Re: performance degredation after upgrade from 9.6 to 12

On Sun, Nov 24, 2019 at 8:52 AM Mariel Cherkassky <
mariel(dot)cherkassky(at)gmail(dot)com> wrote:

> Hey Andrew,
> It seems that changing this parameter worked for me.
> Setting it to zero means that there wont be any parallel workers for one
> query right ?
> Is it something familiar this problem with the gatherers ?
>

Your example would not be using parallel workers anyway, regardless of the
setting of max_parallel_workers_per_gather, so I don't see how changing
this could have worked for you. Unless you mean it worked in your full
test, rather than in your test case. I doubt your test case benchmarking
was very reliable to start with, you only show a single execution and
didn't indicate you had more unshown ones.

If I do more credible benchmarking, I do get a performance regression but
it closer is to 16% than to 3 fold. And it doesn't depend on the setting
of max_parallel_workers_per_gather. I doubt a regression of this size is
even worth investigating.

pgbench -T300 -P5 -f <(echo "select count(*) from test1") -p 9912 -n -M
prepared

Cheers,

Jeff