divide by 0 for Int64 is inconsistent with int64

[hhuuggoo]

In [5]: pd.Series([1], dtype='Int64') / pd.Series([0], dtype='Int64')                                                                                                                                                                                                           
Out[5]:                                                                                                                                                                                                                                                                         
0   NaN                                                                                                                                                                                                                                                                         
dtype: float64     

Where as

In [6]: pd.Series([1], dtype='int64') / pd.Series([0], dtype='int64')                                                                                                                                                                                                           
Out[6]:                                                                                                                                                                                                                                                                         
0    inf                                                                                                                                                                                                                                                                        
dtype: float64        

It's because of this line:

pandas/pandas/core/arrays/integer.py

Line 545 in 0480f4c

mask |= (result == np.inf) | (result == -np.inf)

Is it intentional for the new extension integer type to have different behavior here?

[WillAyd]

These should align though I think the extension type actually handles this more correctly

[jorisvandenbossche]

I am not sure we should change the behaviour of Series with default numpy types, as we then would deviate from numpy behaviour.

Of course for the extension Int type, we have better handling of NA compared to numpy int, so which could be a reason to actually use that here. But, since division will always result in a floating dtype (and never integer), that's not a direct benefit in this case.

We have a pd.options.mode.use_inf_as_na so people can make this in practice to behave as NA, but that is False by default.

[jbrockmendel]

closed by #30183