On Tue, 19 Nov 2019 at 15:08, Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote:
>
> Yes, adding the condition to statext_mcv_clauselist_selectivity() would
> make this go away, and it's about the simplest solution.
>
It's probably worth going a little further, and verifying that
stat_clauses references at least two attributes. We do that further up
for the original clause list, but it may not be true for the filtered
list. For example, given a WHERE clause like
c0 > 0 AND c0 < 10 AND (c0 = 0 OR c1 = 1 OR c2 = 2)
and stats on (c0, c1), stat_clauses would include the first 2 clauses,
but they only reference 1 column, so it would be preferable to not use
the multivariate stats in that case.
> Ideally, we'd be able to improve the statistics matching to recognize
> it has to match all three attributes to match the clause, which in this
> case would mean the OR clause is passed to clause_selectivity, and we do
> some magic with extended statistics there.
>
> I'll see how complex / backpatchable that would be.
>
Yes, that seems like a worthwhile thing to do, but I think it goes
beyond what would normally be back-patched. It would really be a
feature enhancement rather than a bug fix.
Regards,
Dean