ECMWF’s MARS and Metview software introduced the notion of *Fieldset* which is an ordered collection of GRIB message. The *Fieldset* is an abstract concept and can be implemented in many ways. In the case of MARS and Metview, a *Fieldset* is an implemented internally as an array of *Field*, each *Field* being represented by a file path, an offset and a length where the actual GRIB message can be found. Thus, a *Fieldset* can represent an ordered collection of *Field* which are at various locations of several files. *cfgrib* now provides the definition of a ``Field`` and ``Fieldset`` types in the ``cfgrib.abc`` module and additionally a ``MappingFieldset`` for specialised use. The implementations are based on simple python sequences and mappings so that *cfgrib* can build a Dataset for example from something as simple as a list of dicts. Classes that implement the ``Fieldset`` and the ``MappingFieldset`` interface can use the low-level interface ``cfgrib.open_fielset`` to obtain a ``cfgrib.Dataset`` or they can be passed directly to *Xarray*. The simplest *Fieldset* is a list of dictionaries: .. code-block:: python >>> import xarray as xr >>> fieldset = [ ... { ... "gridType": "regular_ll", ... "Nx": 2, ... "Ny": 3, ... "distinctLatitudes": [-10.0, 0.0, 10.0], ... "distinctLongitudes": [0.0, 10.0], ... "paramId": 130, ... "shortName": "t", ... "values": [[1, 2], [3, 4], [5, 6]], ... "dataDate": 20211216, ... "dataTime": 1200, ... } ... ] >>> ds = xr.open_dataset(fieldset, engine="cfgrib") >>> ds Dimensions: (latitude: 3, longitude: 2) Coordinates: time datetime64[ns] ... * latitude (latitude) float64 -10.0 0.0 10.0 * longitude (longitude) float64 0.0 10.0 Data variables: t (latitude, longitude) float32 ... Attributes: Conventions: CF-1.7 history: ... >>> ds.mean() Dimensions: () Coordinates: time datetime64[ns] ... Data variables: t float32 3.5 For example you can implement a dedicated ``Fieldset`` class following this pattern: .. code-block:: python from typing import Iterator from cfgrib import abc class MyFieldset(abc.Fieldset): def __len__(self) -> int: # not used by cfgrib ... def __getitem__(self, item: int) -> abc.Field: ... def __iter__(self) -> Iterator[abc.Field]: ... If ``__getitem__`` and ``__iter__`` implement lazy loading of GRIB fields *cfgrib* and *Xarray* will be able to access larger-than-memory files. In the event a ``Field`` is identified by a more complex *key* than just an sequence *index* developers may implemnt a ``MappingFieldset`` class following this pattern: .. code-block:: python from typing import ItemsView, Iterator from cfgrib import abc class MyFieldset(abc.MappingFieldset[T.Any, abc.Field]): def __len__(self) -> int: # not used by cfgrib ... def __getitem__(self, item: int) -> abc.Field: ... def __iter__(self) -> Iterator[abc.Field]: # not used by cfgrib ... def items() -> ItemsView[T.Any, abc.Field]: ... Again if ``__getitem__`` and ``items`` implement lazy loading of GRIB fields *cfgrib* and *Xarray* will be able to access larger-than-memory files. An example of the ``MappingFieldset`` use is ``cfgrib.messages.FileStream`` that uses the *file offset* as the *key*.