"Nathan Boley" <npboley@gmail.com> writes:
> ... There are two potential problems that I see with this approach:
> 1) It assumes the = is equivalent to <= and >= . This is certainly
> true for real numbers, but is it true for every equality relation that
> eqsel predicts for?
The cases that compute_scalar_stats is used in have that property, since
the < and = operators are taken from the same btree opclass.
> Do people think that the improved estimates would be worth the
> additional overhead?
Your argument seems to consider only columns having a normal
distribution. How badly does it fall apart for non-normal
distributions? (For instance, Zipfian distributions seem to be pretty
common in database work, from what I've seen.)
regards, tom lane