Uses `xarray.core.ops` which is not part of the public xarray API and will vanish

[mikapfl]

Describe the bug

Using the latest upstream (not yet released) version of xarray, we get the error AttributeError: module 'xarray.core' has no attribute 'ops' from various parts of primap2.

Failing Test

 FAILED primap2/tests/csg/test_compose.py::test_compose_simple - AttributeError: module 'xarray.core' has no attribute 'ops'
FAILED primap2/tests/csg/test_compose.py::test_compose_exclude_result - AttributeError: module 'xarray.core' has no attribute 'ops'
FAILED primap2/tests/csg/test_compose.py::test_compose_invalid_strategy_definition - AttributeError: module 'xarray.core' has no attribute 'ops'
FAILED primap2/tests/csg/test_compose.py::test_compose_strategy_skipping - AttributeError: module 'xarray.core' has no attribute 'ops'
FAILED primap2/tests/csg/test_compose.py::test_compose_strategy_all_error - AttributeError: module 'xarray.core' has no attribute 'ops'
FAILED primap2/tests/csg/test_compose.py::test_compose_skip_source - AttributeError: module 'xarray.core' has no attribute 'ops'
FAILED primap2/tests/csg/test_compose.py::test_compose_skip_variable - AttributeError: module 'xarray.core' has no attribute 'ops'
FAILED primap2/tests/csg/test_compose.py::test_compose_skip_entity - AttributeError: module 'xarray.core' has no attribute 'ops'
FAILED primap2/tests/csg/test_compose.py::test_compose_pbar - AttributeError: module 'xarray.core' has no attribute 'ops'
FAILED primap2/tests/csg/test_compose.py::test_compose_sec_cats_missing - AttributeError: module 'xarray.core' has no attribute 'ops'
FAILED primap2/tests/csg/test_compose.py::test_compose_timeseries_trivial - AttributeError: module 'xarray.core' has no attribute 'ops'
FAILED primap2/tests/csg/test_compose.py::test_compose_timeseries_no_match - AttributeError: module 'xarray.core' has no attribute 'ops'
FAILED primap2/tests/csg/test_compose.py::test_compose_timeseries_all_null - AttributeError: module 'xarray.core' has no attribute 'ops'
FAILED primap2/tests/csg/test_strategies.py::test_strategies_conform[substitution] - AttributeError: module 'xarray.core' has no attribute 'ops'
FAILED primap2/tests/csg/test_strategies.py::test_strategies_conform[globalLS] - AttributeError: module 'xarray.core' has no attribute 'ops'
FAILED primap2/tests/csg/test_strategies.py::test_substitution_strategy - AttributeError: module 'xarray.core' has no attribute 'ops'
FAILED primap2/tests/csg/test_strategies.py::test_globalLS_strategy - AttributeError: module 'xarray.core' has no attribute 'ops'

see also https://github.com/primap-community/primap2/actions/runs/13914435902/job/38934794569#step:6:1728

You can see it for yourself locally by running

uv pip install --upgrade --resolution highest --prerelease allow --requirements requirements_upstream_dev.txt
make test

Expected behavior

All tests pass with the upcoming version of xarray.

System (please complete the following information):

• OS: Debian Testing
• Python version 3.12, using xarray main at bb539d8e89fa7601e69fb9dbf849b051c61a3a31

Additional context

The PR in xarray which removed xarray.core.ops: pydata/xarray#10132
The documentation about xarray's public API: https://docs.xarray.dev/en/stable/getting-started-guide/faq.html#what-parts-of-xarray-are-considered-public-api

[mikapfl]

A quick search shows only CSG strategies using this:

> rg xr.core.ops
primap2/csg/_strategies/substitution.py
48:        filled_ts = xr.core.ops.fillna(ts, fill_ts, join="exact")

primap2/csg/_strategies/global_least_squares.py
112:                        filled_ts = xr.core.ops.fillna(ts, fill_ts_harmo, join="exact")
130:                    filled_ts = xr.core.ops.fillna(ts, fill_ts_h, join="exact")

I guess the "official" way to do this is to use ts.fillna(fill_ts, join="exact") instead of using xr.core.ops.fillna. I can't remember anymore why we used xr.core.ops.fillna, can you remember @JGuetschow ?

[mikapfl]

Ah, DataArray.fillna() doesn't support the join argument. Unfortunate. Maybe we just continue using the non-public xarray API and deal with the renaming?

[mikapfl]

It seems that using xr.set_options(arithmetic_join="exact") like this seems to be the official way:

with xr.set_options(arithmetic_join="exact"):
    filled_ts = ts.fillna(fill_ts)

I'll try if this fixes the errors and still does the right thing.

[mikapfl]

@crdanielbusch Do you have time to look into this? I hope the solution is relatively straightforward, as mentioned above, but I'm a bit busy with NDCs. This is not super urgent, we can also just pin the xarray version and will be fine for another 6 months. 😅

[JGuetschow]

That should be a very quick fix as it's only used in one file. I can do it next week unless somebody want's to tackle it earlier.

[crdanielbusch]

I can look into it later today