[FEA] Add optional support for SQL style groupby null in key column handling

[randerzander]

Currently when you run a groupby, any nulls in the key column are dropped from the output:

import cudf

df = cudf.DataFrame()

df['id'] = [0, 1, 1, None, None, 3, 3]
df['val'] = [0, 1, 1, 2, 2, 3, 3]
df.groupby('id').val.sum()

id
0    0
1    2
3    6
Name: val, dtype: int64

This has the same behavior as Pandas, but sometimes it's desired to keep the null values as rows in the output. We should consider adding a dropna param which lets users control this behavior.

We can do this via a workaround with a sentinel and fillna and replace, but this is annoying to manage, and lowers overall workflow perf:

df['id'] = df['id'].fillna(-1)
res = df.groupby('id').val.sum().reset_index()
res['id'] = res['id'].replace(-1, None)
res

  | id | val
-- | -- | --
null | 4
0 | 0
1 | 2
3 | 6

EDIT - Changed suggested param from keep_nulls to dropna to match Pandas's API per Keith's comment.

[jrhemstad]

The C++ implementation already supports this behavior with the ignore_null_keys option:
https://github.com/rapidsai/cudf/blob/branch-0.10/cpp/include/cudf/groupby.hpp#L74

It's just a matter of exposing this in the Python API.

[kkraus14]

Looks like the Pandas community has had discussions surrounding this and the consensus was to use a dropna parameter: pandas-dev/pandas#21669

I'd argue we should do the same.