DataFrame.to_records() converts DatetimeIndex to datetime object, not np.datetime64

[jzwinck]

This code:

df = pd.DataFrame([pd.to_datetime('2017-11-07')], [pd.to_datetime('2017-11-08')])
df.to_records()

Produces a totally different column type for the index vs the data:

rec.array([(datetime.datetime(2017, 11, 8, 0, 0), '2017-11-07T00:00:00.000000000')], 
          dtype=[('index', 'O'), ('0', '<M8[ns]')])

It should produce:

rec.array([('2017-11-08T00:00:00.000000000', '2017-11-07T00:00:00.000000000')], 
          dtype=[('index', '<M8[ns]'), ('0', '<M8[ns]')])

I'm using Pandas 0.20.3, NumPy 1.13.1, and Python 3.5.3.

[jreback]

seems buggy
would appreciate a PR

[jzwinck]

@jreback I just discovered this was broken by #1908 (commit 38333f4) which added the to_records(convert_datetime64) option with the horrible default of True!

It looks like the option was added in order to work around some now-ancient NumPy bug. Setting convert_datetime64=False now works fine (for my test case and for the one in #1908).

What do you think about changing the default from True to False? The current default loses precision and is inefficient.

[jreback]

i would be in favor of deprecating that option entirely

[jzwinck]

Does that require a DeprecationWarning first, then later removal, or should we just remove the option at once?

[jreback]

FutureWarning