Deepak Cherian

There are a bunch of errors if I run mypy in `flox/` A lot seem easy and some seem like bugs =) ``` flox/core.py:69: error: Value of type "Optional[Any]" is...

Closes #80

This could work and be a lot faster. Using pd.cut with an IntervalIndex seems really slow. We might also want to create our own minimal "Index" structure to represent bin-edges

performance benchmarking

1

`bincount`-ed sum is 6x slower than `sum`, and `bincount`-ed `count` is 2x slower than `count` so there are definitely cases where it makes sense to split the dataset early instead...

construct a cost function for the dask graph when evaluating strategies

2

@martindurant's (and @jbednar) idea is that we can examine the dask graph to judge which strategy works best...

How does pandas do it? Or I guess we can just expect the user to pass in numba-decorated functions.

We could support unknown groups in the xarray interface if the user provides a size for "expected groups". The returned xarray object will have an unindexed "groups" coordinate and a...