From: Tom Lane <tgl@sss.pgh.pa.us>
To: Alvaro Herrera <alvherre@2ndquadrant.com>
cc: Andres Freund <andres@anarazel.de>, pgsql-hackers@lists.postgresql.org
Subject: Re: Refactoring IndexPath representation of index conditions
In-reply-to: <20190206013512.GA3101@alvherre.pgsql>
References: <20190206013512.GA3101@alvherre.pgsql>
Comments: In-reply-to Alvaro Herrera <alvherre@2ndquadrant.com>
	message dated "Tue, 05 Feb 2019 22:35:12 -0300"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <26160.1549469066.1@sss.pgh.pa.us>
Date: Wed, 06 Feb 2019 11:04:26 -0500
Message-ID: <26161.1549469066@sss.pgh.pa.us>
Precedence: bulk

Alvaro Herrera <alvherre@2ndquadrant.com> writes:
> On 2019-Feb-02, Tom Lane wrote:
>> Another idea that I looked into is to not create RestrictInfos for
>> derived indexqual clauses, with the hope that that would further
>> reduce the added overhead for the commuted-clause case.  However
>> that crashed and burned when I found out that the extended-stats
>> machinery punts when given a bare clause rather than a RestrictInfo.
>> It could possibly be fixed to not do that, but it looks like the
>> consequences would be extra lookups that'd probably cost more than
>> we saved by omitting the RestrictInfo.  Also, having RestrictInfos
>> means that we can cache selectivity estimates across multiple
>> calls.  I'm not entirely sure how much that matters in this
>> context, but it's probably not negligible.

> Is it reasonable to give ext-stats the option to receive either a
> "plain" clause or a RestrictInfo, and if the former have it construct
> the RestrictInfo and return it?

No, I don't think it'd be sane to have ext-stats modify that data
structure after-the-fact.  Too much risk of trouble (he says while
eyeing the GEQO machinery warily); plus, if we did it like that,
we'd *definitely* be giving up the ability to cache and share
cost/selectivity numbers between ext-stats and other places.

> It seems a pity to waste effort to
> cater for ext-stats, only to be used in the rare case where any
> ext-stats actually exist ... most of the time, it'd be wasted effort.

I'm not sure it's a good idea to design on the assumption that ext-stats
are rare.  I think they'll get more common over time.  Right now that
machinery is hardly built out at all, but it's coming.

			regards, tom lane