issue_comments: 210694521

Yes, my main concern is code bloat. Storing things like the computation graph and command line flags used to invoke a script are certainly useful things to do, and I use versions of this stuff all the time. But they are orthogonal to the labeled data focus of xarray so they belong better in another library.

If you want to take this approach, you might start by using something like dask.imperative, and extracting the task dependencies from the resulting task graph. Or you could even try to work with the full dask graphs created by using dask.array with xarray, but these can get pretty big.

Getting parameters from the calling script is even easier -- just inspect sys.argv and set it as an attribute before saving files (or write your own function). The main complexity here is picking a convention, not implementing it.