issue_comments

However, other_sel in _yield_binary_applied looks like:

python <xarray.DataArray (x: 10)> array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]) Coordinates: * x (x) int64 0 1 2 3 4 5 6 7 8 9 dayofyear int64 1

https://github.com/pydata/xarray/blob/f821fe20595c3700375ccecebf88e01a61444777/xarray/core/groupby.py#L484

This is still relevant. Now raises

python ValueError: 'x' not present in all datasets and coords='different'. Either add 'x' to datasets where it is missing or specify coords='minimal'. I think the issue may be that _dummy_copy does not include the x-dimension: python xr.core.groupby._dummy_copy(d2.groupby('time.dayofyear').mean('time')) yields python <xarray.DataArray ()> array(nan)

Clues about what's going on from pdb: ``` In [23]: import pdb; pdb.pm()

/Users/shoyer/dev/xarray/xarray/core/utils.py(290)getitem() -> return self.mapping[key] (Pdb) u /Users/shoyer/dev/xarray/xarray/core/combine.py(165)<genexpr>() -> for ds in datasets[1:]) (Pdb) u /Users/shoyer/dev/xarray/xarray/core/combine.py(165)differs() -> for ds in datasets[1:]) (Pdb) u /Users/shoyer/dev/xarray/xarray/core/combine.py(168)<genexpr>() -> if k not in concat_over and differs(k)) (Pdb) u /Users/shoyer/dev/xarray/xarray/core/combine.py(167)process_subset_opt() -> concat_new = set(k for k in getattr(datasets[0], subset) (Pdb) u /Users/shoyer/dev/xarray/xarray/core/combine.py(192)_calc_concat_over() -> concat_over.update(process_subset_opt(coords, 'coords')) (Pdb) u /Users/shoyer/dev/xarray/xarray/core/combine.py(212)_dataset_concat() -> concat_over = _calc_concat_over(datasets, dim, data_vars, coords) (Pdb) datasets[-5:] (<xarray.Dataset> Dimensions: (time: 32, x: 10) Coordinates: * time (time) datetime64[ns] 1979-12-28 1980-12-27 1981-12-28 ... * x (x) int64 0 1 2 3 4 5 6 7 8 9 dayofyear int64 362 Data variables: <this-array> (time, x) float64 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ..., <xarray.Dataset> Dimensions: (time: 32, x: 10) Coordinates: * time (time) datetime64[ns] 1979-12-29 1980-12-28 1981-12-29 ... * x (x) int64 0 1 2 3 4 5 6 7 8 9 dayofyear int64 363 Data variables: <this-array> (time, x) float64 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ..., <xarray.Dataset> Dimensions: (time: 32, x: 10) Coordinates: * time (time) datetime64[ns] 1979-12-30 1980-12-29 1981-12-30 ... * x (x) int64 0 1 2 3 4 5 6 7 8 9 dayofyear int64 364 Data variables: <this-array> (time, x) float64 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ..., <xarray.Dataset> Dimensions: (time: 32, x: 10) Coordinates: * time (time) datetime64[ns] 1979-12-31 1980-12-30 1981-12-31 ... * x (x) int64 0 1 2 3 4 5 6 7 8 9 dayofyear int64 365 Data variables: <this-array> (time, x) float64 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ..., <xarray.Dataset> Dimensions: (time: 8) Coordinates: * time (time) datetime64[ns] 1980-12-31 1984-12-31 1988-12-31 ... Data variables: <this-array> (time) float64 nan nan nan nan nan nan nan nan) (Pdb) ```

Yes, we have some logic that is supposed to handle missing groups in groupby, but apparently it's not working properly in this case: https://github.com/pydata/xarray/blob/bcd608101133c0cb84c74d341d22edef71ef4818/xarray/core/groupby.py#L331-L333

This seems to be related somehow to the fact that 1979 isn't a leap year, so we have a missing value to fill in for 366. For example, it works if we switch to 2004 for d2: ``` In [14]: d2 = xr.DataArray(np.zeros((366, 10)), {'time': pd.date_range('2004-01-01', freq='D', periods=366), 'x': ('x', np.ara ...: nge(10))}, dims=['time', 'x'])

In [15]: d1.groupby('time.dayofyear') - d2.groupby('time.dayofyear').mean('time') Out[15]: <xarray.DataArray (time: 12000, x: 10)> array([[ 0., 0., 0., ..., 0., 0., 0.], [ 0., 0., 0., ..., 0., 0., 0.], [ 0., 0., 0., ..., 0., 0., 0.], ..., [ 0., 0., 0., ..., 0., 0., 0.], [ 0., 0., 0., ..., 0., 0., 0.], [ 0., 0., 0., ..., 0., 0., 0.]]) Coordinates: * x (x) int64 0 1 2 3 4 5 6 7 8 9 * time (time) datetime64[ns] 1979-01-01 1979-01-02 1979-01-03 ... dayofyear (time) int64 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 ... ```

@ahuang11 Thanks, I can reproduce on the development version of xarray.

I attempted to reproduce but have not been successful: In [94]: d1 = xr.DataArray(np.zeros(366), [('time', pd.date_range('1979-01-01', freq='D', periods=366))]) In [95]: d2 = xr.DataArray(np.zeros((1000, 10)), {'time': pd.date_range('1979-01-01', freq='D', periods=1000), 'x': ('x', np.arange(10))}, dims=['time', 'x']) In [96]: d1.groupby('time.dayofyear') - d2.groupby('time.dayofyear').mean('time') Out[96]: <xarray.DataArray (time: 366, x: 10)> array([[ 0., 0., 0., ..., 0., 0., 0.], [ 0., 0., 0., ..., 0., 0., 0.], [ 0., 0., 0., ..., 0., 0., 0.], ..., [ 0., 0., 0., ..., 0., 0., 0.], [ 0., 0., 0., ..., 0., 0., 0.], [ 0., 0., 0., ..., 0., 0., 0.]]) Coordinates: * x (x) int64 0 1 2 3 4 5 6 7 8 9 * time (time) datetime64[ns] 1979-01-01 1979-01-02 1979-01-03 ... dayofyear (time) int64 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 ...