issue_comments

I think that the explicit index refactoring would be an opportunity to clarify things a bit. With no concept of a dimension coordinate/ implicit index, ideally we could have the following behavior with a clear distinction between dimensions/coordinates and indexes, following more closely the "explicit is better than implicit" principle:

• set_index / reset_index would only affect the indexes without implicitly renaming any coordinate or dimension.
• actually, setting a new index for a given dimension doesn't make much sense anymore with the new data model. Setting a new index for a given set of variables and/or coordinates makes more sense
• trying to set an index from non-existing data variables or coordinates would raise a KeyError
• .reset_index("x") would not rename the x coordinate to x_

• .set_index(x="b") (where x is a dimension without index/coordinate and b is a variable with dimension x) would be replaced by a more explicit set_index("b").rename({"b": "x"}). Renaming b to x wouldn't even be necessary if we want to use the b coordinate as an index.

• rename_* would only affect the name of coordinates or dimensions (or the name of the DataArray) without implicitly adding or removing any index.

• .set_coords("b").rename({"b": "x"}) in the example above would not implictly create an index for the "x" coordinate

• swap_dims would only affect the existing dimension names without implicitly adding or removing any index and without renaming any coordinate.

This would be in an ideal world, though. More practically, I don't know how we could make a smooth transition.

There's also rename_dims, but only on a Dataset! There's a TODO to deprecate swap_dims in favor of that. Though I'm not sure that's right — swap_dims changes the dim values for another variable's, but rename_dims just changes the name?

swap_dims used as a workaround for switching from one to another coordinate as index on a dimension will no longer be needed after the explicit indexes refactoring. Should we eventually depreciate it, then?

if we do 2 I think we should add a new_dims kwarg to reset_index, otherwise swapping to a dimension without a coordinate is a lot more difficult.

I'm not sure exactly how to rationalize them though.

How about

1. rename_* should only allow renaming to unused names,
2. swap_dims should only work with existing names.
3. set_index should raise an error for the "renaming" and "swapping" cases.

There's also rename_dims, but only on a Dataset! There's a TODO to deprecate swap_dims in favor of that. Though I'm not sure that's right — swap_dims changes the dim values for another variable's, but rename_dims just changes the name?

I agree it's confusing, and I find myself looking up the correct way to do something fairly frequently.

I'm not sure exactly how to rationalize them though.

Without thinking this through in enough detail, Does set_ set the variable to the relevant type and reset_ unsets it to a normal variable? In that framework, I think the only aberrant is swap_dims, which could instead be set_dims?