5734 has greatly improved the performance. Fantastic work @dcherian 👏🏽

```python In [13]: import xarray as xr, pandas as pd, numpy as np

In [14]: ds = xr.Dataset({"a": xr.DataArray(np.r_[np.arange(500.), np.arange(500.)]), ...: "b": xr.DataArray(np.arange(1000.))})

In [15]: ds Out[15]: <xarray.Dataset> Dimensions: (dim_0: 1000) Dimensions without coordinates: dim_0 Data variables: a (dim_0) float64 0.0 1.0 2.0 3.0 4.0 ... 496.0 497.0 498.0 499.0 b (dim_0) float64 0.0 1.0 2.0 3.0 4.0 ... 996.0 997.0 998.0 999.0 ```

```python In [16]: xr.set_options(use_flox=True) Out[16]: <xarray.core.options.set_options at 0x104de21a0>

In [17]: %%timeit ...: ds.groupby("a").mean() ...: ...: 1.5 ms ± 3.3 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)

In [18]: xr.set_options(use_flox=False) Out[18]: <xarray.core.options.set_options at 0x144382350>

In [19]: %%timeit ...: ds.groupby("a").mean() ...: ...: 94 ms ± 715 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) ```

I gave a look to functions such as "np.add.at" which can be highly faster than home-made solution. The aggregate function of the "numpy-groupies" package is even faster (25 x faster than np.add.at in my case). Maybe xarray groupby functionalities can rely on such effective package.

In case anyone gets here by Googling something like "xarray groupby slow" and you loaded data from a netCDF file, be aware that slowness you see in groupby aggregation on a Dataset or DataArray may actually be due not to this issue but to the lazy loading that's done by default. This can be fixed by calling .load() on the Dataset or DataArray. See the Tip about lazy loading at http://xarray.pydata.org/en/stable/io.html#netcdf.

Another approach here (rather than writing something new with Numba) would be to write a pure NumPy engine for groupby that relies on reordering data and np.add.accumulate. This could yield performance within a factor of 2-3x slower than pandas. See this comment for an example: https://github.com/numpy/numpy/issues/7265#issuecomment-198796408

Agreed! If you'd like to make a pull request that would be greatly appreciated

On Sun, Nov 15, 2015 at 10:10 PM, Antony Lee notifications@github.com wrote:

Perhaps worth mentioning in the docs? The difference turned out to be a major bottleneck in my code.

— Reply to this email directly or view it on GitHub https://github.com/xray/xray/issues/659#issuecomment-156925589.

Perhaps worth mentioning in the docs? The difference turned out to be a major bottleneck in my code.

Yes, switching to pandas for these operations is certainly a recommended approach :).

In my case I could just switch to pandas, so I'll leave it as it is for now.

Yes, I'm afraid this is a known issue. Grouped aggregations are currently implemented with a loop in pure Python, which, of course, is pretty slow.

I've done some exploratory work to rewrite them in Numba, which shows some encouraging preliminary results:

``` from numba import guvectorize, jit import pandas as pd import numpy as np

@guvectorize(['(float64[:], int64[:], float64[:])'], '(x),(x),(y)', nopython=True) def _grouped_mean(values, int_labels, target): count = np.zeros(len(target), np.int64) for i in range(len(values)): val = values[i] if not np.isnan(val): lab = int_labels[i] target[lab] += val count[lab] += 1 target /= count

def move_axis_to_end(array, axis): array = np.asarray(array) return np.rollaxis(array, axis, start=array.ndim)

def grouped_mean(values, by, axis=-1): int_labels, uniques = pd.factorize(by, sort=True) values = move_axis_to_end(values, axis) target = np.zeros(values.shape[:-1] + uniques.shape) _grouped_mean(values, int_labels, target) return target, uniques

values = np.random.RandomState(0).rand(int(1e6)) values[::50] = np.nan by = np.random.randint(50, size=int(1e6)) df = pd.DataFrame({'x': values, 'y': by})

np.testing.assert_allclose(grouped_mean(values, by)[0], df.groupby('y')['x'].mean())

%timeit grouped_mean(values, by) # 100 loops, best of 3: 15.3 ms per loop %timeit df.groupby('y').mean() # 10 loops, best of 3: 21.4 ms per loop ```

Unfortunately, I'm unlikely to have time to work on this in the near future. If you or anyone else is interested in taking the lead on this, it would be greatly appreciated!

Note that we can't reuse the routines from pandas because they are only designed for 1D or at most 2D data.