>\\n>>>\\n>>\\n>> --\\n>> Thanks\\n>>\\n>> Hubert Zhang\\n>>\\n>\\n\\n-- \\nThanks\\n\\nHubert Zhang\\n\\nBut it looks like redundant to current GUC configuration and limitswhat do you mean by current GUC configuration? Is that the general block number limit in your patch? If yes, the difference between GUC and pg_diskquota catalog is that pg_diskquota will store different quota limit for the different role, schema or table instead of a single GUC value.On Sat, Sep 22, 2018 at 11:17 AM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer .It's an option. Prefer to consider quota setting store together:CREATE DISK QUOTA way is more nature to store quota setting in a separate pg_diskquota catalogWhile ALTER SET way is more close to store quota setting in pg_class, pg_role, pg_namespace. etc in an integrated way.(Note that here I mean nature/close is not must, ALTER SET could also store in pg_diskquota and vice versa.)I have not a problem with new special table for storing this information. But it looks like redundant to current GUC configuration and limits. Can be messy do some work with ALTER ROLE, and some work via CREATE QUOTE.RegardsPavel Here are some differences I can think of:1 pg_role is a global catalog, not per database level. It's harder to tracker the user's disk usage in the whole clusters(considering 1000+ databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it only tracks the user's disk usage inside the current database.2 using separate pg_diskquota could add more field except for quota limit without adding too many fields in pg_class, e.g. red zone to give the user a warning or the current disk usage of the db objects.On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:Hi all,We redesign disk quota feature based on the comments from Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk quota feature is used to support multi-tenancy environment, different level of database objects could be set a quota limit to avoid over use of disk space. A common case could be as follows: DBA could enable disk quota on a specified database list. DBA could set disk quota limit for tables/schemas/roles in these databases. Separate disk quota worker process will monitor the disk usage for these objects and detect the objects which exceed their quota limit. Queries loading data into these “out of disk quota” tables/schemas/roles will be cancelled.We are currently working at init implementation stage. We would like to propose our idea firstly and get feedbacks from community to do quick iteration.SQL Syntax (How to use disk quota)1 Specify the databases with disk quota enabled in GUC “diskquota_databases” in postgresql.conf and restart the database.2 DBA could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer . 3 Simulate a schema out of quota limit case: suppose table a1 and table a2 are both under schema s1.INSERT INTO a1 SELECT generate_series(1,1000);INSERT INTO a2 SELECT generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);ERROR: schema's disk space quota exceededDROP TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the following components.1. Quota Setting Store is where the disk quota setting to be stored and accessed. We plan to use catalog table pg_diskquota to store these information. pg_diskquota is like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /* diskquota name */ int16 quotatype; /* diskquota type name */ Oid quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /* diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size change of database objects. We plan to use stat collector to detect the ‘active’ table list at initial stage. But stat collector has some limitation on finding the active table which is in a running transaction. Details see TODO section.3. Quota Size Checker is where to calculate the size and compare with quota limit for database objects. According to Pavel’s comment, autovacuum launcher and worker process could be a good reference to disk quota. So we plan to use a disk quota launcher daemon process and several disk quota worker process to finish this work. Launcher process is responsible for starting worker process based on a user defined database list from GUC. Worker process will connect to its target database and monitor the disk usage for objects in this database. In init stage of worker process, it will call calculate_total_relation_size() to calculate the size for each user table. After init stage, worker process will refresh the disk model every N seconds. Refreshing will only recalculate the size of tables in ‘active’ table list, which is generated by Quata Change Detector to minimize the cost.4. Quota Enforcement Operator is where to check for the quota limitation at postgres backend side. We will firstly implement it in ExecCheckRTPerms() as pre-running enforcement. It will check the disk quota of tables being inserted or updated, and report error if table’s or table’s schema’s or table’s owner’s quota limit is exceeded. As a native feature, we plan to add more checkpoint to do running query enforcement. For example, if a disk quota lefts 10MB quota, a query could insert 1GB data. This query could be allowed in pre-running enforcement check, but will be cancelled in running query enforcement check. Therefore, it can improve the accurate of disk quota usage. To achieve this, we plan to add a checkpoint in lower API such as smgr_extened. Hence, the Quota Enforcement Operator will check the disk quota usage when smgr_extened is called. If the quota is over limited, current query will be cancelled. Highlight1. Native feature.Support native Create/Drop Disk Quota SQL statement.New catalog table pg_diskquota to store disk quota setting.2. Auto DML/DDL detection. Table create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change, Schema create/drop and Role create/drop will be detected by disk quota automatically. 3. Low cost disk quota checker.Worker process of disk quota need to refresh the disk usage model every N seconds. Since recalculate the file size using stat() system call is expensive for a large number of files, we use an ‘active’ table list to reduce the real work at each iteration. A basic experiment on our init stage implementation on database with 20K tables shows that the refresh cost is 1% cpu usage and will be finished within 50ms. Todo/LimitationBefore we propose our patch, we plan to enhance it with the following ideas:1. Setting database list with disk quota enabled dynamically without restart database. Since we have the disk quota launcher process, it could detect the new ‘diskquota_databases’ list and start/stop the corresponding disk quota worker process.2. Enforcement when query is running. Considering the case when there is 10MB quota left, but next query will insert 10GB data. Current enforcement design will allow this query to be executed. This is limited by the ‘active’ table detection is generated by stat collector. Postgres backend will only send table stat information to collector only when the transaction ends. We need a new way to detect the ‘active’ table even when this table is being modified inside a running transaction.3. Monitor unlimited number of databases. Current we set the max number of disk quota worker process to be 10 to reduce the affection normal workload. But how about if we want to monitor the disk quota of more than 10 databases? Our solution is to let disk quota launcher to manage a queue of database need to be monitored. And disk quota worker process consuming the queue and refresh the disk usage/quota for this database. After some periods, worker will return the database to the queue, and fetch the top database from queue to process. The period determine the delay of detecting disk quota change. To implement this feature, we need to support a subprocess of postmaster to rebind to another database instead of the database binded in InitPostgres().4. Support active table detection on vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman Flack for the former comments on disk quota feature. Any comments on how to improve disk quota feature are appreciated.On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \\\"implement some quotas on storage level?\\\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \\\"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\\\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n\\n> 2 Keep one worker process for each database. But using a parent/global\\n> quota worker process to manage the lifecycle of database level worker\\n> processes. It could handle the newly created database(avoid restart\\n> database) and save resource when a database is not used. But this needs to\\n> change worker process to be hierarchical. Postmaster becomes the grandfather\\n> of database level worker processes in this case.\\n\\nI am using background workers this way in 9.5 at $work.\\n\\nIn my case, one worker lives forever, wakes up on a set period, and\\nstarts a short-lived worker for every database, waiting for each\\none before starting the next.\\n\\nIt was straightforward to implement. Looking back over the code,\\nI see the global worker assigns its own PID to worker.bgw_notify_pid\\nof each of its children, and also obtains a handle for each child\\nfrom RegisterDynamicBackgroundWorker().\\n\\nI imagine the global quota worker would prefer to start workers\\nfor every database and then just wait for notifications from any\\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\\n\\n-Chap\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\\n\\n-- ThanksHubert Zhang\",\n \"msg_date\": \"Sat, 22 Sep 2018 14:48:03 +0800\",\n \"msg_from\": \"Hubert Zhang <hzhang@pivotal.io>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal for disk quota feature\"\n },\n {\n \"msg_contents\": \"so 22. 9. 2018 v 8:48 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:\\n\\n> But it looks like redundant to current GUC configuration and limits\\n>\\n> what do you mean by current GUC configuration? Is that the general block\\n> number limit in your patch? If yes, the difference between GUC and\\n> pg_diskquota catalog is that pg_diskquota will store different quota limit\\n> for the different role, schema or table instead of a single GUC value.\\n>\\n\\nstorage is not relevant in this moment.\\n\\nI don't see to consistent to sets some limits via SET command, or ALTER X\\nSET, and some other with CREATE QUOTA ON.\\n\\nThe quotas or object limits, resource limits are pretty useful and\\nnecessary, but I don't see these like new type of objects, it is much more\\nsome property of current objects. Because we have one syntax for this\\npurpose I prefer it. Because is not good to have two syntaxes for similar\\npurpose.\\n\\nSo instead CREATE DISC QUATA ON SCHEMA xxx some value I prefer\\n\\nALTER SCHEMA xxx SET disc_quota = xxx;\\n\\nThe functionality is +/- same. But ALTER XX SET was introduce first, and I\\ndon't feel comfortable to have any new syntax for similar purpose\\n\\nRegards\\n\\nPavel\\n\\n\\n\\n\\n\\n>\\n> On Sat, Sep 22, 2018 at 11:17 AM Pavel Stehule <pavel.stehule@gmail.com>\\n> wrote:\\n>\\n>>\\n>>\\n>> pá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io>\\n>> napsal:\\n>>\\n>>> just fast reaction - why QUOTA object?\\n>>>> Isn't ALTER SET enough?\\n>>>> Some like\\n>>>> ALTER TABLE a1 SET quote = 1MB;\\n>>>> ALTER USER ...\\n>>>> ALTER SCHEMA ..\\n>>>> New DDL commans looks like too hard hammer .\\n>>>\\n>>>\\n>>> It's an option. Prefer to consider quota setting store together:\\n>>> CREATE DISK QUOTA way is more nature to store quota setting in a\\n>>> separate pg_diskquota catalog\\n>>> While ALTER SET way is more close to store quota setting in pg_class,\\n>>> pg_role, pg_namespace. etc in an integrated way.\\n>>> (Note that here I mean nature/close is not must, ALTER SET could also\\n>>> store in pg_diskquota and vice versa.)\\n>>>\\n>>\\n>> I have not a problem with new special table for storing this information.\\n>> But it looks like redundant to current GUC configuration and limits. Can be\\n>> messy do some work with ALTER ROLE, and some work via CREATE QUOTE.\\n>>\\n>> Regards\\n>>\\n>> Pavel\\n>>\\n>>\\n>>> Here are some differences I can think of:\\n>>> 1 pg_role is a global catalog, not per database level. It's harder to\\n>>> tracker the user's disk usage in the whole clusters(considering 1000+\\n>>> databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it\\n>>> only tracks the user's disk usage inside the current database.\\n>>> 2 using separate pg_diskquota could add more field except for quota\\n>>> limit without adding too many fields in pg_class, e.g. red zone to give the\\n>>> user a warning or the current disk usage of the db objects.\\n>>>\\n>>> On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com>\\n>>> wrote:\\n>>>\\n>>>>\\n>>>>\\n>>>> pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io>\\n>>>> napsal:\\n>>>>\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>> *Hi all,We redesign disk quota feature based on the comments from\\n>>>>> Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically,\\n>>>>> disk quota feature is used to support multi-tenancy environment, different\\n>>>>> level of database objects could be set a quota limit to avoid over use of\\n>>>>> disk space. A common case could be as follows: DBA could enable disk quota\\n>>>>> on a specified database list. DBA could set disk quota limit for\\n>>>>> tables/schemas/roles in these databases. Separate disk quota worker process\\n>>>>> will monitor the disk usage for these objects and detect the objects which\\n>>>>> exceed their quota limit. Queries loading data into these “out of disk\\n>>>>> quota” tables/schemas/roles will be cancelled.We are currently working at\\n>>>>> init implementation stage. We would like to propose our idea firstly and\\n>>>>> get feedbacks from community to do quick iteration.SQL Syntax (How to use\\n>>>>> disk quota)1 Specify the databases with disk quota enabled in GUC\\n>>>>> “diskquota_databases” in postgresql.conf and restart the database.2 DBA\\n>>>>> could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\\n>>>>> ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\\n>>>>> (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\\n>>>>> ‘3MB’);*\\n>>>>>\\n>>>>\\n>>>> just fast reaction - why QUOTA object?\\n>>>>\\n>>>> Isn't ALTER SET enough?\\n>>>>\\n>>>> Some like\\n>>>>\\n>>>> ALTER TABLE a1 SET quote = 1MB;\\n>>>> ALTER USER ...\\n>>>> ALTER SCHEMA ..\\n>>>>\\n>>>> New DDL commans looks like too hard hammer .\\n>>>>\\n>>>>\\n>>>>\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>> *3 Simulate a schema out of quota limit case: suppose table a1 and\\n>>>>> table a2 are both under schema s1.INSERT INTO a1 SELECT\\n>>>>> generate_series(1,1000);INSERT INTO a2 SELECT\\n>>>>> generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\\n>>>>> generate_series(1,1000);ERROR: schema's disk space quota exceededDROP\\n>>>>> TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\\n>>>>> generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\\n>>>>> following components.1. Quota Setting Store is where the disk quota setting\\n>>>>> to be stored and accessed. We plan to use catalog table pg_diskquota to\\n>>>>> store these information. pg_diskquota is\\n>>>>> like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\\n>>>>> diskquota name */ int16 quotatype; /* diskquota type name */ Oid\\n>>>>> quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\\n>>>>> diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\\n>>>>> MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\\n>>>>> change of database objects. We plan to use stat collector to detect the\\n>>>>> ‘active’ table list at initial stage. But stat collector has some\\n>>>>> limitation on finding the active table which is in a running transaction.