Re: BUG #17804: Assertion failed in pg_stat after fetching from pg_stat_database and switching cache->snapshot

At Thu, 27 Apr 2023 13:00:00 +0300, Alexander Lakhin <exclusion@gmail.com> wrote in
> Hello,
>
> 24.02.2023 19:00, Alexander Lakhin wrote:
> > Hello Kyotaro-san,
> > 24.02.2023 10:08, Kyotaro Horiguchi wrote:
> >> Thanks for the report!  We didn't consider the case that variable is
> >> changed in-transaction.
> >>
> >>
> >> What are your thoughts on this?
> > What if we implement assign_pgstat_fetch_consistency() and execute
> > pgstat_clear_snapshot() in it when the mode switch happens (to avoid
> > possible inconsistency between the GUC and the current pgstat state)?

Thank you for the patch. (Sorry, I forgot about this..)

> Please look at the attached patch, which implements that idea and also
> includes a simple test. The behavior change doesn't affect any of
> existing
> tests and seems natural to me, so I haven't added a note about it in
> the docs.

I tried to keep existing snapshots as long as possible, but
considering how rarely users change the variable and the lack of
necessity to to preserve the snapshots, it's simply easier to clear
them with each variable change.

Nonetheless, I'm a bit concerned about making a direct call to
pgstat_clear_snapshot() from the assign callback, it might be fine for
now, but I worry that it could an issue later on.

So, how about just settin a trigger that causes a snapshot clearing
prior to the next use, like the attached?

As for the test, I can't come up with a better one, but I think the
comment should explain its intention more clearly.

regards.

--
Kyotaro Horiguchi
NTT Open Source Software Center