Empty series dtype discarded by reset_index with MultiIndex

[aquasync]

Code Sample, a copy-pastable example if possible

import pandas as pd

# this is fine
pd.DataFrame({'x': pd.Series([], dtype='datetime64[ns]'), 'y': pd.Series([]), 'z': pd.Series([])}).set_index(['x']).reset_index().dtypes

# this changes 'x' to float64
pd.DataFrame({'x': pd.Series([], dtype='datetime64[ns]'), 'y': pd.Series([]), 'z': pd.Series([])}).set_index(['x', 'y']).reset_index().dtypes

# same behaviour with all missing
pd.DataFrame({'x': pd.Series(['NaT'], dtype='datetime64[ns]'), 'y': pd.Series([2.]), 'z': pd.Series([3.])}).set_index(['x', 'y']).reset_index().dtypes

Problem description

DataFrame.reset_index shouldn't discard the type of the index/index levels when empty or all missing. It correctly handles this for a single level index (eg DatetimeIndex), but doesn't for MultiIndex.

In my case the symptom was a strange TypeError on a join, which I tracked down to the dtype of the series being lost after a reset_index.

The problem may be the branch in _maybe_casted_values, where values is overwritten with an np.empty() if all values are missing. I modified the condition in my copy to if len(mask) and mask.all() which avoids the problem for empty arrays, though not for all-missing.

Output of pd.show_versions()

I've only been able to run this with 0.24.2 so far as 0.25 fails to import on python 3.5.2 (due to this I think) but the relevant code in reset_index looks the same.

INSTALLED VERSIONS

commit: None
python: 3.5.2.final.0
python-bits: 64
OS: Linux
OS-release: 4.4.0-142-generic
machine: x86_64
processor: x86_64
byteorder: little
LC_ALL: None
LANG: en_AU.UTF-8
LOCALE: en_AU.UTF-8

pandas: 0.24.2
pytest: None
pip: 8.1.1
setuptools: 20.7.0
Cython: None
numpy: 1.17.0
scipy: None
pyarrow: None
xarray: None
IPython: 7.3.0
sphinx: None
patsy: None
dateutil: 2.8.0
pytz: 2019.2
blosc: None
bottleneck: 1.2.1
tables: None
numexpr: None
feather: None
matplotlib: 3.0.3
openpyxl: 2.6.2
xlrd: None
xlwt: None
xlsxwriter: 1.1.8
lxml.etree: 3.5.0
bs4: 4.4.1
html5lib: 0.999
sqlalchemy: 1.3.5
pymysql: 0.9.3
psycopg2: 2.6.1 (dt dec pq3 ext lo64)
jinja2: 2.10
s3fs: None
fastparquet: None
pandas_gbq: None
pandas_datareader: None
gcsfs: None

[TomAugspurger]

Does this look to be the same as #19602?

[aquasync]

Ahh indeed it is - not sure why it didn't come up in the related issues. Sorry for the noise.