Can't merge ECMWF perturbed and control forecast together as xarray since release 0.10

[jodemaey]

Hi there,

Since the release 0.10.1, the EUMETNET benchmark plugin is partly broken because apparently cfgrib does not accept anymore to merge the control and perturbed ECMWF forecasts together.

I use the multi source to do that (see here), but when calling to_xarray, I get the following issue:

DatasetBuildError                         Traceback (most recent call last)
File /opt/tljh/user/envs/climetlab_test/lib/python3.9/site-packages/cfgrib/dataset.py:633, in build_dataset_components(index, errors, encode_cf, squeeze, log, read_keys, time_dims, extra_coords)
    632 try:
--> 633     dims, data_var, coord_vars = build_variable_components(
    634         var_index,
    635         encode_cf,
    636         filter_by_keys,
    637         errors=errors,
    638         squeeze=squeeze,
    639         read_keys=read_keys,
    640         time_dims=time_dims,
    641         extra_coords=extra_coords,
    642     )
    643 except DatasetBuildError as ex:
    644     # NOTE: When a variable has more than one value for an attribute we need to raise all
    645     #   the values in the file, not just the ones associated with that variable. See #54.

File /opt/tljh/user/envs/climetlab_test/lib/python3.9/site-packages/cfgrib/dataset.py:471, in build_variable_components(index, encode_cf, filter_by_keys, log, errors, squeeze, read_keys, time_dims, extra_coords)
    460 def build_variable_components(
    461     index: abc.Index[T.Any, abc.Field],
    462     encode_cf: T.Sequence[str] = (),
   (...)
    469     extra_coords: T.Dict[str, str] = {},
    470 ) -> T.Tuple[T.Dict[str, int], Variable, T.Dict[str, Variable]]:
--> 471     data_var_attrs = enforce_unique_attributes(index, DATA_ATTRIBUTES_KEYS, filter_by_keys)
    472     grid_type_keys = GRID_TYPE_MAP.get(index.getone("gridType"), [])

File /opt/tljh/user/envs/climetlab_test/lib/python3.9/site-packages/cfgrib/dataset.py:273, in enforce_unique_attributes(index, attributes_keys, filter_by_keys)
    272         fbks.append(fbk)
--> 273     raise DatasetBuildError("multiple values for key %r" % key, key, fbks)
    274 if values and values[0] not in ("undef", "unknown"):

DatasetBuildError: multiple values for key 'dataType'

During handling of the above exception, another exception occurred:

DatasetBuildError                         Traceback (most recent call last)
Input In [17], in <module>
----> 1 fcs = ds.to_xarray(xarray_open_dataset_kwargs=dict(squeeze=True))
      2 fcs

File ~/climetlab-eumetnet-postprocessing-benchmark/climetlab_eumetnet_postprocessing_benchmark/gridded/training_data_forecasts.py:398, in TrainingDataForecastSurfacePostProcessed.to_xarray(self, **kwargs)
    397 def to_xarray(self, **kwargs):
--> 398     fcs = self.source.to_xarray(**kwargs)
    399     variables = list(fcs.keys())
    400     ds_list = list()

File /opt/tljh/user/envs/climetlab_test/lib/python3.9/site-packages/climetlab/readers/grib/fieldset.py:203, in FieldSet.to_xarray(self, **kwargs)
    185     xarray_open_dataset_kwargs[key] = mix_kwargs(
    186         user=user_xarray_open_dataset_kwargs.pop(key, {}),
    187         default={"errors": "raise"},
   (...)
    190         logging_main_key=key,
    191     )
    192 xarray_open_dataset_kwargs.update(
    193     mix_kwargs(
    194         user=user_xarray_open_dataset_kwargs,
   (...)
    200     )
    201 )
--> 203 result = xr.open_dataset(
    204     self,
    205     **xarray_open_dataset_kwargs,
    206 )
    208 def math_prod(lst):
    209     if not hasattr(math, "prod"):
    210         # python 3.7 does not have math.prod

File /opt/tljh/user/envs/climetlab_test/lib/python3.9/site-packages/xarray/backends/api.py:495, in open_dataset(filename_or_obj, engine, chunks, cache, decode_cf, mask_and_scale, decode_times, decode_timedelta, use_cftime, concat_characters, decode_coords, drop_variables, backend_kwargs, *args, **kwargs)
    483 decoders = _resolve_decoders_kwargs(
    484     decode_cf,
    485     open_backend_dataset_parameters=backend.open_dataset_parameters,
   (...)
    491     decode_coords=decode_coords,
    492 )
    494 overwrite_encoded_chunks = kwargs.pop("overwrite_encoded_chunks", None)
--> 495 backend_ds = backend.open_dataset(
    496     filename_or_obj,
    497     drop_variables=drop_variables,
    498     **decoders,
    499     **kwargs,
    500 )
    501 ds = _dataset_from_backend_dataset(
    502     backend_ds,
    503     filename_or_obj,
   (...)
    510     **kwargs,
    511 )
    512 return ds

