Message-ID: <48D9B01A.5010702@postnewspapers.com.au>
Date: Wed, 24 Sep 2008 11:12:26 +0800
From: Craig Ringer <craig@postnewspapers.com.au>
User-Agent: Thunderbird 2.0.0.16 (X11/20080724)
MIME-Version: 1.0
To: pgsql-performance@postgresql.org
Subject: Re: Chaotically weird execution plan
References: <4f8bd5c90809231753xbcdacd2xfba4d38c2eb34faf@mail.gmail.com>
	<48D9943B.9060801@postnewspapers.com.au>
	<17643.1222225270@sss.pgh.pa.us>
In-Reply-To: <17643.1222225270@sss.pgh.pa.us>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit

Tom Lane wrote:
> Craig Ringer <craig@postnewspapers.com.au> writes:
>> I'd already written: "If you need the test for status = 1, consider a
>> partial index" when I noticed your schema definition:
> 
>>> "comments_created_by" btree (created_by) WHERE status = 1
> 
>> I find it hard to guess why it's having to recheck the WHERE clause
>> given the use of a partial index that should cover that nicely.
> 
> No, that's operating as designed.  A bitmap scan's RECHECK condition
> is only applied when the bitmap has become lossy due to memory
> pressure.  In that case we have to look at each row on each of the pages
> fingered by the index as containing possible matches ... and we'd better
> check the partial-index qual too, since maybe not all the rows on those
> pages will satisfy it.  In a plain indexscan there is no lossiness
> involved and so the partial-index qual need never be rechecked.

Aah. Thanks very much for the explanation of that, the plan now makes sense.

--
Craig Ringer