Inconsistent results due to dask.Cache

[orioltinto]

What happened?

In my workflow, I was opening a dataset from a netcdf, modifying the contents, saving the dataset as a different file and then repeat the process with the same file but a different modification.

It looks like Dask Cache is messing up the results due to using cached information when it shouldn't.

Also, that leds to different results for data_array.values.sum() and data_array.sum().values.

What did you expect to happen?

I expect that when I open a file from the disk, if I access the values from that file these will correspond to what it is actually on the file.

Instead it is reusing data values that should have been discarded when loading the file again.

Minimal Complete Verifiable Example

import numpy as np
import xarray as xr
import dask
from dask.cache import Cache

cache = Cache(2e9)
cache.register()

print(f"Xarray Version: {xr.__version__=}")
print(f"Dask version: {dask.__version__=}")


def file_opener(file_path):
    """
    Open a file and return a dataset object.
    :param file_path:
    :return:
    """
    return xr.open_dataset(file_path, chunks={})


def main():
    # Define names for the files
    file_path = "./dummy_dataset.nc"
    modified_file_path = file_path.replace(".nc", ".modified.nc")

    # Create a synthetic dataset and save it into a file
    data = np.zeros(shape=(4, 3))
    foo = xr.DataArray(data, dims=['time', 'space'])
    ds = foo.to_dataset(name='foo')
    ds.to_netcdf(file_path)

    # Open the file and compute the sum of the values
    for i in range(3):
        print(f"Iteration:{i}")
        print("\tOriginal")
        with file_opener(file_path) as ds:
            print(f"\t\t{ds.foo.sum().values=:2f}")
            print(f"\t\t{ds.foo.values.sum()=:2f}")

            # Modify the dataset and save it as a different file
            ds.foo.values += 1.0
            ds.to_netcdf(modified_file_path)

        # Open the modified file and compute the sum of the values
        print("Modified")
        print(f"\t\t{ds.foo.sum().values=:2f}")
        print(f"\t\t{ds.foo.values.sum()=:2f}")


if __name__ == "__main__":
    main()

MVCE confirmation

• [X] Minimal example — the example is as focused as reasonably possible to demonstrate the underlying issue in xarray.
• [X] Complete example — the example is self-contained, including all data and the text of any traceback.
• [x] Verifiable example — the example copy & pastes into an IPython prompt or Binder notebook, returning the result.
• [X] New issue — a search of GitHub Issues suggests this is not a duplicate.

Relevant log output

The output of running the provided MCV:


Xarray Version: xr.__version__='2022.3.0'
Dask version: dask.__version__='2022.05.2'
Iteration:0
	Original
		ds.foo.sum().values=0.000000
		ds.foo.values.sum()=0.000000
Modified
		ds.foo.sum().values=12.000000
		ds.foo.values.sum()=12.000000
Iteration:1
	Original
		ds.foo.sum().values=0.000000
		ds.foo.values.sum()=12.000000
Modified
		ds.foo.sum().values=24.000000
		ds.foo.values.sum()=24.000000
Iteration:2
	Original
		ds.foo.sum().values=0.000000
		ds.foo.values.sum()=24.000000
Modified
		ds.foo.sum().values=36.000000
		ds.foo.values.sum()=36.000000

Anything else we need to know?

I couldn't run the example in the Binder notebook because it uses Cache from dask.Cache which couldn't be loaded there. However, I could run reproduce the results in a clean environment installing the following dependencies:

• xarray
• dask
• cachey
• netCDF4

Environment

INSTALLED VERSIONS ------------------ commit: None python: 3.9.7 | packaged by conda-forge | (default, Sep 29 2021, 19:20:46) [GCC 9.4.0] python-bits: 64 OS: Linux OS-release: 5.4.0-109-generic machine: x86_64 processor: x86_64 byteorder: little LC_ALL: None LANG: en_US.UTF-8 LOCALE: ('en_US', 'UTF-8') libhdf5: 1.12.0 libnetcdf: 4.7.4

xarray: 2022.3.0 pandas: 1.4.2 numpy: 1.22.4 scipy: 1.8.1 netCDF4: 1.5.8 pydap: None h5netcdf: 1.0.0 h5py: 3.7.0 Nio: None zarr: None cftime: 1.6.0 nc_time_axis: None PseudoNetCDF: None rasterio: None cfgrib: None iris: None bottleneck: None dask: 2022.05.2 distributed: 2022.5.2 matplotlib: 3.5.2 cartopy: 0.20.2 seaborn: 0.11.2 numbagg: None fsspec: 2022.5.0 cupy: None pint: 0.19.2 sparse: None setuptools: 57.4.0 pip: 22.1.2 conda: None pytest: 7.1.2 IPython: None sphinx: None