Michael Paquier <michael@paquier.xyz> writes:
> Yeah. 11 and newer versions have been made even more aggressive with
> the cleanup of orphaned tables in autovacuum, particularly the case
> where a backend reuses the ID of a past session that crashed, leaving
> behind some temporary tables. Perhaps that was the case here?
Off-list, the OP indicated that the problem is actually of the other
category, ie the problematic tables belong to extremely long-lived
sessions that are managed by a connection pooler. I don't quite
understand why the pooler isn't issuing DISCARD ALL between clients,
but anyway it seems that PG is operating as designed here.
regards, tom lane
