The current approach to load data from HDF5 is to load in an h5py or netcdf4 Dataset object and pass that to from_array. This is efficient, but also includes the file pointer within the dask graph. This fails if you need to serialize the graph, such as is necessary in distributed computing.

It might be wise to instead create an explicit da.from_hdf5 function that only stored the filename, datapath, and slice information within the graph.

It would also be nice if this API was exposed up to xarray (cc @shoyer).

cc @rabernat @pwolfram