Hi,
On Fri, 10 May 2024 at 16:21, Nazir Bilal Yavuz <byavuz81@gmail.com> wrote:
>
> Hi,
>
> On Fri, 10 May 2024 at 14:49, Alena Rybakina <lena.ribackina@yandex.ru> wrote:
> >
> > Hi! I could try to check it with the test, but I want to ask you about
> > details, because I'm not sure that I completely understand the test case.
> >
> > You mean that we need to have two backends and on one of them we deleted
> > the tuples before vacuum called the other, do you?
> >
>
> I think triggering a parallel vacuum is enough. I am able to see the
> differences with the following:
>
> You can apply the attached diff file to see the differences between
> the previous version and the patched version. Then, run this query:
>
> CREATE TABLE vacuum_fix (aid int, bid int, cid int) with
> (autovacuum_enabled=false);
> INSERT INTO vacuum_fix SELECT *, *, * FROM generate_series(1, 1000000);
> CREATE INDEX a_idx on vacuum_fix (aid);
> CREATE INDEX b_idx on vacuum_fix (bid);
> CREATE INDEX c_idx on vacuum_fix (cid);
> VACUUM vacuum_fix;
> UPDATE vacuum_fix SET aid = aid + 1;
> VACUUM (VERBOSE, PARALLEL 2) vacuum_fix ;
>
> After that I saw:
>
> INFO: vacuuming "test.public.vacuum_fix"
> INFO: launched 2 parallel vacuum workers for index vacuuming (planned: 2)
> INFO: finished vacuuming "test.public.vacuum_fix": index scans: 1
> ...
> ...
> buffer usage: 29343 hits, 9580 misses in the previous version, 14165
> misses in the patched version, 14262 dirtied
>
> Patched version counts 14165 misses but the previous version counts
> 9580 misses in this specific example.
I am sorry that I showed the wrong thing, this is exactly what is
fixed in this patch. Actually, I do not know how to trigger it;
currently I am looking for it. I will share if anything comes to my
mind.
--
Regards,
Nazir Bilal Yavuz
Microsoft