TechEmpower.FrameworkBenchmarks/frameworks/Python

Python frameworks

The information below contains information specific to Python. For further guidance, review the documentation.

Infrastructure Software Versions

• Python2 2.7.10
• Python3 3.4.3
• PyPy 2.6.0
• nginx

Get Help

Python Experts

• INADA Naoki (@methane) -- Expert of Python and Python's MySQL driver.
• Ludovic Gasc (@GMLudo) -- Expert of AsyncIO.

Python Community

• #python IRC Channel (irc.freenode.net)
• #python-dev IRC Channel (irc.freenode.net)

Python interpreters

(C)Python 3

Newest Python.

It will be most major interpreter for Web Development in this or next year. This is a bit slower than CPython 2, but more simple and productive than it.

(C)Python 2

Legacy Python.

It is most major interpreter for now.

PyPy

PyPy is the fastest Python implementation with JIT.

There is PyPy's Python 3 implementation (PyPy3), but it is not so tuned like PyPy2. So we don't use it for now.

WSGI Servers

Nginx + uWSGI

This is one of fast and realistic way to serve Python web application.

Use unix domain socket between nginx and uWSGI to avoid additional TCP/IP overhead.

Gunicorn + Meinheld

Meinheld is very fast WSGI server.

Since Meinheld is safe against slowloris and support Keep-Alive, you can use it without buffered HTTP reverse proxy (like nginx).

We use meinheld to measure bare (without DB access) performance of framework without overhead of reverse proxying. (plaintext, json test)

Meinheld does not provide worker process management. Gunicorn provide it for Meinheld.

Gunicorn + Tornado

uWSGI + PyPy is difficult to setup. Meinheld doesn't so fast with PyPy because it uses Python/C API heavily. So we use Tornado as HTTP/WSGI server.

It supports keep-alive. So it have nice performance about json or plaintext benchmark.

Before writing new tests.

Don't increase matrix without significant purpose.

We can't maintain matrix of frameworks (5~) * servers (3~) * DBs (3) * Interpreters (3).

If you want to know about performance difference between MySQL and PostgreSQL, there are no need to all frameworks implement both of them. Additionally, Python is not a good language to measure performance of DBs because web application written in Python is slower than DBs.

If you want to benchmark http server, you should not port all tests or frameworks. Simple json or plaintext test is enough.

If your framework uses SQLAlchemy, Flask may be enough to know performance of SQLAlchemy.

Advice when writing new tests

Interpreter

Consider Python 3 first. -- Python 3 will be mainstream for web development soon.

Server

Try Gunicorn + Meinheld first. All WSGI apps in this benchmark uses it. You can compare your framework performance with others.

Files you should provide

You can see Flask's files to know how to write new test.

requirements.txt is standard way to define depending libraries.

install.sh is executed before running test. You should create virtualenv on $TROOT/py2 (or py3 for Python 3 and pypy for PyPy). Then $TROOT/py2/bin/pip install -r requirements.txt. virtualenv is installed on Python 2 and PyPy. Use $IROOT/py3/bin/python3 -m venv $TROOT/py3 for Python 3.

You can set environment variables within install.sh or within setup.sh.

bechmark_config is json file to define test. See here.

setup_py2.sh is used to run test on Python 2. gunicorn_conf.py is configuration for gunicorn. setup_py2.sh and gunicorn_conf.py are written as generic as possible. You may be able to use it with changing wsgi callable. ($PY2_GUNICORN wsgimodule:callable -c gunicorn_conf.py)