Extending pandas

While pandas provides a rich set of methods, containers, and data types, your needs may not be fully satisfied. pandas offers a few options for extending pandas.

Registering custom accessors

Libraries can use the decorators pandas.api.extensions.register_dataframe_accessor(), pandas.api.extensions.register_series_accessor(), and pandas.api.extensions.register_index_accessor(), to add additional “namespaces” to pandas objects. All of these follow a similar convention: you decorate a class, providing the name of attribute to add. The class’s __init__ method gets the object being decorated. For example:

@pd.api.extensions.register_dataframe_accessor("geo")
class GeoAccessor:
    def __init__(self, pandas_obj):
        self._validate(pandas_obj)
        self._obj = pandas_obj

    @staticmethod
    def _validate(obj):
        # verify there is a column latitude and a column longitude
        if "latitude" not in obj.columns or "longitude" not in obj.columns:
            raise AttributeError("Must have 'latitude' and 'longitude'.")

    @property
    def center(self):
        # return the geographic center point of this DataFrame
        lat = self._obj.latitude
        lon = self._obj.longitude
        return (float(lon.mean()), float(lat.mean()))

    def plot(self):
        # plot this array's data on a map, e.g., using Cartopy
        pass

Now users can access your methods using the geo namespace:

>>> ds = pd.DataFrame(
...     {"longitude": np.linspace(0, 10), "latitude": np.linspace(0, 20)}
... )
>>> ds.geo.center
(5.0, 10.0)
>>> ds.geo.plot()
# plots data on a map

This can be a convenient way to extend pandas objects without subclassing them. If you write a custom accessor, make a pull request adding it to our pandas ecosystem page.

We highly recommend validating the data in your accessor’s __init__. In our GeoAccessor, we validate that the data contains the expected columns, raising an AttributeError when the validation fails. For a Series accessor, you should validate the dtype if the accessor applies only to certain dtypes.

Extension types

Note

The pandas.api.extensions.ExtensionDtype and pandas.api.extensions.ExtensionArray APIs were experimental prior to pandas 1.5. Starting with version 1.5, future changes will follow the pandas deprecation policy.

pandas defines an interface for implementing data types and arrays that extend NumPy’s type system. pandas itself uses the extension system for some types that aren’t built into NumPy (categorical, period, interval, datetime with timezone).

Libraries can define a custom array and data type. When pandas encounters these objects, they will be handled properly (i.e. not converted to an ndarray of objects). Many methods like pandas.isna() will dispatch to the extension type’s implementation.

If you’re building a library that implements the interface, please publicize it on Extension data types.

The interface consists of two classes.

ExtensionDtype

A pandas.api.extensions.ExtensionDtype is similar to a numpy.dtype object. It describes the data type. Implementors are responsible for a few unique items like the name.

One particularly important item is the type property. This should be the class that is the scalar type for your data. For example, if you were writing an extension array for IP Address data, this might be ipaddress.IPv4Address.

See the extension dtype source for interface definition.

pandas.api.extensions.ExtensionDtype can be registered to pandas to allow creation via a string dtype name. This allows one to instantiate Series and .astype() with a registered string name, for example 'category' is a registered string accessor for the CategoricalDtype.

See the extension dtype dtypes for more on how to register dtypes.

ExtensionArray

This class provides all the array-like functionality. ExtensionArrays are limited to 1 dimension. An ExtensionArray is linked to an ExtensionDtype via the dtype attribute.

pandas makes no restrictions on how an extension array is created via its __new__ or __init__, and puts no restrictions on how you store your data. We do require that your array be convertible to a NumPy array, even if this is relatively expensive (as it is for Categorical).

They may be backed by none, one, or many NumPy arrays. For example, pandas.Categorical is an extension array backed by two arrays, one for codes and one for categories. An array of IPv6 addresses may be backed by a NumPy structured array with two fields, one for the lower 64 bits and one for the upper 64 bits. Or they may be backed by some other storage type, like Python lists.

See the extension array source for the interface definition. The docstrings and comments contain guidance for properly implementing the interface.

ExtensionArray operator support

By default, there are no operators defined for the class ExtensionArray. There are two approaches for providing operator support for your ExtensionArray:

  1. Define each of the operators on your ExtensionArray subclass.

  2. Use an operator implementation from pandas that depends on operators that are already defined on the underlying elements (scalars) of the ExtensionArray.

Note

Regardless of the approach, you may want to set __array_priority__ if you want your implementation to be called when involved in binary operations with NumPy arrays.

