Content-Type: text/plain; charset=UTF-8
Cc: Joe Conway <mail@joeconway.com>, Josh Berkus <josh@agliodbs.com>,
	PostgreSQL-development <pgsql-hackers@postgresql.org>
Subject: Re: Return of the Solaris vacuum polling problem -- anyone remember
	this?
From: Alvaro Herrera <alvherre@commandprompt.com>
To: Alvaro Herrera <alvherre@commandprompt.com>
In-reply-to: <1281992147-sup-9595@alvh.no-ip.org>
References: <4C69827A.409@agliodbs.com> <4C698CAD.2020001@joeconway.com>
	<4C698D88.6030201@agliodbs.com> <4C69A3D7.8020609@joeconway.com>
	<1281992147-sup-9595@alvh.no-ip.org>
Date: Mon, 16 Aug 2010 17:24:23 -0400
Message-Id: <1281992840-sup-6951@alvh.no-ip.org>
User-Agent: Sup/git
Content-Transfer-Encoding: 8bit

Excerpts from Alvaro Herrera's message of lun ago 16 16:58:31 -0400 2010:

> I suspect that the problem may lie in the "cost_delay rebalance" code in
> autovacuum.

Hmm, so we have this code:

void
AutoVacuumUpdateDelay(void)
{
	if (MyWorkerInfo)
	{
		VacuumCostDelay = MyWorkerInfo->wi_cost_delay;
		VacuumCostLimit = MyWorkerInfo->wi_cost_limit;
	}
}

where the MyWorkerInfo bits come from shared memory and can be modified
by other autovac worker processes.  We could read an incomplete value
into our variables.  But this only makes sense if an "int" variable can
be subject to a partial read/write, which we already assume not to be so
(c.f. GetNewTransactionId).

In any case, if you happen to see this reoccur, could you please attach
GDB to the misbehaving worker and see what VacuumCostDelay and
VacuumCostLimit print out as?

-- 
Álvaro Herrera <alvherre@commandprompt.com>
The PostgreSQL Company - Command Prompt, Inc.
PostgreSQL Replication, Consulting, Custom Development, 24x7 support