On Mon, 2008-01-28 at 10:55 -0500, Christopher Browne wrote:
> Simon Riggs <simon at 2ndquadrant.com> writes:
> > Your earlier graphs highlight the problem very well. Chris mentions that
> > the proposed solution has already been committed, based upon my
> > analysis. 
> >
> > Re-running the tests with the latest version of the software would help
> > confirm that the problem has gone away. Would you be able to repeat the
> > tests to confirm that the response time no longer peaks periodically as
> > it did before?
> 
> I wouldn't be surprised if it still has some periodic peaks that would
> be alleviated by increasing the number of elements processed at a
> time.

In general, you're right and I'd be happy if we added that as a tuning
parameter.

IIRC Cyril's peaks were associated directly with VACUUMs being performed
on the log tables, so we've fixed the cause rather than the symptom by
avoiding the VACUUMs. There might be other reasons why a backlog could
build up though, hence being able to tune the fetch size might allow us
to smooth out other response time peaks that we find later.

-- 
  Simon Riggs
  2ndQuadrant  http://www.2ndQuadrant.com