Robert Haas <robertmhaas@gmail.com> writes:
> As far as I can tell, the focus on trying to estimate the number of
> tuples per bucket is entirely misguided. Supposing the relation is
> mostly unique so that the values don't cluster too much, the right
> answer is (of course) NTUP_PER_BUCKET.
But the entire point of that code is to arrive at a sane estimate
when the inner relation *isn't* mostly unique and *does* cluster.
So I think you're being much too hasty to conclude that it's wrong.
> Because the extra tuples that get thrown into the bucket
> generally don't have the same hash value (or if they did, they would
> have been in the bucket either way...) and get rejected with a simple
> integer comparison, which is much cheaper than
> hash_qual_cost.per_tuple.
Yeah, we are charging more than we ought to for bucket entries that can
be rejected on the basis of hashcode comparisons. The difficulty is to
arrive at a reasonable guess of what fraction of the bucket entries will
be so rejected, versus those that will incur a comparison-function call.
I'm leery of assuming there are no hash collisions, which is what you
seem to be proposing.
regards, tom lane
