Tom Nicholas

@Zac-HD I think this is ready for another review (or a least a reply to the couple of unresolved comments)!

It might take me a little while to get through all this, but this is a great review, thanks again @Zac-HD !

@Zac-HD if I could request one more review please! The two remaining problems for me are: 1) How should we alter the API of `datasets` to make it easier to...

> I think the type Dict[str, xr.Datarray], where the keys are axis names, is sufficient. This is what is currently implemented. Alternatively, we could use an xr.Dataset. As they both...

> Instead, we need to have methods that accept and return vectors. @rabernat do you mean "we need to have additional methods that only accept and return vectors (e.g. `interp_vector`)",...

@jbusecke and I were just chatting and decided that we like the idea of using an API like this for acting on vectors `grid.interp({"X", da_u}, axis="X", other_component={"Y", da_v})`, where the...

This API would also need to be propagated down to `grid.pad`, because we will need the option to pad variable `a` using a piece from variable `b`. This doesn't have...

@rabernat when you say this > The alternative, closer to what we have today in xgcm, is to have separate methods for vectors, as we do today, e.g. diff_2d_vector and...

I think make a new page for vectors.

Data Classes are a good idea. I've been looking for an excuse to try them out and I think the grid class is simple enough for them to make sense....