On 2022-12-28 16:37:27 -0800, Peter Geoghegan wrote:
> On Wed, Dec 28, 2022 at 4:20 PM Andres Freund <andres@anarazel.de> wrote:
> > > Theoretically this is an old issue that dates back to commit
> > > 699bf7d05c, as opposed to an issue in the page-level freezing patch.
> > > But that's not really true in a real practical sense. In practice the
> > > calls to TransactionIdDidCommit() will happen far more frequently
> > > following today's commit 1de58df4fe (since we're using OldestXmin as
> > > the cutoff that gates performing freeze_xmin/freeze_xmax processing --
> > > not FreezeLimit).
> >
> > Hm. But we still only do the check when we actually freeze, rather than just
> > during the pre-check in heap_tuple_should_freeze(). So we'll only incur the
> > increased overhead when we also do more WAL logging etc. Correct?
>
> Yes, that's how it worked up until today's commit 1de58df4fe.
>
> I don't have strong feelings about back patching a fix, but this does
> seem like something that I should fix now, on HEAD.
>
> > Hm. I dimply recall that we had repeated cases where the hint bits were set
> > wrongly due to some of the multixact related bugs. I think I was trying to be
> > paranoid about not freezing stuff in those situations, since it can lead to
> > reviving dead tuples, which obviously is bad.
>
> I think that it's a reasonable check, and I'm totally in favor of
> keeping it (or something very close, at least).
I don't quite follow - one paragraph up you say we should fix something, and
then here you seem to say we should continue not to rely on the hint bits?
