BUG: concat_{categorical, compat} leads to erroneous result on non-ns datetime-EA

[xhochy]

Code to reproduce:

import numpy as np
import pandas as pd

np_datetimes = np.array([datetime.date(2010, 1, 1)], dtype="datetime64[D]")
other = pd.array(["a", "b"], dtype="category")
pd.core.dtypes.concat.concat_categorical([np_datetimes, other])
# outputs:
#   array([Timestamp('1970-01-01 00:00:00.000014610'), 'a', 'b'], dtype=object)
# expected either one of
#   a) array([Timestamp('2010-01-01 00:00:00'), 'a', 'b'], dtype=object)
#   b) array([datetime.date(2010, 1, 1), 'a', 'b'], dtype=object)

This happens as concat_datetime / _convert_datetimelike_to_object assumes that datetimes are nanoseconds only.

[xhochy]

Possible options include support for non-nanosecond timestamps in _convert_datetimelike_to_object or that we directly convert to np.array(…, dtype=object) in concat_categorical.

[jreback]

FYI @xhochy happy to have these issues, but keep in mind that i8 backing of anything datetime related is quite baked in; we will likey need an extended period and a fair amount of effort to generalize this.

I would make a master issue that references these other issues (with check boxes).

[xhochy]

I can make a master issue once I come across more of these.

Also I'm aware that this is non-trivial and I expect nobody but me to implement anything here.

[jbrockmendel]

works on main. could use a test

[urmikakasi]

working on the test