On Feb 13, 2008 4:07 PM, Andrew Sullivan <ajs at crankycanuck.ca> wrote:
> On Wed, Feb 13, 2008 at 12:52:49AM +0100, Dawid Kuroczko wrote:
>
> > For slony-I enabled database you will always get many rollbacks
> > a minute.
>
> > In short: slony's commiting instead of rolling back helps database
> > monitoring.
>
> If you don't collect baselines of your application for what is its normal
> behaviour, I suggest your monitoring plan needs rework.  You still ought to
> be able to see unusual numbers of rollbacks with the tool you want.  It just
> isn't "0".  But your baseline is whatever it is, and if your monitor tool
> can't compare the current behavior to some arbitrary baseline, your tool
> needs work.

Of course my tool has a baseline.  The slaves have constant rate between
39 to 41 rollbacks per minute.  The master calls issues ROLLBACK
between 45 and 60, with average of 50 per minute.  Alarm levels are set
accordingly.

Now, the number of rollbacks on master is closely related to number of
DMLs issued (45-50 : quiet database) (60 -- DML load).

But still I feel this is more like a workaround, especially that since 8.3.x
there should be no difference between commit and rollback for read-only
queries.  Or am I wrong?

   Regards,
      Dawid