BUG: concat of all-nan with empty frame produces object dtype

[jreback]

xref #9149

The resultant has all missing and when concat with an empty frame should result in the same dtype (and not promotion)

In [1]: df_1 = pd.DataFrame({"Row":[0,1,1], "EmptyCol":np.nan, "NumberCol":[1,2,3]})

In [2]: df_2 = pd.DataFrame(columns = df_1.columns)

In [3]: df_concat = pd.concat([df_1, df_2], axis=0)

In [4]: df_1.dtypes
Out[4]: 
EmptyCol     float64
NumberCol      int64
Row            int64
dtype: object

In [6]: df_2.dtypes
Out[6]: 
EmptyCol     object
NumberCol    object
Row          object
dtype: object

In [7]: df_concat.dtypes
Out[7]: 
EmptyCol      object
NumberCol    float64
Row          float64
dtype: object

[tvyomkesh]

Ah just noticed (via xref), that this may be already being looked into by someone else. I will freeze this work here. Sorry for the distraction.

[TomAugspurger]

Just a note that this is handled correctly for Series:

In [18]: a = pd.Series(pd.date_range('2000', periods=4, tz='UTC'))

In [19]: b = pd.Series([])

In [20]: pd.concat([a, b]).dtype
Out[20]: datetime64[ns, UTC]

In [21]: pd.concat([a.to_frame(), b.to_frame()]).dtypes
Out[21]:
0    object
dtype: object

So Out[20] is correct, but the DataFrame version is not.

[mroeschke]

This looks correct from Tom's example now. Could use a test

In [10]: In [18]: a = pd.Series(pd.date_range('2000', periods=4, tz='UTC'))
    ...:
    ...: In [19]: b = pd.Series([])
<ipython-input-10-74472f1f757a>:3: DeprecationWarning: The default dtype for empty Series will be 'object' instead of 'float64' in a future version. Specify a dtype explicitly to silence this warning.
  b = pd.Series([])

In [11]: pd.concat([a.to_frame(), b.to_frame()]).dtypes
Out[11]:
0    datetime64[ns, UTC]
dtype: object

In [12]: pd.__version__
Out[12]: '1.3.0.dev0+1287.g895f0b4022'