How to add a new backend#

Adding a new backend for read support to Xarray does not require to integrate any code in Xarray; all you need to do is:

If you also want to support lazy loading and dask see How to support lazy loading.

Note that the new interface for backends is available from Xarray version >= 0.18 onwards.

You can see what backends are currently available in your working environment with list_engines().

BackendEntrypoint subclassing#

Your BackendEntrypoint sub-class is the primary interface with Xarray, and it should implement the following attributes and methods:

  • the open_dataset method (mandatory)

  • the open_dataset_parameters attribute (optional)

  • the guess_can_open method (optional)

  • the description attribute (optional)

  • the url attribute (optional).

This is what a BackendEntrypoint subclass should look like:

from xarray.backends import BackendEntrypoint

class MyBackendEntrypoint(BackendEntrypoint):
    def open_dataset(
        # other backend specific keyword arguments
        # `chunks` and `cache` DO NOT go here, they are handled by xarray
        return my_open_dataset(filename_or_obj, drop_variables=drop_variables)

    open_dataset_parameters = ["filename_or_obj", "drop_variables"]

    def guess_can_open(self, filename_or_obj):
            _, ext = os.path.splitext(filename_or_obj)
        except TypeError:
            return False
        return ext in {".my_format", ".my_fmt"}

    description = "Use .my_format files in Xarray"

    url = "https://link_to/your_backend/documentation"

BackendEntrypoint subclass methods and attributes are detailed in the following.


The backend open_dataset shall implement reading from file, the variables decoding and it shall instantiate the output Xarray class Dataset.

The following is an example of the high level processing steps:

def open_dataset(
    vars, attrs, coords = my_reader(
    vars, attrs, coords = my_decode_variables(
        vars, attrs, decode_times, decode_timedelta, decode_coords
    )  #  see also conventions.decode_cf_variables

    ds = xr.Dataset(vars, attrs=attrs, coords=coords)

    return ds

The output Dataset shall implement the additional custom method close, used by Xarray to ensure the related files are eventually closed. This method shall be set by using set_close().

The input of open_dataset method are one argument (filename_or_obj) and one keyword argument (drop_variables):

  • filename_or_obj: can be any object but usually it is a string containing a path or an instance of pathlib.Path.

  • drop_variables: can be None or an iterable containing the variable names to be dropped when reading the data.

If it makes sense for your backend, your open_dataset method should implement in its interface the following boolean keyword arguments, called decoders, which default to None:

  • mask_and_scale

  • decode_times

  • decode_timedelta

  • use_cftime

  • concat_characters

  • decode_coords

Note: all the supported decoders shall be declared explicitly in backend open_dataset signature and adding a **kwargs is not allowed.

These keyword arguments are explicitly defined in Xarray open_dataset() signature. Xarray will pass them to the backend only if the User explicitly sets a value different from None. For more details on decoders see Decoders.

Your backend can also take as input a set of backend-specific keyword arguments. All these keyword arguments can be passed to open_dataset() grouped either via the backend_kwargs parameter or explicitly using the syntax **kwargs.

If you don’t want to support the lazy loading, then the Dataset shall contain values as a numpy.ndarray and your work is almost done.


open_dataset_parameters is the list of backend open_dataset parameters. It is not a mandatory parameter, and if the backend does not provide it explicitly, Xarray creates a list of them automatically by inspecting the backend signature.

If open_dataset_parameters is not defined, but **kwargs and *args are in the backend open_dataset signature, Xarray raises an error. On the other hand, if the backend provides the open_dataset_parameters, then **kwargs and *args can be used in the signature. However, this practice is discouraged unless there is a good reasons for using **kwargs or *args.


guess_can_open is used to identify the proper engine to open your data file automatically in case the engine is not specified explicitly. If you are not interested in supporting this feature, you can skip this step since BackendEntrypoint already provides a default guess_can_open() that always returns False.

Backend guess_can_open takes as input the filename_or_obj parameter of Xarray open_dataset(), and returns a boolean.

description and url#

description is used to provide a short text description of the backend. url is used to include a link to the backend’s documentation or code.

