On 4/15/25 15:16, Tomas Vondra wrote:
> On 4/15/25 11:41, Andrei Lepikhov wrote:
>> What do you think about that?
> Not sure we'd want to export GroupVarInfo in the current shape. It was
> meant for a short-lived local state (within a single function), which
> seems quite different from what RestrictInfo does.
>
> In any case, such rework seems out of scope for PG18 - it's much more
> than a fix for the bug.
Of course, it is not for PG 18. I'm working on further GROUP-BY
optimisation where columns are re-arranged in a way to put at the first
position column with maximum ndistinct estimation. Additional work is
needed that may be compensated by distinct caching.
One more ongoing project - correcting ndistinct estimation by passing
through the EquivalenceClass and looking for a minimum ndistinct. It
also needs a kinda cache. That's why I am pondering how to redefine
GroupVarInfo to reduce estimation efforts.
--
regards, Andrei Lepikhov
