Gregory Stark <stark(at)enterprisedb(dot)com> writes:
> It's possible there's some sort of "block-wise correlated" measure
> which would be even better for our needs.
Actually, it seems obvious to me that the correlation measure ought to
ignore within-block ordering, but right now it does not. OTOH it's not
clear how important that is, as on a decent-size table you'll probably
not have more than one sample row in a block anyway.
regards, tom lane