A direct API for storing and loading from HDF5 files could make sense.
Alternatively, and actually more attractive for my use cases (e.g., hooking into niche file formats like netCDF), would be the ability to create a bolt array directly from an indexable object, similarly to dask.array.from_array.