groupby with multiindex behaves differently for series and single-column dataframe

[mcwitt]

This behavior seems strange to me. Starting with a multiindex dataframe with unbalanced levels:

In [3]: pd.__version__
Out[3]: '0.15.2'

In [4]: df = pd.read_csv(StringIO('''
A,B,x
three,a,1
three,b,2
three,c,3
two,a,4
two,b,5
one,a,6'''), index_col=list('AB'))

In [5]: df
Out[5]: 
         x
A     B   
three a  1
      b  2
      c  3
two   a  4
      b  5
one   a  6

Groupby aggregations on this dataframe seem to revert the index to the cross product of the levels, potentially leaving many NAs in the result:

In [6]: df.groupby(level=[0,1]).mean()
Out[6]: 
          x
A     B    
one   a   6
      b NaN
      c NaN
three a   1
      b   2
      c   3
two   a   4
      b   5
      c NaN

But with the series this doesn't occur (no NAs in the result):

In [7]: df.x.groupby(level=[0,1]).mean()
Out[7]: 
A      B
one    a    6
three  a    1
       b    2
       c    3
two    a    4
       b    5
Name: x, dtype: int64

I'm wondering if this is a bug or intended behavior? I haven't been able to find any mention of different behavior in the docs.

[mcwitt]

Related #3835, but looks like the cartesian product index was ruled out as an option there.

[jreback]

this was fixed by #9177, which is in 0.16.0 (released today).

In [3]: df.groupby(level=[0,1]).mean()
Out[3]: 
         x
A     B   
one   a  6
three a  1
      b  2
      c  3
two   a  4
      b  5

In [4]: df.x.groupby(level=[0,1]).mean()
Out[4]: 
A      B
one    a    6
three  a    1
       b    2
       c    3
two    a    4
       b    5
Name: x, dtype: int64