To: Jeremy Palmer <JPalmer@linz.govt.nz>
cc: "pgsql-performance@postgresql.org" <pgsql-performance@postgresql.org>
Subject: Re: Possible to improve query plan? 
In-reply-to: 
 <666FB8D75E95AE42965A0E76A5E5337E06D2C683BF@prdlsmmsg01.ad.linz.govt.nz>
References: 
 <666FB8D75E95AE42965A0E76A5E5337E06D2C683BF@prdlsmmsg01.ad.linz.govt.nz>
Comments: In-reply-to Jeremy Palmer <JPalmer@linz.govt.nz>
	message dated "Mon, 17 Jan 2011 16:21:56 +1300"
Date: Mon, 17 Jan 2011 15:24:26 -0500
Message-ID: <20065.1295295866@sss.pgh.pa.us>
From: Tom Lane <tgl@sss.pgh.pa.us>

Jeremy Palmer <JPalmer@linz.govt.nz> writes:
> I've come to a dead end in trying to get a commonly used query to
> perform better.

> EXPLAIN
> SELECT * FROM (
>     SELECT
>         row_number() OVER (PARTITION BY id ORDER BY _revision_created DESC) as row_number,
>         * 
>     FROM
>         version_crs_coordinate_revision
>     WHERE (
>         (_revision_created <= 16 AND _revision_expired > 16 AND _revision_expired <= 40) OR 
>         (_revision_created > 16 AND _revision_created <= 40)
>     )
> ) AS T 
> WHERE row_number = 1;

If I'm not mistaken, that's a DB2-ish locution for a query with DISTINCT
ON, ie, you're looking for the row with highest _revision_created for
each value of id.  It might perform well on DB2, but it's going to
mostly suck on Postgres --- we don't optimize window-function queries
very much at all at the moment.  Try writing it with DISTINCT ON instead
of a window function, like so:

SELECT DISTINCT ON (id)
        * 
    FROM
        version_crs_coordinate_revision
    WHERE (
        (_revision_created <= 16 AND _revision_expired > 16 AND _revision_expired <= 40) OR 
        (_revision_created > 16 AND _revision_created <= 40)
    )
ORDER BY id, _revision_created DESC;

You could also experiment with various forms of GROUP BY if you're loath
to use any Postgres-specific syntax.

			regards, tom lane