Date: Sat, 18 Feb 2023 12:51:09 -0600
From: Justin Pryzby <pryzby@telsasoft.com>
To: Tomas Vondra <tomas.vondra@enterprisedb.com>
Cc: Andres Freund <andres@anarazel.de>, Greg Stark <stark@mit.edu>,
	Matthias van de Meent <boekewurm+postgres@gmail.com>,
	Zhihong Yu <zyu@yugabyte.com>, pgsql-hackers@lists.postgresql.org
Subject: Re: PATCH: Using BRIN indexes for sorted output
Message-ID: <20230218185109.GH1653@telsasoft.com>
References: 
 <CAEze2Wi7S7bMeUpMyBbnGLrFgcWQTDN1edSXYcd8WixeBj9B_w@mail.gmail.com>
 <87964cda-e440-8759-2838-bc4628072a18@enterprisedb.com>
 <CAM-w4HMirdJ-c4chCZhni_T=9R65z2Zm3Ajg5mBrQ9s6OufZXQ@mail.gmail.com>
 <14568731-ac91-ce2d-1477-0e2504c72775@enterprisedb.com>
 <20221206185143.ivy3afwmkhloksgv@awork3.anarazel.de>
 <4e679562-fd9a-33a3-ce71-cd7a27525f8e@enterprisedb.com>
 <e0b6e395-3ab7-6021-b058-3c5901912b36@enterprisedb.com>
 <20230216161016.GV1653@telsasoft.com>
 <9f73f798-1dac-c75d-d834-aa012ba13b65@enterprisedb.com>
 <8f1374d1-6f4b-ce99-84eb-ffa852c0dadb@enterprisedb.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <8f1374d1-6f4b-ce99-84eb-ffa852c0dadb@enterprisedb.com>
User-Agent: Mutt/1.9.4 (2018-02-28)
Archived-At: 
 <https://www.postgresql.org/message-id/20230218185109.GH1653%40telsasoft.com>
Precedence: bulk

Are (any of) these patches targetting v16 ?

typos:
ar we - we are?
morestly - mostly
interstect - intersect

> + * XXX We don't sort the bins, so just do binary sort. For large number of values
> + * this might be an issue, for small number of values a linear search is fine.

"binary sort" is wrong?

> + * only half of there ranges, thus 1/2. This can be extended to randomly

half of *these* ranges ?

> From 7b3307c27b35ece119feab4891f03749250e454b Mon Sep 17 00:00:00 2001
> From: Tomas Vondra <tomas.vondra@postgresql.org>
> Date: Mon, 17 Oct 2022 18:39:28 +0200
> Subject: [PATCH 01/11] Allow index AMs to build and use custom statistics

I think the idea can also apply to btree - currently, correlation is
considered to be a property of a column, but not an index.  But that
fails to distinguish between a freshly built index, and an index with
out of order heap references, which can cause an index scan to be a lot
more expensive.

I implemented per-index correlation stats way back when:
https://www.postgresql.org/message-id/flat/20160524173914.GA11880%40telsasoft.com

See also:
https://www.postgresql.org/message-id/14438.1512499811@sss.pgh.pa.us

With my old test case:

Index scan is 3x slower than bitmap scan, but index scan is costed as
being cheaper:

postgres=# explain analyze SELECT * FROM t WHERE i>11 AND i<55;
 Index Scan using t_i_idx on t  (cost=0.43..21153.74 rows=130912 width=8) (actual time=0.107..222.737 rows=128914 loops=1)

postgres=# SET enable_indexscan =no;
postgres=# explain analyze SELECT * FROM t WHERE i>11 AND i<55;
 Bitmap Heap Scan on t  (cost=2834.28..26895.96 rows=130912 width=8) (actual time=16.830..69.860 rows=128914 loops=1)

If it's clustered, then the index scan is almost twice as fast, and the
costs are more consistent with the associated time.  The planner assumes
that the indexes are freshly built...

postgres=# CLUSTER t USING t_i_idx ;
postgres=# explain analyze SELECT * FROM t WHERE i>11 AND i<55;
 Index Scan using t_i_idx on t  (cost=0.43..20121.74 rows=130912 width=8) (actual time=0.084..117.549 rows=128914 loops=1)

-- 
Justin