Python Extensions in Pure Rust with Rust-CPython

Python's many advantages come at a cost: execution speed on the "traditional" runtime lags that of other languages by a considerable margin. Python's solution is to expose the runtime to more efficient extensions written C and C++. As noted previously here, Python extensions can also be written in pure Rust through PyO3. But some projects call for greater control. This article describes an alternative that may be useful when a lower-level approach needed.


Rust-CPython is a set of "Rust bindings for the python interpreter." Created in 2015, Rust-CPython was the precursor to the PyO3 project. The main difference between the two projects is implementation: in Rust-CPython developers define classes and functions using declarative macros whereas in PyO3, procedural macros are used. An important secondary difference relates to ownership. In PyO3, the framework owns Python values, whereas Rust-CPython code owns its own values. This distinction gives Rust-CPython users greater flexibility at the cost of some overhead.

Source Code/Example Setup

The complete source code for the examples in this article can be found on GitHub. Each example follows a similar Cargo.toml setup:

# ... package attributes

crate-type = ["cdylib"]

cpython = "0.7"

default = ["python3"]
python3 = ["cpython/python3-sys", "cpython/extension-module"]

The Simplest Module

The core of Rust-CPython is the py_module_initializer! macro. To illustrate its use, consider a module consisting of just a docstring. It can be implemented like so:

use cpython::py_module_initializer;

py_module_initializer! {
    docstring, |py, module| {
        module.add(py, "__doc__", "This module is implemented in Rust.")?;


Note that:

  • docstring is the name of the Python package when imported from Python.
  • py is the Global Interpreter Lock (GIL).
  • module is the module being created.

The block inside the macro adds the attribute __doc__ to the module with the indicated string value.

Compiling with cargo build yields a shared library in target/debug. The filename consists of the package name (defined in Cargo.toml) prefixed with "lib". The filename extension will be platform specific. On MacOS, it should be "dylib". Compilation on MacOS therefore yields the file target/debug/libdocstring.dylib.

Minor changes will be required before using the shared library. On all platforms the "lib" prefix typically needs to be removed, unless a package name like "libdocstring" was used inside py_module_initializer!. MacOS users will also need to swap the .dylib extension for .so. These changes can be accomplished by copying the shared library with a new name and/or extension. On MacOS use:

cp target/debug/libdocstring.dylib target/debug/

It is now possible to use the shared library like any other Python module.

# ./
from target.debug import docstring


Running this program yields the output "This module is implemented in Rust."

You may receive an error along the lines of "ImportError: dynamic module does not define module export function…". This error occurs when the name of the so file does not match the module name provided in the py_module_initializer macro. Be sure that both names are identical. In this case, "docstring" should be used.


Functions are implemented in a similar manner. Conveniently, ordinary Rust functions are used. However, the first argument must be of type Python, and the return type must be PyResult. The py_fn! macro wraps the function added to the module.

use cpython::{py_module_initializer, py_fn, PyResult, Python};

fn greet(_: Python, name: String) -> PyResult<String> {
    Ok(format!("Hello, {}!", name))

py_module_initializer! {
    function, |py, module| {
        module.add(py, "greet", py_fn!(py, greet(string: String)))?;


After building and copying the dynamic library as before, it can be used as an ordinary Python module.

from target.debug.function import greet



Classes are implemented using the py_class! macro. Here a hybrid language mixing Rust and Python syntax is used. The following example illustrates a class that exposes an instance method called greet.

use cpython::{py_class, PyResult, py_module_initializer};

py_class!{pub class Class |py| {
    def __new__(_cls) -> PyResult<Class> {

    def greet(&self, name: String) -> PyResult<String> {
        Ok(format!("Hello, {}!", name))

py_module_initializer! {
    klass, |py, module| {


Compile this extension as before with cargo build, then copy the resulting shared library. The module can then be used as follows.

from target.debug.klass import Class

klass = Class()


Wrapping Rust Types

The py_class! macro can wrap Rust structs containing both accessors and mutators. The following demonstrates the idea based in part on the previous article's PyO3 example.

use std::cell::RefCell;

use cpython::{py_class, py_module_initializer, PyClone, PyObject, PyResult};

type Inner = std::collections::HashSet<u32>;

py_class! {pub class SetIterator |py| {
    data iter: RefCell<std::collections::hash_set::IntoIter<u32>>;

    def __iter__(&self) -> PyResult<Self> {

    def __next__(&self) -> PyResult<Option<u32>> {

py_class! {pub class HashSet |py| {
    data hash_set: RefCell<Inner>;

    def __new__(_cls) -> PyResult<HashSet> {
        HashSet::create_instance(py, RefCell::new(Inner::new()))

    def __len__(&self) -> PyResult<usize> {

    def __contains__(&self, v: u32) -> PyResult<bool> {

    def __iter__(&self) -> PyResult<SetIterator> {

    def add(&self, v: u32) -> PyResult<PyObject> {


py_module_initializer! {
    wrapper, |py, module| {


In analogy with the previous examples, this one can be compiled into a shared library extension and run from Python.

from target.debug.wrapper import HashSet

set = HashSet()


print(f"has 0: {0 in set}")
print(f"has 1: {1 in set}")


for i in set:
    print(f"found {i}")


Most of the functionality resides on HashSet. This type uses an internal data store wrapped in a RefCell. The reason has to do with "interior mutability." In a nutshell, the dynamic environment of the Python runtime breaks Rust's compile-time memory management model. RefCell offers a mechanism to re-instate safe, dynamically-checked borrowing for those situations, like a Python extension. Interior mutability has been covered in an excellent video by Jon Gjengset

This example allows values to not only be added and queried, but also iterated. SetIterator serves this purpose.

The Problem with Iterators

The class HashSet in the previous example uses a helper class called SetIterator. It is produced from a cloned backing Rust set. Although expedient, this solution leads to something that is less flexible or efficient than it could be. Sets containing values that can not be cloned or which can only be cloned expensively would incur potentially unacceptable costs. A better solution would be to delegate to an iterator derived from a borrowed set.

The analogous PyO3 example suffered from the same limitation. As far as I can tell, PyO3 does not offer the low-level access that would be needed to address the problem. Fortunately, Rust-CPython does allow such access. A follow-up article will provide the details. If you'd like to see how this works before that, check out the article by Raphaël Gomès.


Like PyO3, Rust-CPython enables the creation of Python extensions in pure Rust. The main advantage of Rust-CPython is a lower-level API. Use of this API was demonstrated through three simple examples. A fourth example hints at one way to use the greater control afforded by Rust-CPython.