[FEA] Make Groupby.shift work for MultiIndex

[amanlai]

Say I have a MultiIndex Series as follows:

s = cudf.Series(range(4), index=cudf.MultiIndex.from_tuples([('a', 1), ('a', 2), ('b', 1), ('b', 2)], names=['f', 's']))

which looks like:

f   s
a  1    0
    2    1
b  1    2
    2    3
dtype: int64

I want to shift this Series down by the first index level:

s.groupby(level=0).shift(1)

but I get an error:

AttributeError: 'RangeIndex' object has no attribute '_names'

This should produce either

f   s
a  1    NaN
    2    0.0
b  1    NaN
    2    2.0
dtype: float64

or

f   s
a  2    0.0
b  2    2.0
dtype: float64

Right now, I can get my desired output by unstacking the Series, shifting it and then stacking it back:

s.to_frame().unstack('f').shift().rename(columns=lambda x: x[1]).stack()

which is very ugly. I would like cudf to extend groupby.shift to MultiIndex.

[shwina]

Thanks - this seems a bug with our implementation of .shift. Should be a relatively easy fix - we'll try to get to it ASAP!