Error: The defined _FillValue and missing_value are not the same

[cehbrecht]

• clisops version: 0.12.0
• Python version:
• Operating System:

Description

I have run a subsetting operation on a CMIP6 dataset:

• ds_id = c3s-cmip6.CMIP.NCAR.CESM2-WACCM.historical.r1i1p1f1.day.tas.gn.v20190227
• time="2000/2000"
• area="5,49,7,51"

CLISOPS failed with the following error message:

File "/usr/local/anaconda/envs/rook/lib/python3.11/site-packages/clisops/ops/subset.py", line 233, in subset
    return op.process()
           ^^^^^^^^^^^^
  File "/usr/local/anaconda/envs/rook/lib/python3.11/site-packages/clisops/ops/base_operation.py", line 146, in process
    processed_ds = self._remove_redundant_fill_values(processed_ds)
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/anaconda/envs/rook/lib/python3.11/site-packages/clisops/ops/base_operation.py", line 94, in _remove_redundant_fill_values
    raise Exception(
Exception: The defined _FillValue and missing_value for 'tas' are not the same '1.0000000200408773e+20' != '1e+20'.

This error does not happen for all datasets. Our tests didn't cover this case. I suppose the error is caused by the newer xarray and/or cf_xarray version used in clisops v.12.0.

This error does not appear in clisops v0.10.1 used by rook v0.11.0.

What I Did

I have reproduced the error with the following notebook:

https://nbviewer.org/github/roocs/rooki/blob/master/notebooks/errors/cds-error-2023-11-28-subset-cmip6.ipynb

It is using rooki:

wf = ops.Subset(
        ops.Input(
            'tas', ['c3s-cmip6.CMIP.NCAR.CESM2-WACCM.historical.r1i1p1f1.day.tas.gn.v20190227']
        ),
        area="5,49,7,51",
        time="2000/2000"
)
resp = wf.orchestrate()
resp.ok

An example file for testing can be this one: https://data.mips.copernicus-climate.eu/thredds/fileServer/esg_c3s-cmip6/CMIP/NCAR/CESM2-WACCM/historical/r1i1p1f1/day/tas/gn/v20190227/tas_day_CESM2-WACCM_historical_r1i1p1f1_gn_20000101-20091231.nc

[cehbrecht]

The error is caused by this check: https://github.com/roocs/clisops/blob/0b94f66a2628d0ee63a7be8a786264da7175a742/clisops/ops/base_operation.py#L93-L96

This code was added with the regrid operator in release 0.12. So, it is not an issue with cf_xarray dependencies.

[cehbrecht]

@sol1105 I have added a quick-fix in this PR #309. It converts the exception into a warning message. What would be a meaningful solution?

[sol1105]

In #243 I updated your xarray _FillValue workaround (_remove_redundant_fill_values (issues #198 / #224 )). My changes cause this problem.

Since xarray adds NaN as _FillValue, when no _FillValue is defined: if the source data has no missing_value/_FillValue attribute defined: set it to None so xarray would not set it to NaN if both are defined, check if they are the same (it is a data quality check basically that I never thought would trigger for C3S data).

In the case of this dataset, _FillValue is defined as a float (as it should be), while missing_value is defined as double. This is still a flaw in the source data, but none that should prevent processing. One could use numpy.isclose in the check and set both missing_value/_FillValue attributes to the same data type as has the data.

[sol1105]

Update: When xarray opens such a dataset it prints a warning:

xarray/conventions.py:543: SerializationWarning: variable 'tas' has multiple fill values {1e+20, 1e+20}, decoding all values to NaN.

After looking into the xarray code, when writing out the data to disk...

• ... xarray casts the _FillValue / missing_value attributes to the same data type as the data
• ... xarray will fail encoding when the values of both attributes are "not close" (rtol=1e-5, atol=1e-8)

So I guess we can just set if _FillValue != missing_value: FillValue = missing_value with an issued warning and let xarray handle the dtype since it properly decodes the missing values.

[cehbrecht]

fixed by PR #309