Re: possible bug in handling of contrecords in dd38ff28ad (Fix recovery_prefetch with low maintenance_io_concurrency)

On Mon, Jul 3, 2023 at 6:12 AM Tomas Vondra
<tomas(dot)vondra(at)enterprisedb(dot)com> wrote:
> On 7/2/23 04:09, Thomas Munro wrote:
> > When I added that new error I was thinking about that third case. We
> > generally expect to detect the end of WAL replay after a crash with an
> > error ("invalid record length ...: wanted 24, got 0" + several other
> > possibilities), but in this rare case it would be silent. The effect
> > on the first two cases is cosmetic, but certainly annoying. Perhaps I
> > should go ahead and back-patch the attached change, and then we can
> > discuss whether/how we should do a better job of distinguishing "user
> > requested artificial end of decoding" from "unexpectedly ran out of
> > data" for v17?
> >
>
> Yeah, I think that'd be sensible. IMHO we have a habit of scaring users
> with stuff that might be dangerous/bad, but 99% of the time it's
> actually fine and perhaps even expected. It's almost as if we're
> conditioning people to ignore errors.

Done.

There is CF #2490 "Make message at end-of-recovery less scary".
Perhaps we should think about how to classify this type of failure in
the context of that proposal.