Re: Emit fewer vacuum records by reaping removable tuples during pruning

On Wed, Jan 17, 2024 at 3:58 PM Robert Haas <robertmhaas@gmail.com> wrote:
> > Ah, I realize I was not clear. I am now talking about inconsistencies
> > in vacuuming the FSM itself. FreeSpaceMapVacuumRange(). Not updating
> > the freespace map during the course of vacuuming the heap relation.
>
> Fair enough, but I'm still not quite sure exactly what the question
> is. It looks to me like the current code, when there are indexes,
> vacuums the FSM after each round of index vacuuming. When there are no
> indexes, doing it after each round of index vacuuming would mean never
> doing it, so instead we vacuum the FSM every ~8GB. I assume what
> happened here is that somebody decided doing it after each round of
> index vacuuming was the "right thing," and then realized that was not
> going to work if no index vacuuming was happening, and so inserted the
> 8GB threshold to cover that case.

Note that VACUUM_FSM_EVERY_PAGES is applied against the number of
rel_pages "processed" so far -- *including* any pages that were
skipped using the visibility map. It would make a bit more sense if it
was applied against scanned_pages instead (just like
FAILSAFE_EVERY_PAGES has been since commit 07eef53955). In other
words, VACUUM_FSM_EVERY_PAGES is applied against a thing that has only
a very loose relationship with physical work performed/time elapsed.

I tend to suspect that VACUUM_FSM_EVERY_PAGES is fundamentally the
wrong idea. If it's such a good idea then why not apply it all the
time? That is, why not apply it independently of whether nindexes==0
in the current VACUUM operation? (You know, just like with
FAILSAFE_EVERY_PAGES.)

--
Peter Geoghegan