On 28 January 2013 10:47, Heikki Linnakangas <hlinnakangas@vmware.com> wrote:
> There's also some overhead from empty
> buckets when scanning the hash table
Seems like we should measure that overhead. That way we can plot the
cost against number per bucket, which sounds like it has a minima at
1.0, but that doesn't mean its symmetrical about that point. We can
then see where the optimal setting should be.
Having said that the hash bucket estimate is based on ndistinct, which
we know is frequently under-estimated, so it would be useful to err on
the low side.
-- Simon Riggs http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services