Robert Haas <robertmhaas@gmail.com> writes:
> Yeah, and I worry about min and max not being very usable - once they
> get pushed out to extreme values, there's nothing to drag them back
> toward normality except resetting the stats, and that's not something
> we want to encourage people to do frequently. Of course, averages over
> very long sampling intervals may not be too useful anyway, dunno.
Good point, but that doesn't mean that the request is unreasonable.
For min/max, we could possibly address this concern by introducing an
exponential decay over time --- that is, every so often, you take some
small fraction of (max - min) and add that to the running min while
subtracting it from the max. Or some other variant on that theme. There
might be a way to progressively discount old observations for average too,
though I'm not sure exactly how at the moment.
regards, tom lane