Message-ID: <51FB61C8.4090206@dalibo.com>
Date: Fri, 02 Aug 2013 09:37:44 +0200
From: Vik Fearing <vik.fearing@dalibo.com>
User-Agent: Mozilla/5.0 (X11; Linux x86_64;
 rv:17.0) Gecko/20130623 Thunderbird/17.0.7
MIME-Version: 1.0
To: Scott Marlowe <scott.marlowe@gmail.com>
CC: Tom Lane <tgl@sss.pgh.pa.us>,
 "pgsql-performance@postgresql.org" <pgsql-performance@postgresql.org>
Subject: Re: subselect requires offset 0 for good performance.
References: 
 <CAOR=d=2Fq-TzwzRf92BOeeHGqhkPXVG5_zs+_aR690ZxPp3JAg@mail.gmail.com>
 <4522.1375400642@sss.pgh.pa.us>
 <CAOR=d=3UVax7k_nT7Gm+Y2N0=0WZekT2dyOzxWM4wjsryfQL=Q@mail.gmail.com>
In-Reply-To: 
 <CAOR=d=3UVax7k_nT7Gm+Y2N0=0WZekT2dyOzxWM4wjsryfQL=Q@mail.gmail.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk
Sender: pgsql-performance-owner@postgresql.org

On 08/02/2013 03:22 AM, Scott Marlowe wrote:
> On Thu, Aug 1, 2013 at 5:44 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Scott Marlowe <scott.marlowe@gmail.com> writes:
>>> I am running 8.4.15 and can try 8.4.17 if some patch has been applied
>>> to it to address this issue. I just want to know should I
>>> A: upgrade to 8.4.17
>>> or
>>> B: create a self contained test case.
>> A quick look at the release notes shows no planner fixes in 8.4.16 or
>> 8.4.17, so it would be rather surprising if (A) helps.
> OK. I was doing some initial testing and if I select out the 4 columns
> into a test table the query runs fast. If I select all the columns
> into a test table it runs slow, so it appears table width affects
> this. Will have more to report tomorrow on it.

I don't know what your query is, but here's one I was working on
yesterday that shows the problem.  It may not be the smallest test case
possible, but it works.

EXPLAIN ANALYZE
WITH RECURSIVE
x (start_time) AS
(
    SELECT generate_series(1, 1000000)
),
t (time, timeround) AS
(
    SELECT time, time - time % 900000 AS timeround
    FROM (SELECT min(start_time) AS time FROM x) AS tmp
    UNION ALL
    SELECT time, time - time % 900000
    FROM (SELECT (SELECT min(start_time) AS time
                  FROM x
                  WHERE start_time >= t.timeround + 900000)
          FROM t
          WHERE t.time IS NOT NULL OFFSET 0
         ) tmp
)
SELECT count(*) FROM t WHERE time IS NOT NULL;

If you remove the OFFSET 0, you'll see two more subplans (because "time"
is referenced three times).  The difference is much more noticeable if
you make the x CTE its own table.

Vik

PS: This query is emulating a LooseIndexScan.
http://wiki.postgresql.org/wiki/Loose_indexscan


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