[SciPy-dev] time series implementation approach

[Pierre GM]

> implementation could be to make a callback to python, but I'm pretty sure
> Travis has something in the API that would allow this directly from the C
> code (PyUFunc_O_O_method ?)

David,
I think it'd be easier if the C grouping function send back an array that you 
would have to process, for several reasons:
- you have full control on the postprocessing
- you don't have to group your data each time you need a new postprocessing
- masked data are very easy to deal with in python, and I can only assume that 
they're more difficult to deal with in C.

Suppose you have your daily series that you resampled to monthly: with this 
approach, you just group data once, and then can compute your mean and 
variance on the groups. you need a median ? Well, your data's already 
grouped. There were gaps in your data ? Well, you have masked values.

Matt pointed quite justly that this approach doesn't work for some 
pathological cases, but that shouldn't be that common...