NetCDF4 data can be saved as chunks on disk, which has several benefits including efficient reads when using a compatible chunk shape. This is particularly important for files with chunk-based compression (ie all nc4 files with compression) or on HPC and parallel file systems (eg), where IO is typically dominated by the number of reads and chunks-from-disk are often cached. Caches are also common in network data backends such as Thredds OPeNDAP, in which case using disk-compatible chunks will reduce cache pressure as well as latency.

Xarray can use chunks, of course, but as of v0.9 the chunk size has to be specified manually - and the easiest way to discover it is to open the file and look at the _Chunksizes attribute for each variable. I propose that xr.open_dataset (and array, and mfdataset) change their default behaviour.

If Dask is available and chunks=None (the default), chunks should be taken from the file on disk. This may lead to a chunked or unchunked dataset. To force an un-chunked load, users can specify chunks={}, or simple .load() the dataset after opening it.