Yay, mypy is passing now after flox updates. :)

Copy/pasteing the Resample-class removes the rest of the errors. Not especially elegant though, maybe there's a better way?

Ok, this kind of works. But there's a new error because I defined the Data*ResampleBase-classes with a for-loop.

There's a bit of mypy errors now. Maybe just removíng _flox_reduce in _reductions.py is enough?

This is a strict improvement now, well done!

before after ratio [4c865d60] [9b13c974] - 18.7±0.3ms 11.0±0.1ms 0.59 groupby.GroupByDask.time_agg_small_num_groups('sum', 2) - 24.3±1ms 13.8±0.2ms 0.57 groupby.ResampleDask.time_agg_small_num_groups('mean', 2) - 22.0±0.2ms 12.0±0.3ms 0.54 groupby.ResampleDask.time_agg_small_num_groups('mean', 1) - 13.5±0.4ms 7.28±0.3ms 0.54 groupby.GroupByDask.time_agg_small_num_groups('sum', 1) - 33.7±0.4ms 14.2±0.3ms 0.42 groupby.ResampleDask.time_agg_small_num_groups('sum', 2) - 30.7±0.4ms 11.9±0.2ms 0.39 groupby.ResampleDask.time_agg_small_num_groups('sum', 1) - 104±1ms 8.21±0.08ms 0.08 groupby.Resample.time_agg_large_num_groups('sum', 1) - 119±1ms 8.84±0.2ms 0.07 groupby.Resample.time_agg_large_num_groups('sum', 2) - 115±2ms 8.45±0.2ms 0.07 groupby.Resample.time_agg_large_num_groups('mean', 1) - 134±1ms 9.60±0.2ms 0.07 groupby.Resample.time_agg_large_num_groups('mean', 2) - 257±3ms 10.2±0.3ms 0.04 groupby.GroupBy.time_agg_large_num_groups('mean', 2) - 260±5ms 9.99±0.2ms 0.04 groupby.GroupBy.time_agg_large_num_groups('sum', 2) - 190±6ms 6.47±0.2ms 0.03 groupby.GroupBy.time_agg_large_num_groups('mean', 1) - 193±2ms 6.31±0.2ms 0.03 groupby.GroupBy.time_agg_large_num_groups('sum', 1) - 368±5ms 10.9±0.2ms 0.03 groupby.GroupByDask.time_agg_large_num_groups('mean', 2) - 494±2ms 13.7±0.3ms 0.03 groupby.ResampleDask.time_agg_large_num_groups('mean', 2) - 451±8ms 12.3±0.3ms 0.03 groupby.ResampleDask.time_agg_large_num_groups('mean', 1) - 318±3ms 7.13±0.04ms 0.02 groupby.GroupByDask.time_agg_large_num_groups('mean', 1) - 531±6ms 10.6±0.2ms 0.02 groupby.GroupByDask.time_agg_large_num_groups('sum', 2) - 814±10ms 14.0±0.4ms 0.02 groupby.ResampleDask.time_agg_large_num_groups('sum', 2) - 760±8ms 12.0±0.3ms 0.02 groupby.ResampleDask.time_agg_large_num_groups('sum', 1) - 490±6ms 7.13±0.2ms 0.01 groupby.GroupByDask.time_agg_large_num_groups('sum', 1)

Earlier benchmarks:

Maybe it's also on the pint side? Even if numpy_groupies supports the like argument it will crash because pint doesn't support asanyarray.

```python import numpy as np import pint import dask.array as da

pint crashes:

np.asanyarray([1, 2], like=pint.Quantity(1, "s")) Traceback (most recent call last):

File "<ipython-input-11-f8c20802ca08>", line 1, in <module> np.asanyarray([1, 2], like=pint.Quantity(1, "s"))

TypeError: no implementation found for 'numpy.asanyarray' on types that implement array_function: [<class 'pint.quantity.Quantity'>]

dask supports it:

np.asanyarray([1, 2], like=da.array(0)) Out[12]: dask.array<array, shape=(2,), dtype=int32, chunksize=(2,), chunktype=numpy.ndarray> ```

Looking at those unit errors it appears numpy_groupies is forcing the duck arrays to numpy arrays. Perhaps adding a like in the np.asanyarray() will do the trick? python C:\Miniconda\envs\xarray-tests\lib\site-packages\numpy_groupies\utils_numpy.py:199: in input_validation a = np.asanyarray(a) https://github.com/ml31415/numpy-groupies/blob/7a31d1f9bbd51111b4a4d01cf1df01a5b4827e85/numpy_groupies/utils_numpy.py#L199