ENH: groupby(...).agg should only accept reducers

[rhshadrach]

Issues where agg is used with non-reducing functions:

• Behavior of new df.agg, df.transform and df.apply is very inconsistent #18103
• groupby agg with rank and parameter return does not reduce #14741
• BUG: DataFrame.agg with multiple cum functions creates wrong result #35490

I think agg should raise if the function(s) provided are not reducers. This can be tested by if the resulting index is equal to self.grouper.result_index. always reduce regardless of the return value. That is, treat the result of the UDF as if it a scalar (even when it's not).

[WillAyd]

I would be in favor of this - there's a lack of clarity in the API around when agg and apply should or shouldn't be used.

@jreback

[jreback]

yes absolutely agg should error out if the returned shape is incorrect - iow it should be strict (likely can only deprecate for now though)

similarly transform should be strict (i think it is actually)

[rhshadrach]

take

[rhshadrach]

@jreback @WillAyd - Rather than raising, what do you think about instead making it so that agg always aggs regardless of the result type. Some cases:

• DataFrame.agg: always produces a Series whose index is the column labels. It could be that the values of the Series are Series/DataFrames themselves.
• Series.agg: Always produces a single result, but can be a Series/DataFrame.
• groupby(...).agg: with as_index=True, index is always the groups, precisely one row per group. With as_index=False would be the same, but columns instead of index.

This has the benefit of simplifying/clarifying the behavior of agg, and I think makes it more intuitive, while not forbidding a user from getting e.g. a Series full of DataFrames if they so desire.