While looking at the inherited-constraints patch I realized that the
CONSTROID syscache definition fails to list conrelid as a relation OID
column. What this means is that a relcache flush on a given relation
will fail to flush CONSTROID syscache entries for constraints belonging
to that rel.
Now, since there's no mechanism for updating an existing pg_constraint
entry (much less one that depends on a relcache flush signal to get the
update propagated), I can't actually see that there's any visible bug
here. The regular per-catcache-entry flush mechanism will serve to get
rid of entries that actually need to be gotten rid of, such as ones
that are deleted.
So, this should be changed going forward, but does anyone see a need
to back-patch it?
regards, tom lane