issue_comments

@shoyer thanks for looking into this.

I also figured it later that I can just use np.nanmean (or nanmedian) but that function turns out to be much slower than np.mean (or np.median) version. As nans are only happening as the beginning and end of the sequence, is there any efficient way of using nanmean only for those segments and mean for the rest of the processing? My own thought is to have a check for nan in the custom function and apply mean or nanmean depending on the results of that check, but not sure if this can be done more efficiently.

@mrezak Thanks for the report and the clear example!

Certainly this is an annoying inconsistency. I'm trying to figure out whether this is also a bug or not.

I think the difference comes down to how pandas and xarray pass data into the custom function. Pandas passes individual slices, trimming out values outside the window. Xarray passes an N+1 dimensional view of the array data with extra dimension added for the "window offset", with values outside the window filled with NaN: ```python import numpy as np import pandas as pd import xarray

def custom(x, axis=0): print(x) return np.mean(x, axis)

print('pandas example') d = pd.DataFrame(np.random.rand(11,3)) r = d.rolling(10, min_periods=5).apply(custom) print(r.iloc[0:10,:])

print('\nxarray example') xd = d.to_xarray().to_array() r = xd.rolling(index=10, min_periods=5).reduce(custom) print(r[:,0:10]) ``` Output:

Xarray's version is certainly going to be way faster, but it has the downside of treating windows differently. One way to work around this would be to use np.nanmean inside custom instead of np.mean.

cc @jhamman @fujiisoup who worked on this and may have ideas