BUG: Mixed DataFrame with Extension Array incorrect aggregation

[WillAyd]

Surprised to not get an aggregation for "b" with the Int64 dtype

>>> df = pd.DataFrame([["a", 1]], columns=list("ab"))
>>> df.sum()
a    a
b    1
dtype: object
>>> df.astype({"b": "Int64"}).sum()
a    a
dtype: object

[jorisvandenbossche]

This is actually a regression on master, as this "worked" until recently:

In [6]: df.astype({"b": "Int64"}).sum() 
Out[6]: 
a    a
b    1
dtype: object

In [7]: pd.__version__ 
Out[7]: '1.0.3'

There are still many things not correctly working for DataFrame reductions with EA columns, which I am trying to fix in #32867

Now, the specific regression here is related to nansum / maybe to the fact that we now added an IntegerArray.sum method:

In [8]: pd.core.nanops.nansum(pd.Series(pd.array([1])))
...
ValueError: the 'dtype' parameter is not supported in the pandas implementation of sum()

This works in pandas 1.0.3, because we just convert the IntegerArray to a numpy array there, but apparently we stopped doing that.
That's maybe something to fix on its own, but for the DataFrame reduction, the actual fix is to never call nansum, but to call the actual IntegerArray-specific reduction (which is what #32867 is trying to do)

[simonjayhawkins]

This bug only occurs for numeric_only=None, the default.

>>>
>>> df.astype({"b": "Int64"}).sum(numeric_only=True)
b    1
dtype: int64
>>>
>>> df.astype({"b": "Int64"}).sum(numeric_only=False)
a    a
b    1
dtype: object
>>>

[simonjayhawkins]

This works in pandas 1.0.3, because we just convert the IntegerArray to a numpy array there, but apparently we stopped doing that.

it appears #32950 caused the regression

b838508 is the first bad commit
commit b838508
Author: jbrockmendel jbrockmendel@gmail.com
Date: Mon Mar 30 15:43:04 2020 -0700

REF: move mixed-dtype frame_apply check outside of _reduce try/except (#32950)