"on" argument of DataFrame.rolling only works for datetime columns

[bchu]

the on= argument of DataFrame.rolling only works for datetime columns.

df = pd.DataFrame([
    [18, 0],
    [2, 0],
    [1, 0],
    [9, 1],
    [8, 1],
], columns=['value', 'roll'])

df.roll = pd.to_datetime(df.roll, unit='s')
df.rolling('1s', on='roll').value.max()

returns:

0    18.0
1    18.0
2    18.0
3     9.0
4     9.0
Name: value, dtype: float64

as expected.

But

df.rolling(1, on='roll').value.max()

returns:

0    18.0
1     2.0
2     1.0
3     9.0
4     8.0
Name: value, dtype: float64

If this is intentional behavior, I'd be happy to change the docs to note this (the docs currently imply that on= can be used for any column).

[gfyoung]

cc @jreback @jorisvandenbossche

[WillAyd]

Is this just confusion over the size of the window being a frequency versus a fixed size? Both of these results look correct to me

[jreback]

looks ok to me. I think your confusion is that when you have an integer argument basically on doesn't do anything as the roll does not use the index.

In [9]: df.set_index('roll').rolling(2).value.sum()                                                                                                                                                                                                          
Out[9]: 
roll
0     NaN
0    20.0
0     3.0
1    10.0
1    17.0
Name: value, dtype: float64

In [10]: df.rolling(2, on='roll').value.sum()                                                                                                                                                                                                                
Out[10]: 
0     NaN
1    20.0
2     3.0
3    10.0
4    17.0
Name: value, dtype: float64