Robert Haas <robertmhaas@gmail.com> writes:
> Still, maybe we don't have a better option. If it were me, I'd add an
> additional safety valve: use your formula if the percentage of the
> relation scanned is above some threshold where there's unlikely to be
> too much skew. But if the percentage scanned is too small, then don't
> use that formula. Instead, only update relpages/reltuples if the
> relation is now larger; set relpages to the new actual value, and
> scale up reltuples proportionately.
Ah, progress: now we're down to arguing about the size of the fudge
factor ;-). I'll do something involving derating the reliability
when the number is coming from VACUUM.
regards, tom lane