Re: BUG #19483: pg_upgrade fails with orphan records in pg_init_priv catalog table

Laurenz Albe <laurenz.albe@cybertec.at> writes:
> Since there have been very few reports of this problem, the question
> remains if we need this patch at all, or of it should be backpatched.
> My opinion is that it should; every upgrade or restore failure is
> one too many.

I have a more pressing concern: has any performance testing been
done on this?  It looks like it'd be absolutely catastrophic for
pg_dump performance on databases with lots of objects.

The implementation direction I'd been vaguely imagining was for
pg_dump's buildACLCommands() to drop any AclItems that contain
dangling role references (ie, numeric OIDs where a role name
should be).  If the given role name contains any non-digit
characters then it's certainly not dangling, so most of the time
this'd be a very cheap check.  However, if somebody does

	CREATE USER "007";
	GRANT ALL ON TABLE mi6_operations TO "007";

we mustn't get fooled by that.  The backend is doing us no favors by
not making numeric OIDs visibly different from all-digit role names
in AclItems.  In HEAD I'd advocate fixing that on the server side
(as attached), but we can't assume that a back-branch server has such
a fix.  What we could do with an old server is issue a query (once per
pg_dump run) to collect all the valid all-digit role names, which
should surely be a short list in most databases, and then filter
against that within buildACLCommands().

			regards, tom lane