We have a problem in a database running PostgreSQL 15.4 on AWS Aurora where neither AUTOVACUUM nor VACUUM nor even VACUUM FULL do anything to any tables since some time. The processes finish successfully, but the number of dead rows stays the same.
If I run VACUUM VERBOSE manually, I get:
tuples: 0 removed, 61251 remain, 54051 are dead but not yet removable
removable cutoff: 322821406, which was 61069267 XIDs old when operation ended
What's interesting is that the 322821406 xid (?) is the same for every table I checked, even for unrelated ones that are never queried together. There is nothing on that transaction left in pg_stat_activity.
I checked usual suspects: long-running transactions, prepared but not committed transactions, unused physical replication slots - but there's nothing there.
Is there anything non-intrusive (not causing access exclusive lock, like VACUUM FULL, and not involving dropping tables) I can try to fix or debug this?
