On Mon, Feb 8, 2021 at 8:13 PM Hou, Zhijie <houzj.fnst@cn.fujitsu.com> wrote:
>
> > > Did it actually use a parallel plan in your testing?
> > > When I ran these tests with the Parallel INSERT patch applied, it did
> > > not naturally choose a parallel plan for any of these cases.
> >
> > Yes, these cases pick parallel plan naturally on my test environment.
> >
> > postgres=# explain verbose insert into testscan select a from x where
> > a<80000 or (a%2=0 and a>199900000);
> > QUERY PLAN
> > ----------------------------------------------------------------------
> > -----------------------------
> > Gather (cost=4346.89..1281204.64 rows=81372 width=0)
> > Workers Planned: 4
> > -> Insert on public.testscan (cost=3346.89..1272067.44 rows=0
> > width=0)
> > -> Parallel Bitmap Heap Scan on public.x1
> > (cost=3346.89..1272067.44 rows=20343 width=8)
> > Output: x1.a, NULL::integer
> > Recheck Cond: ((x1.a < 80000) OR (x1.a > 199900000))
> > Filter: ((x1.a < 80000) OR (((x1.a % 2) = 0) AND (x1.a >
> > 199900000)))
> > -> BitmapOr (cost=3346.89..3346.89 rows=178808
> > width=0)
> > -> Bitmap Index Scan on x1_a_idx
> > (cost=0.00..1495.19 rows=80883 width=0)
> > Index Cond: (x1.a < 80000)
> > -> Bitmap Index Scan on x1_a_idx
> > (cost=0.00..1811.01 rows=97925 width=0)
> > Index Cond: (x1.a > 199900000)
> >
> > PSA is my postgresql.conf file, maybe you can have a look. Besides, I didn't
> > do any parameters tuning in my test session.
>
> I reproduced this on my machine.
>
> I think we'd better do "analyze" before insert which helps reproduce this easier.
> Like:
>
> -----
> analyze;
> explain analyze verbose insert into testscan select a from x where a<80000 or (a%2=0 and a>199900000);
> -----
OK then.
Can you check if just the underlying SELECTs are run (without INSERT),
is there any performance degradation when compared to a non-parallel
scan?
Regards,
Greg Nancarrow
Fujitsu Australia
