Do you see any issues with Zarr3 compatibility?

[langmore]

Hello. I'm just checking if you think this will be a problem. If so, what work needs to be done. Thanks!

[raspstephan]

I think I will defer this to @shoyer :D

[shoyer]

Not that I know of, but we certainly have not tested this.

[langmore]

I used WBX with Zarr3, without issues, about a month ago. I started using it again, and (with Zarr3 only... Zarr2 is okay) get failures with beam_pipeline_test.py::BeamPipelineTest::test_unaggregated_pipeline The failure happens only with 'Error_5_to_10_days.geopotential' and 'Error_5_to_10_days.2m_temperature' ('SquaredError.geopotential' and 'SquaredError.2m_temperature' are fine). In particular, for these data vars, direct_results is all NaN, and pipeline_results is all 0's.

I've also been using WB2 and xarray_beam (without WB) and haven't seen any issues with Zarr3. So I suspect this is something coming from #96 .

[langmore]

Actually... The issue is a bad test, which, at line 145, selected 'lead_time' in 'both'. Once I properly select only predictions then it passes. The confusion is that, with Zarr2, pipeline_results turns into all 0's (matching direct_results), whereas in Zarr3, it becomes NaN. To make matters more confusing, the behavior of the bias_5_to_10_days results depends on whether mse and rmse are computed!

So the test is broken as-is, but can be fixed with a one line change.

[langmore]

Fixing the test in #103