Skip site navigation (1) Skip section navigation (2)

Strange slowness

From: Russ Neufeld <russn(at)nasuni(dot)com>
To: psycopg(at)postgresql(dot)org
Subject: Strange slowness
Date: 2012-04-30 21:29:48
Message-ID: FB36DC48-1CC4-4FBC-B5D8-9FC8DB36EFCE@nasuni.com (view raw or flat)
Thread:
Lists: psycopg
Hi all,

I'm investigating some occasional slowness in our API servers, and I've hit upon some strange behavior that I can't explain.  I'm hoping someone on this list can help.

Stack is django 1.3.1, Ubuntu 10.04, python 2.6.5, psycopg 2.4.5, postgresql 9.0, pgbouncer 1.4.1.  All of these run on the same machine with the exception of postgresql, which runs on a dedicated db machine.  So, when django creates a connection to the database, it is connecting to localhost and really talking to pgbouncer.

Here's the relevant code from django/db/backends/postgresql_psycopg2/base.py:

	self.connection = Database.connect(**conn_params)
	self.connection.set_client_encoding('UTF8')
	self.connection.set_isolation_level(self.isolation_level)

	cursor = self.connection.cursor()

	cursor.execute("SET TIME ZONE %s", [settings_dict['TIME_ZONE']])

I put timing code between these lines, and found that calling set_client_encoding() was sometimes taking upwards of 6 seconds.  Very strange, but it led me to the fact that our db encoding was mistakenly set to sql_ascii, so I've fixed that.  Now, I'm seeing occasional 6 seconds delays on the next line - set_isolation_level.  We're using defaults, so this is setting the level to 1, or "read committed".  Theoretically that's a noop, but obviously the code needs to check that the value passed in equals the value already set, and then return.  Still, it shouldn't take upwards of 6 seconds to do this at times.  The machines are not swapping, and not doing anything else besides serving some static content.

It seems very strange that after I fixed the db encoding, the timing problem would jump to the next line.  Before I fixed the db encoding, we never saw long times spent in set_isolation_level.  This leads me to believe the issue is something else, perhaps network related.  For example, supposed Database.connect lazily creates a connection, and now that set_client_encoding can short circuit due to the encodings matching, the first time the connection is used is now in set_isolation_level.  This is just wild speculation, but I'm at a loss in explaining what I'm seeing.

Anyone have any ideas about what could be going on, or where to look further?  Thanks,

Russ

Responses

psycopg by date

Next:From: Daniele VarrazzoDate: 2012-04-30 21:46:36
Subject: Re: Strange slowness
Previous:From: Joe AbbateDate: 2012-04-19 00:30:05
Subject: Re: psycopg2+postgres check default charset of db or table

Privacy Policy | About PostgreSQL
Copyright © 1996-2014 The PostgreSQL Global Development Group