Datetimelike index / arrays ignore integer sign and size in astype

[TomAugspurger]

For datetimelike indexes, .astype ignore the sign and size of an integer dtype, and always uses int64

In [1]: import pandas as pd

In [2]: idx = pd.date_range('2000', periods=4)

In [3]: idx.astype('uint16')
Out[3]:
Int64Index([946684800000000000, 946771200000000000, 946857600000000000,
            946944000000000000],
           dtype='int64')

We don't have Index subclasses for each of these types (though we do have uint64 now).

In #24024 we agreed(?) that following the Index was the best behavior for the array class.

In [7]: idx._data.astype('uint16')
Out[7]:
array([946684800000000000, 946771200000000000, 946857600000000000,
       946944000000000000])

(that's also int64).

---

Some questions:

1. Do we want to change this? Note that for "normal" usage of datetimes, this is going to truncate / overflow. This is particularly unfriendly to 32-bit systems, where "int" is "int32".

In [9]: idx._data.astype('int32').astype("int32")
Out[9]: array([1380122624, -476971008, 1960902656,  103809024], dtype=int32)

So maybe we would want safe casting?

In [10]: idx._data.astype('int32').astype("int32", casting="safe")
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-10-2c2a4a677a5c> in <module>
----> 1 idx._data.astype('int32').astype("int32", casting="safe")

TypeError: Cannot cast array from dtype('int64') to dtype('int32') according to the rule 'safe'

2. ExtensionIndex classes are in the not too distant future, so we may be able to relatively easily get Index classes that work with these different integer sizes / signs. So we might be able to properly support this in the future (if we want to).

[jreback]

this is not likely easy to change as we don’t have the Index subclasses to support this

nor is it a big deal ; if/when we have subclasses then can supoort

[jbrockmendel]

This should raise TypeError, no?