issue_comments

I might be wrong, but I think you should be able to implement this by overriding Variable.values.setter in IndexVariable, the same way it is done for Variable.data.setter.

I think what we probably want to do is mark variables used in indexes as having immutable data in xarray's data model, and raise an error for attempts to modify them in-place.

How do I implement this?

The main problem is that when you write ds.time.values = something, you are modifying an attribute of a DataArray that happens to use the same variable as the as the Dataset, but otherwise the DataArray and Dataset are entirely independent. The DataArray does not have a reference to the original dataset, so it can't check or update the indexes.

This happened to work in previous versions of xarray, but only because indexes always mapped directly into the data of an xarray.Variable. But I don't think this is the right pattern to continue in the future, especially because we want to do define indexes over multiple variables (e.g., for MultiIndex).

I think what we probably want to do is mark variables used in indexes as having immutable data in xarray's data model, and raise an error for attempts to modify them in-place.

Is it already decided what the resolution should be? * Giving a warning, as the title of this thread suggests? * Disable setting .values directly for dimensions? * Or making sure that .indexes are updated when .values are set directly

@dcherian Yes, ds.assign_coords does work and that's what I'm using now.

So can ds.assign_coords(..) be seen as the right way and changing the values inplace with ds.time.values = pd.... the wrong way? It all worked fine until changing from version 0.11.3 to 0.12.0.

I've seen this type of code elsewhere: direct assignment to .values and .data of a coordinate variable which then leads to inconsistent objects and confusing behaviour. Is there something we can do to alert users?

Should we raise an error? Is there ever a use here? It might be messy to implement though; the ds.time would have to know it's a coord. Does the bigger index refactor help us at all here?

ds.time.values = pd.to_datetime(ds.time.values) + pd.DateOffset(months=2) By doing this, ds.indexes["time"] is not updated.

@folmerkrikken How about ds.assign_coords(time=pd.to_datetime(ds.time.values) + pd.DateOffset(months=2))

@shoyer I've seen this type of code elsewhere: direct assignment to .values and .data of a coordinate variable which then leads to inconsistent objects and confusing behaviour. Is there something we can do to alert users?