\\n>>>>> Details see TODO section.3. Quota Size Checker is where to calculate the\\n>>>>> size and compare with quota limit for database objects. According to\\n>>>>> Pavel’s comment, autovacuum launcher and worker process could be a good\\n>>>>> reference to disk quota. So we plan to use a disk quota launcher daemon\\n>>>>> process and several disk quota worker process to finish this work. Launcher\\n>>>>> process is responsible for starting worker process based on a user defined\\n>>>>> database list from GUC. Worker process will connect to its target database\\n>>>>> and monitor the disk usage for objects in this database. In init stage of\\n>>>>> worker process, it will call calculate_total_relation_size() to calculate\\n>>>>> the size for each user table. After init stage, worker process will refresh\\n>>>>> the disk model every N seconds. Refreshing will only recalculate the size\\n>>>>> of tables in ‘active’ table list, which is generated by Quata Change\\n>>>>> Detector to minimize the cost.4. Quota Enforcement Operator is where to\\n>>>>> check for the quota limitation at postgres backend side. We will firstly\\n>>>>> implement it in ExecCheckRTPerms() as pre-running enforcement. It will\\n>>>>> check the disk quota of tables being inserted or updated, and report error\\n>>>>> if table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\\n>>>>> As a native feature, we plan to add more checkpoint to do running query\\n>>>>> enforcement. For example, if a disk quota lefts 10MB quota, a query could\\n>>>>> insert 1GB data. This query could be allowed in pre-running enforcement\\n>>>>> check, but will be cancelled in running query enforcement check. Therefore,\\n>>>>> it can improve the accurate of disk quota usage. To achieve this, we plan\\n>>>>> to add a checkpoint in lower API such as smgr_extened. Hence, the Quota\\n>>>>> Enforcement Operator will check the disk quota usage when smgr_extened is\\n>>>>> called. If the quota is over limited, current query will be cancelled.\\n>>>>> Highlight1. Native feature.Support native Create/Drop Disk Quota SQL\\n>>>>> statement.New catalog table pg_diskquota to store disk quota setting.2.\\n>>>>> Auto DML/DDL detection. Table\\n>>>>> create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\\n>>>>> Schema create/drop and Role create/drop will be detected by disk quota\\n>>>>> automatically. 3. Low cost disk quota checker.Worker process of disk quota\\n>>>>> need to refresh the disk usage model every N seconds. Since recalculate the\\n>>>>> file size using stat() system call is expensive for a large number of\\n>>>>> files, we use an ‘active’ table list to reduce the real work at each\\n>>>>> iteration. A basic experiment on our init stage implementation on database\\n>>>>> with 20K tables shows that the refresh cost is 1% cpu usage and will be\\n>>>>> finished within 50ms. Todo/LimitationBefore we propose our patch, we plan\\n>>>>> to enhance it with the following ideas:1. Setting database list with disk\\n>>>>> quota enabled dynamically without restart database. Since we have the disk\\n>>>>> quota launcher process, it could detect the new ‘diskquota_databases’ list\\n>>>>> and start/stop the corresponding disk quota worker process.2. Enforcement\\n>>>>> when query is running. Considering the case when there is 10MB quota left,\\n>>>>> but next query will insert 10GB data. Current enforcement design will allow\\n>>>>> this query to be executed. This is limited by the ‘active’ table detection\\n>>>>> is generated by stat collector. Postgres backend will only send table stat\\n>>>>> information to collector only when the transaction ends. We need a new way\\n>>>>> to detect the ‘active’ table even when this table is being modified inside\\n>>>>> a running transaction.3. Monitor unlimited number of databases. Current we\\n>>>>> set the max number of disk quota worker process to be 10 to reduce the\\n>>>>> affection normal workload. But how about if we want to monitor the disk\\n>>>>> quota of more than 10 databases? Our solution is to let disk quota launcher\\n>>>>> to manage a queue of database need to be monitored. And disk quota worker\\n>>>>> process consuming the queue and refresh the disk usage/quota for this\\n>>>>> database. After some periods, worker will return the database to the queue,\\n>>>>> and fetch the top database from queue to process. The period determine the\\n>>>>> delay of detecting disk quota change. To implement this feature, we need to\\n>>>>> support a subprocess of postmaster to rebind to another database instead of\\n>>>>> the database binded in InitPostgres().4. Support active table detection on\\n>>>>> vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\\n>>>>> are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\\n>>>>> Flack for the former comments on disk quota feature. Any comments on how to\\n>>>>> improve disk quota feature are appreciated.*\\n>>>>>\\n>>>>>\\n>>>>> On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <\\n>>>>> pavel.stehule@gmail.com> wrote:\\n>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>>>>>\\n>>>>>>> Thanks Pavel.\\n>>>>>>> Your patch did enforcement on storage level(md.c or we could also\\n>>>>>>> use smgr_extend). It's straight forward.\\n>>>>>>> But I prefer to implement disk_quota as a feature with following\\n>>>>>>> objectives:\\n>>>>>>> 1 set/alter disk quota setting on different database objects, e.g.\\n>>>>>>> user, database, schema etc. not only a general GUC, but we could set\\n>>>>>>> separate quota limit for a specific objects.\\n>>>>>>> 2 enforcement operator should work at two positions: before query is\\n>>>>>>> running and when query is running. The latter one's implementation maybe\\n>>>>>>> similar to your patch.\\n>>>>>>>\\n>>>>>>\\n>>>>>> The patch was just example. The resource quotes should be more\\n>>>>>> complex - per partition, table, schema, database, user - so GUC are\\n>>>>>> possible, but not very user friendly.\\n>>>>>>\\n>>>>>> Our case is specific, but not too much. The servers are used for\\n>>>>>> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\\n>>>>>> SELECT). We need to solve limits immediately. The implementation is simple,\\n>>>>>> so I did it. Same implementation on database level, or schema level needs\\n>>>>>> some more locks, so it will not be too effective. The resource management\\n>>>>>> can be complex very complex, and I expect so it will be hard work.\\n>>>>>>\\n>>>>>> Regards\\n>>>>>>\\n>>>>>> Pavel\\n>>>>>>\\n>>>>>>\\n>>>>>>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <\\n>>>>>>> pavel.stehule@gmail.com> wrote:\\n>>>>>>>\\n>>>>>>>> Hi\\n>>>>>>>>\\n>>>>>>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>>>>>>>\\n>>>>>>>>> Thanks Chapman.\\n>>>>>>>>> @Pavel, could you please explain more about your second\\n>>>>>>>>> suggestion \\\"implement some quotas on storage level?\\\"\\n>>>>>>>>>\\n>>>>>>>>\\n>>>>>>>> See attached patch - it is very simple - and good enough for our\\n>>>>>>>> purposes.\\n>>>>>>>>\\n>>>>>>>> Regards\\n>>>>>>>>\\n>>>>>>>> Pavel\\n>>>>>>>>\\n>>>>>>>>\\n>>>>>>>>\\n>>>>>>>>> We will not keep the long-lived processes attach to all\\n>>>>>>>>> databases(just like you mentioned servers with thousands of\\n>>>>>>>>> databases)\\n>>>>>>>>> And you are right, we could share ideas with autovacuum process,\\n>>>>>>>>> fork worker processes in need.\\n>>>>>>>>> \\\"autovacuum checks for tables that have had a large number of\\n>>>>>>>>> inserted, updated or deleted tuples. These checks use the statistics\\n>>>>>>>>> collection facility\\\"\\n>>>>>>>>> diskquota process is similar to autovacuum at caring about insert,\\n>>>>>>>>> but the difference is that it also care about vucuum full, truncate and\\n>>>>>>>>> drop. While update and delete may not be interested since no file change\\n>>>>>>>>> happens. So a separate diskquota process is preferred.\\n>>>>>>>>>\\n>>>>>>>>> So if we implemented disk quota as a full native feature, and in\\n>>>>>>>>> the first initial version I prefer to implement the following features:\\n>>>>>>>>> 1 Fork diskquota launcher process under Postmaster serverloop,\\n>>>>>>>>> which is long-lived.\\n>>>>>>>>> 2 Diskquota launcher process is responsible for\\n>>>>>>>>> creating diskquota worker process for every database.\\n>>>>>>>>> 3 DIskquota setting is stored in a separate catalog table for each\\n>>>>>>>>> database.\\n>>>>>>>>> 4 Initialization stage, Diskquota launcher process creates\\n>>>>>>>>> diskquota worker process for all the databases(traverse like\\n>>>>>>>>> autovacuum). Worker process calculates disk usage of db objects and\\n>>>>>>>>> their diskquota setting. If any db object exceeds its quota limit, put them\\n>>>>>>>>> into the blacklist in the shared memory, which will later be used by\\n>>>>>>>>> enforcement operator. Worker process exits when works are done.\\n>>>>>>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\\n>>>>>>>>> process for the database with a large number of insert, copy, truncate,\\n>>>>>>>>> drop etc. or create disk quota statement. Worker process updates the file\\n>>>>>>>>> size for db objects containing the result relation, and compare with the\\n>>>>>>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\\n>>>>>>>>> remove from blacklist vice versa. Worker process exits when works\\n>>>>>>>>> are done and a GUC could control the frequency of worker process restart to\\n>>>>>>>>> a specific database. As you know, this GUC also controls the delay when we\\n>>>>>>>>> do enforcement.\\n>>>>>>>>> 6 Enforcement. When postgres backend executes queries, check the\\n>>>>>>>>> blacklist in shared memory to determine whether the query is allowed(before\\n>>>>>>>>> execute) or need rollback(is executing)?\\n>>>>>>>>>\\n>>>>>>>>> If we implemented disk quota as an extension, we could just use\\n>>>>>>>>> background worker to start diskquota launcher process and use\\n>>>>>>>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\\n>>>>>>>>> processes by the launcher process as suggested by @Chapman.\\n>>>>>>>>> Diskquota setting could be stored in user table in a separate schema for\\n>>>>>>>>> each database(Schema and table created by create extension statement) just\\n>>>>>>>>> like what Heikki has done in pg_quota project. But in this case, we need to\\n>>>>>>>>> create extension for each database before diskquota worker process can be\\n>>>>>>>>> set up for that database.\\n>>>>>>>>>\\n>>>>>>>>> Any comments on the above design and which is preferred, native\\n>>>>>>>>> feature or extension as the POC?\\n>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>> -- Hubert\\n>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <\\n>>>>>>>>> pavel.stehule@gmail.com> wrote:\\n>>>>>>>>>\\n>>>>>>>>>>\\n>>>>>>>>>>\\n>>>>>>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:\\n>>>>>>>>>>\\n>>>>>>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n>>>>>>>>>>>\\n>>>>>>>>>>> > 2 Keep one worker process for each database. But using a\\n>>>>>>>>>>> parent/global\\n>>>>>>>>>>> > quota worker process to manage the lifecycle of database level\\n>>>>>>>>>>> worker\\n>>>>>>>>>>> > processes. It could handle the newly created database(avoid\\n>>>>>>>>>>> restart\\n>>>>>>>>>>> > database) and save resource when a database is not used. But\\n>>>>>>>>>>> this needs to\\n>>>>>>>>>>> > change worker process to be hierarchical. Postmaster becomes\\n>>>>>>>>>>> the grandfather\\n>>>>>>>>>>> > of database level worker processes in this case.\\n>>>>>>>>>>>\\n>>>>>>>>>>> I am using background workers this way in 9.5 at $work.\\n>>>>>>>>>>>\\n>>>>>>>>>>> In my case, one worker lives forever, wakes up on a set period,\\n>>>>>>>>>>> and\\n>>>>>>>>>>> starts a short-lived worker for every database, waiting for each\\n>>>>>>>>>>> one before starting the next.\\n>>>>>>>>>>>\\n>>>>>>>>>>> It was straightforward to implement. Looking back over the code,\\n>>>>>>>>>>> I see the global worker assigns its own PID to\\n>>>>>>>>>>> worker.bgw_notify_pid\\n>>>>>>>>>>> of each of its children, and also obtains a handle for each child\\n>>>>>>>>>>> from RegisterDynamicBackgroundWorker().\\n>>>>>>>>>>>\\n>>>>>>>>>>> I imagine the global quota worker would prefer to start workers\\n>>>>>>>>>>> for every database and then just wait for notifications from any\\n>>>>>>>>>>> of them, but that seems equally straightforward at first glance.\\n>>>>>>>>>>>\\n>>>>>>>>>>\\n>>>>>>>>>> There are servers with thousands databases. Worker per database\\n>>>>>>>>>> is not good idea.\\n>>>>>>>>>>\\n>>>>>>>>>> It should to share ideas, code with autovacuum process.\\n>>>>>>>>>>\\n>>>>>>>>>> Not sure, how to effective implementation based on bg workers can\\n>>>>>>>>>> be. On servers with large set of databases, large set of tables it can\\n>>>>>>>>>> identify too big table too late.\\n>>>>>>>>>>\\n>>>>>>>>>> Isn't better to implement some quotas on storage level?\\n>>>>>>>>>>\\n>>>>>>>>>> Regards\\n>>>>>>>>>>\\n>>>>>>>>>> Pavel\\n>>>>>>>>>>\\n>>>>>>>>>>\\n>>>>>>>>>>\\n>>>>>>>>>>> -Chap\\n>>>>>>>>>>>\\n>>>>>>>>>>>\\n>>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>> --\\n>>>>>>>>> Thanks\\n>>>>>>>>>\\n>>>>>>>>> Hubert Zhang\\n>>>>>>>>>\\n>>>>>>>>\\n>>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>> --\\n>>>>>>> Thanks\\n>>>>>>>\\n>>>>>>> Hubert Zhang\\n>>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>\\n>>>>>\\n>>>>> --\\n>>>>> Thanks\\n>>>>>\\n>>>>> Hubert Zhang\\n>>>>>\\n>>>>\\n>>>\\n>>> --\\n>>> Thanks\\n>>>\\n>>> Hubert Zhang\\n>>>\\n>>\\n>\\n> --\\n> Thanks\\n>\\n> Hubert Zhang\\n>\\n\\nso 22. 9. 2018 v 8:48 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:But it looks like redundant to current GUC configuration and limitswhat do you mean by current GUC configuration? Is that the general block number limit in your patch? If yes, the difference between GUC and pg_diskquota catalog is that pg_diskquota will store different quota limit for the different role, schema or table instead of a single GUC value.storage is not relevant in this moment.I don't see to consistent to sets some limits via SET command, or ALTER X SET, and some other with CREATE QUOTA ON.The quotas or object limits, resource limits are pretty useful and necessary, but I don't see these like new type of objects, it is much more some property of current objects. Because we have one syntax for this purpose I prefer it. Because is not good to have two syntaxes for similar purpose.So instead CREATE DISC QUATA ON SCHEMA xxx some value I preferALTER SCHEMA xxx SET disc_quota = xxx;The functionality is +/- same. But ALTER XX SET was introduce first, and I don't feel comfortable to have any new syntax for similar purposeRegardsPavel On Sat, Sep 22, 2018 at 11:17 AM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer .It's an option. Prefer to consider quota setting store together:CREATE DISK QUOTA way is more nature to store quota setting in a separate pg_diskquota catalogWhile ALTER SET way is more close to store quota setting in pg_class, pg_role, pg_namespace. etc in an integrated way.(Note that here I mean nature/close is not must, ALTER SET could also store in pg_diskquota and vice versa.)I have not a problem with new special table for storing this information. But it looks like redundant to current GUC configuration and limits. Can be messy do some work with ALTER ROLE, and some work via CREATE QUOTE.RegardsPavel Here are some differences I can think of:1 pg_role is a global catalog, not per database level. It's harder to tracker the user's disk usage in the whole clusters(considering 1000+ databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it only tracks the user's disk usage inside the current database.2 using separate pg_diskquota could add more field except for quota limit without adding too many fields in pg_class, e.g. red zone to give the user a warning or the current disk usage of the db objects.On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:Hi all,We redesign disk quota feature based on the comments from Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk quota feature is used to support multi-tenancy environment, different level of database objects could be set a quota limit to avoid over use of disk space. A common case could be as follows: DBA could enable disk quota on a specified database list. DBA could set disk quota limit for tables/schemas/roles in these databases. Separate disk quota worker process will monitor the disk usage for these objects and detect the objects which exceed their quota limit. Queries loading data into these “out of disk quota” tables/schemas/roles will be cancelled.We are currently working at init implementation stage. We would like to propose our idea firstly and get feedbacks from community to do quick iteration.SQL Syntax (How to use disk quota)1 Specify the databases with disk quota enabled in GUC “diskquota_databases” in postgresql.conf and restart the database.2 DBA could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer . 