Benchmark of Django and Python's cached_property decorators (with IO and multithreading)

bench_cached_property_decorator.py

import pyperf
import time
from concurrent.futures import ThreadPoolExecutor
from django.utils.functional import cached_property as dj_cached_property
from functools import cached_property as py_cached_property


# Most web requests have some IO, like database access. Simulate with sleeping
SLEEP_TIME = 0.01


class TestClass:
    @dj_cached_property
    def dj_cached(self):
        time.sleep(SLEEP_TIME)
        return "Test"

    @py_cached_property
    def py_cached(self):
        time.sleep(SLEEP_TIME)
        return "Test"


def test_django():
    t = TestClass()
    t.dj_cached
    t.dj_cached
    t.dj_cached
    t.dj_cached
    t.dj_cached
    t.dj_cached
    t.dj_cached
    t.dj_cached
    t.dj_cached
    t.dj_cached


def test_python():
    t = TestClass()
    t.py_cached
    t.py_cached
    t.py_cached
    t.py_cached
    t.py_cached
    t.py_cached
    t.py_cached
    t.py_cached
    t.py_cached
    t.py_cached


def django_cache(loops):
    range_it = range(loops)
    t0 = pyperf.perf_counter()

    for loop in range_it:
        with ThreadPoolExecutor() as executor:
            executor.submit(test_django)
            executor.submit(test_django)
            executor.submit(test_django)
            executor.submit(test_django)
            executor.submit(test_django)
            executor.submit(test_django)
            executor.submit(test_django)
            executor.submit(test_django)
            executor.submit(test_django)
            executor.submit(test_django)

    return pyperf.perf_counter() - t0


def python_cache(loops):
    range_it = range(loops)
    t0 = pyperf.perf_counter()

    for loop in range_it:
        with ThreadPoolExecutor() as executor:
            executor.submit(test_python)
            executor.submit(test_python)
            executor.submit(test_python)
            executor.submit(test_python)
            executor.submit(test_python)
            executor.submit(test_python)
            executor.submit(test_python)
            executor.submit(test_python)
            executor.submit(test_python)
            executor.submit(test_python)

    return pyperf.perf_counter() - t0


runner = pyperf.Runner()
runner.bench_time_func("Django Cache", django_cache)
runner.bench_time_func("Python Cache", python_cache)

I modified the original benchmark to include IO and multithreading, since that's presumably the context in most Django applications.

After these modifications, the benchmark shows that the perfomance of stdlib's cached_property decorator is worse by an order of magnitude (at least on my laptop):

% python benchmark.py
.....................
Django Cache: Mean +- std dev: 12.8 ms +- 0.2 ms
.....................
Python Cache: Mean +- std dev: 113 ms +- 2 ms

This is because of unsuitable locking implementation in cached_property. See https://bugs.python.org/issue43468

Very interesting benchmark! Any idea how does the implementation in the cached-property package performs?