On 19/9/2024 09:55, Andrei Lepikhov wrote:
> This wrong prediction makes things much worse if the query has more
> upper query blocks.
> His question was: Why not consider the grouping column unique in the
> upper query block? It could improve estimations.
> After a thorough investigation, I discovered that in commit 4767bc8ff2
> most of the work was already done for DISTINCT clauses. So, why not do
> the same for grouping? A sketch of the patch is attached.
> As I see it, grouping in this sense works quite similarly to DISTINCT,
> and we have no reason to ignore it. After applying the patch, you can
> see that prediction has been improved:
>
> Hash Right Join (cost=5.62..162.56 rows=50 width=36)
>
A regression test is added into new version.
The code looks tiny, simple and non-invasive - it will be easy to commit
or reject. So I add it to next commitfest.
--
regards, Andrei Lepikhov