Dataset fixes

Some (model) datasets contain (known) errors that would normally prevent them from being processed correctly by the ESMValTool. The errors can be in the metadata describing the dataset and/or in the actual data. Typical examples of such errors are missing or wrong attributes (e.g. attribute ‘’units’’ says 1e-9 but data are actually in 1e-6), missing or mislabeled coordinates (e.g. ‘’lev’’ instead of ‘’plev’’ or missing coordinate bounds like ‘’lat_bnds’‘) or problems with the actual data (e.g. cloud liquid water only instead of sum of liquid + ice as specified by the CMIP data request).

The ESMValTool can apply on the fly fixes to data sets that have known errors that can be fixed automatically.

Note

CMORization as a fix. As of early 2020, we’ve started implementing CMORization as fixes for observational datasets. Previously, CMORization was an additional function implemented in ESMValTool. This meant that users always had to store 2 copies of their observational data: both raw and CMORized. Implementing CMORization as a fix removes this redundancy, as the fixes are applied ‘on the fly’ when running a recipe. ERA5 is the first dataset for which this ‘CMORization on the fly’ is supported. For more information about CMORization, see: Contributing a CMORizing script for an observational dataset.

Fix structure

Fixes are Python classes stored in esmvalcore/cmor/_fixes/[PROJECT]/[DATASET].py that derive from esmvalcore.cmor._fixes.fix.Fix and are named after the short name of the variable they fix. You can use the name AllVars if you want the fix to be applied to the whole dataset

Warning

Be careful to replace any - with _ in your dataset name. We need this replacement to have proper python module names.

The fixes are automatically loaded and applied when the dataset is preprocessed.

Fixing a dataset

To illustrate the process of creating a fix we are going to construct a new one from scratch for a fictional dataset. We need to fix a CMIPX model called PERFECT-MODEL that is reporting a missing latitude coordinate for variable tas.

Check the output

Next to the error message, you should see some info about the iris cube: size, coordinates. In our example it looks like this:

air_temperature/ (K) (time: 312; altitude: 90; longitude: 180)
    Dimension coordinates:
        time                                     x              -              -
        altitude                                 -              x              -
        longitude                                -              -              x
    Auxiliary coordinates:
        day_of_month                             x              -              -
        day_of_year                              x              -              -
        month_number                             x              -              -
        year                                     x              -              -
    Attributes:
        {'cmor_table': 'CMIPX', 'mip': 'Amon', 'short_name': 'tas', 'frequency': 'mon'})

So now the mistake is clear: the latitude coordinate is badly named and the fix should just rename it.

Create the fix

We start by creating the module file. In our example the path will be esmvalcore/cmor/_fixes/CMIPX/PERFECT_MODEL.py. If it already exists just add the class to the file, there is no limit in the number of fixes we can have in any given file.

Then we have to create the class for the fix deriving from esmvalcore.cmor._fixes.Fix

"""Fixes for PERFECT-MODEL."""
from esmvalcore.cmor.fix import Fix

class tas(Fix):
     """Fixes for tas variable.""""

Next we must choose the method to use between the ones offered by the Fix class:

  • fix_file : should be used only to fix errors that prevent data loading. As a rule of thumb, you should only use it if the execution halts before reaching the checks.

  • fix_metadata : you want to change something in the cube that is not the data (e.g variable or coordinate names, data units).

  • fix_data: you need to fix the data. Beware: coordinates data values are part of the metadata.

In our case we need to rename the coordinate altitude to latitude, so we will implement the fix_metadata method:

"""Fixes for PERFECT-MODEL."""
from esmvalcore.cmor.fix import Fix

class tas(Fix):
    """Fixes for tas variable.""""

    def fix_metadata(self, cubes):
        """
        Fix metadata for tas.

        Fix the name of the latitude coordinate, which is called altitude
        in the original file.
        """"
        # Sometimes Iris will interpret the data as multiple cubes.
        # Good CMOR datasets will only show one but we support the
        # multiple cubes case to be able to fix the errors that are
        # leading to that extra cubes.
        # In our case this means that we can safely assume that the
        # tas cube is the first one
        tas_cube = cubes[0]
        latitude = tas_cube.coord('altitude')

        # Fix the names. Latitude values, units and
        latitude.short_name = 'lat'
        latitude.standard_name = 'latitude'
        latitude.long_name = 'latitude'
        return cubes

This will fix the error. The next time you run ESMValTool you will find that the error is fixed on the fly and, hopefully, your recipe will run free of errors.

Sometimes other errors can appear after you fix the first one because they were hidden by it. In our case, the latitude coordinate could have bad units or values outside the valid range for example. Just extend your fix to address those errors.

