Re: GiST on 64-bit box

Teodor Sigaev <teodor@stack.net> writes:
> My opinion is that second way is right (use GISTENTRY array). But this
> channge requires changes in GiST API: picksplit and union functions
> must retrieve one argument more. Is it possible to make for 7.2.1 or
> such changes must be appyed in TODO for 7.3 ?

Hmm.  If we had any such functions installed in the standard system,
then such a change would mean an initdb, which we couldn't do for 7.2.1.

We could argue that forcing a change in a contrib module isn't an
initdb, but the argument will seem very thin to anyone who has a
running 7.2 database with GIST indexes and wants to update to 7.2.1.
They will have to reinstall their GIST support modules and recreate
their GIST indexes, AFAICS.

On the other hand, changing the signature would be a good thing if the
GIST code were tweaked to check that the referenced function had the
right signature, because that way you could raise an error at runtime
if someone tried to use a non-updated contrib module with an updated
backend.  Without some such check I foresee disasters in the field.

So I think I vote for changing the signature and tweaking initGISTstate
to verify that the number of parameters each function expects is right.

But even with that, it might be argued that we should postpone the
change till 7.3, and just say "sorry folks, GIST doesn't work on
64-bit machines for now".  Is that worse than risking update problems
for existing users of GIST indexes?  Comments anyone?
        regards, tom lane