3 Simulate a schema out of quota limit case: suppose table a1 and table a2 are both under schema s1.INSERT INTO a1 SELECT generate_series(1,1000);INSERT INTO a2 SELECT generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);ERROR: schema's disk space quota exceededDROP TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the following components.1. Quota Setting Store is where the disk quota setting to be stored and accessed. We plan to use catalog table pg_diskquota to store these information. pg_diskquota is like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /* diskquota name */ int16 quotatype; /* diskquota type name */ Oid quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /* diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size change of database objects. We plan to use stat collector to detect the ‘active’ table list at initial stage. But stat collector has some limitation on finding the active table which is in a running transaction. Details see TODO section.3. Quota Size Checker is where to calculate the size and compare with quota limit for database objects. According to Pavel’s comment, autovacuum launcher and worker process could be a good reference to disk quota. So we plan to use a disk quota launcher daemon process and several disk quota worker process to finish this work. Launcher process is responsible for starting worker process based on a user defined database list from GUC. Worker process will connect to its target database and monitor the disk usage for objects in this database. In init stage of worker process, it will call calculate_total_relation_size() to calculate the size for each user table. After init stage, worker process will refresh the disk model every N seconds. Refreshing will only recalculate the size of tables in ‘active’ table list, which is generated by Quata Change Detector to minimize the cost.4. Quota Enforcement Operator is where to check for the quota limitation at postgres backend side. We will firstly implement it in ExecCheckRTPerms() as pre-running enforcement. It will check the disk quota of tables being inserted or updated, and report error if table’s or table’s schema’s or table’s owner’s quota limit is exceeded. As a native feature, we plan to add more checkpoint to do running query enforcement. For example, if a disk quota lefts 10MB quota, a query could insert 1GB data. This query could be allowed in pre-running enforcement check, but will be cancelled in running query enforcement check. Therefore, it can improve the accurate of disk quota usage. To achieve this, we plan to add a checkpoint in lower API such as smgr_extened. Hence, the Quota Enforcement Operator will check the disk quota usage when smgr_extened is called. If the quota is over limited, current query will be cancelled. Highlight1. Native feature.Support native Create/Drop Disk Quota SQL statement.New catalog table pg_diskquota to store disk quota setting.2. Auto DML/DDL detection. Table create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change, Schema create/drop and Role create/drop will be detected by disk quota automatically. 3. Low cost disk quota checker.Worker process of disk quota need to refresh the disk usage model every N seconds. Since recalculate the file size using stat() system call is expensive for a large number of files, we use an ‘active’ table list to reduce the real work at each iteration. A basic experiment on our init stage implementation on database with 20K tables shows that the refresh cost is 1% cpu usage and will be finished within 50ms. Todo/LimitationBefore we propose our patch, we plan to enhance it with the following ideas:1. Setting database list with disk quota enabled dynamically without restart database. Since we have the disk quota launcher process, it could detect the new ‘diskquota_databases’ list and start/stop the corresponding disk quota worker process.2. Enforcement when query is running. Considering the case when there is 10MB quota left, but next query will insert 10GB data. Current enforcement design will allow this query to be executed. This is limited by the ‘active’ table detection is generated by stat collector. Postgres backend will only send table stat information to collector only when the transaction ends. We need a new way to detect the ‘active’ table even when this table is being modified inside a running transaction.3. Monitor unlimited number of databases. Current we set the max number of disk quota worker process to be 10 to reduce the affection normal workload. But how about if we want to monitor the disk quota of more than 10 databases? Our solution is to let disk quota launcher to manage a queue of database need to be monitored. And disk quota worker process consuming the queue and refresh the disk usage/quota for this database. After some periods, worker will return the database to the queue, and fetch the top database from queue to process. The period determine the delay of detecting disk quota change. To implement this feature, we need to support a subprocess of postmaster to rebind to another database instead of the database binded in InitPostgres().4. Support active table detection on vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman Flack for the former comments on disk quota feature. Any comments on how to improve disk quota feature are appreciated.On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \\\"implement some quotas on storage level?\\\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \\\"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\\\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n\\n> 2 Keep one worker process for each database. But using a parent/global\\n> quota worker process to manage the lifecycle of database level worker\\n> processes. It could handle the newly created database(avoid restart\\n> database) and save resource when a database is not used. But this needs to\\n> change worker process to be hierarchical. Postmaster becomes the grandfather\\n> of database level worker processes in this case.\\n\\nI am using background workers this way in 9.5 at $work.\\n\\nIn my case, one worker lives forever, wakes up on a set period, and\\nstarts a short-lived worker for every database, waiting for each\\none before starting the next.\\n\\nIt was straightforward to implement. Looking back over the code,\\nI see the global worker assigns its own PID to worker.bgw_notify_pid\\nof each of its children, and also obtains a handle for each child\\nfrom RegisterDynamicBackgroundWorker().\\n\\nI imagine the global quota worker would prefer to start workers\\nfor every database and then just wait for notifications from any\\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\\n\\n-Chap\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\\n\\n-- ThanksHubert Zhang\",\n \"msg_date\": \"Sat, 22 Sep 2018 09:02:25 +0200\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal for disk quota feature\"\n },\n {\n \"msg_contents\": \">\\n> The quotas or object limits, resource limits are pretty useful and\\n> necessary, but I don't see these like new type of objects, it is much more\\n> some property of current objects. Because we have one syntax for this\\n> purpose I prefer it. Because is not good to have two syntaxes for similar\\n> purpose.\\n\\nSCHEMA and TABLE are OK for me, But as I mentioned before, ROLE is a\\nspecial case when using ALTER SET at this moment.\\nTABLE and SCHEMA are both database level, e.g. pg_class and pg_namespace\\nboth residents in one database. But ROLE is cluster-level. They don't\\nbelong to a database. ALTER ROLE XXX SET disk_quota = xxx means to set the\\nquota for the user on all the databases in the first glance. But in our\\nfirst stage design, ROLE's quota is bind to a specific database. E.g. Role\\nJack could have 10GB quota on database A and 2GB quota on database B.\\n\\nSQL syntax is not hard to modify, I don't think this should block the main\\ndesign of disk quota feature. Is there any comment on the design and\\narchitecture? If no, we'll firstly submit our patch and involve more\\ndiscussion?\\n\\nOn Sat, Sep 22, 2018 at 3:03 PM Pavel Stehule <pavel.stehule@gmail.com>\\nwrote:\\n\\n>\\n>\\n> so 22. 9. 2018 v 8:48 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:\\n>\\n>> But it looks like redundant to current GUC configuration and limits\\n>>\\n>> what do you mean by current GUC configuration? Is that the general block\\n>> number limit in your patch? If yes, the difference between GUC and\\n>> pg_diskquota catalog is that pg_diskquota will store different quota limit\\n>> for the different role, schema or table instead of a single GUC value.\\n>>\\n>\\n> storage is not relevant in this moment.\\n>\\n> I don't see to consistent to sets some limits via SET command, or ALTER X\\n> SET, and some other with CREATE QUOTA ON.\\n>\\n> The quotas or object limits, resource limits are pretty useful and\\n> necessary, but I don't see these like new type of objects, it is much more\\n> some property of current objects. Because we have one syntax for this\\n> purpose I prefer it. Because is not good to have two syntaxes for similar\\n> purpose.\\n>\\n> So instead CREATE DISC QUATA ON SCHEMA xxx some value I prefer\\n>\\n> ALTER SCHEMA xxx SET disc_quota = xxx;\\n>\\n> The functionality is +/- same. But ALTER XX SET was introduce first, and I\\n> don't feel comfortable to have any new syntax for similar purpose\\n>\\n> Regards\\n>\\n> Pavel\\n>\\n>\\n>\\n>\\n>\\n>>\\n>> On Sat, Sep 22, 2018 at 11:17 AM Pavel Stehule <pavel.stehule@gmail.com>\\n>> wrote:\\n>>\\n>>>\\n>>>\\n>>> pá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io>\\n>>> napsal:\\n>>>\\n>>>> just fast reaction - why QUOTA object?\\n>>>>> Isn't ALTER SET enough?\\n>>>>> Some like\\n>>>>> ALTER TABLE a1 SET quote = 1MB;\\n>>>>> ALTER USER ...\\n>>>>> ALTER SCHEMA ..\\n>>>>> New DDL commans looks like too hard hammer .\\n>>>>\\n>>>>\\n>>>> It's an option. Prefer to consider quota setting store together:\\n>>>> CREATE DISK QUOTA way is more nature to store quota setting in a\\n>>>> separate pg_diskquota catalog\\n>>>> While ALTER SET way is more close to store quota setting in pg_class,\\n>>>> pg_role, pg_namespace. etc in an integrated way.\\n>>>> (Note that here I mean nature/close is not must, ALTER SET could also\\n>>>> store in pg_diskquota and vice versa.)\\n>>>>\\n>>>\\n>>> I have not a problem with new special table for storing this\\n>>> information. But it looks like redundant to current GUC configuration and\\n>>> limits. Can be messy do some work with ALTER ROLE, and some work via CREATE\\n>>> QUOTE.\\n>>>\\n>>> Regards\\n>>>\\n>>> Pavel\\n>>>\\n>>>\\n>>>> Here are some differences I can think of:\\n>>>> 1 pg_role is a global catalog, not per database level. It's harder to\\n>>>> tracker the user's disk usage in the whole clusters(considering 1000+\\n>>>> databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it\\n>>>> only tracks the user's disk usage inside the current database.\\n>>>> 2 using separate pg_diskquota could add more field except for quota\\n>>>> limit without adding too many fields in pg_class, e.g. red zone to give the\\n>>>> user a warning or the current disk usage of the db objects.\\n>>>>\\n>>>> On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com>\\n>>>> wrote:\\n>>>>\\n>>>>>\\n>>>>>\\n>>>>> pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io>\\n>>>>> napsal:\\n>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>> *Hi all,We redesign disk quota feature based on the comments from\\n>>>>>> Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically,\\n>>>>>> disk quota feature is used to support multi-tenancy environment, different\\n>>>>>> level of database objects could be set a quota limit to avoid over use of\\n>>>>>> disk space. A common case could be as follows: DBA could enable disk quota\\n>>>>>> on a specified database list. DBA could set disk quota limit for\\n>>>>>> tables/schemas/roles in these databases. Separate disk quota worker process\\n>>>>>> will monitor the disk usage for these objects and detect the objects which\\n>>>>>> exceed their quota limit. Queries loading data into these “out of disk\\n>>>>>> quota” tables/schemas/roles will be cancelled.We are currently working at\\n>>>>>> init implementation stage. We would like to propose our idea firstly and\\n>>>>>> get feedbacks from community to do quick iteration.SQL Syntax (How to use\\n>>>>>> disk quota)1 Specify the databases with disk quota enabled in GUC\\n>>>>>> “diskquota_databases” in postgresql.conf and restart the database.2 DBA\\n>>>>>> could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\\n>>>>>> ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\\n>>>>>> (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\\n>>>>>> ‘3MB’);*\\n>>>>>>\\n>>>>>\\n>>>>> just fast reaction - why QUOTA object?\\n>>>>>\\n>>>>> Isn't ALTER SET enough?\\n>>>>>\\n>>>>> Some like\\n>>>>>\\n>>>>> ALTER TABLE a1 SET quote = 1MB;\\n>>>>> ALTER USER ...\\n>>>>> ALTER SCHEMA ..\\n>>>>>\\n>>>>> New DDL commans looks like too hard hammer .\\n>>>>>\\n>>>>>\\n>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>> *3 Simulate a schema out of quota limit case: suppose table a1 and\\n>>>>>> table a2 are both under schema s1.INSERT INTO a1 SELECT\\n>>>>>> generate_series(1,1000);INSERT INTO a2 SELECT\\n>>>>>> generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\\n>>>>>> generate_series(1,1000);ERROR: schema's disk space quota exceededDROP\\n>>>>>> TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\\n>>>>>> generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\\n>>>>>> following components.1. Quota Setting Store is where the disk quota setting\\n>>>>>> to be stored and accessed. We plan to use catalog table pg_diskquota to\\n>>>>>> store these information. pg_diskquota is\\n>>>>>> like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\\n>>>>>> diskquota name */ int16 quotatype; /* diskquota type name */ Oid\\n>>>>>> quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\\n>>>>>> diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\\n>>>>>> MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\\n>>>>>> change of database objects. We plan to use stat collector to detect the\\n>>>>>> ‘active’ table list at initial stage. But stat collector has some\\n>>>>>> limitation on finding the active table which is in a running transaction.\\n>>>>>> Details see TODO section.3. Quota Size Checker is where to calculate the\\n>>>>>> size and compare with quota limit for database objects. According to\\n>>>>>> Pavel’s comment, autovacuum launcher and worker process could be a good\\n>>>>>> reference to disk quota. So we plan to use a disk quota launcher daemon\\n>>>>>> process and several disk quota worker process to finish this work. Launcher\\n>>>>>> process is responsible for starting worker process based on a user defined\\n>>>>>> database list from GUC. Worker process will connect to its target database\\n>>>>>> and monitor the disk usage for objects in this database. In init stage of\\n>>>>>> worker process, it will call calculate_total_relation_size() to calculate\\n>>>>>> the size for each user table. After init stage, worker process will refresh\\n>>>>>> the disk model every N seconds. Refreshing will only recalculate the size\\n>>>>>> of tables in ‘active’ table list, which is generated by Quata Change\\n>>>>>> Detector to minimize the cost.4. Quota Enforcement Operator is where to\\n>>>>>> check for the quota limitation at postgres backend side. We will firstly\\n>>>>>> implement it in ExecCheckRTPerms() as pre-running enforcement. It will\\n>>>>>> check the disk quota of tables being inserted or updated, and report error\\n>>>>>> if table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\\n>>>>>> As a native feature, we plan to add more checkpoint to do running query\\n>>>>>> enforcement. For example, if a disk quota lefts 10MB quota, a query could\\n>>>>>> insert 1GB data. This query could be allowed in pre-running enforcement\\n>>>>>> check, but will be cancelled in running query enforcement check. Therefore,\\n>>>>>> it can improve the accurate of disk quota usage. To achieve this, we plan\\n>>>>>> to add a checkpoint in lower API such as smgr_extened. Hence, the Quota\\n>>>>>> Enforcement Operator will check the disk quota usage when smgr_extened is\\n>>>>>> called. If the quota is over limited, current query will be cancelled.\\n>>>>>> Highlight1. Native feature.Support native Create/Drop Disk Quota SQL\\n>>>>>> statement.New catalog table pg_diskquota to store disk quota setting.2.\\n>>>>>> Auto DML/DDL detection. Table\\n>>>>>> create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\\n>>>>>> Schema create/drop and Role create/drop will be detected by disk quota\\n>>>>>> automatically. 