Re: using extended statistics to improve join estimates

On 17/6/2024 18:10, Tomas Vondra wrote:
> Let me quickly go through the original parts - most of this is already
> in the "review" patches, but it's better to quote the main points here
> to start a discussion. I'll omit some of the smaller suggestions, so
> please look at the 'review' patches.
>
>
> v20240617-0001-Estimate-joins-using-extended-statistics.patch
>
> - rewords a couple comments, particularly for statext_find_matching_mcv
>
> - a couple XXX comments about possibly stale/inaccurate comments
> v20240617-0054-clauselist_selectivity_hook.patch
>
> - I believe this does not work with the earlier patch that removed
> estimatedclaused bitmap from the "try" function.
This patch set is too big to eat at once - it's just challenging to
invent examples and counterexamples. Can we see these two patches in the
master and analyse further improvements based on that?

Some thoughts:
You remove verRelid. I have thought about replacing this value with
RelOptInfo, which would allow extensions (remember selectivity hook) to
know about the underlying path tree.

The first patch is generally ok, and I vote for having it in the master.
However, the most harmful case I see most reports about is parameterised
JOIN on multiple anded clauses. In that case, we have a scan filter on
something like the below:
x = $1 AND y = $2 AND ...
As I see, current patch doesn't resolve this issue currently.

--
regards, Andrei Lepikhov