Operations between DataFrame and ExtensionArray loses EA dtype

[TomAugspurger]

What's the expected behavior here?

In [9]: a = pd.array([1, None, 2, 3], dtype='Int64')

In [10]: df = pd.DataFrame([[1, 2, 3, 4]])

In [11]: (a + df).dtypes
Out[11]:
0      int64
1    float64
2      int64
3      int64
dtype: object

Would we expect the eventual op to dispatch to the ExtensionArray, preserving the Int64Dtype for each?

I believe this is different from #28506

[jbrockmendel]

Would you be surprised if I told you that this would be easy to fix if we could reshape a to (1, N)?

I'm pretty sure a + df ends up going through dispatch_to_series, where it does something like:

new_data = {i: df.iloc[:, i] + a[i] for i in range(len(df.columns))}

Each a[i] is a scalar, so loses its dtype. If we could reshape to (1, N) this would instead be:

new_data = {i: df.iloc[:, i] + a[:, i] for i in range(len(df.columns))}

and each a[:, i] would be an IntegerArrray so the dtype would be retained.

[TomAugspurger]

Would it also be solved by doing a[[i]]? Then you have an (N,) Series and a (1,) array, which should broadcast fine.

[jbrockmendel]

Yah, I think that works for now while we operate column-wise, but stops working once we operate block-wise

[TomAugspurger]

Mmm, why's that? If an EA is involved, then we know that the input to ExtensionArray.__add__ is 1D, right?

[jbrockmendel]

If an EA is involved, then we know that the input to ExtensionArray.add is 1D, right?

In that scenario we're back to operating column-wise.