3. Low cost disk quota checker.Worker process of disk quota\\n>>>>>> need to refresh the disk usage model every N seconds. Since recalculate the\\n>>>>>> file size using stat() system call is expensive for a large number of\\n>>>>>> files, we use an ‘active’ table list to reduce the real work at each\\n>>>>>> iteration. A basic experiment on our init stage implementation on database\\n>>>>>> with 20K tables shows that the refresh cost is 1% cpu usage and will be\\n>>>>>> finished within 50ms. Todo/LimitationBefore we propose our patch, we plan\\n>>>>>> to enhance it with the following ideas:1. Setting database list with disk\\n>>>>>> quota enabled dynamically without restart database. Since we have the disk\\n>>>>>> quota launcher process, it could detect the new ‘diskquota_databases’ list\\n>>>>>> and start/stop the corresponding disk quota worker process.2. Enforcement\\n>>>>>> when query is running. Considering the case when there is 10MB quota left,\\n>>>>>> but next query will insert 10GB data. Current enforcement design will allow\\n>>>>>> this query to be executed. This is limited by the ‘active’ table detection\\n>>>>>> is generated by stat collector. Postgres backend will only send table stat\\n>>>>>> information to collector only when the transaction ends. We need a new way\\n>>>>>> to detect the ‘active’ table even when this table is being modified inside\\n>>>>>> a running transaction.3. Monitor unlimited number of databases. Current we\\n>>>>>> set the max number of disk quota worker process to be 10 to reduce the\\n>>>>>> affection normal workload. But how about if we want to monitor the disk\\n>>>>>> quota of more than 10 databases? Our solution is to let disk quota launcher\\n>>>>>> to manage a queue of database need to be monitored. And disk quota worker\\n>>>>>> process consuming the queue and refresh the disk usage/quota for this\\n>>>>>> database. After some periods, worker will return the database to the queue,\\n>>>>>> and fetch the top database from queue to process. The period determine the\\n>>>>>> delay of detecting disk quota change. To implement this feature, we need to\\n>>>>>> support a subprocess of postmaster to rebind to another database instead of\\n>>>>>> the database binded in InitPostgres().4. Support active table detection on\\n>>>>>> vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\\n>>>>>> are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\\n>>>>>> Flack for the former comments on disk quota feature. Any comments on how to\\n>>>>>> improve disk quota feature are appreciated.*\\n>>>>>>\\n>>>>>>\\n>>>>>> On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <\\n>>>>>> pavel.stehule@gmail.com> wrote:\\n>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>>>>>>\\n>>>>>>>> Thanks Pavel.\\n>>>>>>>> Your patch did enforcement on storage level(md.c or we could also\\n>>>>>>>> use smgr_extend). It's straight forward.\\n>>>>>>>> But I prefer to implement disk_quota as a feature with following\\n>>>>>>>> objectives:\\n>>>>>>>> 1 set/alter disk quota setting on different database objects, e.g.\\n>>>>>>>> user, database, schema etc. not only a general GUC, but we could set\\n>>>>>>>> separate quota limit for a specific objects.\\n>>>>>>>> 2 enforcement operator should work at two positions: before query\\n>>>>>>>> is running and when query is running. The latter one's implementation maybe\\n>>>>>>>> similar to your patch.\\n>>>>>>>>\\n>>>>>>>\\n>>>>>>> The patch was just example. The resource quotes should be more\\n>>>>>>> complex - per partition, table, schema, database, user - so GUC are\\n>>>>>>> possible, but not very user friendly.\\n>>>>>>>\\n>>>>>>> Our case is specific, but not too much. The servers are used for\\n>>>>>>> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\\n>>>>>>> SELECT). We need to solve limits immediately. The implementation is simple,\\n>>>>>>> so I did it. Same implementation on database level, or schema level needs\\n>>>>>>> some more locks, so it will not be too effective. The resource management\\n>>>>>>> can be complex very complex, and I expect so it will be hard work.\\n>>>>>>>\\n>>>>>>> Regards\\n>>>>>>>\\n>>>>>>> Pavel\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <\\n>>>>>>>> pavel.stehule@gmail.com> wrote:\\n>>>>>>>>\\n>>>>>>>>> Hi\\n>>>>>>>>>\\n>>>>>>>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>>>>>>>>\\n>>>>>>>>>> Thanks Chapman.\\n>>>>>>>>>> @Pavel, could you please explain more about your second\\n>>>>>>>>>> suggestion \\\"implement some quotas on storage level?\\\"\\n>>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>> See attached patch - it is very simple - and good enough for our\\n>>>>>>>>> purposes.\\n>>>>>>>>>\\n>>>>>>>>> Regards\\n>>>>>>>>>\\n>>>>>>>>> Pavel\\n>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>>> We will not keep the long-lived processes attach to all\\n>>>>>>>>>> databases(just like you mentioned servers with thousands of\\n>>>>>>>>>> databases)\\n>>>>>>>>>> And you are right, we could share ideas with autovacuum process,\\n>>>>>>>>>> fork worker processes in need.\\n>>>>>>>>>> \\\"autovacuum checks for tables that have had a large number of\\n>>>>>>>>>> inserted, updated or deleted tuples. These checks use the statistics\\n>>>>>>>>>> collection facility\\\"\\n>>>>>>>>>> diskquota process is similar to autovacuum at caring about\\n>>>>>>>>>> insert, but the difference is that it also care about vucuum full, truncate\\n>>>>>>>>>> and drop. While update and delete may not be interested since no file\\n>>>>>>>>>> change happens. So a separate diskquota process is preferred.\\n>>>>>>>>>>\\n>>>>>>>>>> So if we implemented disk quota as a full native feature, and in\\n>>>>>>>>>> the first initial version I prefer to implement the following features:\\n>>>>>>>>>> 1 Fork diskquota launcher process under Postmaster serverloop,\\n>>>>>>>>>> which is long-lived.\\n>>>>>>>>>> 2 Diskquota launcher process is responsible for\\n>>>>>>>>>> creating diskquota worker process for every database.\\n>>>>>>>>>> 3 DIskquota setting is stored in a separate catalog table for\\n>>>>>>>>>> each database.\\n>>>>>>>>>> 4 Initialization stage, Diskquota launcher process creates\\n>>>>>>>>>> diskquota worker process for all the databases(traverse like\\n>>>>>>>>>> autovacuum). Worker process calculates disk usage of db objects and\\n>>>>>>>>>> their diskquota setting. If any db object exceeds its quota limit, put them\\n>>>>>>>>>> into the blacklist in the shared memory, which will later be used by\\n>>>>>>>>>> enforcement operator. Worker process exits when works are done.\\n>>>>>>>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\\n>>>>>>>>>> process for the database with a large number of insert, copy, truncate,\\n>>>>>>>>>> drop etc. or create disk quota statement. Worker process updates the file\\n>>>>>>>>>> size for db objects containing the result relation, and compare with the\\n>>>>>>>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\\n>>>>>>>>>> remove from blacklist vice versa. Worker process exits when\\n>>>>>>>>>> works are done and a GUC could control the frequency of worker process\\n>>>>>>>>>> restart to a specific database. As you know, this GUC also controls the\\n>>>>>>>>>> delay when we do enforcement.\\n>>>>>>>>>> 6 Enforcement. When postgres backend executes queries, check the\\n>>>>>>>>>> blacklist in shared memory to determine whether the query is allowed(before\\n>>>>>>>>>> execute) or need rollback(is executing)?\\n>>>>>>>>>>\\n>>>>>>>>>> If we implemented disk quota as an extension, we could just use\\n>>>>>>>>>> background worker to start diskquota launcher process and use\\n>>>>>>>>>> RegisterDynamicBackgroundWorker() to fork child diskquota worker\\n>>>>>>>>>> processes by the launcher process as suggested by @Chapman.\\n>>>>>>>>>> Diskquota setting could be stored in user table in a separate schema for\\n>>>>>>>>>> each database(Schema and table created by create extension statement) just\\n>>>>>>>>>> like what Heikki has done in pg_quota project. But in this case, we need to\\n>>>>>>>>>> create extension for each database before diskquota worker process can be\\n>>>>>>>>>> set up for that database.\\n>>>>>>>>>>\\n>>>>>>>>>> Any comments on the above design and which is preferred, native\\n>>>>>>>>>> feature or extension as the POC?\\n>>>>>>>>>>\\n>>>>>>>>>>\\n>>>>>>>>>> -- Hubert\\n>>>>>>>>>>\\n>>>>>>>>>>\\n>>>>>>>>>>\\n>>>>>>>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <\\n>>>>>>>>>> pavel.stehule@gmail.com> wrote:\\n>>>>>>>>>>\\n>>>>>>>>>>>\\n>>>>>>>>>>>\\n>>>>>>>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>\\n>>>>>>>>>>> :\\n>>>>>>>>>>>\\n>>>>>>>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n>>>>>>>>>>>>\\n>>>>>>>>>>>> > 2 Keep one worker process for each database. But using a\\n>>>>>>>>>>>> parent/global\\n>>>>>>>>>>>> > quota worker process to manage the lifecycle of database\\n>>>>>>>>>>>> level worker\\n>>>>>>>>>>>> > processes. It could handle the newly created database(avoid\\n>>>>>>>>>>>> restart\\n>>>>>>>>>>>> > database) and save resource when a database is not used. But\\n>>>>>>>>>>>> this needs to\\n>>>>>>>>>>>> > change worker process to be hierarchical. Postmaster becomes\\n>>>>>>>>>>>> the grandfather\\n>>>>>>>>>>>> > of database level worker processes in this case.\\n>>>>>>>>>>>>\\n>>>>>>>>>>>> I am using background workers this way in 9.5 at $work.\\n>>>>>>>>>>>>\\n>>>>>>>>>>>> In my case, one worker lives forever, wakes up on a set period,\\n>>>>>>>>>>>> and\\n>>>>>>>>>>>> starts a short-lived worker for every database, waiting for each\\n>>>>>>>>>>>> one before starting the next.\\n>>>>>>>>>>>>\\n>>>>>>>>>>>> It was straightforward to implement. Looking back over the code,\\n>>>>>>>>>>>> I see the global worker assigns its own PID to\\n>>>>>>>>>>>> worker.bgw_notify_pid\\n>>>>>>>>>>>> of each of its children, and also obtains a handle for each\\n>>>>>>>>>>>> child\\n>>>>>>>>>>>> from RegisterDynamicBackgroundWorker().\\n>>>>>>>>>>>>\\n>>>>>>>>>>>> I imagine the global quota worker would prefer to start workers\\n>>>>>>>>>>>> for every database and then just wait for notifications from any\\n>>>>>>>>>>>> of them, but that seems equally straightforward at first glance.\\n>>>>>>>>>>>>\\n>>>>>>>>>>>\\n>>>>>>>>>>> There are servers with thousands databases. Worker per database\\n>>>>>>>>>>> is not good idea.\\n>>>>>>>>>>>\\n>>>>>>>>>>> It should to share ideas, code with autovacuum process.\\n>>>>>>>>>>>\\n>>>>>>>>>>> Not sure, how to effective implementation based on bg workers\\n>>>>>>>>>>> can be. On servers with large set of databases, large set of tables it can\\n>>>>>>>>>>> identify too big table too late.\\n>>>>>>>>>>>\\n>>>>>>>>>>> Isn't better to implement some quotas on storage level?\\n>>>>>>>>>>>\\n>>>>>>>>>>> Regards\\n>>>>>>>>>>>\\n>>>>>>>>>>> Pavel\\n>>>>>>>>>>>\\n>>>>>>>>>>>\\n>>>>>>>>>>>\\n>>>>>>>>>>>> -Chap\\n>>>>>>>>>>>>\\n>>>>>>>>>>>>\\n>>>>>>>>>>>\\n>>>>>>>>>>\\n>>>>>>>>>>\\n>>>>>>>>>> --\\n>>>>>>>>>> Thanks\\n>>>>>>>>>>\\n>>>>>>>>>> Hubert Zhang\\n>>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>\\n>>>>>>>>\\n>>>>>>>> --\\n>>>>>>>> Thanks\\n>>>>>>>>\\n>>>>>>>> Hubert Zhang\\n>>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>> --\\n>>>>>> Thanks\\n>>>>>>\\n>>>>>> Hubert Zhang\\n>>>>>>\\n>>>>>\\n>>>>\\n>>>> --\\n>>>> Thanks\\n>>>>\\n>>>> Hubert Zhang\\n>>>>\\n>>>\\n>>\\n>> --\\n>> Thanks\\n>>\\n>> Hubert Zhang\\n>>\\n>\\n\\n-- \\nThanks\\n\\nHubert Zhang\\n\\nThe quotas or object limits, resource limits are pretty useful and necessary, but I don't see these like new type of objects, it is much more some property of current objects. Because we have one syntax for this purpose I prefer it. Because is not good to have two syntaxes for similar purpose.SCHEMA and TABLE are OK for me, But as I mentioned before, ROLE is a special case when using ALTER SET at this moment.TABLE and SCHEMA are both database level, e.g. pg_class and pg_namespace both residents in one database. But ROLE is cluster-level. They don't belong to a database. ALTER ROLE XXX SET disk_quota = xxx means to set the quota for the user on all the databases in the first glance. But in our first stage design, ROLE's quota is bind to a specific database. E.g. Role Jack could have 10GB quota on database A and 2GB quota on database B. SQL syntax is not hard to modify, I don't think this should block the main design of disk quota feature. Is there any comment on the design and architecture? If no, we'll firstly submit our patch and involve more discussion?On Sat, Sep 22, 2018 at 3:03 PM Pavel Stehule <pavel.stehule@gmail.com> wrote:so 22. 9. 2018 v 8:48 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:But it looks like redundant to current GUC configuration and limitswhat do you mean by current GUC configuration? Is that the general block number limit in your patch? If yes, the difference between GUC and pg_diskquota catalog is that pg_diskquota will store different quota limit for the different role, schema or table instead of a single GUC value.storage is not relevant in this moment.I don't see to consistent to sets some limits via SET command, or ALTER X SET, and some other with CREATE QUOTA ON.The quotas or object limits, resource limits are pretty useful and necessary, but I don't see these like new type of objects, it is much more some property of current objects. Because we have one syntax for this purpose I prefer it. Because is not good to have two syntaxes for similar purpose.So instead CREATE DISC QUATA ON SCHEMA xxx some value I preferALTER SCHEMA xxx SET disc_quota = xxx;The functionality is +/- same. But ALTER XX SET was introduce first, and I don't feel comfortable to have any new syntax for similar purposeRegardsPavel On Sat, Sep 22, 2018 at 11:17 AM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer .It's an option. Prefer to consider quota setting store together:CREATE DISK QUOTA way is more nature to store quota setting in a separate pg_diskquota catalogWhile ALTER SET way is more close to store quota setting in pg_class, pg_role, pg_namespace. etc in an integrated way.(Note that here I mean nature/close is not must, ALTER SET could also store in pg_diskquota and vice versa.)I have not a problem with new special table for storing this information. But it looks like redundant to current GUC configuration and limits. Can be messy do some work with ALTER ROLE, and some work via CREATE QUOTE.RegardsPavel Here are some differences I can think of:1 pg_role is a global catalog, not per database level. It's harder to tracker the user's disk usage in the whole clusters(considering 1000+ databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it only tracks the user's disk usage inside the current database.2 using separate pg_diskquota could add more field except for quota limit without adding too many fields in pg_class, e.g. red zone to give the user a warning or the current disk usage of the db objects.On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com> wrote:pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:Hi all,We redesign disk quota feature based on the comments from Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically, disk quota feature is used to support multi-tenancy environment, different level of database objects could be set a quota limit to avoid over use of disk space. A common case could be as follows: DBA could enable disk quota on a specified database list. DBA could set disk quota limit for tables/schemas/roles in these databases. Separate disk quota worker process will monitor the disk usage for these objects and detect the objects which exceed their quota limit. Queries loading data into these “out of disk quota” tables/schemas/roles will be cancelled.We are currently working at init implementation stage. We would like to propose our idea firstly and get feedbacks from community to do quick iteration.SQL Syntax (How to use disk quota)1 Specify the databases with disk quota enabled in GUC “diskquota_databases” in postgresql.conf and restart the database.2 DBA could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);just fast reaction - why QUOTA object?Isn't ALTER SET enough?Some likeALTER TABLE a1 SET quote = 1MB;ALTER USER ...ALTER SCHEMA ..New DDL commans looks like too hard hammer . 