Currently (v1.11.1.0), the treatment of HDF5 chunking is a bit inadequate:

• Chunking can only be set on a per-Data object basis
• Chunking can only be defined by explicitly setting the chunks shape on each axis
• Chunking is ignored in an output file unless native compression is on
• Chunks from an input file are not stored

A more comprehensive and flexible API is needed:

• cfdm.write should chunk by default, and have a keywork argument (hdf5_chunks) to configure the default chunking.
• cfdm.read should, by default, store HDF5 chunking on the returned data, so that it will be used when when writing out to a new netCDF4 file.
• Setting a HDF5 chunking strategy should be more intuitive. E.g. it should be easy to "chunk the time axis by 12 elements, leaving all other axes unchunked": f.nc_set_hdf_chunksizes({'T': 12})
• Setting HDF5 chunksizes follows the Dask API for defining its computaitonal chunk sizes. E.g. f.nc_set_hdf_chunksizes("8 MiB")

PR to follow.