Should I use 'allowed' instead of 'parallelized' if my apply_ufunc are simply a variety of numpy functions?

[ahuang11]

""" http://xarray.pydata.org/en/stable/dask.html
For the majority of NumPy functions that are already wrapped by Dask, it’s usually a better idea to use the pre-existing dask.array function, by using either a pre-existing xarray methods or apply_ufunc() with dask='allowed'. Dask can often have a more efficient implementation that makes use of the specialized structure of a problem, unlike the generic speedups offered by dask='parallelized'.
"""

If my function passed to xr.apply_ufunc only encompasses numpy functions (e.g. np.mean), but does not explicitly call dask.array functions (e.g. dask.array.mean), will the above still apply; i.e. should I use dask='allowed' rather than 'parallelized'?

Function similar to:
https://github.com/xarray-contrib/xskillscore/blob/master/xskillscore/core/np_deterministic.py#L527-L540

[dcherian]

Some of that advice needs to be updated.

Things like np.mean will dispatch to dask.array.mean with recent enough versions of dask/numpy (see NEP-18). So dask="allowed" will work. This is easy to check: if the call to apply_ufunc ends up computing it means some function is not compatible with dask arrays. Either avoid that function or use dask="parallelized" if appropriate (the function can be applied independently on each block of the dask array, and does not need communication between blocks).