To: Drew Wilson <drewmwilson@gmail.com>
cc: pgsql-performance@postgresql.org
Subject: Re: how to efficiently update tuple in many-to-many relationship? 
In-reply-to: <FCEF7DE2-913F-4C36-B117-8363EDE39BF3@gmail.com> 
References: <233A9987-ABA5-4491-A304-FDA1F8FCD962@gmail.com>
	<14019.1176155022@sss.pgh.pa.us>
	<22512E5C-A60F-43E7-A370-257F7AF956FD@gmail.com>
	<16374.1176171185@sss.pgh.pa.us>
	<D14BB823-DA04-453C-AD70-248B34F1307B@gmail.com>
	<25200.1176213242@sss.pgh.pa.us>
	<FCEF7DE2-913F-4C36-B117-8363EDE39BF3@gmail.com>
Comments: In-reply-to Drew Wilson <drewmwilson@gmail.com>
	message dated "Tue, 10 Apr 2007 08:37:31 -0700"
Date: Tue, 10 Apr 2007 12:08:14 -0400
Message-ID: <26925.1176221294@sss.pgh.pa.us>
From: Tom Lane <tgl@sss.pgh.pa.us>

Drew Wilson <drewmwilson@gmail.com> writes:
> If I understand the EXPLAIN ANALYZE results below, it looks like the  
> time spent applying the "set is_public = true" is much much more than  
> the fetch. I don't see any triggers firing.

Nope, there aren't any.  8.2 is smart enough to bypass firing FK
triggers on UPDATE if the key columns didn't change.  Of course that
check takes a certain amount of time, but I don't think it's a large
amount.  So basically we're looking at index update and WAL logging
as the major costs here, I think.

[ thinks for a bit... ]  Part of the problem might be that the working
set for updating all the indexes is too large.  What do you have
shared_buffers set to, and can you increase it?

Oh, and checkpoint_segments at 8 is probably still not nearly enough;
if you have disk space to spare, try something like 30 (which'd eat
about 1Gb of disk space).  It might be educational to set
checkpoint_warning to 300 or so and watch the log to see how often
checkpoints happen during the update --- you want them at least a couple
minutes apart even during max load.

Also, bumping up wal_buffers a little might help some, for large updates
like this.  I've heard people claim that values as high as 64 are helpful.

			regards, tom lane