On Dec12, 2013, at 19:29 , Tom Lane <tgl@sss.pgh.pa.us> wrote:
> However ... where this thread started was not about trying to reduce
> the remaining statistical imperfections in our existing sampling method.
> It was about whether we could reduce the number of pages read for an
> acceptable cost in increased statistical imperfection.
True, but Jeff's case shows that even the imperfections of the current
sampling method are larger than what the n_distinct estimator expects.
Making it even more biased will thus require us to rethink how we
obtain a n_distinct estimate or something equivalent. I don't mean that
as an argument against changing the sampling method, just as something
to watch out for.
best regards,
Florian Pflug