issue_comments

I want to apply a function over a given axis. This functionality is well defined and not covered by pipe (well, you can select the axis INSIDE the function passed, but it's bad practice because that means the function has an extra parameter to select the axis, not very KISS imo)

Not stale! Why not follow the pd convention of map vs apply for dataarray vs dataset?

In order to maintain a list of currently relevant issues, we mark issues as stale after a period of inactivity If this issue remains relevant, please comment here; otherwise it will be marked as closed automatically

Actually, I think that pipe should default to raw=False if there even is to be such a parameter. the reason is that one usually uses pipe to chain together functions, each of which usually expect a DataArray and "downcasting" to ndarray often breaks such a chain. If you insist on having one method behave as if raw=True, then I think it should be apply in order to be constsent with the python apply function which simply applies the function and nothing more.

@burnpanck I missed your comments from a few weeks ago. Yes, please do make PRs to update the docs.

I am OK with adding DataArray.apply for consistency and discoverability, even if it's basically an alias of DataArray.pipe. And if we add a raw argument I suppose it should probably default to raw=False for both .apply methods.

I think #1130 is related. I also think that apply is somewhat synonymous to pipe and is a lot more understandable for people without a pandas background. It would also be more consistent to have them both named the same on both Dataset and DataArray.

As for the consistency concern, I wouldn't have expected that to be a big issue. I'd argue that most functions mapping np.ndarray -> np.ndarray will not mind receiving a DataArray instead. On the other hand, functions mapping DataArray -> np.ndarray would seldom prefer to receive the raw np.ndarray. So I see no use to the raw parameter (but then again, I do not know pandas and their use-case), such that my hypotetical DataArray.apply and the existing DataArray.pipe are essentially the same.

Aha! For my use-case, DataArray.pipe is perfectly fine, I just didn't know about it. I have to admit that I know nothing about pandas. Before I learned about xarray, pandas was not interesting to me at all. My datasets are often high-dimensional which does not work well with pandas' orientation towards (one-dimensional) collections of observations. In that sense, I could rather relabel this issue (or create a new one) as a documentation problem. The API reference does not indicate the existence of DataArray.pipe at all (only Dataset.pipe, even though that one mentions it works on DataArrays too). Also, there could possibly be a see-also link to pipe from apply. Shall I have a go at a PR?

We could try to get in a simple version of this before #964.

My only concern is that the functionality here is slightly different from Dataset.apply, which takes functions that map DataArray -> DataArray or DataArray -> numpy.ndarray.

For DataArray.apply, we really want something that maps numpy.ndarray -> numpy.ndarray, which is now inconsistent. I guess we could pass the original DataArray to the function, but we already have a .pipe method for that.

This is also a concern for #964, because any new xarray.apply function would have similar consistency issues with Dataset.apply.

Two possible solutions, neither of which is fully satisfying: - New keyword argument raw to Dataset.apply, defaulting to False. Works like the raw argument to DataFrame.apply. If raw=True, then Dataset.apply passes unlabeled arrays to the provided function, like DataArray.apply. This makes the difference a little less jarring. - Pick a new name for one of these uses of apply, e.g., apply_raw for this use case. xarray.apply_raw or DataArray.apply_raw is pretty verbose, though.