DataFrame.sum() with mix of timedelta and IntegerNA returns NA for IntNA

[TomAugspurger]

In [43]: df = pd.DataFrame({"A": pd.to_timedelta([0, 1, None]), "B": pd.array([1, 2, None], dtype='Int32')})

In [44]: df.sum()
Out[44]:
A   00:00:00.000000
B               NaT
dtype: timedelta64[ns]

cc @jbrockmendel

Not a blocker for the RC, but is a blocker for the release.

[TomAugspurger]

Looking into this now. It's kinda messy, but I think there are maybe two bugs.

In [4]: df = pd.DataFrame({'A': pd.Series([0, None], dtype='timedelta64[ns]'), 'B': pd.array([0, None], dtype='Int64')})

In [5]: df.sum()
> /Users/taugspurger/sandbox/pandas/pandas/core/apply.py(306)wrap_results()
-> if len(results) > 0 and is_sequence(results[0]):
(Pdb) results
{0: Timedelta('0 days 00:00:00'), 1: nan}

That looks fishy. results[1] (for the IntNA column) should be 0, not NaN. If we fix that, we'd fix the output. But that would leave a second buggy case where you have all-NA.

(Pdb) n
> /Users/taugspurger/sandbox/pandas/pandas/core/apply.py(311)wrap_results()
-> result = self.obj._constructor_sliced(results)
(Pdb) n
> /Users/taugspurger/sandbox/pandas/pandas/core/apply.py(312)wrap_results()
-> result.index = self.res_index
(Pdb) result
0   0 days
1      NaT
dtype: timedelta64[ns]

I think we can fix this after the RC.

[TomAugspurger]

@jreback OK to move off 1.0? You added it back in July.

[mroeschke]

This looks fixed on master. Could use a test

In [1]: In [43]: df = pd.DataFrame({"A": pd.to_timedelta([0, 1, None]), "B": pd.
   ...: array([1, 2, None], dtype='Int32')})
   ...:
   ...: In [44]: df.sum()
Out[1]:
A    0 days 00:00:00.000000001
B                            3
dtype: object