安西 直也 <anzai-naoya@mxu.nes.nec.co.jp> writes:
> I have checked latest source code.
> But, backward scan doesn't work correctly...
[ pokes at that... ] Hmm, the patches I applied a couple days ago
assumed that we are stepping forward or back from a place where the
WHERE clauses are satisfied. But in this example, the system just
applies _bt_endpoint to descend to the right-hand end of the index,
since there is no upper-bound qual with which to do anything different.
So we start from a place where the clauses aren't satisfied. That also
means that we haven't really fixed the original performance complaint:
there could be lots of nulls to be stepped over before we reach the
first matching row.
I think that the right fix for this is probably to make
_bt_preprocess_keys explicitly generate the "id is not null" qual that's
implied by "id > 0", so that it will have what amounts to a range
condition on the index contents (since for NULLS LAST, "id is not null"
amounts to "id is less than null", as it were). Then, instead of applying
_bt_endpoint, it will use the less-than key to descend the btree to the
last non-null entry, and we'll be good for both correctness and
performance.
I don't see any big problem in doing this in HEAD, but it's getting past
what seems like a sane back-patch. So probably we should revert the
back-branch versions of the prior patch, and just say that the
performance problem is only going to be addressed in HEAD.
regards, tom lane