RE: Partial aggregates pushdown - Mailing list pgsql-hackers

From Fujii.Yuki@df.MitsubishiElectric.co.jp"
Subject RE: Partial aggregates pushdown
Date
Msg-id OS3PR01MB6660FD42509A170E5A64AC699538A@OS3PR01MB6660.jpnprd01.prod.outlook.com
Whole thread Raw
In response to Re: Partial aggregates pushdown  (Alexander Pyhalov <a.pyhalov@postgrespro.ru>)
Responses RE: Partial aggregates pushdown
List pgsql-hackers
Hi Mr.Pyhalov.

> From: Alexander Pyhalov <a.pyhalov@postgrespro.ru>
> Sent: Friday, July 14, 2023 10:40 PM
> 1) In foreign_join_ok() should we set fpinfo->user if
> fpinfo->check_partial_aggregate_support is set like it's done for 
> fpinfo->use_remote_estimate? It seems we can end up with fpinfo->user 
> fpinfo->=
> NULL if use_remote_estimate is not set.
You are right. I will modify this patch according to your advice.
Thank you for advice.

> 2) It seeems we found an additional issue with original patch, which 
> is present in current one. I'm attaching a patch which seems to fix 
> it, but I'm not quite sure in it.
Thank you for pointing out the issue.
If a query's group-by clause contains variable based expression(not variable)
and the query's select clause contains another expression,
the partial aggregate could be unsafe to push down.

An example of such queries:
SELECT (b/2)::numeric, avg(a), max(a), count(*) FROM pagg_tab GROUP BY b/2

Your patch disables partial aggregate pushdown for such queries.
I'll see if we can modify the patch to safely do a partial aggregate pushdown for such queries as well.
Such a query expects the variable in the select clause expression to be included in the target of the grouped rel
(let see make_partial_grouping_target), 
but the original groupby clause has no reference to this variable,
this seems to be the direct cause(let see foreign_grouping_ok). 
I will examine whether a safe pushdown can be achieved by matching the
groupby clause information referenced by foreign_grouping_ok with the grouped rel target information.

Sincerely yours,
Yuuki Fujii

--
Yuuki Fujii
Information Technology R&D Center Mitsubishi Electric Corporation

pgsql-hackers by date:

Previous
From: Peter Smith
Date:
Subject: Re: [PATCH] Reuse Workers and Replication Slots during Logical Replication
Next
From: Amit Kapila
Date:
Subject: Re: logicalrep_message_type throws an error