Skip to content


Subversion checkout URL

You can clone with
Download ZIP
Tree: 217112dfee
Fetching contributors…

Cannot retrieve contributors at this time

1279 lines (958 sloc) 30.6 KB



The structure of this tutorial assumes an intermediate level knowledge of Python but not much else. No knowledge of concurrency is expected. The goal is to give you the tools you need to get going with gevent, help you tame your existing concurrency problems and start writing asynchronous applications today.


In chronological order of contribution: Stephen Diehl Jérémy Bethmont sww Bruno Bigras David Ripton Travis Cline Boris Feld youngsterxyf Eddie Hebert Alexis Metaireau Daniel Velkov

This is a collaborative document published under MIT license. Have something to add? See a typo? Fork and issue a pull request Github. Any and all contributions are welcome.



The primary pattern used in gevent is the Greenlet, a lightweight coroutine provided to Python as a C extension module. Greenlets all run inside of the OS process for the main program but are scheduled cooperatively.

Only one greenlet is ever running at any given time.

This differs from any of the real parallelism constructs provided by multiprocessing or threading libraries which do spin processes and POSIX threads which are scheduled by the operating system and are truly parallel.

Synchronous & Asynchronous Execution

The core idea of concurrency is that a larger task can be broken down into a collection of subtasks whose and scheduled to run simultaneously or asynchronously, instead of one at a time or synchronously. A switch between the two subtasks is known as a context switch.

A context switch in gevent is done through yielding. In this case example we have two contexts which yield to each other through invoking gevent.sleep(0).

[[[cog import gevent

def foo(): print('Running in foo') gevent.sleep(0) print('Explicit context switch to foo again')

def bar(): print('Explicit context to bar') gevent.sleep(0) print('Implicit context switch back to bar')

gevent.joinall([ gevent.spawn(foo), gevent.spawn(bar), ]) ]]] [[[end]]]

It is illuminating to visualize the control flow of the program or walk through it with a debugger to see the context switches as they occur.

Greenlet Control Flow

The real power of gevent comes when we use it for network and IO bound functions which can be cooperatively scheduled. Gevent has taken care of all the details to ensure that your network libraries will implicitly yield their greenlet contexts whenever possible. I cannot stress enough what a powerful idiom this is. But maybe an example will illustrate.

In this case the select() function is normally a blocking call that polls on various file descriptors.

[[[cog import time import gevent from gevent import select

start = time.time() tic = lambda: 'at %1.1f seconds' % (time.time() - start)

def gr1(): # Busy waits for a second, but we don't want to stick around... print('Started Polling: ', tic())[], [], [], 2) print('Ended Polling: ', tic())

def gr2(): # Busy waits for a second, but we don't want to stick around... print('Started Polling: ', tic())[], [], [], 2) print('Ended Polling: ', tic())

def gr3(): print("Hey lets do some stuff while the greenlets poll, at", tic()) gevent.sleep(1)

gevent.joinall([ gevent.spawn(gr1), gevent.spawn(gr2), gevent.spawn(gr3), ]) ]]] [[[end]]]

Another somewhat synthetic example defines a task function which is non-deterministic (i.e. its output is not guaranteed to give the same result for the same inputs). In this case the side effect of running the function is that the task pauses its execution for a random number of seconds.

[[[cog import gevent import random

def task(pid): """ Some non-deterministic task """ gevent.sleep(random.randint(0,2)*0.001) print('Task', pid, 'done')

def synchronous(): for i in range(1,10): task(i)

def asynchronous(): threads = [gevent.spawn(task, i) for i in xrange(10)] gevent.joinall(threads)

print('Synchronous:') synchronous()

print('Asynchronous:') asynchronous() ]]] [[[end]]]

In the synchronous case all the tasks are run sequentially, which results in the main programming blocking ( i.e. pausing the execution of the main program ) while each task executes.

The important parts of the program are the gevent.spawn which wraps up the given function inside of a Greenlet thread. The list of initialized greenlets are stored in the array threads which is passed to the gevent.joinall function which blocks the current program to run all the given greenlets. The execution will step forward only when all the greenlets terminate.

