On Sat, Apr 26, 2008 at 12:10 AM, Brendan Jurd <direvus@gmail.com> wrote:
> Has anyone had a close look at how hard it would be allow just the
> "add to the end" capability?
If the OIDs haven't wrapped around since the enum was created, it's
trivial. If they have, well someone with more OID-fu than me will have
to explain what we'd have to do to guarantee getting a new OID higher
than a certain value. And if your highest enum oid happens to be
4^32-1, you're in trouble :).
I wonder if it's worth revisiting the decision to save enums on disk
as oids. The very first idea that I had was to have an enum value as
the combination of both an enum id and the ordinal value. We would
presumably make both say 16bits so we could still be be passed by
value. This would restrict us to 2^16 enum types per database and 2^16
values per type, but if anyone is getting within an order of magnitude
of either limit I'd be very interested in seeing what they're doing.
The nice thing about the above approach is that we could space out the
ordinal values so as to allow people to insert a fair number of extra
values before forcing a rewrite of the table. The only thing we really
couldn't handle that way would be reordering - we'd need an extra
layer of indirection for that, which would have some performance
penalties. None of the standard operators for enums require a syscache
lookup currently, only I/O does,
Cheers
Tom
