On Wed, Jun 07, 2006 at 04:04:33PM -0400, Tom Lane wrote:
> Martijn van Oosterhout <kleptog@svana.org> writes:
> > Anyway, as a test, if you take the approach that the measurement at
> > item X only applies to the tuples immediately preceding it, for the
> > data you posted you get a result of 0.681148 seconds. How long did that
> > query run that produced that data?
<snip>
> The above idea won't fix it anyway, only move the failure cases around.
Well, if we're trying to make a system that never ever produces strange
looking results, then yes, we have to drop it. No matter how you
sample, there's going to be a failure mode somewhere.
If you are prepared to live with a certain margin (it will be within X%
of the real value Y% of the time) then I think that's an acheivable
goal (I'm not saying that's necessarily what we have now). You could
always give people the choice of disabling sampling if it looks wierd,
but that just gives people more knobs to twiddle and get upset about.
Have a nice day,
--
Martijn van Oosterhout <kleptog@svana.org> http://svana.org/kleptog/
> From each according to his ability. To each according to his ability to litigate.