What happened: Looks like when I use map_blocks with functions that operate on a subset of variables, the computation is much faster if I subsample the dataset first.

What you expected to happen: In the example below, I wouldn't expect such a difference in computation time.

Minimal Complete Verifiable Example:

python import xarray as xr ds = xr.tutorial.open_dataset("air_temperature") ds["foo"] = xr.DataArray(ds["time"].values, dims="time") ds = ds.chunk(dict(time=1, lon=-1, lat=-1))

python def func(obj): return obj[["foo"]]

python %%timeit ds.map_blocks(func).compute()

17.3 s ± 179 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

```python %%timeit

Subsample the dataset before calling map_blocks

func(ds).map_blocks(func).compute() ```

5.63 s ± 175 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

Anything else we need to know?:

Environment:

Output of <tt>xr.show_versions()</tt>

INSTALLED VERSIONS ------------------ commit: None python: 3.9.2 | packaged by conda-forge | (default, Feb 21 2021, 05:02:46) [GCC 9.3.0] python-bits: 64 OS: Linux OS-release: 3.10.0-1062.18.1.el7.x86_64 machine: x86_64 processor: x86_64 byteorder: little LC_ALL: None LANG: en_US.UTF-8 LOCALE: en_US.UTF-8 libhdf5: 1.10.6 libnetcdf: 4.7.4 xarray: 0.17.0 pandas: 1.2.3 numpy: 1.20.1 scipy: 1.6.1 netCDF4: 1.5.6 pydap: None h5netcdf: None h5py: None Nio: None zarr: None cftime: 1.4.1 nc_time_axis: None PseudoNetCDF: None rasterio: None cfgrib: None iris: None bottleneck: None dask: 2021.03.0 distributed: 2021.03.0 matplotlib: 3.3.4 cartopy: None seaborn: None numbagg: None pint: None setuptools: 49.6.0.post20210108 pip: 21.0.1 conda: None pytest: 6.2.2 IPython: 7.21.0 sphinx: None