For the first approach, you define selected operators, e.g., __add__, __le__, etc. that you want your ExtensionArray subclass to support.

The second approach assumes that the underlying elements (i.e., scalar type) of the ExtensionArray have the individual operators already defined. In other words, if your ExtensionArray named MyExtensionArray is implemented so that each element is an instance of the class MyExtensionElement, then if the operators are defined for MyExtensionElement, the second approach will automatically define the operators for MyExtensionArray.

A mixin class, ExtensionScalarOpsMixin supports this second approach. If developing an ExtensionArray subclass, for example MyExtensionArray, can simply include ExtensionScalarOpsMixin as a parent class of MyExtensionArray, and then call the methods _add_arithmetic_ops() and/or _add_comparison_ops() to hook the operators into your MyExtensionArray class, as follows:

from pandas.api.extensions import ExtensionArray, ExtensionScalarOpsMixin


class MyExtensionArray(ExtensionArray, ExtensionScalarOpsMixin):
    pass


MyExtensionArray._add_arithmetic_ops()
MyExtensionArray._add_comparison_ops()

Note

Since pandas automatically calls the underlying operator on each element one-by-one, this might not be as performant as implementing your own version of the associated operators directly on the ExtensionArray.

For arithmetic operations, this implementation will try to reconstruct a new ExtensionArray with the result of the element-wise operation. Whether or not that succeeds depends on whether the operation returns a result that’s valid for the ExtensionArray. If an ExtensionArray cannot be reconstructed, an ndarray containing the scalars returned instead.

For ease of implementation and consistency with operations between pandas and NumPy ndarrays, we recommend not handling Series and Indexes in your binary ops. Instead, you should detect these cases and return NotImplemented. When pandas encounters an operation like op(Series, ExtensionArray), pandas will

  1. unbox the array from the Series (Series.array)

  2. call result = op(values, ExtensionArray)

  3. re-box the result in a Series

NumPy universal functions

Series implements __array_ufunc__. As part of the implementation, pandas unboxes the ExtensionArray from the Series, applies the ufunc, and re-boxes it if necessary.

If applicable, we highly recommend that you implement __array_ufunc__ in your extension array to avoid coercion to an ndarray. See the NumPy documentation for an example.

As part of your implementation, we require that you defer to pandas when a pandas container (Series, DataFrame, Index) is detected in inputs. If any of those is present, you should return NotImplemented. pandas will take care of unboxing the array from the container and re-calling the ufunc with the unwrapped input.

Testing extension arrays

We provide a test suite for ensuring that your extension arrays satisfy the expected behavior. To use the test suite, you must provide several pytest fixtures and inherit from the base test class. The required fixtures are found in https://github.com/pandas-dev/pandas/blob/main/pandas/tests/extension/conftest.py.

To use a test, subclass it:

from pandas.tests.extension import base


class TestConstructors(base.BaseConstructorsTests):
    pass

See https://github.com/pandas-dev/pandas/blob/main/pandas/tests/extension/base/__init__.py for a list of all the tests available.

Compatibility with Apache Arrow

An ExtensionArray can support conversion to / from pyarrow arrays (and thus support for example serialization to the Parquet file format) by implementing two methods: ExtensionArray.__arrow_array__ and ExtensionDtype.__from_arrow__.

The ExtensionArray.__arrow_array__ ensures that pyarrow knowns how to convert the specific extension array into a pyarrow.Array (also when included as a column in a pandas DataFrame):

class MyExtensionArray(ExtensionArray):
    ...

    def __arrow_array__(self, type=None):
        # convert the underlying array values to a pyarrow Array
        import pyarrow

        return pyarrow.array(..., type=type)

The ExtensionDtype.__from_arrow__ method then controls the conversion back from pyarrow to a pandas ExtensionArray. This method receives a pyarrow Array or ChunkedArray as only argument and is expected to return the appropriate pandas ExtensionArray for this dtype and the passed values:

class ExtensionDtype:
    ...

    def __from_arrow__(self, array: pyarrow.Array/ChunkedArray) -> ExtensionArray:
        ...

See more in the Arrow documentation.

Those methods have been implemented for the nullable integer and string extension dtypes included in pandas, and ensure roundtrip to pyarrow and the Parquet file format.

Subclassing pandas data structures

Warning

There are some easier alternatives before considering subclassing pandas data structures.

  1. Extensible method chains with pipe

  2. Use composition. See here.

  3. Extending by registering an accessor

  4. Extending by extension type

This section describes how to subclass pandas data structures to meet more specific needs. There are two points that need attention:

  1. Override constructor properties.

  2. Define original properties

Note

You can find a nice example in geopandas project.

