support dtype kwarg for grouped reductions

[dcherian]

No description provided.

[max-sixty]

Is .sum(dtype=foo) equivalent to calling .astype(foo) on the input? On the result?

[dcherian]

No, dtype is the dtype of the accumulator, so for example, you might want to use dtype=np.float64 to more accurately reduce np.float32 values.

Here's a PR with useful tests and discussion: xarray-contrib/flox#131

[max-sixty]

I see. I think this is possible, though not trivial — currently we have a list of candidate signatures which numba will JIT compile for. We could generate new ones at runtime, but I'm not sure how numba handles a many-to-one mapping of inputs-to-outputs; so possibly we'd need to compile a new function...

[dcherian]

Ah I think this is lower priority. We can easily raise if dtype is provided and ask them to use another "engine"

[max-sixty]

We can easily raise if dtype is provided and ask them to use another "engine"

Yes or to convert first (albeit requires a copy)

[max-sixty]

One question out of #128:

To what extent do we want to convert types? Currently we always have the same input type as output type... ...except for bool which we convert to ints (ref #128 )

But often that's not what we want — e.g. it is useful to have a float mean of ints, rather than an "int mean" (sum and then integer division).

A few choices:

• Always keep types — current state (except for bools...)
• Always convert to float64 — pandas' approach IIUC
• Allow passing a dtype arg — great but requires some work

Unless folks disagree, I think we'll stay at the current state and see if we get traction and folks ask for a change...

[dcherian]

To what extent do we want to convert types?

I think we can just ask the user to do so...

I think we'll stay at the current state and see if we get traction and folks ask for a change

IMO this is OK.

A few choices:
Always keep types — current state (except for bools...)
Always convert to float64 — pandas' approach IIUC
Allow passing a dtype arg — great but requires some work

Perhaps @shoyer has an opinion