Re: Design of pg_stat_subscription_workers vs pgstats

On Tue, Feb 22, 2022 at 8:02 PM Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:
>
> On Tue, Feb 22, 2022 at 6:53 PM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> >
>
> > 3. Can we send error stats pgstat_report_stat() as that will be called
> > via proc_exit() path. We can set the phase (apply/sync) in
> > apply_error_callback_arg and then use that to send the appropriate
> > message. I think this will obviate the need for try..catch.
>
> If we use pgstat_report_stat() to send subscription stats messages,
> all processes end up going through that path. It might not bring
> overhead in practice but I'd like to avoid it.
>

I am not sure about overhead but I see another problem if we use that
approach. In the exit path, logicalrep_worker_onexit() will get called
before pgstat_report_stat() and that will clear the
MyLogicalRepWorker->subid, so we won't know the id for which to send
stats. So, the way patch is doing seems reasonable to me unless
someone has better ideas.

--
With Regards,
Amit Kapila.