On Fri, Jun 07, 2019 at 03:58:43PM -0700, Andres Freund wrote:
> Do we need to move the orphan temp cleanup code into database vacuums or
> such?
When entering into the vacuum() code path for an autovacuum, only one
relation at a time is processed, and we have prior that extra
processing related to toast relations when selecting the relations to
work on, or potentially delete orphaned temp tables. For a manual
vacuum, we finish by deciding which relation to process in
get_all_vacuum_rels(), so the localized processing is a bit different
than what's done in do_autovacuum() when scanning pg_class for
relations.
Technically, I think that it would work to give up on the gathering of
the orphaned OIDs in a gathering and let them be gathered in the list
of items to vacuum, and then put the deletion logic down to
vacuum_rel(). However, there is a take: for autovacuum we gather the
orphaned entries and the other relations to process, then drop all the
orphaned OIDs, and finally vacuum/analyze the entries collected. So
if you put the deletion logic down into vacuum_rel() then we won't be
able to drop orphaned tables before working on a database, which would
be bad if we know about an orphaned set, but autovacuum works for a
long time on other legit entries first.
--
Michael