Commit 052026c9b9 made heap_vacuum_rel call RelationGetNumberOfBlocks
before it calls vacuum_get_cutoffs -- it swapped the order. This is
wrong, as explained by an intact comment above the call to
vacuum_get_cutoffs.
In short, there is now a brief window during which the relation can be
extended that'll allow heap pages with tuple headers < VACUUM's
OldestXmin to be created, which are overlooked by that same VACUUM
(they're beyond the same VACUUM's rel_pages). As a result of all this,
VACUUM might advance pg_class.relfrozenxid to a later/younger value
than a remaining/unfrozen XID value from one of these unscanned heap
pages.
--
Peter Geoghegan