issue_comments: 704631854

I see this as complementary to @alexamici's proposal in https://github.com/pydata/xarray/issues/4309#issuecomment-697952300, which I also like. I guess the main difference is moving decode_cf into the explicit signature of open_dataset, rather than leaving it in **kwargs.

Auto-completion is one reason to prefer a class, but enforced error checking and consistency between backends in the data model are also good reasons. In particular, it is important that users get an error if they mis-spell an argument name, e.g., open_dataset(path, decode_times=False) vs open_dataset(path, decode_time=False). We can definitely achieve this with putting decoding options into a dict, too, but we would need to be carefully to always validate the set of dict keys.

I guess cfgrib is an example of a backend with its own CF decoding options? This is indeed a tricky design question. I don't know if it's possible to make xarray.open_dataset() directly extensible in this way -- this could still be a reason for a user to use a backend-specific cfgrib.open_dataset() function.