Good news:
I think we've got GiST working (somewhat anyways), as we found
gistKeyIsEQ(giststate, 0, datum, currdatum) in gistutil.c does the trick of comparing two datums. I swear most of our trouble is just finding our way around the postgres codebase, but we're getting there little by little. We've gone back to revisit hash to see if we can figure it out now that we understand a little bit about GiST, but we can't find an equivelent function in hash for the KeyIsEQ().
So two questions really. The first is if such a function exists for hash. The second is that nbtree and Neil Conways work a few years ago (
http://archives.postgresql.org/pgsql-patches/2003-09/msg00252.php) use the scan and scankey stuff, but we're having trouble understanding how these work. Is there some documentation on using these correctly (outside of just looking at nbtree code)? Thanks so much for the help folks!
-Matt
On 2/13/07, Teodor Sigaev <teodor@sigaev.ru> wrote: > indexes, then it must use operator number so-and-so for equality. But
> there are lots of GiST opclasses that don't include equality at all; we
> can't break that case.
There is a GiST support function for equality of keys, in btree_gist it's named
as gbt_*_same. Equality function has support number 7 and is used for stored keys.
But the real issue in unique GiST index is unique :). First, the algorithm of
insertion doesn't compare indexed keys on leaf page at all. Values on the same
page are compared only when page is splitting (picksplit support method).
Second, GiST implementation supports only unordered trees (btree_gist is a some
kind of emulation) and it cannot guarantee that equal keys will be close in
index. That's related to picksplit and gistpenalty method problem/optimization
and data set.
--
Teodor Sigaev E-mail: teodor@sigaev.ru
WWW: http://www.sigaev.ru/
