Hi,
I didn't quite get to responding in depth, but I wanted to at least respond to
one point today.
On 2022-01-25 20:27:07 +0900, Masahiko Sawada wrote:
> > The pgstat entries are quite wide (292 bytes), because of the error message
> > stored. That's nearly twice the size of PgStat_StatTabEntry. And as far as I
> > can tell, once there was an error, we'll just keep the stats entry around
> > until the subscription is dropped.
>
> We can drop the particular statistics by
> pg_stat_reset_subscription_worker() function.
Only if either the user wants to drop all stats, or somehow knows the oids of
already dropped tables...
> > Why isn't this just storing data in pg_subscription_rel?
>
> These need to be updated on error which means for a failed xact and we
> don't want to update the system catalog in that state.
Rightly so! In fact, I'm concerned with sending a pgstats message in that
state as well. But: You don't need to. Just abort the current transaction,
start a new one, and update the state.
> There will be some challenges in a case where updating pg_subscription_rel
> also failed too (what to report to the user, etc.). And moreover, we don't
> want to consume space for temporary information in the system catalog.
You're consuming resources in a *WAY* worse way right now. The stats file gets
constantly written out, and quite often read back by backends. In contrast to
parts of pg_subscription_rel or such that data can't be removed from
shared_buffers under pressure.
Greetings,
Andres Freund