Message-ID: <42EA5E64.6050209@beccati.com>
Date: Fri, 29 Jul 2005 18:50:44 +0200
From: Matteo Beccati <php@beccati.com>
User-Agent: Mozilla Thunderbird 1.0.2 (Windows/20050317)
MIME-Version: 1.0
To: pgsql-performance@postgresql.org
Subject: ltree <@ operator selectivity causes very slow plan
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit

Hi,

I'm happily using ltree since a long time, but I'm recently having 
troubles because of ltree <@ operator selectivity that is causing very 
bad planner choices.

An example of slow query is:

SELECT
   batch_id,
   b.t_stamp AS t_stamp,
   objects,
   CASE WHEN sent IS NULL THEN gw_batch_sent(b.batch_id) ELSE sent END 
AS sent
FROM
   gw_users u JOIN gw_batches b USING (u_id)
WHERE
   u.tree <@ '1041' AND
   b.t_stamp >= 'today'::date - '7 days'::interval AND
   b.t_stamp < 'today'
ORDER BY
   t_stamp DESC;

I've posted the EXPLAIN ANALYZE output here for better readability: 
http://rafb.net/paste/results/NrCDMs50.html

As you may see, disabling nested loops makes the query lightning fast.


The problem is caused by the fact that most of the records of gw_users 
match the "u.tree <@ '1041'" condition:

SELECT COUNT(*) FROM gw_users;
  count
-------
   5012

SELECT COUNT(*) FROM gw_users WHERE tree <@ '1041';
  count
-------
   4684

Is there anything I can do apart from disabling nested loops?


Best regards
--
Matteo Beccati
http://phpadsnew.com
http://phppgads.com