3 Simulate a schema out of quota limit case: suppose table a1 and table a2 are both under schema s1.INSERT INTO a1 SELECT generate_series(1,1000);INSERT INTO a2 SELECT generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);ERROR: schema's disk space quota exceededDROP TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the following components.1. Quota Setting Store is where the disk quota setting to be stored and accessed. We plan to use catalog table pg_diskquota to store these information. pg_diskquota is like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /* diskquota name */ int16 quotatype; /* diskquota type name */ Oid quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /* diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size change of database objects. We plan to use stat collector to detect the ‘active’ table list at initial stage. But stat collector has some limitation on finding the active table which is in a running transaction. Details see TODO section.3. Quota Size Checker is where to calculate the size and compare with quota limit for database objects. According to Pavel’s comment, autovacuum launcher and worker process could be a good reference to disk quota. So we plan to use a disk quota launcher daemon process and several disk quota worker process to finish this work. Launcher process is responsible for starting worker process based on a user defined database list from GUC. Worker process will connect to its target database and monitor the disk usage for objects in this database. In init stage of worker process, it will call calculate_total_relation_size() to calculate the size for each user table. After init stage, worker process will refresh the disk model every N seconds. Refreshing will only recalculate the size of tables in ‘active’ table list, which is generated by Quata Change Detector to minimize the cost.4. Quota Enforcement Operator is where to check for the quota limitation at postgres backend side. We will firstly implement it in ExecCheckRTPerms() as pre-running enforcement. It will check the disk quota of tables being inserted or updated, and report error if table’s or table’s schema’s or table’s owner’s quota limit is exceeded. As a native feature, we plan to add more checkpoint to do running query enforcement. For example, if a disk quota lefts 10MB quota, a query could insert 1GB data. This query could be allowed in pre-running enforcement check, but will be cancelled in running query enforcement check. Therefore, it can improve the accurate of disk quota usage. To achieve this, we plan to add a checkpoint in lower API such as smgr_extened. Hence, the Quota Enforcement Operator will check the disk quota usage when smgr_extened is called. If the quota is over limited, current query will be cancelled. Highlight1. Native feature.Support native Create/Drop Disk Quota SQL statement.New catalog table pg_diskquota to store disk quota setting.2. Auto DML/DDL detection. Table create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change, Schema create/drop and Role create/drop will be detected by disk quota automatically. 3. Low cost disk quota checker.Worker process of disk quota need to refresh the disk usage model every N seconds. Since recalculate the file size using stat() system call is expensive for a large number of files, we use an ‘active’ table list to reduce the real work at each iteration. A basic experiment on our init stage implementation on database with 20K tables shows that the refresh cost is 1% cpu usage and will be finished within 50ms. Todo/LimitationBefore we propose our patch, we plan to enhance it with the following ideas:1. Setting database list with disk quota enabled dynamically without restart database. Since we have the disk quota launcher process, it could detect the new ‘diskquota_databases’ list and start/stop the corresponding disk quota worker process.2. Enforcement when query is running. Considering the case when there is 10MB quota left, but next query will insert 10GB data. Current enforcement design will allow this query to be executed. This is limited by the ‘active’ table detection is generated by stat collector. Postgres backend will only send table stat information to collector only when the transaction ends. We need a new way to detect the ‘active’ table even when this table is being modified inside a running transaction.3. Monitor unlimited number of databases. Current we set the max number of disk quota worker process to be 10 to reduce the affection normal workload. But how about if we want to monitor the disk quota of more than 10 databases? Our solution is to let disk quota launcher to manage a queue of database need to be monitored. And disk quota worker process consuming the queue and refresh the disk usage/quota for this database. After some periods, worker will return the database to the queue, and fetch the top database from queue to process. The period determine the delay of detecting disk quota change. To implement this feature, we need to support a subprocess of postmaster to rebind to another database instead of the database binded in InitPostgres().4. Support active table detection on vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman Flack for the former comments on disk quota feature. Any comments on how to improve disk quota feature are appreciated.On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Pavel.Your patch did enforcement on storage level(md.c or we could also use smgr_extend). It's straight forward.But I prefer to implement disk_quota as a feature with following objectives:1 set/alter disk quota setting on different database objects, e.g. user, database, schema etc. not only a general GUC, but we could set separate quota limit for a specific objects. 2 enforcement operator should work at two positions: before query is running and when query is running. The latter one's implementation maybe similar to your patch.The patch was just example. The resource quotes should be more complex - per partition, table, schema, database, user - so GUC are possible, but not very user friendly.Our case is specific, but not too much. The servers are used for multidimensional analyses - and some tables can grow too fast (COPY, INSERT SELECT). We need to solve limits immediately. The implementation is simple, so I did it. Same implementation on database level, or schema level needs some more locks, so it will not be too effective. The resource management can be complex very complex, and I expect so it will be hard work.RegardsPavelOn Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:Hi2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:Thanks Chapman.@Pavel, could you please explain more about your second suggestion \\\"implement some quotas on storage level?\\\"See attached patch - it is very simple - and good enough for our purposes. RegardsPavel We will not keep the long-lived processes attach to all databases(just like you mentioned servers with thousands of databases) And you are right, we could share ideas with autovacuum process, fork worker processes in need. \\\"autovacuum checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the statistics collection facility\\\"diskquota process is similar to autovacuum at caring about insert, but the difference is that it also care about vucuum full, truncate and drop. While update and delete may not be interested since no file change happens. So a separate diskquota process is preferred.So if we implemented disk quota as a full native feature, and in the first initial version I prefer to implement the following features:1 Fork diskquota launcher process under Postmaster serverloop, which is long-lived.2 Diskquota launcher process is responsible for creating diskquota worker process for every database.3 DIskquota setting is stored in a separate catalog table for each database.4 Initialization stage, Diskquota launcher process creates diskquota worker process for all the databases(traverse like autovacuum). Worker process calculates disk usage of db objects and their diskquota setting. If any db object exceeds its quota limit, put them into the blacklist in the shared memory, which will later be used by enforcement operator. Worker process exits when works are done.5 Running stage, Diskquota launcher process creates diskquota worker process for the database with a large number of insert, copy, truncate, drop etc. or create disk quota statement. Worker process updates the file size for db objects containing the result relation, and compare with the diskquota setting. Again, if exceeds quota limit, put them into blacklist, remove from blacklist vice versa. Worker process exits when works are done and a GUC could control the frequency of worker process restart to a specific database. As you know, this GUC also controls the delay when we do enforcement.6 Enforcement. When postgres backend executes queries, check the blacklist in shared memory to determine whether the query is allowed(before execute) or need rollback(is executing)?If we implemented disk quota as an extension, we could just use background worker to start diskquota launcher process and use RegisterDynamicBackgroundWorker() to fork child diskquota worker processes by the launcher process as suggested by @Chapman. Diskquota setting could be stored in user table in a separate schema for each database(Schema and table created by create extension statement) just like what Heikki has done in pg_quota project. But in this case, we need to create extension for each database before diskquota worker process can be set up for that database.Any comments on the above design and which is preferred, native feature or extension as the POC? -- HubertOn Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net>:On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n\\n> 2 Keep one worker process for each database. But using a parent/global\\n> quota worker process to manage the lifecycle of database level worker\\n> processes. It could handle the newly created database(avoid restart\\n> database) and save resource when a database is not used. But this needs to\\n> change worker process to be hierarchical. Postmaster becomes the grandfather\\n> of database level worker processes in this case.\\n\\nI am using background workers this way in 9.5 at $work.\\n\\nIn my case, one worker lives forever, wakes up on a set period, and\\nstarts a short-lived worker for every database, waiting for each\\none before starting the next.\\n\\nIt was straightforward to implement. Looking back over the code,\\nI see the global worker assigns its own PID to worker.bgw_notify_pid\\nof each of its children, and also obtains a handle for each child\\nfrom RegisterDynamicBackgroundWorker().\\n\\nI imagine the global quota worker would prefer to start workers\\nfor every database and then just wait for notifications from any\\nof them, but that seems equally straightforward at first glance.There are servers with thousands databases. Worker per database is not good idea.It should to share ideas, code with autovacuum process.Not sure, how to effective implementation based on bg workers can be. On servers with large set of databases, large set of tables it can identify too big table too late.Isn't better to implement some quotas on storage level?RegardsPavel\\n\\n-Chap\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\\n\\n\\n-- ThanksHubert Zhang\\n\\n-- ThanksHubert Zhang\\n\\n-- ThanksHubert Zhang\",\n \"msg_date\": \"Mon, 24 Sep 2018 21:56:08 +0800\",\n \"msg_from\": \"Hubert Zhang <hzhang@pivotal.io>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal for disk quota feature\"\n },\n {\n \"msg_contents\": \"Hi All,\\n\\nWe prepare a WIP patch which is implemented based on our redesigned\\nproposal. This patch is for demonstration our ideas and currently can\\nsupport create/delete disk quota for table/schema/user.\\n\\nHere is the usage,\\n--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------\\n1 Specify the databases with disk quota enabled in GUC\\n“diskquota_databases” in postgresql.conf and restart the database.\\n\\n2 DBA could set disk quota limit for table/schema/role.\\n\\nCREATE DISK QUOTA tablea1 ON TABLE a1 with (quota = ‘1MB’);\\nCREATE DISK QUOTA roleu1 ON USER u1 with (quota = ‘1GB’);\\nCREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota = ‘3MB’);\\n\\n3 Simulate a schema out of quota limit case: suppose table a1 and table a2\\nare both under schema s1.\\n\\nINSERT INTO a1 SELECT generate_series(1,1000);\\nINSERT INTO a2 SELECT generate_series(1,3000000);\\nSELECT pg_sleep(5)\\n\\nINSERT INTO a1 SELECT generate_series(1,1000);\\nERROR: schema's disk space quota exceeded\\n\\nDROP TABLE a2;\\nSELECT pg_sleep(5)\\n\\nINSERT INTO a1 SELECT generate_series(1,1000);\\nINSERT 0 1000\\n------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------\\n\\nThis patch works well on small size of databases (tens of databases) as\\neach database in GUC diskquota_databases will have a background worker\\nprocess for monitoring purpose.\\n\\nThanks very much for Heikki's suggestion and comments.\\nAll comments are very welcome!\\n\\nOn Mon, Sep 24, 2018 at 9:56 PM Hubert Zhang <hzhang@pivotal.io> wrote:\\n\\n> The quotas or object limits, resource limits are pretty useful and\\n>> necessary, but I don't see these like new type of objects, it is much more\\n>> some property of current objects. Because we have one syntax for this\\n>> purpose I prefer it. Because is not good to have two syntaxes for similar\\n>> purpose.\\n>\\n> SCHEMA and TABLE are OK for me, But as I mentioned before, ROLE is a\\n> special case when using ALTER SET at this moment.\\n> TABLE and SCHEMA are both database level, e.g. pg_class and pg_namespace\\n> both residents in one database. But ROLE is cluster-level. They don't\\n> belong to a database. ALTER ROLE XXX SET disk_quota = xxx means to set the\\n> quota for the user on all the databases in the first glance. But in our\\n> first stage design, ROLE's quota is bind to a specific database. E.g. Role\\n> Jack could have 10GB quota on database A and 2GB quota on database B.\\n>\\n> SQL syntax is not hard to modify, I don't think this should block the\\n> main design of disk quota feature. Is there any comment on the design and\\n> architecture? If no, we'll firstly submit our patch and involve more\\n> discussion?\\n>\\n> On Sat, Sep 22, 2018 at 3:03 PM Pavel Stehule <pavel.stehule@gmail.com>\\n> wrote:\\n>\\n>>\\n>>\\n>> so 22. 9. 2018 v 8:48 odesílatel Hubert Zhang <hzhang@pivotal.io> napsal:\\n>>\\n>>> But it looks like redundant to current GUC configuration and limits\\n>>>\\n>>> what do you mean by current GUC configuration? Is that the general block\\n>>> number limit in your patch? If yes, the difference between GUC and\\n>>> pg_diskquota catalog is that pg_diskquota will store different quota limit\\n>>> for the different role, schema or table instead of a single GUC value.\\n>>>\\n>>\\n>> storage is not relevant in this moment.\\n>>\\n>> I don't see to consistent to sets some limits via SET command, or ALTER X\\n>> SET, and some other with CREATE QUOTA ON.\\n>>\\n>> The quotas or object limits, resource limits are pretty useful and\\n>> necessary, but I don't see these like new type of objects, it is much more\\n>> some property of current objects. Because we have one syntax for this\\n>> purpose I prefer it. Because is not good to have two syntaxes for similar\\n>> purpose.\\n>>\\n>> So instead CREATE DISC QUATA ON SCHEMA xxx some value I prefer\\n>>\\n>> ALTER SCHEMA xxx SET disc_quota = xxx;\\n>>\\n>> The functionality is +/- same. But ALTER XX SET was introduce first, and\\n>> I don't feel comfortable to have any new syntax for similar purpose\\n>>\\n>> Regards\\n>>\\n>> Pavel\\n>>\\n>>\\n>>\\n>>\\n>>\\n>>>\\n>>> On Sat, Sep 22, 2018 at 11:17 AM Pavel Stehule <pavel.stehule@gmail.com>\\n>>> wrote:\\n>>>\\n>>>>\\n>>>>\\n>>>> pá 21. 9. 2018 v 16:21 odesílatel Hubert Zhang <hzhang@pivotal.io>\\n>>>> napsal:\\n>>>>\\n>>>>> just fast reaction - why QUOTA object?\\n>>>>>> Isn't ALTER SET enough?\\n>>>>>> Some like\\n>>>>>> ALTER TABLE a1 SET quote = 1MB;\\n>>>>>> ALTER USER ...\\n>>>>>> ALTER SCHEMA ..\\n>>>>>> New DDL commans looks like too hard hammer .\\n>>>>>\\n>>>>>\\n>>>>> It's an option. Prefer to consider quota setting store together:\\n>>>>> CREATE DISK QUOTA way is more nature to store quota setting in a\\n>>>>> separate pg_diskquota catalog\\n>>>>> While ALTER SET way is more close to store quota setting in pg_class,\\n>>>>> pg_role, pg_namespace. etc in an integrated way.\\n>>>>> (Note that here I mean nature/close is not must, ALTER SET could also\\n>>>>> store in pg_diskquota and vice versa.)\\n>>>>>\\n>>>>\\n>>>> I have not a problem with new special table for storing this\\n>>>> information. But it looks like redundant to current GUC configuration and\\n>>>> limits. Can be messy do some work with ALTER ROLE, and some work via CREATE\\n>>>> QUOTE.\\n>>>>\\n>>>> Regards\\n>>>>\\n>>>> Pavel\\n>>>>\\n>>>>\\n>>>>> Here are some differences I can think of:\\n>>>>> 1 pg_role is a global catalog, not per database level. It's harder to\\n>>>>> tracker the user's disk usage in the whole clusters(considering 1000+\\n>>>>> databases). So the semantic of CREATE DISK QUOTA ON USER is limited: it\\n>>>>> only tracks the user's disk usage inside the current database.\\n>>>>> 2 using separate pg_diskquota could add more field except for quota\\n>>>>> limit without adding too many fields in pg_class, e.g. red zone to give the\\n>>>>> user a warning or the current disk usage of the db objects.