NaN is converted to strings when reassigning a column with .loc

[remidomingues]

Hello team,

Here is a simple working exemple of value assignment on a dataframe:

df = pd.DataFrame({'A': [np.nan, np.nan, 'b']})
df['A'].loc[[0, 1]] = ['a', np.nan]
df.notna()
Out[126]: 
       A
0   True
1  False
2   True

However, if we use loc to assign the entire column with an array including NaNs, all values are converted to strings, breaking notna():

df = pd.DataFrame({'A': [np.nan, np.nan, 'b']})
df['A'].loc[[0, 1, 2]] = ['a', np.nan, np.nan]
df['A'].notna()                                                                                                                                                                                   
Out[133]: 
0    True
1    True
2    True
Name: A, dtype: bool

[TomAugspurger]

Thanks for the report, probably a bug.

Note that this is following the numpy behavior

In [23]: np.array(['a', np.nan])
Out[23]: array(['a', 'nan'], dtype='<U3')

I suspect that when df['A'] is presently an object-dtype column, we don't convert the incoming list-like with the right sanitizer.

Note that this wouldn't be a problem with a proper StringDtype: #27949

[jreback]

#28176 might fix this actuallly

[jbrockmendel]

Works on master, could use a test.

[NumberPiOso]

take