apply_ufunc and Datasets with variables without the core dimension

[dcherian]

Is your feature request related to a problem?

Consider this example

ds = xr.Dataset({"a": ("x", [1, 2, 3]), "b": ("y", [1, 2, 3])})
xr.apply_ufunc(np.mean, ds, input_core_dims=[["x"]])

This raises

ValueError: operand to apply_ufunc has required core dimensions ['x'], but some of these dimensions are absent on an input variable: ['x']

because core dimension x is missing on variable b. This behaviour makes it annoying to use apply_ufunc on Datasets.

Describe the solution you'd like

Add a new kwarg to apply_ufunc called missing_core_dim that controls how to handle variables without all input core dimensions. This kwarg could take one of two values:

1. "raise" - raise an error, current behaviour
2. "copy" - skip applying the function and copy the variable from input to output.

Describe alternatives you've considered

No response

Additional context

No response

[headtr1ck]

I would like to have a "ignore" option as well, such that I can do input_core_dims=[["x","y"]] in your example.

[dcherian]

@shoyer what do you think of adding such a kwarg?

[dcherian]

The logic in Dataset.reduce is a bit nuanced, it keeps coordinate variables but runs the function on variables without core dimensions.

https://github.com/pydata/xarray/blob/e2b6f3468ef829b8a83637965d34a164bf3bca78/xarray/core/dataset.py#L6740-L6751

[shoyer]

Belatedly -- I like this idea!

[max-sixty]

Check out #8138 for a draft impl!