On Tue, Apr 7, 2020 at 5:17 PM Masahiko Sawada <masahiko.saw...@2ndquadrant.com> wrote: > > On Tue, 7 Apr 2020 at 18:29, Masahiko Sawada > <masahiko.saw...@2ndquadrant.com> wrote: > > > > On Tue, 7 Apr 2020 at 17:42, Amit Kapila <amit.kapil...@gmail.com> wrote: > > > > > > On Tue, Apr 7, 2020 at 1:30 PM Masahiko Sawada > > > <masahiko.saw...@2ndquadrant.com> wrote: > > > > > > > > Buffer usage statistics seem correct. The small differences would be > > > > catalog lookups Peter mentioned. > > > > > > > > > > Agreed, but can you check which part of code does that lookup? I want > > > to see if we can avoid that from buffer usage stats or at least write > > > a comment about it, otherwise, we might have to face this question > > > again and again. > > > > Okay, I'll check it. > > > > I've checked the buffer usage differences when parallel btree index creation. > > TL;DR; > > During tuple sorting individual parallel workers read blocks of > pg_amproc and pg_amproc_fam_proc_index to get the sort support > function. The call flow is like: > > ParallelWorkerMain() > _bt_parallel_scan_and_sort() > tuplesort_begin_index_btree() > PrepareSortSupportFromIndexRel() > FinishSortSupportFunction() > get_opfamily_proc() >
Thanks for the investigation. I don't see we can do anything special about this. In an ideal world, this should be done once and not for each worker but I guess it doesn't matter too much. I am not sure if it is worth adding a comment for this, what do you think? -- With Regards, Amit Kapila. EnterpriseDB: http://www.enterprisedb.com