API: Series/DataFrame logical op NaN propagation

[sinhrks]

The same as #13806, but for discuss differences with DataFrame.

pd.Series([True, False, np.nan]) | pd.Series([True, False, True])
#0     True
#1    False
#2    False
# dtype: bool

pd.Series([True, False, True]) | pd.Series([True, False, np.nan])
#0     True
#1    False
#2     True
# dtype: bool

pd.DataFrame([True, False, np.nan]) | pd.DataFrame([True, False, True])
#        0
#0   True
#1  False
#2    NaN

pd.DataFrame([True, False, True]) | pd.DataFrame([True, False, np.nan])
#        0
#0   True
#1  False
#2    NaN

There looks to be 2 points:

1. NaN comparison result

As @TomAugspurger described in #13806, the basic rule is below. Based on this, DataFrame result is incorrect.

In [40]: float('nan') or 1
Out[40]: nan

In [41]: 1 or float('nan')
Out[41]: 1

2. filling NaN result

Series fills NaN included in the result with False. DataFrame should also?

[mroeschke]

This looks fixed on master. Could use a test:

In [96]: pd.Series([True, False, np.nan]) | pd.Series([True, False, True])
    ...:
Out[96]:
0     True
1    False
2    False
dtype: bool

In [97]: pd.Series([True, False, True]) | pd.Series([True, False, np.nan])
    ...:
Out[97]:
0     True
1    False
2     True
dtype: bool

In [98]: pd.DataFrame([True, False, np.nan]) | pd.DataFrame([True, False, True])
    ...:
Out[98]:
       0
0   True
1  False
2  False

In [99]: pd.DataFrame([True, False, True]) | pd.DataFrame([True, False, np.nan])
    ...:
Out[99]:
       0
0   True
1  False
2   True

In [100]: pd.__version__
Out[100]: '0.26.0.dev0+593.g9d45934af'