Finishing

Chances are that you are not the only one that wants to use that dataset and variable. Other users could take advantage of your fixes as soon as possible. Please, create a separated pull request for the fix and submit it.

It will also be very helpful if you just scan a couple of other variables from the same dataset and check if they share this error. In case that you find that it is a general one, you can change the fix name to AllVars so it gets executed for all variables in the dataset. If you find that this is shared only by a handful of similar vars you can just make the fix for those new vars derive from the one you just created:

"""Fixes for PERFECT-MODEL."""
from esmvalcore.cmor.fix import Fix

class tas(Fix):
    """Fixes for tas variable.""""

    def fix_metadata(self, cubes):
        """
        Fix metadata for tas.

        Fix the name of the latitude coordinate, which is called altitude
        in the original file.
        """"
        # Sometimes Iris will interpret the data as multiple cubes.
        # Good CMOR datasets will only show one but we support the
        # multiple cubes case to be able to fix the errors that are
        # leading to that extra cubes.
        # In our case this means that we can safely assume that the
        # tas cube is the first one
        tas_cube = cubes[0]
        latitude = tas_cube.coord('altitude')

        # Fix the names. Latitude values, units and
        latitude.short_name = 'lat'
        latitude.standard_name = 'latitude'
        latitude.long_name = 'latitude'
        return cubes


class ps(tas):
    """Fixes for ps variable."""

Common errors

The above example covers one of the most common cases: variables / coordinates that have names that do not match the expected. But there are some others that use to appear frequently. This section describes the most common cases.

Bad units declared

It is quite common that a variable declares to be using some units but the data is stored in another. This can be solved by overwriting the units attribute with the actual data units.

def fix_metadata(self, cubes):
    cube.units = 'real_units'

Detecting this error can be tricky if the units are similar enough. It also has a good chance of going undetected until you notice strange results in your diagnostic.

For the above example, it can be useful to access the variable definition and associated coordinate definitions as provided by the CMOR table. For example:

def fix_metadata(self, cubes):
    cube.units = self.vardef.units

To learn more about what is available in these definitions, see: esmvalcore.cmor.table.VariableInfo and esmvalcore.cmor.table.CoordinateInfo.

Coordinates missing

Another common error is to have missing coordinates. Usually it just means that the file does not follow the CF-conventions and Iris can therefore not interpret it.

If this is the case, you should see a warning from the ESMValTool about discarding some cubes in the fix metadata step. Just before that warning you should see the full list of cubes as read by Iris. If that list contains your missing coordinate you can create a fix for this model:

def fix_metadata(self, cubes):
    coord_cube = cubes.extract_strict('COORDINATE_NAME')
    # Usually this will correspond to an auxiliary coordinate
    # because the most common error is to forget adding it to the
    # coordinates attribute
    coord = iris.coords.AuxCoord(
        coord_cube.data,
        var_name=coord_cube.var_name,
        standard_name=coord_cube.standard_name,
        long_name=coord_cube.long_name,
        units=coord_cube.units,
    }

    # It may also have bounds as another cube
    coord.bounds = cubes.extract_strict('BOUNDS_NAME').data

    data_cube = cubes.extract_strict('VAR_NAME')
    data_cube.add_aux_coord(coord, DIMENSIONS_INDEX_TUPLE)
    return [data_cube]

Customizing checker strictness

The data checker classifies its issues using four different levels of severity. From highest to lowest:

  • CRITICAL: issues that most of the time will have severe consequences.

  • ERROR: issues that usually lead to unexpected errors, but can be safely

    ignored sometimes.

  • WARNING: something is not up to the standard but is unlikely to have

    consequences later.

  • DEBUG: any info that the checker wants to communicate. Regardless of

    checker strictness, those will always be reported as debug messages.

Users can have control about which levels of issues are interpreted as errors, and therefore make the checker fail or warnings or debug messages. For this purpose there is an optional command line option –check-level that can take a number of values, listed below from the lowest level of strictness to the highest:

  • ignore: all issues, regardless of severity, will be reported as

    warnings. Checker will never fail. Use this at your own risk.

  • relaxed: only CRITICAL issues are treated as errors. We recommend not to

    rely on this mode, although it can be useful if there are errors preventing the run that you are sure you can manage on the diagnostics or that will not affect you.

  • default: fail if there are any CRITICAL or ERROR issues (DEFAULT); Provides

    a good measure of safety.

  • strict: fail if there are any warnings, this is the highest level of

    strictness. Mostly useful for checking datasets that you have produced, to be sure that future users will not be distracted by inoffensive warnings.