On Tue, 2011-10-11 at 13:22 -0400, Robert Haas wrote:
> The real issue is that the costing estimates need to be accurate, and
> that's where the rubber hits the road. Otherwise, even if we pick the
> right way to scan the table, we may do silly things up the line when
> we go to start constructing the join order. I think we need to beef
> up ANALYZE to gather statistics on the fraction of the pages that are
> marked all-visible, or maybe VACUUM should gather that information.
> The trouble is that if we VACUUM and then ANALYZE, we'll often get
> back a value very close to 100%, but then the real value may diminish
> quite a bit before the next auto-analyze fires. I think if we can
> figure out what to do about that problem we'll be well on our way...
Can you send stats messages to keep track when you unset a bit in the
VM? That might allow it to be more up-to-date.
Regards,Jeff Davis