Feature request: Create a Datatree instead of a Dataset when there are groups

[bzah]

Feature request

When I open a ncml that contains groups and sub-groups, the result is flatten: each variable is added to a single "root" Dataset and the initial hierarchy is lost.
This behavior is different from xarray open_dataset() which will only open the root group (but not subgroups) and, given the group="/data" kwarg, would only open the group "/data".

It would be nice if we could keep the hierarchy in either a DataArray encoding property (like the parent group path of the variable) or, even better, to output a Datatree which would contain all groups and subgroups.

Or should we wait until datatree is merged into xarray before considering this feature ?

Ref

• datatree : https://github.com/xarray-contrib/datatree
• open_dataset doc: https://docs.xarray.dev/en/stable/generated/xarray.open_dataset.html

[huard]

My objective is for open_ncml to behave as much as possible as open_dataset.

I don't think I've ever used netCDF groups, so I'll defer to your judgement regarding implementation.

[bzah]

My objective is for open_ncml to behave as much as possible as open_dataset.

Then we could add a group kwarg to open_ncml and only open the selected group, by default opening the root group just like open_dataset.

As I work with datasets with many layered groups, I would also be interested if we can keep the current behavior, maybe with a special flag argument all_group, default to False, which would order to open every group and add the paths in encoding or attrs of the DataArrays.

It might take a long time before Datatree is merged into xarray, hence the above proposal would be a satisfactory solution (for me) until it is merged.

• I'm willing to work on this (not now, but soon).