Behavior of new df.agg, df.transform and df.apply is very inconsistent

[memeplex]

This looks really messy, it's not difficult to find out inconsistent behavior in all three methods. It seems to me that the dataframe implementation is too liberally blurring the distinction between agg, transform and apply.

In: df
Out: 
   x  y  z
0  1  1  4
1  1  2  5
2  2  3  6
3  2  1  4
4  3  2  5
5  3  3  6

In: gb = df.groupby('x')

1. dataframe.apply accepts lists and dicts while groupby.apply doesn't:

In: df.apply(['sum', 'mean'])
Out: 
         x     y     z
sum   12.0  12.0  30.0
mean   2.0   2.0   5.0

In: gb.apply(['sum', 'mean'])
TypeError    

2. dataframe.transform disallows aggregations while groupby.transforms broadcasts them to the original shape:

In: df.transform('sum')
ValueError: transforms cannot produce aggregated results

In: gb.transform('sum')
Out: 
   y   z
0  3   9
1  3   9
2  4  10
3  4  10
4  5  11
5  5  11

3. dataframe.agg allows non-aggregations while groupby.agg doesn't:

In: df.agg(lambda x: x)
Out: 
   x  y  z
0  1  1  4
1  1  2  5
2  2  3  6
3  2  1  4
4  3  2  5
5  3  3  6

In: gb.agg(lambda x: x)
ValueError: cannot copy sequence with size 3 to array axis with dimension 2

[gfyoung]

@memeplex : Thanks for reporting this! Just for reference, did you notice this "inconsistent" behavior in previous versions of pandas, or has this surfaced just now (you use the word "new" in the title)?

[gfyoung]

Regarding df.apply, we don't actually say that we accept lists of functions, so that behavior is somewhat accidental. Nevertheless, I don't see why we couldn't implement for groupby's?

Regarding df.transform, your groupby transform call should have worked according to documentation, so that looks like a bug to me.

Regarding df.agg, that looks like a bug to me as well. While lambda x: x is vacuously an aggregation, it should work for groupby as well AFAICT.

[memeplex]

I have tested it with 0.21 only but transform and agg were acquired by df in 0.20 so the behavior can't be older than that. It's not surprising given the fact that the current implementation mostly delegates everything to _aggregate.

I'm not sure allowing apply to receive a full mapper as part of its "public" signature is a good idea since the difference between the three methods is becoming almost impossible to describe (e.g. I don't see how the "more flexible" role the traditional groupby documentation assigns to apply is relevant anymore for dfs).