This would need to eagerly evaluated, or perhaps return an imperative value, because the result shape is not known apriori.

This sort of thing could be useful for pydata/xarray#815 if you want to mask out a small region of a very large array. For example, imagine you have a tiled digital elevation model dataset at high resolution covering the globe, and now you want to extract out the region corresponding to California. There are certainly more intelligent indexing strategies for geospatial data, but I could see something like ds.sel_where(ds.region_enum == CALIFORNIA) being convenient.