Re: BUG #17973: Reinit of pgstats entry for dropped DB can break autovacuum daemon

Hi,

On 2023-06-20 15:41:22 +0900, Michael Paquier wrote:
> On Mon, Jun 19, 2023 at 10:45:23AM -0700, Andres Freund wrote:
> > I think it'd take a fair amount of work to track these stats in a more useful
> > manner for the startup process, by virtue of it effectively being connected to
> > multiple databases. We'd need to track
> > pgStatBlockReadTime/pgStatBlockWriteTime on a per-database level, which
> > wouldn't be easy to do without increasing overhead.
>
> Is it really necessary to do this much amount of work for the scope of
> this issue, though?

No! That'd just be for actually making db-level stats for the startup process
work. In other words, I don't think there's anything we could break, because
it already doesn't work :)

> Relying on MyDatabaseId to control if these updates should happen does not
> look like the right move to me, to be honest, because this can be used to
> update shared stats.

Shared stats only make sense when you know what non-shared stats are. At the
moment that afaict only works (and afaict only ever has worked) for database
connected backends.

I guess we could start reporting everything to the shared entry if we aren't
connected to a database, but that doesn't really seem like an improvement.

> In the pgstat shutdown callback, shouldn't we try to check if the database
> entry exists and/or has been dropped and just do nothing in this case?

I don't think that'd be a good approach.

Greetings,

Andres Freund