Dean Rasheed <dean.a.rasheed@gmail.com> writes:
> On 17 October 2010 16:49, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> [ scratches head... ] �And where does it get that number of elements
>> from, if not by doing the same work that would allow it to fill the
>> array completely? �Something seems ill-designed here.
> Hmm. That's coming from a new column added to pg_type (typnlabels).
> Perhaps that's not safe though. Are there potential race conditions
> there?
I knew I shoulda read this patch ;-). That seems a lot more invasive
than this feature justifies. And I share your qualms about whether it's
race-condition-proof. We don't have very much locking on pg_type
entries, so making a hard assumption about consistency between two
different catalogs seems pretty risky.
The way I'd be inclined to design this is that altering an enum doesn't
change its pg_type entry at all, just add another row to pg_enum.
When first needing to compare values of an enum, load up the typcache
entry for it. This involves scanning all the entries for that type OID
in pg_enum, and determining from that whether you can compare the easy
way or not. If not, build the array that tells you how to sort, and put
it in the typcache entry.
The missing piece in this is how to determine when the typcache entry
has to be flushed. That should be driven by sinval signalling. There
are two different ways you could do it:
1. Watch for SI events on pg_enum. The problem here is we don't have
a syscache on pg_enum, and there's no obvious other reason to want one.
Also I think you'd have to flush *all* enum typcache entries, since you
couldn't always tell which one had been modified.
2. Watch for SI events on the pg_type row. However, since according
to what I said above an ALTER TYPE wouldn't actually modify the pg_type
row, you'd need to have the ALTER send out a "dummy" SI event. This
is not hard to do --- we have the concept of dummy inval events on
relations already --- but it's a bit ugly because of the risk of
omission. But the number of places that would need to do this seems
small, so I don't think that risk is large.
On balance I like the second idea better.
regards, tom lane