Override constructor properties

Each data structure has several constructor properties for returning a new data structure as the result of an operation. By overriding these properties, you can retain subclasses through pandas data manipulations.

There are 3 possible constructor properties to be defined on a subclass:

  • DataFrame/Series._constructor: Used when a manipulation result has the same dimension as the original.

  • DataFrame._constructor_sliced: Used when a DataFrame (sub-)class manipulation result should be a Series (sub-)class.

  • Series._constructor_expanddim: Used when a Series (sub-)class manipulation result should be a DataFrame (sub-)class, e.g. Series.to_frame().

Below example shows how to define SubclassedSeries and SubclassedDataFrame overriding constructor properties.

class SubclassedSeries(pd.Series):
    @property
    def _constructor(self):
        return SubclassedSeries

    @property
    def _constructor_expanddim(self):
        return SubclassedDataFrame


class SubclassedDataFrame(pd.DataFrame):
    @property
    def _constructor(self):
        return SubclassedDataFrame

    @property
    def _constructor_sliced(self):
        return SubclassedSeries
>>> s = SubclassedSeries([1, 2, 3])
>>> type(s)
<class '__main__.SubclassedSeries'>

>>> to_framed = s.to_frame()
>>> type(to_framed)
<class '__main__.SubclassedDataFrame'>

>>> df = SubclassedDataFrame({"A": [1, 2, 3], "B": [4, 5, 6], "C": [7, 8, 9]})
>>> df
   A  B  C
0  1  4  7
1  2  5  8
2  3  6  9

>>> type(df)
<class '__main__.SubclassedDataFrame'>

>>> sliced1 = df[["A", "B"]]
>>> sliced1
   A  B
0  1  4
1  2  5
2  3  6

>>> type(sliced1)
<class '__main__.SubclassedDataFrame'>

>>> sliced2 = df["A"]
>>> sliced2
0    1
1    2
2    3
Name: A, dtype: int64

>>> type(sliced2)
<class '__main__.SubclassedSeries'>

Define original properties

To let original data structures have additional properties, you should let pandas know what properties are added. pandas maps unknown properties to data names overriding __getattribute__. Defining original properties can be done in one of 2 ways:

  1. Define _internal_names and _internal_names_set for temporary properties which WILL NOT be passed to manipulation results.

  2. Define _metadata for normal properties which will be passed to manipulation results.

Below is an example to define two original properties, “internal_cache” as a temporary property and “added_property” as a normal property

class SubclassedDataFrame2(pd.DataFrame):

    # temporary properties
    _internal_names = pd.DataFrame._internal_names + ["internal_cache"]
    _internal_names_set = set(_internal_names)

    # normal properties
    _metadata = ["added_property"]

    @property
    def _constructor(self):
        return SubclassedDataFrame2
>>> df = SubclassedDataFrame2({"A": [1, 2, 3], "B": [4, 5, 6], "C": [7, 8, 9]})
>>> df
   A  B  C
0  1  4  7
1  2  5  8
2  3  6  9

>>> df.internal_cache = "cached"
>>> df.added_property = "property"

>>> df.internal_cache
cached
>>> df.added_property
property

# properties defined in _internal_names is reset after manipulation
>>> df[["A", "B"]].internal_cache
AttributeError: 'SubclassedDataFrame2' object has no attribute 'internal_cache'

# properties defined in _metadata are retained
>>> df[["A", "B"]].added_property
property

Plotting backends

Starting in 0.25 pandas can be extended with third-party plotting backends. The main idea is letting users select a plotting backend different than the provided one based on Matplotlib. For example:

>>> pd.set_option("plotting.backend", "backend.module")
>>> pd.Series([1, 2, 3]).plot()

This would be more or less equivalent to:

>>> import backend.module
>>> backend.module.plot(pd.Series([1, 2, 3]))

The backend module can then use other visualization tools (Bokeh, Altair,…) to generate the plots.

Libraries implementing the plotting backend should use entry points to make their backend discoverable to pandas. The key is "pandas_plotting_backends". For example, pandas registers the default “matplotlib” backend as follows.

# in setup.py
setup(  # noqa: F821
    ...,
    entry_points={
        "pandas_plotting_backends": [
            "matplotlib = pandas:plotting._matplotlib",
        ],
    },
)

More information on how to implement a third-party plotting backend can be found at https://github.com/pandas-dev/pandas/blob/main/pandas/plotting/__init__.py#L1.