Re: 64-bit XIDs in deleted nbtree pages

On Mon, Mar 1, 2021 at 1:40 PM Peter Geoghegan <pg(at)bowt(dot)ie> wrote:
> > Since it seems not a bug I personally think we don't need to do
> > anything for back branches. But if we want not to trigger an index
> > scan by vacuum_cleanup_index_scale_factor, we could change the default
> > value to a high value (say, to 10000) so that it can skip an index
> > scan in most cases.
>
> One reason to remove vacuum_cleanup_index_scale_factor in the back
> branches is that it removes any need to fix the
> "IndexVacuumInfo.num_heap_tuples is inaccurate outside of
> btvacuumcleanup-only VACUUMs" bug -- it just won't matter if
> btm_last_cleanup_num_heap_tuples is inaccurate anymore. (I am still
> not sure about backpatch being a good idea, though.)

Attached is v8 of the patch series, which has new patches. No real
changes compared to v7 for the first patch, though.

There are now two additional prototype patches to remove the
vacuum_cleanup_index_scale_factor GUC/param along the lines we've
discussed. This requires teaching VACUUM ANALYZE about when to trust
VACUUM cleanup to set the statistics (that's what v8-0002* does).

The general idea for VACUUM ANALYZE in v8-0002* is to assume that
cleanup-only VACUUMs won't set the statistics accurately -- so we need
to keep track of this during VACUUM (in case it's a VACUUM ANALYZE,
which now needs to know if index vacuuming was "cleanup only" or not).
This is not a new thing for hash indexes -- they never did anything in
the cleanup-only case (hashvacuumcleanup() just returns NULL). And now
nbtree does the same thing (usually). Not all AMs will, but the new
assumption is much better than the one it replaces.

I thought of another existing case that violated the faulty assumption
made by VACUUM ANALYZE (which v8-0002* fixes): VACUUM's INDEX_CLEANUP
feature (which was added to Postgres 12 by commit a96c41feec6) is
another case where VACUUM does nothing with indexes. VACUUM ANALYZE
mistakenly considers that index vacuuming must have run and set the
pg_class statistics to an accurate value (more accurate than it is
capable of). But with INDEX_CLEANUP we won't even call
amvacuumcleanup().

--
Peter Geoghegan