Re: WIP: extensible enums

Robert Haas <robertmhaas(at)gmail(dot)com> writes:
> On Oct 23, 2010, at 7:12 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> I've been working through this patch. It occurs to me that there's a
>> fairly serious problem with the current implementation of insertion of
>> new values within the bounds of the current sort ordering. Namely, that
>> it does that by reassigning the enumsortorder values of pre-existing
>> rows. That creates a race condition:

> It strikes me that this is merely one facet of our failure to do proper locking on DDL objects other than relations, and that this would be as good a time as any to start fixing it. ISTM that ALTER TYPE should grab a self-excluding lock just as ALTER TABLE already does.

The point of all the design thrashing we've been doing here is to
*avoid* taking locks while comparing enum OIDs. So I'm not impressed
with this proposal. (A self-exclusive lock to prevent concurrent
ALTER TYPEs might be a good idea, but I don't want to block enum
comparisons too.)

I did just think of a possible solution that would work with the
updating implementation: readers of pg_enum could use an MVCC snapshot
instead of SnapshotNow while loading their caches. I'm not certain
offhand how unpleasant this'd be at the C-code level, but it should
be possible. I still prefer the idea of not changing rows once they're
inserted, though --- doing so could possibly also cause transient
failures in, eg, enum_in/enum_out, since those depend on syscaches that
are loaded with SnapshotNow.

regards, tom lane