From: Tom Lane <tgl@sss.pgh.pa.us>
To: Gustavo Rezende Montesino <gustavo.montesino@trtsp.jus.br>
cc: Clailson <clailson.dba@gmail.com>, phillip@couto.in,
        "pgsql-performance@postgresql.org" <pgsql-performance@postgresql.org>
Subject: Re: Optimization inner join
In-reply-to: <5880BDEA.9050804@trtsp.jus.br>
References: <7ad1128e-7fdb-9976-823c-d4abdf554fa2@gmail.com>
 <497e3581-bcce-4efc-b281-6b17a93479c4@mtasv.net>
 <2cd4d430-54ad-4ab8-baae-b597c4a18c43@gmail.com>
 <5880BDEA.9050804@trtsp.jus.br>
Comments: In-reply-to Gustavo Rezende Montesino
 <gustavo.montesino@trtsp.jus.br>
	message dated "Thu, 19 Jan 2017 11:23:54 -0200"
MIME-Version: 1.0
Content-Type: text/plain; charset="UTF-8"
Content-ID: <16726.1484835221.1@sss.pgh.pa.us>
Content-Transfer-Encoding: quoted-printable
Date: Thu, 19 Jan 2017 09:13:41 -0500
Message-ID: <16727.1484835221@sss.pgh.pa.us>
Precedence: bulk
Sender: pgsql-performance-owner@postgresql.org

Gustavo Rezende Montesino <gustavo.montesino@trtsp.jus.br> writes:
> Being the client in question, I would like to make a little remark: What=
=20
> we thought could be optimized here at first is on the row estimate of=20
> the index scan; which could take null_frac into account. To put things=20
> into perspective, our similar case in production has a table with 6=20
> million lines where only 9.5k aren=C2=B4t null for the join field, an the=
=20
> over-estimation is throwing away good plans (like ~150ms execution time)=
=20
> in favor of pretty bad ones (~80s execution time).

Please provide a concrete test case for that.  AFAIK the null fraction
should be accounted for in join size estimates.  Here's a little test
case showing that it is:

regression=3D# create table t1 as select generate_series(1,1000000) as f1;
SELECT 1000000
regression=3D# analyze t1;
ANALYZE
regression=3D# create table t2 as select generate_series(1,1000000) as f1;
SELECT 1000000
regression=3D# analyze t2;
ANALYZE
regression=3D# explain select * from t1,t2 where t1.f1=3Dt2.f1;
                               QUERY PLAN=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20
------------------------------------------------------------------------
 Hash Join  (cost=3D30832.00..70728.00 rows=3D1000000 width=3D8)
   Hash Cond: (t1.f1 =3D t2.f1)
   ->  Seq Scan on t1  (cost=3D0.00..14425.00 rows=3D1000000 width=3D4)
   ->  Hash  (cost=3D14425.00..14425.00 rows=3D1000000 width=3D4)
         ->  Seq Scan on t2  (cost=3D0.00..14425.00 rows=3D1000000 width=3D=
4)
(5 rows)

regression=3D# insert into t2 select null from generate_series(1,1000000);
INSERT 0 1000000
regression=3D# analyze t2;
ANALYZE
regression=3D# explain select * from t1,t2 where t1.f1=3Dt2.f1;
                               QUERY PLAN=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20
------------------------------------------------------------------------
 Hash Join  (cost=3D30832.00..95727.00 rows=3D1000000 width=3D8)
   Hash Cond: (t2.f1 =3D t1.f1)
   ->  Seq Scan on t2  (cost=3D0.00..27862.00 rows=3D2000000 width=3D4)
   ->  Hash  (cost=3D14425.00..14425.00 rows=3D1000000 width=3D4)
         ->  Seq Scan on t1  (cost=3D0.00..14425.00 rows=3D1000000 width=3D=
4)
(5 rows)

The join size estimate is still correct even though it knows there are
many more rows in t2.

As for inserting a not-null test at the scan level, I'm not exactly
convinced that it's a win:

regression=3D# \timing
Timing is on.
regression=3D# select count(*) from t1,t2 where t1.f1=3Dt2.f1;
  count=20=20
---------
 1000000
(1 row)

Time: 562.914 ms
regression=3D# select count(*) from t1,t2 where t1.f1=3Dt2.f1 and t2.f1 is =
not null;
  count=20=20
---------
 1000000
(1 row)

Time: 564.896 ms

[ ftr, these times are best-of-three-trials ]

It's possible that in the case where an explicit sort has to be inserted,
reducing the amount of data passing through the sort would be worth doing;
but in the general case that's unproven.

			regards, tom lane


--=20
Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-performance