DEPR: groupby with as_index=False doesn't add grouper as column when passing a Series as group key

[jorisvandenbossche]

I don't know if this is strictly speaking a but (the documentation doesn't really specify it), but I would have expected a different result:

In [28]: df = pd.DataFrame({"a": [1, 2, 3], "b": [4, 5, 6]})

In [29]: key = pd.Series(['foo', 'foo', 'bar'], name="key")

In [30]: df.groupby(key).sum()
Out[30]: 
     a  b
key      
bar  3  6
foo  3  9

In [31]: df.groupby(key, as_index=False).sum()
Out[31]: 
   a  b
0  3  6
1  3  9

(running with current main, 2.0.0.dev0+532.gdec9be2a5c)

When your key is a column in the DataFrame, specifying as_index=False will move the group key values to a column in the resulting dataframe:

In [34]: df["key"] = key

In [35]: df.groupby("key", as_index=False).sum()
Out[35]: 
   key  a  b
0  bar  3  6
1  foo  3  9

I would have expected the same behaviour if you pass the key as a Series.

cc @rhshadrach

This is related to #49450 (comment), where I encountered this behaviour for the case where the Series is actually a column of the original DataFrame (when changing this Series to be a view, and not an identical object).

[jorisvandenbossche]

Hmm, this seems to be a duplicate of #40591, which is mentioning the user guide (I only looked in the reference docs, where this is indeed no mentioned)

[rhshadrach]

Assuming that the implementation/performance isn't a barrier here (I can't imagine why it would be), +1 on always including the groupers when as_index=False regardless if they are "in-axis" or not. Also, while #40591 may have preceded this, I would recommend closing that is as a duplicate of this as this one is more concise.

[rhshadrach]

I've confirmed the implementation is rather straightforward here, but the deprecation less so. It is easy to reliably emit a warning the behavior will change when more groupers will be included in the future, however there is no way for users to adopt future behavior early, preserve old behavior when the change is made, nor silence the warning.

I think we also can't call this purely a bug fix, it does appear in the documentation:

Passing as_index=False will return the groups that you are aggregating over, if they are named columns.

I wonder if this is a case to make a breaking change without deprecation. A few opinionated thoughts:

• It's hard to imagine preferring only certain types of groupers be included in the result and not others
• My experience is that most groupers are in-axis

[rhshadrach]

Actually, we can instruct users to use in-axis groupers in a deprecation warning. I'm in favor of deprecating this properly.