Re: [BUG]: the walsender does not update its IO statistics until it exits

Hi,

On Wed, Feb 26, 2025 at 03:37:10PM +0900, Michael Paquier wrote:
> On Tue, Feb 25, 2025 at 01:42:08PM +0000, Bertrand Drouvot wrote:
> > Now we can see that the numbers increased for the relation object and that we
> > get non zeros numbers for the wal object too (which makes fully sense).
> > 
> > With the attached patch applied, we would get the same numbers already in
> > step 4. (means the stats are flushed without the need to wait for the walsender
> > to exit).
> 
> @@ -2793,6 +2794,12 @@ WalSndLoop(WalSndSendDataCallback send_data)
>          if (pq_flush_if_writable() != 0)
>              WalSndShutdown();
>  
> +        /*
> +         * Report IO statistics
> +         */
> +        pgstat_flush_io(false);
> +        (void) pgstat_flush_backend(false, PGSTAT_BACKEND_FLUSH_IO);
> +
>          /* If nothing remains to be sent right now ... */
>          if (WalSndCaughtUp && !pq_is_send_pending())
>          {
> 
> That's bad, worse for a logical WAL sender, because it means that we
> have no idea what kind of I/O happens in this process until it exits,
> and logical WAL senders could loop forever, since v16 where we've
> begun tracking I/O.

Yeah... And while the example shared up-thread is related to logical walsender,
the same issue exists for a physical walsender.

OTOH, It's also great to see that the new stats that have been added (the WAL
ones) helped to spot the issue.

> A non-forced periodic flush like you are proposing here sounds OK to
> me,

Thanks for looking at it!

> but the position of the flush could be positioned better in the
> loop.  If there is a SIGUSR2 (aka got_SIGUSR2 is true), WAL senders
> would shut down,

That's true for a physical walsender but I'm not sure it is for a logical
walsender (due to the "sentPtr == replicatedPtr" check in WalSndDone()).

> so it seems rather pointless to do a flush just
> before exiting the process in WalSndDone(), no?  I'd suggest to move
> the flush attempt closer to where we wait for some activity, just
> after WalSndKeepaliveIfNecessary().

Yeah I think that makes sense, done that way in the attached.

Speaking about physical walsender, I moved the test to 001_stream_rep.pl instead
(would also fail without the fix).

Regards,

-- 
Bertrand Drouvot
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com