The important fact to notice is that the order of execution in the async case is essentially random and that the total execution time in the async case is much less than the sync case. In fact the maximum time for the synchronous case to complete is when each tasks pauses for 2 seconds resulting in a 20 seconds for the whole queue. In the async case the maximum runtime is roughly 2 seconds since none of the tasks block the execution of the others.

A more common use case, fetching data from a server asynchronously, the runtime of fetch() will differ between requests given the load on the remote server.

import gevent.monkey

import gevent
import urllib2
import simplejson as json

def fetch(pid):
    response = urllib2.urlopen('')
    result =
    json_result = json.loads(result)
    datetime = json_result['datetime']

    print 'Process ', pid, datetime
    return json_result['datetime']

def synchronous():
    for i in range(1,10):

def asynchronous():
    threads = []
    for i in range(1,10):
        threads.append(gevent.spawn(fetch, i))

print 'Synchronous:'

print 'Asynchronous:'


As mentioned previously, greenlets are deterministic. Given the same configuration of greenlets and the same set of inputs and they always produce the same output. For example lets spread a task across a multiprocessing pool compared to a gevent pool.

import time

def echo(i):
    return i

# Non Deterministic Process Pool

from multiprocessing.pool import Pool

p = Pool(10)
run1 = [a for a in p.imap_unordered(echo, xrange(10))]
run2 = [a for a in p.imap_unordered(echo, xrange(10))]
run3 = [a for a in p.imap_unordered(echo, xrange(10))]
run4 = [a for a in p.imap_unordered(echo, xrange(10))]

print( run1 == run2 == run3 == run4 )

# Deterministic Gevent Pool

from gevent.pool import Pool

p = Pool(10)
run1 = [a for a in p.imap_unordered(echo, xrange(10))]
run2 = [a for a in p.imap_unordered(echo, xrange(10))]
run3 = [a for a in p.imap_unordered(echo, xrange(10))]
run4 = [a for a in p.imap_unordered(echo, xrange(10))]

print( run1 == run2 == run3 == run4 )


Even though gevent is normally deterministic, sources of non-determinism can creep into your program when you begin to interact with outside services such as sockets and files. Thus even though green threads are a form of "deterministic concurrency", they still can experience some of the same problems that POSIX threads and processes experience.

The perennial problem involved with concurrency is known as a race condition. Simply put is when two concurrent threads / processes depend on some shared resource but also attempt to modify this value. This results in resources whose values become time-dependent on the execution order. This is a problem, and in general one should very much try to avoid race conditions since they result program behavior which is globally non-deterministic.*

The best approach to this is to simply avoid all global state all times. Global state and import-time side effects will always come back to bite you!

Spawning Threads

gevent provides a few wrappers around Greenlet initialization. Some of the most common patterns are:

[[[cog import gevent from gevent import Greenlet

def foo(message, n): """ Each thread will be passed the message, and n arguments in its initialization. """ gevent.sleep(n) print(message)

Initialize a new Greenlet instance running the named function


thread1 = Greenlet.spawn(foo, "Hello", 1)

Wrapper for creating and runing a new Greenlet from the named

function foo, with the passed arguments

thread2 = gevent.spawn(foo, "I live!", 2)

Lambda expressions

thread3 = gevent.spawn(lambda x: (x+1), 2)

threads = [thread1, thread2, thread3]

Block until all threads complete.

gevent.joinall(threads) ]]] [[[end]]]

In addition to using the base Greenlet class, you may also subclass Greenlet class and overload the _run method.

[[[cog import gevent from gevent import Greenlet

class MyGreenlet(Greenlet):

def __init__(self, message, n):
    self.message = message
    self.n = n

def _run(self):

g = MyGreenlet("Hi there!", 3) g.start() g.join() ]]] [[[end]]]

Greenlet State

Like any other segment of code, Greenlets can fail in various ways. A greenlet may fail to throw an exception, fail to halt or consume too many system resources.

