From: Tomas Vondra <tomas.vondra@2ndquadrant.com>
To: Dean Rasheed <dean.a.rasheed@gmail.com>
Cc: Bruce Momjian <bruce@momjian.us>
Cc: Alvaro Herrera <alvherre@2ndquadrant.com>
Cc: Andres Freund <andres@anarazel.de>
Cc: Thomas Munro <thomas.munro@enterprisedb.com>
Cc: Mark Dilger <hornschnorter@gmail.com>
Cc: Adrien Nayrat <adrien.nayrat@dalibo.com>
Cc: Pg Hackers <pgsql-hackers@postgresql.org>
Subject: Re: [HACKERS] PATCH: multivariate histograms and MCV lists
Date: Tue, 17 Jul 2018 15:03:28 +0200
Message-ID: <c2614611-fd99-124f-c9aa-0fef3859c78f@2ndquadrant.com> (raw)
In-Reply-To: <CAEZATCX04NOAqJm7+vb_jyrMxDo44Pzx8xd_z5rRH7snsLqxEA@mail.gmail.com>
References: <49348328-1a02-7e45-e50f-e1d4d0243c5c@2ndquadrant.com>
<b423b2fe-8aff-d64c-152a-8c77cf45faa7@2ndquadrant.com>
<CAEZATCW=081sEDERoty8knpjwjnCQP+zCh-bFSb8xVxT8KKwHQ@mail.gmail.com>
<7d12b5d1-ac77-67e3-c67e-7fcbd3753e9c@2ndquadrant.com>
<abfb7b66-af22-ac32-a74d-3fcf9bdb5c5c@2ndquadrant.com>
<20180407141201.GA10394@momjian.us>
<CAEZATCXYHJoocBq3yqEG_BNDgLuvDiCxcK9Tg-qbQXrs9pnnHg@mail.gmail.com>
<4416a567-4452-c5fd-7117-5cc7827b5fad@2ndquadrant.com>
<f52637bd-f761-f91b-c8ac-85e0c159807f@2ndquadrant.com>
<CAEZATCVa-86yMzwVYVnrvfagO07YpXPR5X1uisvVasGHd3k-4A@mail.gmail.com>
<d6e5e15d-3f53-239a-7064-bb99f037a0da@2ndquadrant.com>
<CAEZATCUtvWzQ8BPgNcY8u6tnyAf6cvVoVC6RRdqbMuDNHfuS2g@mail.gmail.com>
<f0625b09-bf50-5703-6241-51aea96a3f6a@2ndquadrant.com>
<CAEZATCWORz=bXEPVJNxK49Ws9zeBHqsEdd3JsrrZWMvQ8oXuSA@mail.gmail.com>
<80bc98b2-1c43-1fe8-b316-d447bb68cff2@2ndquadrant.com>
<CAEZATCX04NOAqJm7+vb_jyrMxDo44Pzx8xd_z5rRH7snsLqxEA@mail.gmail.com>
On 07/17/2018 11:09 AM, Dean Rasheed wrote:
> On 16 July 2018 at 21:55, Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote:
>>
...
>>
>> So, how would the proposed algorithm work? Let's start with "a=1":
>>
>> sel(a=1) = 0.1508
>>
>> I don't see much point in applying the two "b" clauses independently (or
>> how would it be done, as it's effectively a single clause):
>>
>> sel(b=1 or b=2) = 0.0673
>>
>> And we get
>>
>> total_sel = sel(a=1) * sel(b=1 or b=2) = 0.0101
>>
>> From the multivariate MCV we get
>>
>> mcv_sel = 0.0399
>>
>> And finally
>>
>> total_sel = Max(total_sel, mcv_sel) = 0.0399
>>
>> Which is great, but I don't see how that actually helped anything? We
>> still only have the estimate for the ~7% covered by the MCV list, and
>> not the remaining non-MCV part.
>>
>
> Right. If these are the only stats available, and there are just 2
> top-level clauses like this, it either returns the MCV estimate, or
> the old univariate estimate (whichever is larger). It avoids
> over-estimating, but will almost certainly under-estimate when the MCV
> matches are incomplete.
>
Yeah :-(
In my experience under-estimates tend to have much worse consequences
(say a nested loop chosen by under-estimate vs. hash join chosen by
over-estimate). This certainly influenced some of the choices I've made
in this patch (extrapolation to non-MCV part is an example of that), but
I agree it's not particularly scientific approach and I'd very much want
something better.
>
>> I could do the same thing for the second query, but the problem there is
>> actually exactly the same - extrapolation from MCV to non-MCV part
>> roughly doubles the estimate.
>>
>> So unless I'm applying your algorithm incorrectly, this does not seem
>> like a very promising direction :-(
>>
>
> You could be right. Actually it's the order dependence with more than
> 2 top-level clauses that bothers me most about this algorithm. It's
> also not entirely obvious how to include histogram stats in this
> scheme.
>
I think for inequalities that's fairly simple - histograms work pretty
well for that, and I have a hunch that replacing the 0.5 estimate for
partially-matching buckets with conver_to_scalar-like logic and the
geometric mean (as you proposed) will work well enough.
For equalities it's going to be hard. The only thing I can think of at
the moment is checking if there are any matching buckets at all, and
using that to decide whether to extrapolate the MCV selectivity to the
non-MCV part or not (or perhaps to what part of the non-MCV part).
> A different approach that I have been thinking about is, in
> mcv_update_match_bitmap(), attempt to work out the probability of all
> the clauses matching and it not being an MCV value. For example, given
> a clause like a=1 whose univariate estimate we know (0.1508 in the
> above example), knowing that the MCV values account for 0.0222+0.0177
> of that, the remainder must be from non-MCV values. So we could say
> that the probability that a=1 and it not being and MCV is
> 0.1508-0.0222-0.0177 = 0.1109. So then the question is could we
> combine these non-MCV probabilities to give an estimate of how many
> non-MCV values we need to worry about? I've not fully thought that
> through, but it might be useful.
Could we use it to derive some upper boundaries on the non-MCV part?
> The problem is, it's still likely to
> over-estimate when the MCV matches fully cover all possibilities, and
> I still don't see a way to reliably detect that case.
>
I guess we can use a histogram to limit the over-estimates, as explained
above. It may not be 100% reliable as it depends on the sample and how
exactly the buckets are formed, but it might help.
But perhaps these over-estimates are a particularly serious issue? When
you already get matches in a MCV, the number of matching rows is going
to be pretty significant. If you over-estimate 2x, so what? IMHO that's
still pretty accurate estimate.
regards
--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Reply to all the recipients using the --to and --cc options:
reply via email
To: pgsql-hackers@postgresql.org
Cc: tomas.vondra@2ndquadrant.com, dean.a.rasheed@gmail.com, bruce@momjian.us, alvherre@2ndquadrant.com, andres@anarazel.de, thomas.munro@enterprisedb.com, hornschnorter@gmail.com, adrien.nayrat@dalibo.com
Subject: Re: [HACKERS] PATCH: multivariate histograms and MCV lists
In-Reply-To: <c2614611-fd99-124f-c9aa-0fef3859c78f@2ndquadrant.com>
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
This inbox is served by DDX for PostgreSQL; see mirroring instructions
for how to clone and mirror all data and code used for this inbox