\\n>>>>>\\n>>>>> On Fri, Sep 21, 2018 at 8:01 PM Pavel Stehule <pavel.stehule@gmail.com>\\n>>>>> wrote:\\n>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>> pá 21. 9. 2018 v 13:32 odesílatel Hubert Zhang <hzhang@pivotal.io>\\n>>>>>> napsal:\\n>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>> *Hi all,We redesign disk quota feature based on the comments from\\n>>>>>>> Pavel Stehule and Chapman Flack. Here are the new design.OverviewBasically,\\n>>>>>>> disk quota feature is used to support multi-tenancy environment, different\\n>>>>>>> level of database objects could be set a quota limit to avoid over use of\\n>>>>>>> disk space. A common case could be as follows: DBA could enable disk quota\\n>>>>>>> on a specified database list. DBA could set disk quota limit for\\n>>>>>>> tables/schemas/roles in these databases. Separate disk quota worker process\\n>>>>>>> will monitor the disk usage for these objects and detect the objects which\\n>>>>>>> exceed their quota limit. Queries loading data into these “out of disk\\n>>>>>>> quota” tables/schemas/roles will be cancelled.We are currently working at\\n>>>>>>> init implementation stage. We would like to propose our idea firstly and\\n>>>>>>> get feedbacks from community to do quick iteration.SQL Syntax (How to use\\n>>>>>>> disk quota)1 Specify the databases with disk quota enabled in GUC\\n>>>>>>> “diskquota_databases” in postgresql.conf and restart the database.2 DBA\\n>>>>>>> could set disk quota limit for table/schema/role.CREATE DISK QUOTA tablea1\\n>>>>>>> ON TABLE a1 with (quota = ‘1MB’);CREATE DISK QUOTA roleu1 ON USER u1 with\\n>>>>>>> (quota = ‘1GB’);CREATE DISK QUOTA schemas1 ON SCHEMA s1 with (quota =\\n>>>>>>> ‘3MB’);*\\n>>>>>>>\\n>>>>>>\\n>>>>>> just fast reaction - why QUOTA object?\\n>>>>>>\\n>>>>>> Isn't ALTER SET enough?\\n>>>>>>\\n>>>>>> Some like\\n>>>>>>\\n>>>>>> ALTER TABLE a1 SET quote = 1MB;\\n>>>>>> ALTER USER ...\\n>>>>>> ALTER SCHEMA ..\\n>>>>>>\\n>>>>>> New DDL commans looks like too hard hammer .\\n>>>>>>\\n>>>>>>\\n>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>> *3 Simulate a schema out of quota limit case: suppose table a1 and\\n>>>>>>> table a2 are both under schema s1.INSERT INTO a1 SELECT\\n>>>>>>> generate_series(1,1000);INSERT INTO a2 SELECT\\n>>>>>>> generate_series(1,3000000);SELECT pg_sleep(5)INSERT INTO a1 SELECT\\n>>>>>>> generate_series(1,1000);ERROR: schema's disk space quota exceededDROP\\n>>>>>>> TABLE a2;SELECT pg_sleep(5)INSERT INTO a1 SELECT\\n>>>>>>> generate_series(1,1000);INSERT 0 1000ArchitectureDisk quota has the\\n>>>>>>> following components.1. Quota Setting Store is where the disk quota setting\\n>>>>>>> to be stored and accessed. We plan to use catalog table pg_diskquota to\\n>>>>>>> store these information. pg_diskquota is\\n>>>>>>> like:CATALOG(pg_diskquota,6122,DiskQuotaRelationId){ NameData quotaname; /*\\n>>>>>>> diskquota name */ int16 quotatype; /* diskquota type name */ Oid\\n>>>>>>> quotatargetoid; /* diskquota target db object oid*/ int32 quotalimit; /*\\n>>>>>>> diskquota size limit in MB*/ int32 quotaredzone; /* diskquota redzone in\\n>>>>>>> MB*/} FormData_pg_diskquota;2. Quota Change Detector is the monitor of size\\n>>>>>>> change of database objects. We plan to use stat collector to detect the\\n>>>>>>> ‘active’ table list at initial stage. But stat collector has some\\n>>>>>>> limitation on finding the active table which is in a running transaction.\\n>>>>>>> Details see TODO section.3. Quota Size Checker is where to calculate the\\n>>>>>>> size and compare with quota limit for database objects. According to\\n>>>>>>> Pavel’s comment, autovacuum launcher and worker process could be a good\\n>>>>>>> reference to disk quota. So we plan to use a disk quota launcher daemon\\n>>>>>>> process and several disk quota worker process to finish this work. Launcher\\n>>>>>>> process is responsible for starting worker process based on a user defined\\n>>>>>>> database list from GUC. Worker process will connect to its target database\\n>>>>>>> and monitor the disk usage for objects in this database. In init stage of\\n>>>>>>> worker process, it will call calculate_total_relation_size() to calculate\\n>>>>>>> the size for each user table. After init stage, worker process will refresh\\n>>>>>>> the disk model every N seconds. Refreshing will only recalculate the size\\n>>>>>>> of tables in ‘active’ table list, which is generated by Quata Change\\n>>>>>>> Detector to minimize the cost.4. Quota Enforcement Operator is where to\\n>>>>>>> check for the quota limitation at postgres backend side. We will firstly\\n>>>>>>> implement it in ExecCheckRTPerms() as pre-running enforcement. It will\\n>>>>>>> check the disk quota of tables being inserted or updated, and report error\\n>>>>>>> if table’s or table’s schema’s or table’s owner’s quota limit is exceeded.\\n>>>>>>> As a native feature, we plan to add more checkpoint to do running query\\n>>>>>>> enforcement. For example, if a disk quota lefts 10MB quota, a query could\\n>>>>>>> insert 1GB data. This query could be allowed in pre-running enforcement\\n>>>>>>> check, but will be cancelled in running query enforcement check. Therefore,\\n>>>>>>> it can improve the accurate of disk quota usage. To achieve this, we plan\\n>>>>>>> to add a checkpoint in lower API such as smgr_extened. Hence, the Quota\\n>>>>>>> Enforcement Operator will check the disk quota usage when smgr_extened is\\n>>>>>>> called. If the quota is over limited, current query will be cancelled.\\n>>>>>>> Highlight1. Native feature.Support native Create/Drop Disk Quota SQL\\n>>>>>>> statement.New catalog table pg_diskquota to store disk quota setting.2.\\n>>>>>>> Auto DML/DDL detection. Table\\n>>>>>>> create/update/insert/delete/vacuum/truncate/drop/schema_change/owner_change,\\n>>>>>>> Schema create/drop and Role create/drop will be detected by disk quota\\n>>>>>>> automatically. 3. Low cost disk quota checker.Worker process of disk quota\\n>>>>>>> need to refresh the disk usage model every N seconds. Since recalculate the\\n>>>>>>> file size using stat() system call is expensive for a large number of\\n>>>>>>> files, we use an ‘active’ table list to reduce the real work at each\\n>>>>>>> iteration. A basic experiment on our init stage implementation on database\\n>>>>>>> with 20K tables shows that the refresh cost is 1% cpu usage and will be\\n>>>>>>> finished within 50ms. Todo/LimitationBefore we propose our patch, we plan\\n>>>>>>> to enhance it with the following ideas:1. Setting database list with disk\\n>>>>>>> quota enabled dynamically without restart database. Since we have the disk\\n>>>>>>> quota launcher process, it could detect the new ‘diskquota_databases’ list\\n>>>>>>> and start/stop the corresponding disk quota worker process.2. Enforcement\\n>>>>>>> when query is running. Considering the case when there is 10MB quota left,\\n>>>>>>> but next query will insert 10GB data. Current enforcement design will allow\\n>>>>>>> this query to be executed. This is limited by the ‘active’ table detection\\n>>>>>>> is generated by stat collector. Postgres backend will only send table stat\\n>>>>>>> information to collector only when the transaction ends. We need a new way\\n>>>>>>> to detect the ‘active’ table even when this table is being modified inside\\n>>>>>>> a running transaction.3. Monitor unlimited number of databases. Current we\\n>>>>>>> set the max number of disk quota worker process to be 10 to reduce the\\n>>>>>>> affection normal workload. But how about if we want to monitor the disk\\n>>>>>>> quota of more than 10 databases? Our solution is to let disk quota launcher\\n>>>>>>> to manage a queue of database need to be monitored. And disk quota worker\\n>>>>>>> process consuming the queue and refresh the disk usage/quota for this\\n>>>>>>> database. After some periods, worker will return the database to the queue,\\n>>>>>>> and fetch the top database from queue to process. The period determine the\\n>>>>>>> delay of detecting disk quota change. To implement this feature, we need to\\n>>>>>>> support a subprocess of postmaster to rebind to another database instead of\\n>>>>>>> the database binded in InitPostgres().4. Support active table detection on\\n>>>>>>> vacuum full and vacuum analyze. Currently vacuum full and vacuum analyze\\n>>>>>>> are not tracked by stat collector.Thanks to Heikki, Pavel Stehule,Chapman\\n>>>>>>> Flack for the former comments on disk quota feature. Any comments on how to\\n>>>>>>> improve disk quota feature are appreciated.*\\n>>>>>>>\\n>>>>>>>\\n>>>>>>> On Mon, Sep 3, 2018 at 12:05 PM, Pavel Stehule <\\n>>>>>>> pavel.stehule@gmail.com> wrote:\\n>>>>>>>\\n>>>>>>>>\\n>>>>>>>>\\n>>>>>>>> 2018-09-03 3:49 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>>>>>>>\\n>>>>>>>>> Thanks Pavel.\\n>>>>>>>>> Your patch did enforcement on storage level(md.c or we could also\\n>>>>>>>>> use smgr_extend). It's straight forward.\\n>>>>>>>>> But I prefer to implement disk_quota as a feature with following\\n>>>>>>>>> objectives:\\n>>>>>>>>> 1 set/alter disk quota setting on different database objects, e.g.\\n>>>>>>>>> user, database, schema etc. not only a general GUC, but we could set\\n>>>>>>>>> separate quota limit for a specific objects.\\n>>>>>>>>> 2 enforcement operator should work at two positions: before query\\n>>>>>>>>> is running and when query is running. The latter one's implementation maybe\\n>>>>>>>>> similar to your patch.\\n>>>>>>>>>\\n>>>>>>>>\\n>>>>>>>> The patch was just example. The resource quotes should be more\\n>>>>>>>> complex - per partition, table, schema, database, user - so GUC are\\n>>>>>>>> possible, but not very user friendly.\\n>>>>>>>>\\n>>>>>>>> Our case is specific, but not too much. The servers are used for\\n>>>>>>>> multidimensional analyses - and some tables can grow too fast (COPY, INSERT\\n>>>>>>>> SELECT). We need to solve limits immediately. The implementation is simple,\\n>>>>>>>> so I did it. Same implementation on database level, or schema level needs\\n>>>>>>>> some more locks, so it will not be too effective. The resource management\\n>>>>>>>> can be complex very complex, and I expect so it will be hard work.\\n>>>>>>>>\\n>>>>>>>> Regards\\n>>>>>>>>\\n>>>>>>>> Pavel\\n>>>>>>>>\\n>>>>>>>>\\n>>>>>>>>> On Sun, Sep 2, 2018 at 8:44 PM, Pavel Stehule <\\n>>>>>>>>> pavel.stehule@gmail.com> wrote:\\n>>>>>>>>>\\n>>>>>>>>>> Hi\\n>>>>>>>>>>\\n>>>>>>>>>> 2018-09-02 14:18 GMT+02:00 Hubert Zhang <hzhang@pivotal.io>:\\n>>>>>>>>>>\\n>>>>>>>>>>> Thanks Chapman.\\n>>>>>>>>>>> @Pavel, could you please explain more about your second\\n>>>>>>>>>>> suggestion \\\"implement some quotas on storage level?\\\"\\n>>>>>>>>>>>\\n>>>>>>>>>>\\n>>>>>>>>>> See attached patch - it is very simple - and good enough for our\\n>>>>>>>>>> purposes.\\n>>>>>>>>>>\\n>>>>>>>>>> Regards\\n>>>>>>>>>>\\n>>>>>>>>>> Pavel\\n>>>>>>>>>>\\n>>>>>>>>>>\\n>>>>>>>>>>\\n>>>>>>>>>>> We will not keep the long-lived processes attach to all\\n>>>>>>>>>>> databases(just like you mentioned servers with thousands of\\n>>>>>>>>>>> databases)\\n>>>>>>>>>>> And you are right, we could share ideas with autovacuum process,\\n>>>>>>>>>>> fork worker processes in need.\\n>>>>>>>>>>> \\\"autovacuum checks for tables that have had a large number of\\n>>>>>>>>>>> inserted, updated or deleted tuples. These checks use the statistics\\n>>>>>>>>>>> collection facility\\\"\\n>>>>>>>>>>> diskquota process is similar to autovacuum at caring about\\n>>>>>>>>>>> insert, but the difference is that it also care about vucuum full, truncate\\n>>>>>>>>>>> and drop. While update and delete may not be interested since no file\\n>>>>>>>>>>> change happens. So a separate diskquota process is preferred.\\n>>>>>>>>>>>\\n>>>>>>>>>>> So if we implemented disk quota as a full native feature, and\\n>>>>>>>>>>> in the first initial version I prefer to implement the following features:\\n>>>>>>>>>>> 1 Fork diskquota launcher process under Postmaster serverloop,\\n>>>>>>>>>>> which is long-lived.\\n>>>>>>>>>>> 2 Diskquota launcher process is responsible for\\n>>>>>>>>>>> creating diskquota worker process for every database.\\n>>>>>>>>>>> 3 DIskquota setting is stored in a separate catalog table for\\n>>>>>>>>>>> each database.\\n>>>>>>>>>>> 4 Initialization stage, Diskquota launcher process creates\\n>>>>>>>>>>> diskquota worker process for all the databases(traverse like\\n>>>>>>>>>>> autovacuum). Worker process calculates disk usage of db objects and\\n>>>>>>>>>>> their diskquota setting. If any db object exceeds its quota limit, put them\\n>>>>>>>>>>> into the blacklist in the shared memory, which will later be used by\\n>>>>>>>>>>> enforcement operator. Worker process exits when works are done.\\n>>>>>>>>>>> 5 Running stage, Diskquota launcher process creates diskquota worker\\n>>>>>>>>>>> process for the database with a large number of insert, copy, truncate,\\n>>>>>>>>>>> drop etc. or create disk quota statement. Worker process updates the file\\n>>>>>>>>>>> size for db objects containing the result relation, and compare with the\\n>>>>>>>>>>> diskquota setting. Again, if exceeds quota limit, put them into blacklist,\\n>>>>>>>>>>> remove from blacklist vice versa. Worker process exits when\\n>>>>>>>>>>> works are done and a GUC could control the frequency of worker process\\n>>>>>>>>>>> restart to a specific database. As you know, this GUC also controls the\\n>>>>>>>>>>> delay when we do enforcement.\\n>>>>>>>>>>> 6 Enforcement. When postgres backend executes queries, check the\\n>>>>>>>>>>> blacklist in shared memory to determine whether the query is allowed(before\\n>>>>>>>>>>> execute) or need rollback(is executing)?\\n>>>>>>>>>>>\\n>>>>>>>>>>> If we implemented disk quota as an extension, we could just use\\n>>>>>>>>>>> background worker to start diskquota launcher process and use\\n>>>>>>>>>>> RegisterDynamicBackgroundWorker() to fork child diskquota\\n>>>>>>>>>>> worker processes by the launcher process as suggested by @Chapman.\\n>>>>>>>>>>> Diskquota setting could be stored in user table in a separate schema for\\n>>>>>>>>>>> each database(Schema and table created by create extension statement) just\\n>>>>>>>>>>> like what Heikki has done in pg_quota project. But in this case, we need to\\n>>>>>>>>>>> create extension for each database before diskquota worker process can be\\n>>>>>>>>>>> set up for that database.\\n>>>>>>>>>>>\\n>>>>>>>>>>> Any comments on the above design and which is preferred, native\\n>>>>>>>>>>> feature or extension as the POC?\\n>>>>>>>>>>>\\n>>>>>>>>>>>\\n>>>>>>>>>>> -- Hubert\\n>>>>>>>>>>>\\n>>>>>>>>>>>\\n>>>>>>>>>>>\\n>>>>>>>>>>> On Fri, Aug 31, 2018 at 3:32 AM, Pavel Stehule <\\n>>>>>>>>>>> pavel.stehule@gmail.com> wrote:\\n>>>>>>>>>>>\\n>>>>>>>>>>>>\\n>>>>>>>>>>>>\\n>>>>>>>>>>>> 2018-08-30 16:22 GMT+02:00 Chapman Flack <chap@anastigmatix.net\\n>>>>>>>>>>>> >:\\n>>>>>>>>>>>>\\n>>>>>>>>>>>>> On 08/30/2018 09:57 AM, Hubert Zhang wrote:\\n>>>>>>>>>>>>>\\n>>>>>>>>>>>>> > 2 Keep one worker process for each database. But using a\\n>>>>>>>>>>>>> parent/global\\n>>>>>>>>>>>>> > quota worker process to manage the lifecycle of database\\n>>>>>>>>>>>>> level worker\\n>>>>>>>>>>>>> > processes. It could handle the newly created database(avoid\\n>>>>>>>>>>>>> restart\\n>>>>>>>>>>>>> > database) and save resource when a database is not used. But\\n>>>>>>>>>>>>> this needs to\\n>>>>>>>>>>>>> > change worker process to be hierarchical. Postmaster becomes\\n>>>>>>>>>>>>> the grandfather\\n>>>>>>>>>>>>> > of database level worker processes in this case.\\n>>>>>>>>>>>>>\\n>>>>>>>>>>>>> I am using background workers this way in 9.5 at $work.\\n>>>>>>>>>>>>>\\n>>>>>>>>>>>>> In my case, one worker lives forever, wakes up on a set\\n>>>>>>>>>>>>> period, and\\n>>>>>>>>>>>>> starts a short-lived worker for every database, waiting for\\n>>>>>>>>>>>>> each\\n>>>>>>>>>>>>> one before starting the next.\\n>>>>>>>>>>>>>\\n>>>>>>>>>>>>> It was straightforward to implement. Looking back over the\\n>>>>>>>>>>>>> code,\\n>>>>>>>>>>>>> I see the global worker assigns its own PID to\\n>>>>>>>>>>>>> worker.bgw_notify_pid\\n>>>>>>>>>>>>> of each of its children, and also obtains a handle for each\\n>>>>>>>>>>>>> child\\n>>>>>>>>>>>>> from RegisterDynamicBackgroundWorker().\\n>>>>>>>>>>>>>\\n>>>>>>>>>>>>> I imagine the global quota worker would prefer to start workers\\n>>>>>>>>>>>>> for every database and then just wait for notifications from\\n>>>>>>>>>>>>> any\\n>>>>>>>>>>>>> of them, but that seems equally straightforward at first\\n>>>>>>>>>>>>> glance.\\n>>>>>>>>>>>>>\\n>>>>>>>>>>>>\\n>>>>>>>>>>>> There are servers with thousands databases. Worker per database\\n>>>>>>>>>>>> is not good idea.\\n>>>>>>>>>>>>\\n>>>>>>>>>>>> It should to share ideas, code with autovacuum process.\\n>>>>>>>>>>>>\\n>>>>>>>>>>>> Not sure, how to effective implementation based on bg workers\\n>>>>>>>>>>>> can be. On servers with large set of databases, large set of tables it can\\n>>>>>>>>>>>> identify too big table too late.\\n>>>>>>>>>>>>\\n>>>>>>>>>>>> Isn't better to implement some quotas on storage level?\\n>>>>>>>>>>>>\\n>>>>>>>>>>>> Regards\\n>>>>>>>>>>>>\\n>>>>>>>>>>>> Pavel\\n>>>>>>>>>>>>\\n>>>>>>>>>>>>\\n>>>>>>>>>>>>\\n>>>>>>>>>>>>> -Chap\\n>>>>>>>>>>>>>\\n>>>>>>>>>>>>>\\n>>>>>>>>>>>>\\n>>>>>>>>>>>\\n>>>>>>>>>>>\\n>>>>>>>>>>> --\\n>>>>>>>>>>> Thanks\\n>>>>>>>>>>>\\n>>>>>>>>>>> Hubert Zhang\\n>>>>>>>>>>>\\n>>>>>>>>>>\\n>>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>>\\n>>>>>>>>> --\\n>>>>>>>>> Thanks\\n>>>>>>>>>\\n>>>>>>>>> Hubert Zhang\\n>>>>>>>>>\\n>>>>>>>>\\n>>>>>>>>\\n>>>>>>>\\n>>>>>>>\\n>>>>>>> --\\n>>>>>>> Thanks\\n>>>>>>>\\n>>>>>>> Hubert Zhang\\n>>>>>>>\\n>>>>>>\\n>>>>>\\n>>>>> --\\n>>>>> Thanks\\n>>>>>\\n>>>>> Hubert Zhang\\n>>>>>\\n>>>>\\n>>>\\n>>> --\\n>>> Thanks\\n>>>\\n>>> Hubert Zhang\\n>>>\\n>>\\n>\\n> --\\n> Thanks\\n>\\n> Hubert Zhang\\n>\\n\\n\\n-- \\nRegards,\\nHaozhou\",\n \"msg_date\": \"Wed, 26 Sep 2018 18:10:13 +0800\",\n \"msg_from\": \"Haozhou Wang <hawang@pivotal.io>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal for disk quota feature\"\n }\n]","isConversational":false}}}],"truncated":false,"partial":false},"paginationData":{"pageIndex":471,"numItemsPerPage":100,"numTotalItems":58371,"offset":47100,"length":100}},"jwt":"eyJhbGciOiJFZERTQSJ9.eyJyZWFkIjp0cnVlLCJwZXJtaXNzaW9ucyI6eyJyZXBvLmNvbnRlbnQucmVhZCI6dHJ1ZX0sImlhdCI6MTc3MjYxNDc3OCwic3ViIjoiL2RhdGFzZXRzL3dhbnNoZW5sL3Bnc3FsLWhhY2tlcnMiLCJleHAiOjE3NzI2MTgzNzgsImlzcyI6Imh0dHBzOi8vaHVnZ2luZ2ZhY2UuY28ifQ.bYOyP7I5j5D-70wKCxMoLkAI6Ql7bXh-NhQFfN2f5OX0I5-p6jq54YrN8vDEy87HiXxiwa8mYRjXEAZsw3M8Cw","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": "Hello!\n\nIn an environment where we control the host system and all installed\nextensions, we need to allow postgresql non-superuser to install all of\nthem, without opening gaps that will let this user gain superuser\nprivileges. We have a sample solution to add a new default role\npg_creat... |
