cumsum sums the groupby column

[hayd]

It shouldn't sum the groupby'd col (in fact index col should be the index, if groupby as_index).

In [13]: df = pd.DataFrame([[1, 2, np.nan], [1, np.nan, 9], [3, 4, 9]], columns=['A', 'B', 'C'])

In [14]: g = df.groupby('A')

In [16]: g.cumsum()
Out[16]: 
   A   B   C
0  1   2 NaN
1  2 NaN   9
2  3   4   9

[3 rows x 3 columns]

Nature of it being dispatch. Should fix up for 0.14 possibly along with some other whitelisted groupby functions.

[jorisvandenbossche]

What would be the expected output? Something like this?:

In [29]: g.cumsum
Out[29]:
       B   C
  A
0 1   2 NaN
1   NaN   9
2 3   4   9

And should it then also be the case for cumcount if as_index=True?

[hayd]

@jorisvandenbossche RE the index of cumcount, possibly yes it should respect as_index... I think it's debatable if this would ever be desired though... the main problem however is it's slow (I don't think efficient way to append index to index to make MI) and this is the default. I had thought I had posted about this somewhere but can't find issue...

I think so, though like I say I think we need to have a discussion about as_index (there are at least three different ways used in groupby atm)... I had a partially filled in issue about it from a week or so ago... :s will look at it again after the weekend and try to post it. It's kinda a mess and some conventions are of dubious value (e.g. that of head)

[jorisvandenbossche]

Yes, you did :-) Here: #4646 (comment)
And it is indeed, for cumcount, maybe in some way more consistent to also return a MI, but I also think you mostly wouldn't want it.

[hayd]

I think should add some UserWarnings in 0.14 about this kind of behaviour, link to #5755.

[jreback]

@hayd you have anything in the works about this? push to 0.15 otherwise

[hayd]

I think I do, hope to get in the week.

[jreback]

ping!

[jreback]

I think this is closed by #7000, maybe just add a test?

[hayd]

Weirdly with the above example we don't have A as the index!

In [4]: df = pd.DataFrame([[1, 2, np.nan], [1, np.nan, 9], [3, 4, 9]], columns=['A', 'B', 'C'])

In [5]: g = df.groupby('A')

In [6]: g.cumsum()  # should have A as index
Out[6]:
    B   C
0   2 NaN
1 NaN   9
2   4   9

In [7]: g = df.groupby('A', as_index=False)  # this is correct

In [8]: g.cumsum()
Out[8]:
   A   B   C
0  1   2 NaN
1  2 NaN   9
2  3   4   9

[hayd]

Ah wait, this is a feature! Coool!

[jreback]

hmm...the index should have a named index (as A)...let me fix