"Joel Jacobson" <joel@compiler.org> writes:
> On Fri, Feb 27, 2026, at 20:15, Tom Lane wrote:
>> Joel, do you want to run this to ground, and in particular
>> see if that way of fixing it passes your sanity tests?
> Challenge accepted!
Thanks!
> [...hours later...]
> My conclusion is that we still need to move avgfreq
> computation, like I suggested.
Hmm ... doesn't this contradict your argument that avgfreq and
mcv_freq need to be calculated on the same basis? Admittedly
that was just a heuristic, but I'm not seeing why it's wrong.
> The reason for this is that estfract is calculated as:
> estfract = 1.0 / ndistinct;
> where ndistinct has been adjusted to account for restriction clauses.
> Therefore, we must also use the adjusted avgfreq when adjusting
> estfract here:
It feels like that might end up double-counting the effects of
the restriction clauses.
Anyway, we all seem to agree that s/rel->rows/rel->tuples/ is the
correct fix for a newly-introduced bug. I'm inclined to proceed
by committing that fix (along with any regression test fallout)
and then investigating the avgfreq change as an independent matter.
regards, tom lane
