Support operations with pandas Offset objects

[max-sixty]

Is your feature request related to a problem? Please describe.

Currently xarray objects containting datetimes don't operate with pandas' offset objects:

times = pd.date_range("2000-01-01", freq="6H", periods=10)
ds = xr.Dataset(
    {
        "foo": (["time", "x", "y"], np.random.randn(10, 5, 3)),
        "bar": ("time", np.random.randn(10), {"meta": "data"}),
        "time": times,
    }
)
ds.attrs["dsmeta"] = "dsdata"
ds.resample(time="24H").mean("time").time + to_offset("8H")

raises:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-29-f9de46fe6c54> in <module>
----> 1 ds.resample(time="24H").mean("time").time + to_offset("8H")

/usr/local/lib/python3.8/site-packages/xarray/core/dataarray.py in func(self, other)
   2763 
   2764             variable = (
-> 2765                 f(self.variable, other_variable)
   2766                 if not reflexive
   2767                 else f(other_variable, self.variable)

/usr/local/lib/python3.8/site-packages/xarray/core/variable.py in func(self, other)
   2128             with np.errstate(all="ignore"):
   2129                 new_data = (
-> 2130                     f(self_data, other_data)
   2131                     if not reflexive
   2132                     else f(other_data, self_data)

TypeError: unsupported operand type(s) for +: 'numpy.ndarray' and 'pandas._libs.tslibs.offsets.Hour'

This is an issue because pandas resampling has deprecated loffset — from our test suite:

xarray/tests/test_dataset.py::TestDataset::test_resample_loffset
  /Users/maximilian/workspace/xarray/xarray/tests/test_dataset.py:3844: FutureWarning: 'loffset' in .resample() and in Grouper() is deprecated.

  >>> df.resample(freq="3s", loffset="8H")

  becomes:

  >>> from pandas.tseries.frequencies import to_offset
  >>> df = df.resample(freq="3s").mean()
  >>> df.index = df.index.to_timestamp() + to_offset("8H")

    ds.bar.to_series().resample("24H", loffset="-12H").mean()

...and so we'll need to support something like this in order to maintain existing behavior.

Describe the solution you'd like
I'm not completely sure; I think probably supporting the operations between xarray objects containing datetime objects and pandas' offset objects.

[dcherian]

For resample, we could add it to .indexes["time"] which is pd.DatetimeIndex so the addition should work.

I agree with generally supporting adding pandas offset objects.

EDIT: to clarify, ds["time"] = ds.indexes["time"] + to_offset(loffset) should do what we want

[max-sixty]

Yes, good point, that would be easy and probably sufficient for most use-cases.