"Simon Riggs" <simon@2ndquadrant.com> writes:
> I'm slightly worried though since that seems to have changed from 8.2,
> which I oprofiled over Christmas.
If you were testing a case with wider rows than Heikki tested, you'd see
less impact --- the cost of the old way was O(N^2) in the number of
tuples that fit on a page, so the behavior gets rapidly worse as you get
down to smaller tuple sizes. (Come to think of it, the cmin/cmax
collapse would be a factor here too.)
regards, tom lane