Hello Tom,
>> Here is a attempt at extending --scale so that it can be given a size.
>
> I do not actually find this to be a good idea. It's going to be
> platform-dependent, or not very accurate, or both, and thereby
> contribute to confusion by making results less reproducible.
I have often wanted to have such an option for testing, with criterion
like "within shared_buffers", "within memory", "twice the available
memory", to look for behavioral changes in some performance tests.
I you want reproducible (for some definition of reproducible) and
accurate, you can always use scale with a number. The report provides the
actual scale used anyway, so providing the size is just a convenience for
the initialization phase. I agree that it cannot be really exact.
Would it be more acceptable with some clear(er)/explicit caveat?
> Plus, what do we do if the backend changes table representation in
> some way that invalidates Kaarel's formula altogether?
Then the formula (a simple linear regression, really) should have to be
updated?
> More confusion would be inevitable.
There is no much confusion when the "scale" is reported. As for confusion,
a performance tests is influenced by dozen of parameters anyway.
Now if you do not want such a feature, you can mark it as rejected, and we
will keep on trying to guess or look for the formula till the end of
time:-)
--
Fabien.