Rolling operations loose chunking with dask and bottleneck

[ScottWales]

Code Sample, a copy-pastable example if possible

A "Minimal, Complete and Verifiable Example" will make it much easier for maintainers to help you:
http://matthewrocklin.com/blog/work/2018/02/28/minimal-bug-reports

import bottleneck
import xarray
import dask

data = dask.array.ones((100,), chunks=(10,))
da = xarray.DataArray(data, dims=['time'])

rolled = da.rolling(time=15).mean()

# Expect the 'rolled' dataset to be chunked approximately the same as 'data',
# however there is only one chunk in 'rolled' instead of 10
assert len(rolled.chunks[0]) > 1

Problem description

Rolling operations loose chunking over the rolled dimension when using dask datasets with bottleneck installed, which is a problem for large datasets where we don't want to load the entire thing.

The issue appears to be caused by xarray.core.dask_array_ops.dask_rolling_wrapper calling dask.array.overlap.overlap on a DataArray instead of a Dask array. Possibly #2940 is related?

Expected Output

Chunks should be preserved through .rolling().mean()

Output of xr.show_versions()

INSTALLED VERSIONS ------------------ commit: None python: 3.6.7 | packaged by conda-forge | (default, Feb 28 2019, 09:07:38) [GCC 7.3.0] python-bits: 64 OS: Linux OS-release: 3.10.0-862.14.4.el6.x86_64 machine: x86_64 processor: x86_64 byteorder: little LC_ALL: en_AU.utf8 LANG: C LOCALE: en_AU.UTF-8 libhdf5: 1.10.4 libnetcdf: 4.6.2

xarray: 0.12.1
pandas: 0.24.2
numpy: 1.16.3
scipy: 1.2.1
netCDF4: 1.5.0.1
pydap: installed
h5netcdf: None
h5py: 2.9.0
Nio: None
zarr: 2.3.1
cftime: 1.0.3.4
nc_time_axis: 1.2.0
PseudonetCDF: None
rasterio: None
cfgrib: None
iris: 2.2.0
bottleneck: 1.2.1
dask: 1.2.0
distributed: 1.27.1
matplotlib: 3.0.3
cartopy: 0.17.0
seaborn: 0.9.0
setuptools: 41.0.1
pip: 19.1
conda: None
pytest: 4.4.1
IPython: 7.5.0
sphinx: None

[ScottWales]

Looks like there is already a fix in #2942