C Extensions & Cython

The CPython C API

A C extension compiles against Python.h and exports one initializer:

#define PY_SSIZE_T_CLEAN
#include <Python.h>

static PyObject *sum_range(PyObject *self, PyObject *args) {
    long n;
    if (!PyArg_ParseTuple(args, "l", &n)) return NULL;
    long s = 0;
    for (long i = 0; i <= n; i++) s += i;
    return PyLong_FromLong(s);
}

static PyMethodDef methods[] = {
    {"sum_range", sum_range, METH_VARARGS, "sum 0..n"},
    {NULL, NULL, 0, NULL}
};

PyMODINIT_FUNC PyInit_fastmod(void) {
    return PyModuleDef_Init(&(PyModuleDef){PyModuleDef_HEAD_INIT,
        "fastmod", NULL, -1, methods});
}

Rules that prevent crashes: every Py* call returning an object hands you a new reference you must Py_DECREF; NULL plus a set exception is the error path. Ship it with setuptools Extension("fastmod", ["fastmod.c"]).

Cython and Release Builds

Cython turns annotated .pyx files into C — most of the speed, fraction of the pain:

# fast.pyx
def clamp(double x, double lo, double hi):
    cdef double r = x if x < hi else hi    # typed locals skip object boxing
    return r if r > lo else lo

# setup.py
from setuptools import setup
from Cython.Build import cythonize
setup(ext_modules=cythonize("fast.pyx"))

Typed variables, cdef functions taking C types, and with nogil: blocks (for pure compute after references are released) remove interpreter dispatch per operation. Compile with -O3; profile first — NumPy vectorization often wins without any C.

The GIL: native code still holds it, so a CPU-bound C function blocks other threads. Release it around pure computation with Py_BEGIN_ALLOW_THREADS / Py_END_ALLOW_THREADS, but only when no Python objects are touched.

ctypes, Pitfalls, and Interviews

No compiler needed for existing shared libraries:

import ctypes
lib = ctypes.CDLL("libz.so")          # or zlib via cdll.LoadLibrary
lib.compressBound.argtypes = [ctypes.c_ulong]
lib.compressBound.restype = ctypes.c_ulong   # declare both — defaults assume int

Failure modes to name in interviews:

• Missing argtypes/restype: default int truncation of pointers and 64-bit returns → segfaults, not exceptions
• Refcount leaks in C: Py_INCREF without DECREF grows the heap invisibly — watch sys.getrefcount in tests
• Portability: extensions must be built per platform/Python ABI; use cibuildwheel to produce manylinux/musllinux/macOS wheels
• Choose the cheapest tool: array/NumPy vectorization first, Cython second, raw C last; cffi sits between ctypes and the C API when you need both speed and sanity