On Thu, Aug 04, 2022 at 06:32:38AM -0400, David Steele wrote: > On 8/4/22 04:06, Kyotaro Horiguchi wrote: > >At Wed, 3 Aug 2022 23:24:56 -0700, Noah Misch <n...@leadboat.com> wrote in > >>>>I think in this case a HINT might be sufficient to at least keep people > >>>>from > >>>>wasting time tracking down a problem that has already been fixed. > >> > >>Here's a patch to add that HINT. I figure it's better to do this before > >>next > >>week's minor releases. In the absence of objections, I'll push this around > >>2022-08-05 14:00 UTC. > > > >+1 > > Looks good to me as well.
Thanks for reviewing. > >>>>However, there is another issue [1] that might argue for a back patch if > >>>>this patch (as I believe) would fix the issue. > >>> > >>>>[1] > >>>>https://www.postgresql.org/message-id/CAHJZqBDxWfcd53jm0bFttuqpK3jV2YKWx%3D4W7KxNB4zzt%2B%2BqFg%40mail.gmail.com > >>> > >>>That makes sense. Each iteration of the restartpoint recycle loop has a > >>>1/N > >>>chance of failing. Recovery adds >N files between restartpoints. Hence, > >>>the > >>>WAL directory grows without bound. Is that roughly the theory in mind? > >> > >>On further reflection, I don't expect it to happen that way. Each failure > >>message is LOG-level, so the remaining recycles still happen. (The race > >>condition can yield an ERROR under PreallocXlogFiles(), but recycling is > >>already done at that point.) > > > >I agree to this. > > Hmmm, OK. We certainly have a fairly serious issue here, i.e. pg_wal growing > without bound. Even if we are not sure what is causing it, how confident are > we that the patches applied to v15 would fix it? I'll say 7%; certainly too low to stop investigating. The next things I'd want to see are: - select name, setting, source from pg_settings where setting <> boot_val; - log_checkpoints log entries, and other log entries having the same PID If the theory about checkpoint skipping holds, there should be a period where the volume of WAL replayed greatly exceeds max_wal_size, yet 0-1 restartpoints begin and 0 restartpoints complete.