[
{
"msg_contents": "Hi,\n\nrunning sqlsmith on REL_11_STABLE at 1b9d1b08fe for a couple hours\nyielded the previously-unseen internal error \"partition missing from\nsubplans\". It is readily reproducible on the regression database with\nthe following query:\n\nselect * from public.fk_partitioned_fk as sample_0... |
[
{
"msg_contents": "\nAll the remaining items have been moved to the next commitfest, in some \ncases with changed status.\n\n\nThanks to all the authors, reviewers and committers.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, R... |
[
{
"msg_contents": "Hello pgdev,\n\nThis patch rework & clarifies pgbench internal state machine.\n\nIt was indirectly triggered by Heikki review of pgbench tap tests \n(https://commitfest.postgresql.org/19/1306/), and by Marina's patch about \ntx retry on some errors (https://commitfest.postgresql.org/19/1645/)... |
[
{
"msg_contents": "Hi,\n\nsqlsmith caused another internal error while testing REL_11_STABLE at\n1b9d1b08fe. The query below on the regression DB yields \"plan should\nnot reference subplan's variable\" for me.\n\nregards,\nAndreas\n\ndelete from public.prt1_l\nwhere\nEXISTS (\n select\n from\n public... |
[
{
"msg_contents": "In a moment of idle curiosity, I tried to build PG --with-libxml\non NetBSD-current (well, mostly current, from May or so).\nThe configure script blew up, complaining that it couldn't execute\na test program. Investigation showed that xml2-config reports this:\n\n$ xml2-config --libs\n-Wl,-R... |
[
{
"msg_contents": "Hello,\n\nTo scratch an itch, I have been working on teaching TidScan how to do\nrange queries, i.e. those using >=, <, BETWEEN, etc. This means we\ncan write, for instance,\n\n SELECT * FROM t WHERE ctid >= '(1000,0)' AND ctid < '(2000,0)';\n\ninstead of resorting to the old trick:\n\n ... |
[
{
"msg_contents": "Hi all,\n(In CC are the folks who have reviewed the first patch versions, Nathan\nand Horiguchi-san.) \n\nAfter TRUNCATE and REINDEX, here is the third and last thread I am\nspawning for the previous thread \"Canceling authentication due to\ntimeout aka Denial of Service Attack\":\nhttps://ww... |
[
{
"msg_contents": "While poking at that xml issue I tried to compare the memory usage of\nthe xml queries with vaguely comparable regexp_split_to_table queries,\nand found that I could not do so; the regexp queries simply did not\ncomplete in any sensible timeframe.\n\nInvestigation showed that in UTF8 (or othe... |
[
{
"msg_contents": "Hi!\n\nThis subject was already raised multiple times [1], [2], [3]. In\nshort, our LWLock implementation has pathological behavior when shared\nlockers constitute a continuous flood. In this case exclusive lock\nwaiters are not guaranteed to eventually get the lock. When shared\nlock is h... |
[
{
"msg_contents": "I noticed that the word \"procedure\" appears in at least a few places\nin the v11 docs as a not-quite-apt synonym of \"function\". For example,\nhttps://www.postgresql.org/docs/11/static/plpgsql-trigger.html talks\nabout \"PL/pgSQL Trigger Procedures\" which are actually all functions\nin pr... |
[
{
"msg_contents": "All the regexp functions blow up with \"invalid memory alloc request\"\nerrors when the input string exceeds 256MB in length. This restriction\ndoes not seem to be documented anywhere that I could see.\n\n(Also for regexp_split* and regexp_matches, there's a limit of 64M total\nmatches, which... |
[
{
"msg_contents": "On Thu, Nov 30, 2017 at 11:09 AM, Michael Paquier\n<michael.paquier@gmail.com> wrote:\n> On Tue, Oct 24, 2017 at 5:54 AM, Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n>> Yeah, I was thinking the commit is relevant with this issue but as\n>> Amit mentioned this error is emitted by DROP SCHE... |
[
{
"msg_contents": "Hello,\nI used PostgreSQL 9.3 but I executed upgrade few days ago.\nNow, I am using 10.4 PostgreSQL and:\ndoctrine/annotations v1.2.7\ndoctrine/cache v1.4.2\ndoctrine/collections v1.3.0\ndoctrine/common v2.7.3\ndoctrin... |
[
{
"msg_contents": "Hi\n\nI would to introduce new flag for routines - PRIVATE. Routines with this\nflag can be called only from other routines assigned with same schema.\nBecause these routines are not available for top queries, we can hide these\nfunctions from some outputs like \\df ..\n\nExample:\n\nCREATE S... |
[
{
"msg_contents": "It seems to me that it would make sense if InsertPgAttributeTuple() were\nto set attcacheoff to -1 instead of taking it from the caller.\n\nInsertPgAttributeTuple() is the interface between in-memory tuple\ndescriptors and on-disk pg_attribute, so it makes sense to give it the\njob of resetti... |
[
{
"msg_contents": "The pq_sendfloat4 and 8 prototypes are duplicated.\n\nChristoph",
"msg_date": "Tue, 14 Aug 2018 18:55:37 +0200",
"msg_from": "Christoph Berg <myon@debian.org>",
"msg_from_op": true,
"msg_subject": "[patch] Duplicated pq_sendfloat4/8 prototypes"
},
{
"msg_contents": "Ch... |
[
{
"msg_contents": "Hello hackers,\n\nAt the Unconference in Ottawa this year, I pitched the idea of moving\ncomponents off of SLRU and on to the buffer cache. The motivation\nbehind the idea was three fold:\n\n * Improve performance by eliminating fixed sized caches, simplistic\n scan and eviction algorithm... |
[
{
"msg_contents": "I noticed that src/port/snprintf.c still follows the old (pre-C99)\nsemantics for what to return from snprintf() after buffer overrun.\nThis seems bad for a couple of reasons:\n\n* It results in potential performance loss in psnprintf and\nappendPQExpBufferVA, since those have to fly blind ab... |
[
{
"msg_contents": "Hi,\n\nI found some improvements in Client Applications in /src/bin/scripts when I\nresumed development of progress monitor for cluster command.\n\nAttached patch gives the following query a semicolon for readability.\n\n s/SELECT pg_catalog.set_config ('search_path', '', false)/\n SELE... |
[
{
"msg_contents": "1. Why do the files holding the spilled transaction data in reorderbuffer.c\nhave a \".snap\" suffix?\n\n2. Those files can get quite large. Would it be reasonable to store them in\nanother directory (e.g. pg_replslot_tmp) so that they can be placed on\nanother mount point? It would also simp... |
[
{
"msg_contents": "hi all! I discovered today that the REFRESH MATERIALIZED VIEW \ndocumentation doesn't mention that only the owner (or a superuser) may \nactually perform the refresh operation. This patch adds a note to that \neffect.\n\nDian Fay",
"msg_date": "Wed, 15 Aug 2018 19:46:49 -0400",
"msg_f... |
[
{
"msg_contents": "Hi, Hackers,\n\nThe attached small patch updates src/backend/utils/errcode.txt.\nIn the comment of this file, a list of file names generated from this file is written.\nHowever, the file name of the 'errcodes.sgml' file is incorrect.\nThe attached patch fixes the incorrect path of errcodes.sg... |
[
{
"msg_contents": "Hi hackers,\n\nWhile contemplating plans to use new kinds of OIDs in buffer tags, I\nnoticed $SUBJECT. It says:\n\n * Check for wraparound of the OID counter. We *must* not return 0\n * (InvalidOid); and as long as we have to check that, it seems a good\n * idea to skip over everything belo... |
[
{
"msg_contents": "This is the continuation of the discussion at:\n\nhttps://www.postgresql.org/message-id/7500.1531920772%40sss.pgh.pa.us\n\nActually, for more background on what I've written below, reading this\nemail in the same discussion would help:\n\nhttps://www.postgresql.org/message-id/4114.1531674142@... |
[
{
"msg_contents": "Hi,\n\nAlmost all client applications and extensions will show \"Options\" and\n\"Connection options\" sections when running with help option (--help).\nHowever, \"oid2name\" was different, there is only the Options section.\nFor example,\n\n$ vacuumlo --help\n vacuumlo removes unreferenc... |
[
{
"msg_contents": "The lack of LargeObject-support makes it unusable for installations using it.\n \nAny plans to support LOs in logical decoding so we can replicate databases \nusing them?\n \n--\nAndreas Joseph Krogh",
"msg_date": "Thu, 16 Aug 2018 13:31:14 +0200 (CEST)",
"msg_from": "Andreas Joseph K... |
[
{
"msg_contents": "Hi,\n\nI am trying to change cstore_fdw to scan in parallel. FDW interface provide\n'EstimateDSMForeignScan' for required shared memory size to scan in parallel.\n\nHow to estimate the shared memory size required for parallel scan?\n\nthanks.\n-- \nMasayuki Takahashi\n\n",
"msg_date": "Th... |
[
{
"msg_contents": "FETCH FIRST with PERCENT option is SQL standard that state limit count to\nbe specified in a percentage in addition to specify it in exact count and\nlisted as one of Major features simply not implemented yet in recent wiki\npage [1].\n\nI implemented it by executing the query plan twice. One... |
[
{
"msg_contents": "Hi,\n\nI found that has_table_privilege returns an error when a table is specified\nby schema-qualified name and the user doen't have privilege for its schema.\n\n postgres=> select has_table_privilege('myschema.tbl','select');\n ERROR: permission denied for schema myschema\n\nI think that t... |
[
{
"msg_contents": "Hello,\n\nPlease find below a submission of a patch to the PostgreSQL documentation.\n\nYou can also find the patch in this git commit:\nhttps://gitlab.com/paulrbr/postgresql/commit/024d3870450df6dcdc69bddbe2de46084b73e3a2.diff\n\n============\ncommit 024d3870450df6dcdc69bddbe2de46084b73e3a2\... |
[
{
"msg_contents": "Hi, hackers!\n\nI propose the patch for fix one small code defect.\nThe XLogReadRecord() function reads the pages of a WAL segment that \ncontain a WAL-record. Then it creates a readRecordBuf buffer in private \nmemory of a backend and copy record from the pages to the readRecordBuf \nbuffer.... |
[
{
"msg_contents": "Hello hackers,\n\n\n\nThe attached patch adds a new option work_mem to postgres_fdw contrib module.\n\nPreviously, it was impossible to change the setting of work_mem for remote session with connection by postgres_fdw.\n\nBy adding this option to the CREATE SERVER statement, you can also chan... |
[
{
"msg_contents": "Follow up on a patch and discussion with Tom, currently integer parsing on \nkeywords in libpq is quite loose, resulting in trailing garbage being \nignored and allowing to hide bugs, eg:\n\n sh> psql \"connect_timeout=2,port=5433\"\n\nThe timeout is set to 2, and the port directive is sile... |
[
{
"msg_contents": "Hi,\n\nIn [1] , it was shown that the partition tuples are needlessly formed\nand deformed during tuple conversion (do_convert_tuple), when the same\noperation can be done using tuple slots. This is because the input\nslot might already have a deformed tuple.\n\nAttached is a patch tup_conver... |
[
{
"msg_contents": "Over in the what-about-%m thread, we speculated about replacing the\nplatform's *printf functions if they didn't support %m, which would\nbasically mean using src/port/snprintf.c on all non-glibc platforms,\nrather than only on Windows as happens right now (ignoring some\nobsolete platforms w... |
[
{
"msg_contents": "Hi,\n\nI Initially pointed out here[1] that running REFRESH MATERIALIZED VIEW as a\nnon-superuser or table owner yields the following message:\n\n test=> REFRESH MATERIALIZED VIEW blah;\n ERROR: must be owner of relation blah\n\nThe error message should say \"...owner of materialized vi... |
[
{
"msg_contents": "Hi developers,\n\nI’m currently building a Postgres C extension that fetch data from a Postgres table.\nSince the table can be large, in order to prevent memory overrun, I use SPI_cursor_fetch to fetch chunks of data. The result rows are saved in SPITupleTable* SPI_tuptable and attributes are... |
[
{
"msg_contents": "Hello hackers,\n\nThis is a late summary of pg performance farm in gsoc. Although it is not\nyet perfect, but it has began to take shape.\n\n1. The implementation of the basic test results upload interface to ensure\nthat the upload operation for the atomic level (including different client\n... |
[
{
"msg_contents": "Since the introduction of the CATALOG_VARLEN stuff, the fixed size of\npg_attribute is exactly sizeof(FormData_pg_attribute), so the ancient\nmechanism to track the fixed size manually using\nATTRIBUTE_FIXED_PART_SIZE can be removed.\n\n-- \nPeter Eisentraut http://www.2ndQuadran... |
[
{
"msg_contents": "Hello monters and hackers,\n\nThis is a late summary of pg performance farm in gsoc. Although it is not\nyet perfect, but it has began to take shape.\n\n1. The implementation of the basic test results upload interface to ensure\nthat the upload operation for the atomic level (including differ... |
[
{
"msg_contents": "Hi,\n\n\"For examples of usage look at simplehash.c ...\"\n\nThere is no such file in the tree. Maybe this should say tidbitmap.c?\n\n-- \nThomas Munro\nhttp://www.enterprisedb.com\n\n",
"msg_date": "Mon, 20 Aug 2018 16:38:20 +1200",
"msg_from": "Thomas Munro <thomas.munro@enterprise... |
[
{
"msg_contents": "Currently in PG, the precedence of = and <> is supposed to be equal, and\nthe precedence of unary - is very high.\n\nSo, (true=1 between 1 and 1) parses as ((true)=(1 between 1 and 1)),\nand (true=-1 between 1 and 1) parses as ((true)=((-1) between 1 and 1)).\nAll good so far.\n\n(true<>-1 be... |
[
{
"msg_contents": "I believe that we should be delaying the PlannerInfo's\ntotal_table_pages calculation until after constraint exclusion and\npartition pruning have taken place. Doing this calculation before we\ndetermine which relations we don't need to scan can lead to\nincorrectly applying random_page_cost ... |
[
{
"msg_contents": "Hello devs,\n\nWhile reviewing various patches by Tom which are focussing on libpq \nmulti-host behavior,\n\n \thttps://commitfest.postgresql.org/19/1749/\n \thttps://commitfest.postgresql.org/19/1752/\n\nit occured to me that there are a few more problems with the \ndocumentation, the host/h... |
[
{
"msg_contents": "The attached patch factors out the CREATE INDEX CONCURRENTLY code that\nwaits for transactions with older snapshots to finish into a new\nfunction WaitForOlderSnapshots().\n\nThis refactoring was part of a previously posted REINDEX CONCURRENTLY\npatch. But this code is now also appearing as ... |
[
{
"msg_contents": "Hi all;\n\nI am looking at trying to make two modifications to the PostgreSQL FDW and\nwould like feedback on this before I do.\n\n1. INSERTMETHOD=[insert|copy] option on foreign table.\n\nOne significant limitation of the PostgreSQL FDW is that it does a prepared\nstatement insert on each r... |
[
{