These attributes are surfaced when a user prints BackendEntrypoint. If description or url are not defined, an empty string is returned.


The decoders implement specific operations to transform data from on-disk representation to Xarray representation.

A classic example is the “time” variable decoding operation. In NetCDF, the elements of the “time” variable are stored as integers, and the unit contains an origin (for example: “seconds since 1970-1-1”). In this case, Xarray transforms the pair integer-unit in a numpy.datetime64.

The standard coders implemented in Xarray are:

Xarray coders all have the same interface. They have two methods: decode and encode. The method decode takes a Variable in on-disk format and returns a Variable in Xarray format. Variable attributes no more applicable after the decoding, are dropped and stored in the Variable.encoding to make them available to the encode method, which performs the inverse transformation.

In the following an example on how to use the coders decode method:

In [1]: var = xr.Variable(
   ...:     dims=("x",), data=np.arange(10.0), attrs={"scale_factor": 10, "add_offset": 2}
   ...: )

In [2]: var
<xarray.Variable (x: 10)>
array([0., 1., 2., 3., 4., 5., 6., 7., 8., 9.])
    scale_factor:  10
    add_offset:    2

In [3]: coder = xr.coding.variables.CFScaleOffsetCoder()

In [4]: decoded_var = coder.decode(var)

In [5]: decoded_var
<xarray.Variable (x: 10)>
[10 values with dtype=float64]

In [6]: decoded_var.encoding
Out[6]: {'scale_factor': 10, 'add_offset': 2}

Some of the transformations can be common to more backends, so before implementing a new decoder, be sure Xarray does not already implement that one.

The backends can reuse Xarray’s decoders, either instantiating the coders and using the method decode directly or using the higher-level function decode_cf_variables() that groups Xarray decoders.

In some cases, the transformation to apply strongly depends on the on-disk data format. Therefore, you may need to implement your own decoder.

An example of such a case is when you have to deal with the time format of a grib file. grib format is very different from the NetCDF one: in grib, the time is stored in two attributes dataDate and dataTime as strings. Therefore, it is not possible to reuse the Xarray time decoder, and implementing a new one is mandatory.

Decoders can be activated or deactivated using the boolean keywords of Xarray open_dataset() signature: mask_and_scale, decode_times, decode_timedelta, use_cftime, concat_characters, decode_coords. Such keywords are passed to the backend only if the User sets a value different from None. Note that the backend does not necessarily have to implement all the decoders, but it shall declare in its open_dataset interface only the boolean keywords related to the supported decoders.

How to register a backend#

Define a new entrypoint in your (or setup.cfg) with:

  • group: xarray.backends

  • name: the name to be passed to open_dataset() as engine

  • object reference: the reference of the class that you have implemented.

You can declare the entrypoint in using the following syntax:

        "xarray.backends": ["my_engine=my_package.my_module:MyBackendEntryClass"],

in setup.cfg:

xarray.backends =
    my_engine = my_package.my_module:MyBackendEntryClass

See for more information

If you are using Poetry for your build system, you can accomplish the same thing using “plugins”. In this case you would need to add the following to your pyproject.toml file:

"my_engine" = "my_package.my_module:MyBackendEntryClass"

See for more information on Poetry plugins.

How to support lazy loading#

If you want to make your backend effective with big datasets, then you should support lazy loading. Basically, you shall replace the numpy.ndarray inside the variables with a custom class that supports lazy loading indexing. See the example below:

backend_array = MyBackendArray()
data = indexing.LazilyIndexedArray(backend_array)
var = xr.Variable(dims, data, attrs=attrs, encoding=encoding)


  • LazilyIndexedArray is a class provided by Xarray that manages the lazy loading.

  • MyBackendArray shall be implemented by the backend and shall inherit from BackendArray.

BackendArray subclassing#

The BackendArray subclass shall implement the following method and attributes:

  • the __getitem__ method that takes in input an index and returns a NumPy array

  • the shape attribute

  • the dtype attribute.

