Interestingly changing the return type of .dims to a set breaks several tests, which seem to rely on dimension order. These tests should be changed, as they should never have relied on dimension order, and it might be easier to change them if we had #5733 implemented.

This feels a bit quick to me because this change had a pretty big blast radius. Shall we wait till September or so?

Happy to wait.

I just noticed that in the docstring for Dataset.transpose it says

the dataset dimensions themselves will remain in fixed (sorted) order.

Does that mean that returning a frozenset is not correct? And I should change it to return a sorted tuple instead??

Does that mean that returning a frozenset is not correct? And I should change it to return a sorted tuple instead??

I think we've discussed it before and decided that the unordered nature of set was a good hint to users to not depend on the ordering for anything.

Okay that's what I would want too.

But that does mean that Dataset.transpose will not change Dataset.dims... It can at most change the dims of the Variables wrapped by that Dataset.

But that does mean that Dataset.transpose will not change Dataset.dims... It can at most change the dims of the Variables wrapped by that Dataset.

This sounds fine to me. You can't rely on Dataset.dims for a dimensional ordering in the general case anyway.