Message-ID: <29bd1f4f-6b69-df43-1f3a-6ef1fc501657@enterprisedb.com>
Date: Mon, 24 Jul 2023 11:56:24 +0200
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.12.0
Subject: Re: POC: GROUP BY optimization
To: Andrey Lepikhov <a.lepikhov@postgrespro.ru>,
 PostgreSQL Developers <pgsql-hackers@lists.postgresql.org>,
 Tom Lane <tgl@sss.pgh.pa.us>
Cc: Zhihong Yu <zyu@yugabyte.com>, Teodor Sigaev <teodor@sigaev.ru>,
 Tomas Vondra <tomas.vondra@2ndquadrant.com>,
 David Rowley <dgrowleyml@gmail.com>, "a.rybakina"
 <a.rybakina@postgrespro.ru>,
 =?UTF-8?B?0JHQtdC70Y/Qu9C+0LIg0JTQsNC80LjRgCDQndCw0LjQu9C10LLQuNGH?=
 <d.belyalov@postgrespro.ru>
References: <CA+q6zcVRrd-z4YZ4M43ccst7aGL9==w5r1fionRWhP9ot6mybQ@mail.gmail.com>
 <dc933c66-41b9-8907-327e-f6361a690654@postgrespro.ru>
 <4ee5b15c-1ce0-59c7-8d4a-a1fd68ee0d12@postgrespro.ru>
 <721b3ff9-f214-f5d4-86c7-bae515bbec18@enterprisedb.com>
 <7bea462e-8c3e-0f8c-c7d5-f4daa12f3baf@postgrespro.ru>
 <9af89543-3d17-4c98-59de-1daa5bceeb06@enterprisedb.com>
 <9408e450-60c6-6fbc-d5c4-467bb0abfe67@postgrespro.ru>
 <44472925-f6c6-e8de-6f72-0451458532c4@enterprisedb.com>
 <CALNJ-vQp3WNV-kKuGsjSOhjRfdECA1Dbxa-vjAgOmQ4f4Qaw-Q@mail.gmail.com>
 <7a80a207-5e8e-b80c-476c-2d290b0526e9@enterprisedb.com>
 <65941e87-86c1-c09c-8ad5-8e102aebe8ee@enterprisedb.com>
 <CAApHDvq65=9Ro+hLX1i9ugWEiNDvHrBibAO7ARcTnf38_JE+UQ@mail.gmail.com>
 <CAApHDvoEF0GsqmBuCPXb6f12jPwz8EeLG-oYsOSc5zrJtKZHOg@mail.gmail.com>
 <7256159a-bb83-7ef3-c9c9-b2188a29aaf8@postgrespro.ru>
 <60610df1-c32f-ebdf-e58c-7a664431f452@enterprisedb.com>
 <3d30ae10-5c35-63c4-577c-d2e2479e15fd@postgrespro.ru>
Content-Language: en-US
From: Tomas Vondra <tomas.vondra@enterprisedb.com>
In-Reply-To: <3d30ae10-5c35-63c4-577c-d2e2479e15fd@postgrespro.ru>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Archived-At: <https://www.postgresql.org/message-id/29bd1f4f-6b69-df43-1f3a-6ef1fc501657%40enterprisedb.com>
Precedence: bulk

On 7/24/23 04:10, Andrey Lepikhov wrote:
> On 20/7/2023 18:46, Tomas Vondra wrote:
>> On 7/20/23 08:37, Andrey Lepikhov wrote:
>>> On 3/10/2022 21:56, Tom Lane wrote:
>>>> Revert "Optimize order of GROUP BY keys".
>>>>
>>>> This reverts commit db0d67db2401eb6238ccc04c6407a4fd4f985832 and
>>>> several follow-on fixes.
>>>> ...
>>>> Since we're hard up against the release deadline for v15, let's
>>>> revert these changes for now.  We can always try again later.
>>>
>>> It may be time to restart the project. As a first step, I rebased the
>>> patch on the current master. It wasn't trivial because of some latest
>>> optimizations (a29eab, 1349d27 and 8d83a5d).
>>> Now, Let's repeat the review and rewrite the current path according to
>>> the reasons uttered in the revert commit.
>>
>> I think the fundamental task is to make the costing more reliable, and
>> the commit message 443df6e2db points out a couple challenges in this
>> area. Not sure how feasible it is to address enough of them ...
>>
>> 1) procost = 1.0 - I guess we could make this more realistic by doing
>> some microbenchmarks and tuning the costs for the most expensive cases.
>>
>> 2) estimating quicksort comparisons - This relies on ndistinct
>> estimates, and I'm not sure how much more reliable we can make those.
>> Probably not much :-( Not sure what to do about this, the only thing I
>> can think of is to track "reliability" of the estimates and only do the
>> reordering if we have high confidence in the estimates. That means we'll
>> miss some optimization opportunities, but it should limit the risk.
> I read up on the history of this thread.
> As I see, all the problems mentioned above can be beaten by excluding
> the new cost model at all. We can sort GROUP BY columns according to the
> 'ndistinct' value.
> I see the reason for introducing the cost model in [1]. The main
> supporting point here is that with this patch, people couldn't optimize
> the query by themselves, organizing the order of the columns in a more
> optimal way. But now we have at least the GUC to switch off the
> behaviour introduced here. Also, some extensions, like the well-known
> pg_hint_plan, can help with automation.

I think the main concern is that if we reorder the group keys and get it
wrong, it's a regression. If that happens often (due to costing based on
poor stats), it's a problem. Yes, there's a GUC, but that's a rather
blunt instrument, unfortunately.

> So, how about committing of the undoubted part of the feature and
> working on the cost model in a new thread?
> 

But Tom's commit message says this:

    Worse, to arrive at estimates of the number of calls made to the
    lower-order-column comparison functions, the code needs to make
    estimates of the numbers of distinct values of multiple columns,
    which are necessarily even less trustworthy than per-column stats.

so I'm not sure this really counts as "undoubted".


regards

-- 
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company