BUG: maybe_infer_to_datetimelike incorrect on 2D inputs

[jbrockmendel]

Moved from discussion in #23730

>>> ts = pd.Timestamp.now().tz_localize('US/Pacific')
>>> data = np.array([[ts, ts, ts], [ts, ts, ts]])
>>> pd.DataFrame(data)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "pandas/core/frame.py", line 416, in __init__
    copy=copy)
  File "pandas/core/frame.py", line 574, in _init_ndarray
    return create_block_manager_from_blocks([values], [columns, index])
  File "pandas/core/internals/managers.py", line 1665, in create_block_manager_from_blocks
    mgr = BlockManager(blocks, axes)
  File "pandas/core/internals/managers.py", line 118, in __init__
    self=self.ndim))
AssertionError: Number of Block dimensions (1) must equal number of axes (2)

This is because maybe_infer_to_datetimelike returns a DatetimeIndex (and not something 2D)

>>> from pandas.core.frame import maybe_infer_to_datetimelike
>>> maybe_infer_to_datetimelike(data)
DatetimeIndex(['2018-11-28 16:07:44.050944-08:00',
               '2018-11-28 16:07:44.050944-08:00',
               '2018-11-28 16:07:44.050944-08:00',
               '2018-11-28 16:07:44.050944-08:00',
               '2018-11-28 16:07:44.050944-08:00',
               '2018-11-28 16:07:44.050944-08:00'],
              dtype='datetime64[ns, US/Pacific]', freq=None)

cc @h-vetinari since you're already on a roll in this neighborhood

[mroeschke]

I'm guessing this is specific to the timezone aware case and perpetuated by the fact that the DatetimeTZ block only stores 1D arrays.

[h-vetinari]

Sorry I can't be of help here, haven't made it to maybe_infer_to_datetimelike yet (and I feel I have only the most rudimentary overview of the datetime/timedelta code)...

[mroeschke]

This looks to be fixed on master; was this fixed by #26825 or your EA refactoring @jbrockmendel?

In [16]: >>> ts = pd.Timestamp.now().tz_localize('US/Pacific')
    ...: >>> data = np.array([[ts, ts, ts], [ts, ts, ts]])
    ...: >>> pd.DataFrame(data)
Out[16]:
                                 0                                1                                2
0 2019-07-04 17:54:16.485999-07:00 2019-07-04 17:54:16.485999-07:00 2019-07-04 17:54:16.485999-07:00
1 2019-07-04 17:54:16.485999-07:00 2019-07-04 17:54:16.485999-07:00 2019-07-04 17:54:16.485999-07:00

In [18]: pd.__version__
Out[18]: '0.25.0rc0'

[jbrockmendel]

26825 fixed the DataFrame constructor error. Haven't looked at if maybe_infer_to_datetimelike needs separate attention