Arbitrary fill value

[hameerabbasi]

As requested in #1.

[hameerabbasi]

@shoyer How important is this to XArray? While I find that implementing fill values is trivial algorithmically, a few lines go into every single function in the whole library, and may lead to a lot of corner cases and bugs. I'm debating whether it's worth it to implement this in practice.

[shoyer]

I do think that sparse arrays in xarray be a very interesting alternative to using MultiIndex in pandas, especially when coupled with tools like dask for scaling to large datasets.

For this to work in any reasonable way, support for NaN/NA as a fill-value will be much pretty essential.
Most xarray operations automatically insert or skip NaN values (in data alignment/aggregations), so NaN is much more natural and useful fill-value than zero. Pandas's sparse data-structures support flexible fill-values defaulting to NaN for this same reason.

For computation oriented use-cases, I'm sure sparse would get a bit of use with xarray even if fill-values were always zero. But for analytics (which I think is the most exciting use-case), fill-values of NaN would be pretty much essential.