Xarray supports different type of Indexing and selecting data, that can be grouped in three types of indexes BasicIndexer, OuterIndexer and VectorizedIndexer. This implies that the implementation of the method __getitem__ can be tricky. In order to simplify this task, Xarray provides a helper function, explicit_indexing_adapter(), that transforms all the input indexer types (basic, outer, vectorized) in a tuple which is interpreted correctly by your backend.

This is an example BackendArray subclass implementation:

from xarray.backends import BackendArray

class MyBackendArray(BackendArray):
    def __init__(
        # other backend specific keyword arguments
        self.shape = shape
        self.dtype = dtype
        self.lock = lock

    def __getitem__(
        self, key: xarray.core.indexing.ExplicitIndexer
    ) -> np.typing.ArrayLike:
        return indexing.explicit_indexing_adapter(

    def _raw_indexing_method(self, key: tuple) -> np.typing.ArrayLike:
        # thread safe method that access to data on disk
        with self.lock:
            return item

Note that BackendArray.__getitem__ must be thread safe to support multi-thread processing.

The explicit_indexing_adapter() method takes in input the key, the array shape and the following parameters:

  • indexing_support: the type of index supported by raw_indexing_method

  • raw_indexing_method: a method that shall take in input a key in the form of a tuple and return an indexed numpy.ndarray.

For more details see IndexingSupport and Indexing examples.

In order to support Dask Distributed and multiprocessing, BackendArray subclass should be serializable either with Pickle or cloudpickle. That implies that all the reference to open files should be dropped. For opening files, we therefore suggest to use the helper class provided by Xarray CachingFileManager.

Indexing examples#


In the BASIC indexing support, numbers and slices are supported.


In [7]: # () shall return the full array
   ...: backend_array._raw_indexing_method(())
Out[7]: array([[0, 1, 2, 3], [4, 5, 6, 7], [8, 9, 10, 11]])

In [8]: # shall support integers
   ...: backend_array._raw_indexing_method(1, 1)
Out[8]: 5

In [9]: # shall support slices
   ...: backend_array._raw_indexing_method(slice(0, 3), slice(2, 4))
Out[9]: array([[2, 3], [6, 7], [10, 11]])


The OUTER indexing shall support number, slices and in addition it shall support also lists of integers. The the outer indexing is equivalent to combining multiple input list with itertools.product():

In [10]: backend_array._raw_indexing_method([0, 1], [0, 1, 2])
Out[10]: array([[0, 1, 2], [4, 5, 6]])

# shall support integers
In [11]: backend_array._raw_indexing_method(1, 1)
Out[11]: 5


The OUTER_1VECTOR indexing shall supports number, slices and at most one list. The behaviour with the list shall be the same of OUTER indexing.

If you support more complex indexing as explicit indexing or numpy indexing, you can have a look to the implementation of Zarr backend and Scipy backend, currently available in backends module.

Preferred chunk sizes#

To potentially improve performance with lazy loading, the backend may define for each variable the chunk sizes that it prefers—that is, sizes that align with how the variable is stored. (Note that the backend is not directly involved in Dask chunking, because Xarray internally manages chunking.) To define the preferred chunk sizes, store a mapping within the variable’s encoding under the key "preferred_chunks" (that is, var.encoding["preferred_chunks"]). The mapping’s keys shall be the names of dimensions with preferred chunk sizes, and each value shall be the corresponding dimension’s preferred chunk sizes expressed as either an integer (such as {"dim1": 1000, "dim2": 2000}) or a tuple of integers (such as {"dim1": (1000, 100), "dim2": (2000, 2000, 2000)}).

Xarray uses the preferred chunk sizes in some special cases of the chunks argument of the open_dataset() and open_mfdataset() functions. If chunks is a dict, then for any dimensions missing from the keys or whose value is None, Xarray sets the chunk sizes to the preferred sizes. If chunks equals "auto", then Xarray seeks ideal chunk sizes informed by the preferred chunk sizes. Specifically, it determines the chunk sizes using dask.array.core.normalize_chunks() with the previous_chunks argument set according to the preferred chunk sizes.