I am trying to apply an averaging function to multi year chunks of monthly model data. At the core the function performs a weighted average (and then some coordinate manipulations). I am using resample(time='1AS') and then try to map my custom function onto the data (see example below). Without actually loading the data, this step is prohibitively long in my workflow (20-30min depending on the model). Is there a way to apply this step completely lazily, like in the case where a simple non-weighted .mean() is used?

```python from dask.diagnostics import ProgressBar import xarray as xr import numpy as np

simple customized weighted mean function

def mean_func(ds): return ds.weighted(ds.weights).mean('time')

example dataset

t = xr.cftime_range(start='2000', periods=1000, freq='1AS') weights = xr.DataArray(np.random.rand(len(t)),dims=['time'], coords={'time':t}) data = xr.DataArray(np.random.rand(len(t)),dims=['time'], coords={'time':t, 'weights':weights}) ds = xr.Dataset({'data':data}).chunk({'time':1}) ds ```

Using resample with a simple mean works without any computation being triggered: python with ProgressBar(): ds.resample(time='3AS').mean('time')

But when I do the same step with my custom function, there are some computations showing up

python with ProgressBar(): ds.resample(time='3AS').map(mean_func) [########################################] | 100% Completed | 0.1s [########################################] | 100% Completed | 0.1s [########################################] | 100% Completed | 0.1s [########################################] | 100% Completed | 0.1s I am quite sure these are the same kind of computations that make my real-world workflow so slow.

I also confirmed that this not happening when I do not use resample first with ProgressBar(): mean_func(ds) this does not trigger a computation either. So this must be somehow related to resample? I would be happy to dig deeper into this, if somebody with more knowledge could point me to the right place.

Environment: