On Sun, Mar 1, 2026, at 22:12, Tom Lane wrote:
> What is more sensible I think is just to clamp estfract to be at least
> mcv_freq always, dispensing with a whole bunch of the complication
> here. We actually don't need avgfreq at all, and therefore not
> stanullfrac, and also the ad-hoc range clamp at the bottom no longer
> seems necessary. Interestingly, this also makes the logic more
> nearly like the "isdefault" early exit:
>
> *bucketsize_frac = (Selectivity) Max(0.1, *mcv_freq);
>
> which IIRC was added a lot later than the original logic.
Nice simplification.
> So I end with the attached draft patch. Very interestingly,
> neither this version nor your lets-calculate-avgfreq-later
> patch change any regression tests at all compared to git HEAD.
LGTM.
> Maybe we should try to add a case that does change.
Yes, I think that would be good.
Attached patch adds a new test that does change.
/Joel