ENH: add groupby(..).sort_values()

might be useful to expose this as a top-level groupby method for intra-group sorting. Anyone have a good usecase for this? We could make this much faster that the current impl (not that this is an issue, though potentially it could be).

In [10]: df = DataFrame({'A': [1, 1, 2, 2, 2], 'B': [2, 1, 3, 3, 1], 'C':[1, 1, 2, 1, 1]})

In [11]: df
Out[11]: 
   A  B  C
0  1  2  1
1  1  1  1
2  2  3  2
3  2  3  1
4  2  1  1

In [12]: df.groupby('A').apply(lambda x: x.sort_values(['B', 'C']))
Out[12]: 
     A  B  C
A           
1 1  1  1  1
  0  1  2  1
2 4  2  1  1
  3  2  3  1
  2  2  3  2

[jacquespeeters]

Because the sum of all sorting’ complexity inside groups is lower than the sorting’ complexity on the whole dataset. groupby complexity doesn’t change.

There is a huge performance difference when dataset is big, and groups small. It is right now slower, not because of the complexity, but because of apply().

A good way to compare both operations complexity would be :

df.groupby('story_id').apply(lambda x: x.sort_values(by = 'relevance', ascending = False))
df.apply(lambda x: x.sort_values(by = 'relevance', ascending = False)).groupby('story_id')

[jorisvandenbossche]

And again, why don’t you use df.sort_values(['story_id', 'relevance']) ?