From: Tom Lane <tgl@sss.pgh.pa.us>
To: Stefan Andreatta <s.andreatta@synedra.com>
cc: pgsql-performance@postgresql.org
Subject: Re: anti-join with small table via text/varchar cannot estimate rows
 correctly
In-reply-to: <4de52c93-68e1-7c4b-86ff-4dc17ae400bb@synedra.com>
References: <4de52c93-68e1-7c4b-86ff-4dc17ae400bb@synedra.com>
Comments: In-reply-to Stefan Andreatta <s.andreatta@synedra.com>
	message dated "Wed, 01 Mar 2017 23:00:06 +0100"
Date: Wed, 01 Mar 2017 20:06:58 -0500
Message-ID: <1454.1488416818@sss.pgh.pa.us>
Precedence: bulk
Sender: pgsql-performance-owner@postgresql.org

Stefan Andreatta <s.andreatta@synedra.com> writes:
> The same anti-join using the text fields, however estimates just 1 
> resulting row, while there are still of course 9,999 of them:

> =# explain analyze
>       select tmp_san_1.id
>       from tmp_san_1
>         left join tmp_san_2 on tmp_san_1.text = tmp_san_2.text
>       where tmp_san_2.id is null;

That is not an anti-join.  To make it one, you have to constrain the RHS
join column to be IS NULL, not some random other column.  Note the join
type isn't getting shown as Anti:

>   Hash Left Join  (cost=1.02..192.53 rows=1 width=4) (actual time=0.020..3.091 rows=9999 loops=1)

As written, the query could return some rows that weren't actually
antijoin rows, ie tmp_san_1.text *did* have a match in tmp_san_2,
but that row chanced to have a null value of id.

Possibly the planner could be smarter about estimating for this case,
but it doesn't look much like a typical use-case to me.

			regards, tom lane


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