Subject: Re: PostgreSQL as a local in-memory cache
From: Brad Nicholson <bnichols@ca.afilias.info>
To: Bruce Momjian <bruce@momjian.us>
Cc: Jignesh Shah <jkshah@gmail.com>, Tom Lane <tgl@sss.pgh.pa.us>,
	Robert Haas <robertmhaas@gmail.com>,
	Dimitri Fontaine <dfontaine@hi-media.com>, Josh Berkus <josh@agliodbs.com>,
	postgres performance list <pgsql-performance@postgresql.org>
In-Reply-To: <201006300139.o5U1dHN12285@momjian.us>
References: <201006300139.o5U1dHN12285@momjian.us>
Content-Type: text/plain; charset="UTF-8"
Date: Wed, 30 Jun 2010 10:30:57 -0400
Message-ID: <1277908257.19899.67.camel@bnicholson-desktop>
Mime-Version: 1.0
Content-Transfer-Encoding: 7bit

On Tue, 2010-06-29 at 21:39 -0400, Bruce Momjian wrote:
> Jignesh Shah wrote:
> > On Tue, Jun 29, 2010 at 2:45 PM, Bruce Momjian <bruce@momjian.us> wrote:
> > > Tom Lane wrote:
> > >> Bruce Momjian <bruce@momjian.us> writes:
> > >> >>> I asked on IRC and was told it is true, and looking at the C code it
> > >> >>> looks true. ?What synchronous_commit = false does is to delay writing
> > >> >>> the wal buffers to disk and fsyncing them, not just fsync, which is
> > >> >>> where the commit loss due to db process crash comes from.
> > >>
> > >> >> Ah, I see. ?Thanks.
> > >>
> > >> > I am personally surprised it was designed that way; ?I thought we would
> > >> > just delay fsync.
> > >>
> > >> That would require writing and syncing to be separable actions. ?If
> > >> you're using O_SYNC or similar, they aren't.
> > >
> > > Ah, very good point. ?I have added a C comment to clarify why this is
> > > the current behavior; ?attached and applied.
> > >
> > > --
> > > ?Bruce Momjian ?<bruce@momjian.us> ? ? ? ?http://momjian.us
> > > ?EnterpriseDB ? ? ? ? ? ? ? ? ? ? ? ? ? ? http://enterprisedb.com
> > 
> > 
> > Though has anybody seen a behaviour where synchronous_commit=off is
> > slower than synchronous_commit=on  ? Again there are two cases here
> > one with O_* flag and other with f*sync flags. But I had seen that
> > behavior with PostgreSQL 9.0 beta(2 I think) though havent really
> > investigated it much yet .. (though now I dont remember which
> > wal_sync_method flag) . Just curious if anybody has seen that
> > behavior..
> 
> I have trouble believing how synchronous_commit=off could be slower than
> 'on'.
> 

I wonder if it could be contention on wal buffers?

Say I've turned synchronous_commit off, I drive enough traffic fill up
my wal_buffers.  I assume that we would have to start writing buffers
down to disk before allocating to the new process.

-- 
Brad Nicholson  416-673-4106
Database Administrator, Afilias Canada Corp.