To: =?ISO-8859-2?Q?Sz=FBcs_G=E1bor?= <surrano@gmail.com>
Cc: "pgsql-performance@postgresql.org" <pgsql-performance@postgresql.org>
Subject: Re: fake condition causes far better plan 
In-reply-to: <430B379A.7090204@gmail.com> 
References: <430B379A.7090204@gmail.com>
Comments: In-reply-to =?ISO-8859-2?Q?Sz=FBcs_G=E1bor?= <surrano@gmail.com>
	message dated "Tue, 23 Aug 2005 16:50:02 +0200"
Date: Tue, 23 Aug 2005 11:14:19 -0400
Message-ID: <2638.1124810059@sss.pgh.pa.us>
From: Tom Lane <tgl@sss.pgh.pa.us>

=?ISO-8859-2?Q?Sz=FBcs_G=E1bor?= <surrano@gmail.com> writes:
> [ bad query plan ]

Most of the problem is here:

>                       ->  Index Scan using muvelet_vonalkod_ny_idopont on 
> muvelet_vonalkod_ny ny  (cost=0.00..1351.88 rows=24649 width=4) (actual 
> time=0.161..10.735 rows=3943 loops=1)
>                             Index Cond: (idopont >= (now() - 
> ('00:00:00'::interval + ('1 days'::text)::interval)))

(BTW, you lied about the query, because this index condition doesn't
match anything in the given query text.)

Pre-8.0 releases aren't capable of making useful statistical estimates
for conditions involving nonconstant subexpressions, so you get a
badly-mistaken row count estimate that leads to a poor choice of plan.

If you can't update to 8.0, the best answer is to do the date arithmetic
on the client side.  Another way is to create an allegedly-immutable
function along the lines of "ago(interval) returns timestamptz" to hide
the now() call --- this is dangerous but sometimes it's the easiest answer.
See the archives.

			regards, tom lane