API/DEPR: casting bool-indexer to slice in dt64/td64/period

[jbrockmendel]

We use special logic for dt64/td64/period dtypes that makes view/copy behavior different from other dtypes:

dti = pd.date_range("2016-01-01", periods=4, tz="US/Pacific")
key = np.array([True, True, False, False])

ser1 = pd.Series(dti._data)
ser2 = pd.Series(range(4))

res1 = ser1[key]
res2 = ser2[key]

>>> res1._values._data.base is None
False
>>> res2._values.base is None
True

cc @jorisvandenbossche IIRC you advocated not doing this special casing.

[jbrockmendel]

cc @jorisvandenbossche i havent thought of any nice ways to deprecate this (s.t. we can say "do X to get the future behavior, do Y to get the old behavior"). thoughts?

[jorisvandenbossche]

Hmm, I also can't think of any way to deprecate this ..
It could maybe be done within the more general copy/view discussion, but then it's not something on the short term.

I am inclined to say that in this case it might be fine to simply change (because it's a behaviour that only comes up in rare value-dependent cases, and when it happens can lead to silent bugs).

[jbrockmendel]

I checked what happened if I remove this behavior: 26 test failures, all about a freq mismatch.

[jorisvandenbossche]

Because the slice keeps the freq per definition, and when doing a mask it always looses the freq?

[jbrockmendel]

That was my thought too, but when i tried patching get_getitem_freq to account for that it didnt help

[jbrockmendel]

update: got it working, was patching incorrectly in one place