Datatree does not inherit non-dimension coordinates

[erooke]

What is your issue?

When defining a data tree only dimension coordinates are inherited. For example

import numpy as np
import xarray as xr

N = 10


ds_1 = xr.Dataset(
    coords={
        "a": ("a", np.arange(N)),
        "a_meta": ("a", np.arange(N) * 3),
    }
)

ds_2 = xr.Dataset(data_vars={"data": ("a", np.arange(N) * 5)})

dt = xr.DataTree.from_dict(
    {
        "/": ds_1,
        "/data": ds_2,
    }
)

print(dt["/data"])

Results in:

<xarray.DataTree 'data'>
Group: /data
    Dimensions:  (a: 10)
    Inherited coordinates:
      * a        (a) int64 80B 0 1 2 3 4 5 6 7 8 9
    Data variables:
        data     (a) int64 80B 0 5 10 15 20 25 30 35 40 45

I could not find any mention in the docs that this is the expected behavior. The user guide does not make any sort of distinction and just refers to "coordinates", same with the data structures entry on inheritance.

I'm not sure if this is a bug or undocumented behavior but as a new user after reading the above docs I expected a_meta to be inherited.

System Info

• python version: 3.13.5
• xarray version: 2025.4.0

[erooke]

Of course after opening this issues I come across this note in the docs:

Xarray’s terminology differs from the CF terminology, where the “dimension coordinates” are called “coordinate variables”, and the “non-dimension coordinates” are called “auxiliary coordinate variables” (see GH1295 for more details).

I'm now reading the docs as if "coordinate variables" is referring to the "CF terminology" and not the "xarray terminology". Which, if true, I think makes this a documentation issue and not a bug.

[ianhi]

Coordinates, dimensions and their associated indices are propagated from downward starting from the root node to all descendent nodes.

I agree this is directly counter to the documented behavior. Re CF conventions, there are non-cf users of xarray so i think ideally we should follow xarray terminology.

I think the relevant line is this one which explicitly only inherits variables with an index:

xarray/xarray/core/datatree.py

Lines 580 to 588 in eed12c4

	@property
	def _coord_variables(self) -> ChainMap[Hashable, Variable]:
	# ChainMap is incorrected typed in typeshed (only the first argument
	# needs to be mutable)
	# https://github.com/python/typeshed/issues/8430
	return ChainMap(
	self._node_coord_variables,
	*(p._node_coord_variables_with_index for p in self.parents), # type: ignore[arg-type]
	)

[erooke]

I'm not too familiar with either xarray or CF so the following question is out of ignorance. Is there a reason for this restriction?

Edit, after some git history spelunking there is a reason for this restriction found in #9555 and #9475 (comment) so this is purely a documentation issue

I have a multiscale data set where one of the dimensions is participants in a study. I have a lot of metadata about the participants I would like to have available at the leaf nodes. Inheriting non-coordinate dimensions was the most obvious seeming solution. However there may be a more canonical way of handling such a case.

[TomNicholas]

Edit, after some git history spelunking there is a reason for this restriction found in #9555 and #9475 (comment) so this is purely a documentation issue

That's correct. The documentation definitely needs to be improved.

However there may be a more canonical way of handling such a case.

Unfortunately I don't think there is. You can add indexes to your metadata coordinates, but that's pretty clunky too.

cc'ing @shoyer in case he has any other ideas.