Xarray v0.12.2 (29 June 2019) breaking metpy No module named 'xarray.core.accessors'

[atmosmattbrewer]

I just created a new environment and installed the latest versions of xarray and metpy. The latest version of xarray seems to break metpy and when importing metpy I got this error:

xarray version: 0.12.2 py_0 conda-forge
metpy version: 0.10.0 py37_1001 conda-forge

Rolling back to xarray v0.12.1 fixes this issue.

[zbruick]

Yeah I just saw this pop up on one our of workshop notebooks that I was updating. I'll get a PR up on this shortly.

[jthielen]

It looks like this was broken by pydata/xarray#2991, which moved the DatetimeAccessor from xarray.core.accessors to xarray.core.accessor_dt. I'm not sure whether that accessor was part of the public API or not. Although, now with xarray v0.12.2, the strftime() monkey patch is no longer needed.

So, this will probably have to change to a conditional import for xarray <0.12.2? Or should the minimum xarray verison be bumped to 0.12.2 and the monkey patch removed? (Sorry, forgot that xarray v0.12 dropped Python 2, and MetPy still needs it for this next v0.11 release.)

[akrherz]

seems like this is an unwanted API break for xarray? cc'ing @shoyer

[brian-rose]

Just found the same error when updating some climlab notebooks. Downgrading to xarray 0.12.1 resolved the problem.

[shoyer]

Sorry for the trouble here! We should clarify that everything inside xarray.core is considered an internal/private API. It would be wise to vendor copies of the other internal imports from xarray.core that I see in metpy/xarray.py.

[dopplershift]

@shoyer No worries. Considering I put in the original PR, I should have seen this coming.

Is xarray.core being internal noted somewhere that I should have seen? Regardless the vendoring is a good idea since we're really only re-using some utility functions.

[shoyer]

Is xarray.core being internal noted somewhere that I should have seen? Regardless the vendoring is a good idea since we're really only re-using some utility functions.

No, it's just a loose convention, e.g., everything in pandas.core and numpy.core is also internal API. I'll add an explicit note to our API docs.