[REVIEW] Fix bug in dask-cudf caused by inconsistency with Pandas empty groupby result

[thomcom]

For some reason I decided to throw an exception when the groupby result was empty. This doesn't correspond with Pandas, which returns an empty DataFrame containing only the non by columns. This feature is important to dask-cudf in metadata relating to the groupbys. @ayushdg proposed the below solution which gives 100% test passes in dask-cudf and also cudf.

This should be considered a pre-release bugfix to MultiIndex and should go into branch-0.7 today if it checks out.

Fixes rapidsai/dask-cudf#231

[mrocklin]

For the future, I tend to parametrize these as follows:

@pytest.mark.parametrize('func', [
    lambda df: df.groupby(['x', 'y']).sum(),
    ...
])
def test_empty_groupby(func):
    pdf = pd.DataFrame({'x': [], 'y': []})
    gdf = cudf.from_pandas(pdf)
    assert_eq(func(pdf), func(gdf))

Just a style thing though. Nothing that should slow down this PR.

[mrocklin]

@thomcom is there someone we can ping to take a look at and review this?

[mrocklin]

@thomcom is there someone we can ping to take a look at and review this?

Maybe @shwina ?

[rlratzel]

@thomcom any idea about these failing CI checks? check-style should at least be an easy fix.

[thomcom]

It isn't ready

[ayushdg]

Got the slice issue working. Not sure if I'm doing the right thing though. Can discuss this tomorrow @thomcom .
multiindex.txt

[thomcom]

Can't wait to see your solution @ayushdg

[shwina]

@thomcom is there someone we can ping to take a look at and review this?

Maybe @shwina ?

Happy to review when this is ready

[thomcom]

Closing this in favor of @ayushdg's #1659. Tons of credit go to @quasiben and @ayushdg.