"msg_contents": "Hello, I found a typo in guc.c.\n\n> {\"enable_parallel_hash\", PGC_USERSET, QUERY_TUNING_METHOD,\n> - gettext_noop(\"Enables the planner's user of parallel hash plans.\"),\n> + gettext_noop(\"Enables the planner's use of parallel hash plans.\"),\n\nThe \"user\" should be \"use... |
[
{
"msg_contents": "Hi,\n\nWhile testing something that I needed to ensure a generic plan was\nbeing used, during editing postgresql.conf I couldn't quite remember\nthe exact spelling of the option to do that. I think the valid\noptions for the setting should be listed in the sample config file.\n\nPatch attach... |
[
{
"msg_contents": "The following bug has been logged on the website:\n\nBug reference: 15346\nLogged by: Alexander Kukushkin\nEmail address: cyberdemn@gmail.com\nPostgreSQL version: 9.6.10\nOperating system: Ubuntu\nDescription: \n\nWe run quite a lot of clusters and from time to tim... |
[
{
"msg_contents": "Hello ,\n\nDid you take a look into the patches from the previous mail ?\nWhat do you think about my proposal ?\nCould you please say if there are any objections in regard to it? \n\nThank you in advance, \nAndrei Yahorau\n\n\n\nFrom: Andrei Yahorau/IBA\nTo: pgsql-hackers@postgresql.org... |
[
{
"msg_contents": "Hi all,\n\nI’m creating a C extension to write data from the Postgres to a text file . Since the table can be very large, I only want to process one chunk of data per time to keep the memory stable. To achieve this, I use SPI cursor(https://www.postgresql.org/docs/current/static/spi-spi-cur... |
[
{
"msg_contents": "Continuing along on my adventures of making performance improvements\nfor partitioned tables, I discovered that getrelid is quite slow in\nInitPlan() during UPDATEs and DELETEs when there are many\nresultRelations to process. A List is a pretty poor choice for this\ndata structure since we a... |
[
{
"msg_contents": "Hi!\n\n\nI have noticed the novice users are stuck trying to create tablespace \nover a directory whose owner is not the system postgres user. They \nobserved the message \"could not set permissions on directory ...: \npermission denied\".\n\nI want to add patch that prints hint to set requir... |
[
{
"msg_contents": "Some minor comments on the JIT explain output. For example:\n\n Seq Scan on tenk1 (cost=0.00..483.00 rows=3333 width=244) (actual\ntime=5.564..8.755 rows=4960 loops=1)\n Filter: (unique1 > unique2)\n Rows Removed by Filter: 5040\n Planning Time: 0.857 ms\n JIT:\n Functions: 2\n Gene... |
[
{
"msg_contents": "Hi.\n\nIt seems that the following commit missed updating ddl.sgml:\n\ncommit 5220bb7533f9891b1e071da6461d5c387e8f7b09\nAuthor: Heikki Linnakangas <heikki.linnakangas@iki.fi>\nDate: Thu Jul 19 13:49:43 2018 +0300\n\n Expand run-time partition pruning to work with MergeAppend\n\nAttached ... |
[
{
"msg_contents": "Hi,\n\nI'm going to propose libpq debug log for analysis of queries on the application side.\nI think that it is useful to determine whether the cause is on the application side or the server side when a slow query occurs. \n\nThe provided information is \"date and time\" at which execution o... |
[
{
"msg_contents": "Hi,\n\nWhen working on other patch[1], I found there are almost same\nfunctions, texttoQualifiedNameList() and stringToQualifiedNameList().\nThe only difference is the argument type, text or char*. I don't know\nwhy these functions are defined seperately, but I think the former \nfunction can... |
[
{
"msg_contents": "Hi All,\n\nThe redirect link for applicable_roles view is missing in the\ndocumentation page for enabled_roles view. Attached is the patch that\nadds the same. Please refer to the following URL for details.\n\nhttps://www.postgresql.org/docs/devel/static/infoschema-enabled-roles.html\n\n-- \n... |
[
{
"msg_contents": "Hi All,\n\nCurrently, table_privileges view in information_schema.sql doesn't\nshow privileges on materialized views for currently enabled roles. As\nper the documentation-[1], it should be showing the all privileges\ngranted on tables and views (the documentation doesn't says it has to\nbe n... |
[
{
"msg_contents": "Hi,\n\nIn a recent commit [1] I added a static inline which castoroides\ndislikes:\ncc -D_STDC_C99= -Xa -g -m64 -xarch=native -xdepend -xO4 -xprefetch=auto,explicit pg_waldump.o compat.o xlogreader.o rmgrdesc.o brindesc.o clogdesc.o committsdesc.o dbasedesc.o genericdesc.o gindesc.o gistdesc.... |
[
{
"msg_contents": "I'm curious about this option:\n\n -r RELFILENODE check only relation with specified relfilenode\n\nbut there is no facility to specify a database.\n\nAlso, referring to the relfilenode of a mapped relation seems a bit\ninaccurate.\n\nMaybe reframing this in terms of the file name of... |
[
{
"msg_contents": "Hello All,\n\nOne of our customers reported a situation where one of the many segments\nbacking a table went missing. They don't know why that happened and we\ncouldn't determine if this could be a PostgreSQL bug or FS bug or simply an\noperator error.\n\nThe very first errors seen in the log... |
[
{
"msg_contents": "Commit 8ecdefc26 reminded me of something I'd intended to whine about\nand forgot. port.h has this in its USE_REPL_SNPRINTF stanza:\n\n/*\n * The GCC-specific code below prevents the pg_attribute_printf above from\n * being replaced, and this is required because gcc doesn't know anythi... |
[
{
"msg_contents": "I started seeing the following compile failures in head today on current\nDebian Jessie using the default gcc version 4.9.2 (Debian\n4.9.2-10+deb8u1):\n\n\tpostgres.c: In function ‘exec_parse_message’:\n\tpostgres.c:1368:3: error: ‘for’ loop initial declarations are only allowed in C99 or C11... |
[
{
"msg_contents": "WalSndLoop() does this, simplifying considerably:\n\n\tfor (;;)\n\t{\n\t\t/* does: last_reply_timestamp = GetCurrentTimestamp() */\n\t\tProcessRepliesIfAny(); \n\t\tsend_data(); /* e.g. XLogSendPhysical(), which calls XLogRead() */\n\t\tWalSndCheckTimeOut(GetCurrentTimestamp());\n\t}\n\nA c... |
[
{
"msg_contents": "I noticed one while looking at partprune.c and found the others with grep.\n\n-- \n David Rowley http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services",
"msg_date": "Sun, 26 Aug 2018 23:42:48 +1200",
"msg_from": "David Rowley <david.... |
[
{
"msg_contents": "Hello hackers,\n\nit is possible to start bgworker with bgw_start_time =\nBgWorkerStart_PostmasterStart, which will be started immediately after\npostmaster.\n\nBut if you try to do a fast shutdown while postmaster still in the\npmState == PM_STARTUP, bgworker will never get SIGTERM and postm... |
[
{
"msg_contents": "I noticed while testing the fix for Alexander Kukushkin's complaint that\nif the syslogger process detects a problem (e.g., failure to open a new\nlog file) while log_destination is set to \"csvlog\", the error report\nactually goes to THREE places. It gets spit out to csvlog, *and* to the\n... |
[
{
"msg_contents": "While answering someone's question I was motivated to try this:\n\nregression=# create type comp_type as (f1 int, f2 int);\nCREATE TYPE\nregression=# create function foo(int) returns int language plpgsql as '\ndeclare z comp_type;\nbegin\n z.f1 = $1; z.f2 = $1+1;\n return z.f1 + z.f2;\nend'... |
[
{
"msg_contents": "During the development of an unrelated feature, I have experienced\nfailures in a pg_dump test case, specifically\n\nt/002_pg_dump.pl ....... 1825/6052\n# Failed test 'defaults_parallel: should not dump COPY\nfk_reference_test_table second'\n# at t/002_pg_dump.pl line 3454.\n\nThis test s... |
[
{
"msg_contents": "Here are three patches to add more detailed error location support to\nsome parts of CREATE TABLE (defaults, check constraints, partition\nspecifications) as well as the COPY command. Examples can be seen in\nthe regression test output.\n\nThe first two are low-hanging fruit, since all the i... |
[
{
"msg_contents": "Whenever some pg_attribute field is added or changed, a lot of\nrepetitive changes all over the code are necessary. Here is a small\nchange to remove one such place.\n\nWe have two code paths for initializing the tuple descriptor for a new\nindex: For a normal index, we copy the tuple descri... |
[
{
"msg_contents": "So I started poking at the idea of sorting by size during parallel\nrestore instead of sorting pg_dump's TOC that way. While investigating\njust where to do that, I discovered that, using the regression database\nas test case, restore_toc_entries_parallel() finds these objects to\nbe *immedi... |
[
{
"msg_contents": "Having quit Amazon, where I was doing Postgres development, I've started looking at various things I might work on for fun. One thought is to start with something easy like the scalability of GetSnapshotData(). :-)\n\n\nI recently found it interesting to examine performance while running near... |
[
{
"msg_contents": "Hello,\n\nI'm curious about something which may qualify as a stupid question.\n\nWhat bad thing would happen if we used OIDs directly as hash values in\ninternal hash tables (that is, instead of uint32_hash() we'd use\nuint32_identity(), or somehow optimise it away entirely, as you can\nsee i... |
[
{
"msg_contents": "I've attached a patch to fix a typo in typcache.c. I ended up also\nrephrasing the sentence since \"information about data types that is\",\nalmost made me also want to change \"is\" to \"are\", since \"types\" is\nplural. That would have been a mistake since it's talking about the\ninformati... |
[
{
"msg_contents": "Hi,\nI'm in a situation where we quite often generate more WAL than we can\narchive. The thing is - archiving takes long(ish) time but it's\nmulti-step process and includes talking to remote servers over network.\n\nI tested that simply by running archiving in parallel I can easily get\n2-3 t... |
[
{
"msg_contents": "This is reproducible with PG11 and PG12:\n\ninitdb -k data\npostgres -D data\n\nmake installcheck\n# shut down postgres with Ctrl-C\n\npg_verify_checksums data\n\npg_verify_checksums: checksum verification failed in file\n\"data/base/16384/28647\", block 65: calculated checksum DC70 but expec... |
[
{
"msg_contents": "contrib/cube has an arbitrary limit of 100 on the number of dimensions\nin a cube, but it actually enforces that only in cube_in and\ncube_enlarge, with the other cube creation functions happy to create\ncubes of more dimensions.\n\nI haven't actually tested, but this implies that one can cre... |
[
{
"msg_contents": "Hi,\nI sent already an email about this topic to pgsql-admins but I think that\nit might be more relevant to this mailing list. I'm trying to investigate a\ncorruption that happened on a machine of one of our clients.\n\nA little background :\n-os Centos 6.6\n-PostgreSQL v9.2.5\n-I was asked ... |
[
{
"msg_contents": "Hello\n\nIn the process of fixing a bug, I inadvertently broke ABI compatibility\nin all released branches that sport logical decoding.\n\n[TL;DR: I propose to set up abidiff.postgresql.org.]\n\nThis is the commit:\n\nAuthor: Alvaro Herrera <alvherre@alvh.no-ip.org>\nBranch: master Release: R... |
[
{
"msg_contents": "Parallel pg_restore generally assumes that the archive file is telling it\nthe truth about data dependencies; in particular, that a TABLE DATA item\nnaming a particular target table is loading data into exactly that table.\n--load-via-partition-root creates a significant probability that that... |
[
{
"msg_contents": "There is a lot of version dependent code in pg_dump now, especially\nper-version queries. The way they were originally built was that we\njust have an entirely separate query per version. But as the number of\nversions and the size of the query grows, this becomes unwieldy.\n\nSo I tried, a... |
[
{
"msg_contents": "Hello.\n\nIn the discussion about cube's dimention limit [1], I found that\nthe error messages looks strange.\n\nhttps://www.postgresql.org/message-id/F0E1A404-A495-4F38-B817-06355B537E88@yandex-team.ru\n\n> postgres=# create table y as select cube(array(SELECT random() as a FROM generate_se... |
[
{
"msg_contents": "I have found some dying code in PL/Python.\n\nThe simple slicing API (sq_slice, sq_ass_slice) has been deprecated\nsince Python 2.0 and has been removed altogether in Python 3, so we can\nremove those functions from the PLyResult class. Instead, the non-slice\nmapping functions mp_subscript ... |
[
{
"msg_contents": "Here is a patch to change some struct initializations to use C99-style\ndesignated initializers. These are just a few particularly egregious\ncases that were hard to read and write, and error prone because of many\nsimilar adjacent types.\n\n(The PL/Python changes currently don't compile wit... |
[
{
"msg_contents": "While trying to debug a recent bug report on hash indexes [1], I\nnoticed that pg_verify_checksums don't work on Windows (or at least in\nmy environment).\n\ninitdb -k ..\\..\\data\npg_verify_checksums.exe ..\\..\\Data\npg_verify_checksums: short read of block 0 in file\n\"..\\..\\Data/globa... |
[
{
"msg_contents": "It is more or less well known that the planner doesn't perform well with\nmore than a few hundred partitions even when only a handful of partitions\nare ultimately included in the plan. Situation has improved a bit in PG\n11 where we replaced the older method of pruning partitions one-by-one... |
[
{
"msg_contents": "I've seen this message now a handful of times recently. It seems to happen\novernight, during a maintenance job which reindex things, including system\ncatalog indices.\n\nIt's easy to reproduce error under 10.5, but not under 10.3 nor 10.4.\n\nwhile :; do for a in pg_class_oid_index pg_clas... |
[
{
"msg_contents": "Hello all.\n\nbackground worker can use SPI to read a database, but it can call\nBackgroundWorkerInitializeConnection(dbname) only once.\n\nI wonder if there is a way to let a child process of postmaster to access\nall the databases one by one?\n\n\n-- \nThanks\n\nHubert Zhang\n\nHello all.ba... |
[
{
"msg_contents": "Hello,\n\nTo test postgres 11, I still regularly run series of short sessions of \npgbench-over-logical-replication (basically the same thing that I used \nlast year [1] - now in a perl incarnation). Most of the time the \nreplication is stable and finishes correctly but sometimes (rarely) I... |
[
{
"msg_contents": "Hi,\n\nI'm trying to upgrade an old PG 9.4 cluster to PG 10, and it's failing on\ncreating an index that uses the unaccent(text) function.\n\nThat function is part of the unaccent extension, which is installed in the\nold DB cluster. I expect pg_upgrade to create that extension as part of the... |
[
{
"msg_contents": "Stumble upon this formatting issue while reading the pg_upgrade docs.\n\nMartín",
"msg_date": "Wed, 29 Aug 2018 20:02:54 -0700",
"msg_from": "=?UTF-8?q?Mart=C3=ADn_Fern=C3=A1ndez?= <fmartin91@gmail.com>",
"msg_from_op": true,
"msg_subject": "[PATCH] Fix formatting in pg_upgrad... |
[
{
"msg_contents": "Hi,\n\nAttached is a patch to fix comments of IndexInfo. ii_KeyAttrNumbers was\nrenamed to ii_IndexAttrNumbers and ii_Am was added but these are\nnot reflected to the comment.\n\nRegards,\n-- \nYugo Nagata <nagata@sraoss.co.jp>",
"msg_date": "Thu, 30 Aug 2018 13:48:31 +0900",
"msg_fr... |
[
{
"msg_contents": "Hi hackers,\n\nI am really confused. If my conclusions are correct, then nobody ever \ntried to use replication slots at Windows!\nThe function RestoreSlotFromDisk in slot.c contains the following code:\n\nstatic void\nRestoreSlotFromDisk(const char *name)\n{\n ReplicationSlotOnDisk cp;\... |
[
{
"msg_contents": "Hi,\n\nI noticed that pg_verify_checksums computes bogus checksums if I compile\nit with '-O2 -Wall' but without -fno-strict-aliasing. Also I am getting\na compile warning then:\n\n|src/bin/pg_verify_checksums$ make CFLAGS='-g -O2 -Wall'\n[...]\n|gcc -g -O2 -Wall -I../../../src/include \n| -... |
[
{
"msg_contents": "Hi,\n\nHere is a small patch to remove $SUBJECT: s/has contains/contains/\n\nBest regards,\nEtsuro Fujita",
"msg_date": "Thu, 30 Aug 2018 20:27:52 +0900",
"msg_from": "Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>",
"msg_from_op": true,
"msg_subject": "Extra word in src/backend/... |
[
{
"msg_contents": "Hi!\n\nOur customer have a bad plan problem, which could be reduced to the\nfollowing example.\n\ncreate table t1 (id int primary key, k int);\ncreate table t2 (id int);\n\ninsert into t1 (select i, i from generate_series(1,1000000) i);\ninsert into t2 (select 0 from generate_series(1,100)i);... |
[
{
"msg_contents": "Hi all,\nWe want to introduce disk quota feature into Postgres.\n\n*Why disk quota*\n*In a multi-tenant environment, there is a requirement to limit the disk\nquota that database/schema/table can be written or a user can consume for\ndifferent organizations.*\n*Meanwhile, other databases such... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.