Re: POC, WIP: OR-clause support for indexes

On Fri, Nov 29, 2024 at 9:54 AM Alexander Korotkov <aekorotkov@gmail.com> wrote:
> On Fri, Nov 29, 2024 at 7:51 AM Alena Rybakina
> <a.rybakina@postgrespro.ru> wrote:
> >
> > On 29.11.2024 03:04, Alexander Korotkov wrote:
> > > On Thu, Nov 28, 2024 at 9:33 PM Alena Rybakina
> > > <a.rybakina@postgrespro.ru> wrote:
> > >> On 28.11.2024 22:28, Ranier Vilela wrote:
> > >>
> > >> Em qui., 28 de nov. de 2024 às 16:03, Alena Rybakina <a.rybakina@postgrespro.ru> escreveu:
> > >>> Hi! Thank you for the case.
> > >>>
> > >>> On 28.11.2024 21:00, Alexander Lakhin wrote:
> > >>>> Hello Alexander,
> > >>>>
> > >>>> 21.11.2024 09:34, Alexander Korotkov wrote:
> > >>>>> I'm going to push this if no objections.
> > >>>> Please look at the following query, which triggers an error after
> > >>>> ae4569161:
> > >>>> SET random_page_cost = 1;
> > >>>> CREATE TABLE tbl(u UUID);
> > >>>> CREATE INDEX idx ON tbl USING HASH (u);
> > >>>> SELECT COUNT(*) FROM tbl WHERE u = '00000000000000000000000000000000' OR
> > >>>>    u = '11111111111111111111111111111111';
> > >>>>
> > >>>> ERROR:  XX000: ScalarArrayOpExpr index qual found where not allowed
> > >>>> LOCATION:  ExecIndexBuildScanKeys, nodeIndexscan.c:1625
> > >>>>
> > >>>>
> > >>> I found out what the problem is index scan method was not generated. We
> > >>> need to check this during OR clauses for SAOP transformation.
> > >>>
> > >>> There is a patch to fix this problem.
> > >> Hi.
> > >> Thanks for the quick fix.
> > >>
> > >> But I wonder if it is not possible to avoid all if the index is useless?
> > >> Maybe moving your fix to the beginning of the function?
> > >>
> > >> diff --git a/src/backend/optimizer/path/indxpath.c b/src/backend/optimizer/path/indxpath.c
> > >> index d827fc9f4d..5ea0b27d01 100644
> > >> --- a/src/backend/optimizer/path/indxpath.c
> > >> +++ b/src/backend/optimizer/path/indxpath.c
> > >> @@ -3248,6 +3248,10 @@ match_orclause_to_indexcol(PlannerInfo *root,
> > >>    Assert(IsA(orclause, BoolExpr));
> > >>    Assert(orclause->boolop == OR_EXPR);
> > >>
> > >> + /* Ignore index if it doesn't support index scans */
> > >> + if(!index->amsearcharray)
> > >> + return NULL;
> > >> +
> > >>
> > >> Agree. I have updated the patch
> > >>
> > >>    /*
> > >>    * Try to convert a list of OR-clauses to a single SAOP expression. Each
> > >>    * OR entry must be in the form: (indexkey operator constant) or (constant
> > >>
> > >> The test bug:
> > >> EXPLAIN SELECT COUNT(*) FROM tbl WHERE u = '00000000000000000000000000000000' OR u = '11111111111111111111111111111111';
> > >>                                                              QUERY PLAN
> > >> ----------------------------------------------------------------------------------------------------------------------------------
> > >>   Aggregate  (cost=12.46..12.47 rows=1 width=8)
> > >>     ->  Bitmap Heap Scan on tbl  (cost=2.14..12.41 rows=18 width=0)
> > >>           Recheck Cond: ((u = '00000000-0000-0000-0000-000000000000'::uuid) OR (u = '11111111-1111-1111-1111-111111111111'::uuid))
> > >>           ->  BitmapOr  (cost=2.14..2.14 rows=18 width=0)
> > >>                 ->  Bitmap Index Scan on idx  (cost=0.00..1.07 rows=9 width=0)
> > >>                       Index Cond: (u = '00000000-0000-0000-0000-000000000000'::uuid)
> > >>                 ->  Bitmap Index Scan on idx  (cost=0.00..1.07 rows=9 width=0)
> > >>                       Index Cond: (u = '11111111-1111-1111-1111-111111111111'::uuid)
> > >> (8 rows)
> > > I slightly revised the fix and added similar check to
> > > group_similar_or_args().  Could you, please, review that before
> > > commit?
> > >
> > I agree with changes. Thank you!
>
> Andrei, Alena, thank you for the feedback. Pushed!

I think we should give some more attention to the patch enabling OR to
SAOP transformation for joins (first time posted in [1]).  I think we
tried to only work with Const and Param, because we were previously
working during parse stage. So, at that stage if we have the clause
like "a.x = 1 OR a.x = b.x OR b.x = 2", then we don't know if we
should transform it into "a.x = ANY(1, b.x) OR b.x = 2" or into "a.x
=1 OR b.x = ANY(a.x, 2)". But if we do the transformation during the
index matching, we would actually be able to try the both and select
the best.

The revised patch is attached.  Most notably it revises
group_similar_or_args() to have the same notion of const-ness as
others.  In that function we split potential index key and constant
early to save time on enumerating all possible index keys.  But it
appears to be possible to split by relids bitmapsets: index key should
use our relid, while const shouldn't.  Other that that, comments,
commit message and naming are revised.

Links.
1. https://www.postgresql.org/message-id/CAPpHfdu9QJ%3DGbua3CUUH2KKG_8urakJTen4JD47PGh9wWP%3DQxQ%40mail.gmail.com

------
Regards,
Alexander Korotkov
Supabase