Date: Tue, 20 Apr 2010 11:20:30 -0700
From: David Kerr <dmk@mr-paradox.net>
To: Nikolas Everett <nik9000@gmail.com>
Cc: Robert Haas <robertmhaas@gmail.com>, pgsql-performance@postgresql.org
Subject: Re: Very high effective_cache_size == worse performance?
Message-ID: <20100420182030.GA53489@mr-paradox.net>
References: <20100420173936.GA50886@mr-paradox.net>
	<h2w603c8f071004201044tbc8ce8fenb846b7643a2b9376@mail.gmail.com>
	<20100420180351.GA51940@mr-paradox.net>
	<w2ld4e11e981004201112v3571f350zaaa0112d1e035097@mail.gmail.com>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <w2ld4e11e981004201112v3571f350zaaa0112d1e035097@mail.gmail.com>
User-Agent: Mutt/1.4.2.3i

On Tue, Apr 20, 2010 at 02:12:15PM -0400, Nikolas Everett wrote:
- On Tue, Apr 20, 2010 at 2:03 PM, David Kerr <dmk@mr-paradox.net> wrote:
- 
- > that thought occured to me while I was testing this. I ran a vacuumdb -z
- > on my database during the load and it didn't impact performance at all.
- >
- > Incidentally the code is written to work like this :
- >
- > while (read X lines in file){
- > Process those lines.
- > write lines to DB.
- > }
- >
- > So i would generally expect to get the benefits of the updated staticis
- > once the loop ended. no?  (would prepared statements affect that possibly?)
- >
- > Also, while I was debugging the problem, I did load a 2nd file into the DB
- > ontop of one that had been loaded. So the statistics almost certinaly
- > should
- > have been decent at that point.
- >
- > I did turn on log_min_duration_statement but that caused performance to be
- > unbearable,
- > but i could turn it on again if it would help.
- >
- > Dave
- 
- 
- You can absolutely use copy if you like but you need to use a non-standard
- jdbc driver:  kato.iki.fi/sw/db/postgresql/jdbc/copy/.  I've used it in the
- past and it worked.
- 
- Is the whole thing going in in one transaction?  I'm reasonably sure
- statistics aren't kept for uncommited transactions.
- 
- For inserts the prepared statements can only help.  For selects they can
- hurt because eventually the JDBC driver will turn them into back end
- prepared statements that are only planned once.  The price here is that that
- plan may not be the best plan for the data that you throw at it.
- 
- What was log_min_duration_statement logging that it killed performance?
- 
- --Nik

Good to know about the jdbc-copy. but this is a huge project and the load is 
just one very very tiny component, I don't think we could introduce anything
new to assist that.

It's not all in one tx. I don't have visibility to the code to determine how 
it's broken down, but most likely each while loop is a tx.

I set it to log all statements (i.e., = 0.). that doubled the load time from 
~15 to ~30 hours. I could, of course, be more granular if it would be helpful.

Dave