On Tue, 2 Dec 2025 at 20:26, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>
> I experimented with the attached patch, which is very incomplete;
> I just carried it far enough to be able to run performance checks on
> the modified code, and so all the binary statistics aggregates except
> corr() are broken. I observe about 2% slowdown on this test case:
I played around with having just 2 extra array elements, constX and
constY equal to the common value if all the values are the same, and
NaN otherwise. For me, that was slightly faster, which I put down to
floating point comparison being faster than converting back and forth
between floating point and integer. Either way, it seems like a
difference that no one is likely to notice.
Doing it that way does lead to one difference though: all-NaN inputs
leads to a NaN result, whereas your patch produces NULL for that case.
Regards,
Dean
