Repository navigation
DataTree.map_over_datasets behavior can be obscure sometimes #11813
Description
What is your issue?
Consider this code:
import numpy as np
import pandas as pd
import xarray as xr
da = xr.DataArray(
data=[1., np.nan, 3.],
coords={"datetime": pd.date_range("2026-10-26", periods=3, freq="D")},
name="up",
)
dtree = xr.DataTree(dataset=da.to_dataset())
dtree.map_over_datasets(lambda ds: ds.interpolate_na(dim="datetime"))AttributeError: DatasetView objects are not to be initialized directly.This will not run because map_over_datasets will iterate over DatasetView objects, even though the documentation says explicitly that the parameter func expects Dataset:
https://docs.xarray.dev/en/latest/generated/xarray.DataTree.map_over_datasets.html
But if you dig in xarray/core/datatree.py and have a look at DatasetView, this is stated in an exception that I didn't encounter:
def __setitem__(self, key, val) -> None:
raise AttributeError(
"Mutation of the DatasetView is not allowed, please use `.__setitem__` on the wrapping DataTree node, "
"or use `dt.to_dataset()` if you want a mutable dataset. If calling this from within `map_over_datasets`,"
"use `.copy()` first to get a mutable version of the input dataset."
)So the solution is to run lambda ds: ds.copy().interpolate_na(dim="datetime"). Not intuitive at all, expecially if one has been using map_over_datasets with other functions without a problem. Shouldn't this be more clear in the documentation?
Activity
mohamd764 commentedon Oct 27, 2026
I hit the same thing on xarray 2026.9.0 (Python 3.13, Linux), and I think it's a bug more than a docs gap.
interpolate_nadoesn't mutate the input. It goes through_apply_over_vars_with_diminxarray/core/missing.py, which builds the result withInside
map_over_datasets,type(self)isDatasetView, and its__init__always raises. Other methods I tried on a view (mean,fillna,differentiate) return a plainDatasetand work fine.ffill,bfill,polyfitandcurvefitfail the same way. The last two use the same pattern inxarray/computation/fit.py(type(obj)(...)). All of these are still there on main (aee3dc5).As a quick check I monkeypatched
_apply_over_vars_with_dimto buildDataset(coords=self.coords, attrs=self.attrs)instead, andinterpolate_naandffillthen worked insidemap_over_datasets. I haven't run the test suite with that change, so it's just a pointer. Until it's fixed,ds.copy()first is the workaround, as you found.