Coordinates.dims: only include dimensions found in coordinate variables.

[benbovy]

• Closes ds.coords.dims can display dims not present in the coordinate vars #9466 (leaving the addition of a "dimensions" section to the Coordinates repr for a follow-up PR)
• Tests added
• User visible changes (including notable bug fixes) are documented in whats-new.rst
• New functions/methods are listed in api.rst

Two comments:

• To me it looks rather like a bug fix. Do others view this as a breaking change?
• Dataset.coords.dims, DataArray.coords.dims and DataTree.coords.dims now all calculate the dimensions on the fly (no cache), which may have an impact of performance. Not sure those properties are heavily used, though (not much reports apart from ds.coords.dims can display dims not present in the coordinate vars #9466).

[TomNicholas]

The test failures here look real?

[TomNicholas]

Suggested change

	def test_coords_dims(self) -> None:
	# https://github.com/pydata/xarray/issues/9466
	def test_coords_dims(self) -> None:

[TomNicholas]

Should probably also add a similar test to the one you used for DataArray here too

[benbovy]

The test failures here look real?

This is caused by this line

xarray/xarray/core/dataarray.py

Line 453 in 6e9414d

dims = getattr(data, "dims", getattr(coords, "dims", None))

where DataArray dimensions, if not explicitly set via the constructor argument, may be ultimately set from coords.dims. However, if I replace getattr(coords, "dims", None) by None, other tests fails.

This is why this PR may be considered as both a bug fix and a breaking change. Apparently quite a few tests (and maybe 3rd-party code?) rely on the dimensions of the object wrapped by the Coordinates proxy object (passed via the coords argument) to set the dimension names of the new DataArray.

On the main branch:

>>> import numpy as np
>>> import xarray as xr
>>> da = xr.DataArray([[1, 2], [3, 4]], coords={"x": [0, 1]}, dims=("x", "y"))
>>> values = np.array([[10, 20], [30, 40]])

Creating a new DataArray by directly passing the Coordinates proxy object (success, dimension name "y" magically discovered from da.coords.dims pointing to da.dims!!):

>>> coords = da.coords
>>> xr.DataArray(values, coords=coords) 
<xarray.DataArray (x: 2, y: 2)> Size: 32B
array([[10, 20],
       [30, 40]])
Coordinates:
  * x        (x) int64 16B 0 1
Dimensions without coordinates: y

Creating a new DataArray by passing a new "standalone" Coordinates object (error, the coordinates have only one "x" dimension):

>>> coords = xr.Coordinates(da.coords)
>>> xr.DataArray(values, coords=coords)
ValueError: different number of dimensions on data and dims: 2 vs 1

This seems very confusing to me and I think that we should deprecate the succeeding behavior above.