Fix off-by-one in aggregations

[WillAyd]

ref #55713 (comment)

I'm not super familiar with this function so maybe there is a more intelligent way of doing this. But the segfault occurs when self.window_size == 0 because of these lines:

end = np.arange(1 + offset, num_values + 1 + offset, step, dtype="int64")
start = end - self.window_size

In the segfaulting example, num_values is 20 for an array with 20 elements. end ends up becoming 1-20 as does start with the window_size of 0. Later code then tries to index the 20th index in array, hence the address violation

[mroeschke]

Wouldn't this result in the same resulting offset as before? i.e.

In [1]: 0 - 1 // 2
Out[1]: 0

[WillAyd]

In your code 1 // 2 is getting evaluated first. The parentheses matter

>>> (0 - 1) // 2
-1

[WillAyd]

I verified this fixed the segfault via a local run of #55102 - hopefully can make progress soon to getting that up in CI since this is now 2 releases in a row where we have a segfault when cutting the release.

#55151 is a big step towards getting UBSAN/ASAN set up as the datetime code issues a ton of errors, and by design those tools stop at the first error encountered. If you have any time to help review that and push through would appreciate it

[mroeschke]

Thanks @WillAyd (for fixing my math mistake 😅 )