Re: Vacuum ERRORs out considering freezing dead tuples from before OldestXmin

On Mon, Jun 24, 2024 at 1:05 PM Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
>
> On Mon, Jun 24, 2024 at 12:43 PM Peter Geoghegan <pg(at)bowt(dot)ie> wrote:
> > The problem here is that OldestXmin is supposed to be more
> > conservative than vistest, which it almost always is, except in this
> > one edge case. I don't think that plugging that hole changes the basic
> > fact that there is one source of truth about what *needs* to be
> > pruned. There is such a source of truth: OldestXmin.
>
> Well, another approach could be to make it so that OldestXmin actually
> is always more conservative than vistest rather than almost always.

For the purposes of pruning, we are effectively always using the more
conservative of the two with this patch.

Are you more concerned about having a single horizon for pruning or
about having a horizon that does not move backwards after being
established at the beginning of vacuuming the relation?

Right now, in master, we do use a single horizon when determining what
is pruned -- that from GlobalVisState. OldestXmin is only used for
freezing and full page visibility determinations. Using a different
horizon for pruning by vacuum than freezing is what is causing the
error on master.

> I agree with you that letting the pruning horizon move forward during
> vacuum is desirable. I'm just wondering if having the vacuum code need
> to know a second horizon is really the best way to address that.

I was thinking about this some more and I realized I don't really get
why we think using GlobalVisState for pruning will let us remove more
tuples in the common case.

I had always thought it was because the vacuuming backend's
GlobalVisState will get updated periodically throughout vacuum and so,
assuming the oldest running transaction changes, our horizon for
vacuum would change. But, in writing this repro, it is actually quite
hard to get GlobalVisState to update. Our backend's RecentXmin needs
to have changed. And there aren't very many places where we take a new
snapshot after starting to vacuum a relation. One of those is at the
end of index vacuuming, but that can only affect the pruning horizon
if we have to do multiple rounds of index vacuuming. Is that really
the case we are thinking of when we say we want the pruning horizon to
move forward during vacuum?

- Melanie