File /opt/tljh/user/envs/climetlab_test/lib/python3.9/site-packages/cfgrib/xarray_plugin.py:99, in CfGribBackend.open_dataset(self, filename_or_obj, mask_and_scale, decode_times, concat_characters, decode_coords, drop_variables, use_cftime, decode_timedelta, lock, indexpath, filter_by_keys, read_keys, encode_cf, squeeze, time_dims, errors, extra_coords)
     77 def open_dataset(
     78     self,
     79     filename_or_obj: T.Union[str, abc.MappingFieldset[T.Any, abc.Field]],
   (...)
     96     extra_coords: T.Dict[str, str] = {},
     97 ) -> xr.Dataset:
---> 99     store = CfGribDataStore(
    100         filename_or_obj,
    101         indexpath=indexpath,
    102         filter_by_keys=filter_by_keys,
    103         read_keys=read_keys,
    104         encode_cf=encode_cf,
    105         squeeze=squeeze,
    106         time_dims=time_dims,
    107         lock=lock,
    108         errors=errors,
    109         extra_coords=extra_coords,
    110     )
    111     with xr.core.utils.close_on_error(store):
    112         vars, attrs = store.load()  # type: ignore

File /opt/tljh/user/envs/climetlab_test/lib/python3.9/site-packages/cfgrib/xarray_plugin.py:39, in CfGribDataStore.__init__(self, filename, lock, **backend_kwargs)
     37 else:
     38     opener = dataset.open_fieldset
---> 39 self.ds = opener(filename, **backend_kwargs)

File /opt/tljh/user/envs/climetlab_test/lib/python3.9/site-packages/cfgrib/dataset.py:730, in open_fieldset(fieldset, indexpath, filter_by_keys, read_keys, time_dims, extra_coords, computed_keys, log, **kwargs)
    728 index = messages.FieldsetIndex.from_fieldset(fieldset, index_keys, computed_keys)
    729 filtered_index = index.subindex(filter_by_keys)
--> 730 return open_from_index(filtered_index, read_keys, time_dims, extra_coords, **kwargs)

File /opt/tljh/user/envs/climetlab_test/lib/python3.9/site-packages/cfgrib/dataset.py:706, in open_from_index(index, read_keys, time_dims, extra_coords, **kwargs)
    699 def open_from_index(
    700     index: abc.Index[T.Any, abc.Field],
    701     read_keys: T.Sequence[str] = (),
   (...)
    704     **kwargs: T.Any,
    705 ) -> Dataset:
--> 706     dimensions, variables, attributes, encoding = build_dataset_components(
    707         index, read_keys=read_keys, time_dims=time_dims, extra_coords=extra_coords, **kwargs
    708     )
    709     return Dataset(dimensions, variables, attributes, encoding)

File /opt/tljh/user/envs/climetlab_test/lib/python3.9/site-packages/cfgrib/dataset.py:654, in build_dataset_components(index, errors, encode_cf, squeeze, log, read_keys, time_dims, extra_coords)
    652         fbks.append(fbk)
    653         error_message += "\n    filter_by_keys=%r" % fbk
--> 654     raise DatasetBuildError(error_message, key, fbks)
    655 short_name = data_var.attributes.get("GRIB_shortName", "paramId_%d" % param_id)
    656 var_name = data_var.attributes.get("GRIB_cfVarName", "unknown")

DatasetBuildError: multiple values for unique key, try re-open the file with one of:
    filter_by_keys={'dataType': 'pf'}
    filter_by_keys={'dataType': 'cf'}

So the problem seems to be that it has become more strict about the multiple values for keys, using this function enforce_unique_attributes.

cfgrib version is 0.9.10.

How can I resolve this issue? Can I tell him to be more forgiving about the attributes uniqueness?

Thank you in advance.

---

To reproduce, simply install the plugin

pip install climetlab climetlab-eumetnet-postprocessing-benchmark

and run the notebook https://github.com/Climdyn/climetlab-eumetnet-postprocessing-benchmark/tree/main/notebooks/demo_ensemble_forecasts.ipynb .

[floriankrb]

I see. I reproduced the issue, thanks for the detailed report.

The reason is that your grib files (from mutiple sources) are not compatible.

Previous versions of (0.9.9 and below), we were sending each grib separately to cfgrib. But now, for multi source of grib, climetlab is doing the indexing, computing offsets etc. and presenting to cfgrib something that looks like a unique grib file. If you open such a grib file cfgrib fails with this error.

Nevertheless, we do want to support this use case nicely. I will work on it during the next few days.

[floriankrb]

We would like to make cfgrib ignore the value of the inconsistent key (namely "dataType" which is sometimes 'pf' and sometimes 'cf'). This is not yet implemented in cfgrib. I added it in climetlab and opened an issue to have it implemented in cfgrib.

Would you mind checking that this solves the issue?

1) Use the current develop branch of climetlab (commit ee697ff or later) 2) Change your TrainingDataForecastSurfacePostProcessed.to_xarray() to add the option to ignore the failing keys as follow:

        fcs = self.source.to_xarray(
            xarray_open_dataset_kwargs={"backend_kwargs": {"ignore_keys": ["dataType"]}},
            **kwargs,
        )

[jodemaey]

It seems to work now indeed:

Does that mean that we need to overload the to_xarray routine everywhere this might happen?

[floriankrb]

Yes, the current default behaviour of cfgrib is to raise an error when the GRIB are not consistent. The rationale is that we really do not want to change the grib file on the fly without the user deciding it explicitly. So, yes, the option "ignore_keys" needs to be added everywhere you need it.

[jodemaey]

Ok thanks, I think that this issue can be closed when there is a new release with the fix then.

[floriankrb]

version 0.10.3 released.

[jodemaey]

After implementing this modification, everything seems fine and I will release a new version of the plugin. Thanks for your help.