The internal state of a greenlet is generally a time-dependent parameter. There are a number of flags on greenlets which let you monitor the state of the thread

  • started -- Boolean, indicates whether the Greenlet has been started.
  • ready() -- Boolean, indicates whether the Greenlet has halted
  • successful() -- Boolean, indicates whether the Greenlet has halted and not thrown an exception
  • value -- arbitrary, the value returned by the Greenlet
  • exception -- exception, uncaught exception instance thrown inside the greenlet

[[[cog import gevent

def win(): return 'You win!'

def fail(): raise Exception('You fail at failing.')

winner = gevent.spawn(win) loser = gevent.spawn(fail)

print(winner.started) # True print(loser.started) # True

Exceptions raised in the Greenlet, stay inside the Greenlet.

try: gevent.joinall([winner, loser]) except Exception as e: print('This will never be reached')

print(winner.value) # 'You win!' print(loser.value) # None

print(winner.ready()) # True print(loser.ready()) # True

print(winner.successful()) # True print(loser.successful()) # False

The exception raised in fail, will not propogate outside the

greenlet. A stack trace will be printed to stdout but it

will not unwind the stack of the parent.


It is possible though to raise the exception again outside

raise loser.exception

or with


]]] [[[end]]]

Program Shutdown

Greenlets that fail to yield when the main program receives a SIGQUIT may hold the program's execution longer than expected. This results in so called "zombie processes" which need to be killed from outside of the Python interpreter.

A common pattern is to listen SIGQUIT events on the main program and to invoke gevent.shutdown before exit.

import gevent
import signal

def run_forever():

if __name__ == '__main__':
    gevent.signal(signal.SIGQUIT, gevent.shutdown)
    thread = gevent.spawn(run_forever)


Timeouts are a constraint on the runtime of a block of code or a Greenlet.

import gevent
from gevent import Timeout

seconds = 10

timeout = Timeout(seconds)

def wait():

except Timeout:
    print 'Could not complete'

Or with a context manager in a with a statement.

import gevent
from gevent import Timeout

time_to_wait = 5 # seconds

class TooLong(Exception):

with Timeout(time_to_wait, TooLong):

In addition, gevent also provides timeout arguments for a variety of Greenlet and data stucture related calls. For example:

[[[cog import gevent from gevent import Timeout

def wait(): gevent.sleep(2)

timer = Timeout(1).start() thread1 = gevent.spawn(wait)

try: thread1.join(timeout=timer) except Timeout: print('Thread 1 timed out')


timer = Timeout.start_new(1) thread2 = gevent.spawn(wait)

try: thread2.get(timeout=timer) except Timeout: print('Thread 2 timed out')


try: gevent.with_timeout(1, wait) except Timeout: print('Thread 3 timed out')

]]] [[[end]]]

Data Structures


Events are a form of asynchronous communication between Greenlets.

import gevent
from gevent.event import AsyncResult

a = AsyncResult()

def setter():
    After 3 seconds set wake all threads waiting on the value of

def waiter():
    After 3 seconds the get call will unblock.
    a.get() # blocking
    print 'I live!'


A extension of the Event object is the AsyncResult which allows you to send a value along with the wakeup call. This is sometimes called a future or a deferred, since it holds a reference to a future value that can be set on an arbitrary time schedule.

import gevent
from gevent.event import AsyncResult
a = AsyncResult()

def setter():
    After 3 seconds set the result of a.

def waiter():
    After 3 seconds the get call will unblock after the setter
    puts a value into the AsyncResult.
    print a.get()



Queues are ordered sets of data that have the usual put / get operations but are written in a way such that they can be safely manipulated across Greenlets.

For example if one Greenlet grabs an item off of the queue, the same item will not grabbed by another Greenlet executing simultaneously.

[[[cog import gevent from gevent.queue import Queue

tasks = Queue()

def worker(n): while not tasks.empty(): task = tasks.get() print('Worker %s got task %s' % (n, task)) gevent.sleep(0)

print('Quitting time!')

def boss(): for i in xrange(1,25): tasks.put_nowait(i)


gevent.joinall([ gevent.spawn(worker, 'steve'), gevent.spawn(worker, 'john'), gevent.spawn(worker, 'nancy'), ]) ]]] [[[end]]]

Queues can also block on either put or get as the need arises.

Each of the put and get operations has a non-blocking counterpart, put_nowait and get_nowait which will not block, but instead raise either gevent.queue.Empty or gevent.queue.Full in the operation is not possible.

In this example we have the boss running simultaneously to the workers and have a restriction on the Queue that it can contain no more than three elements. This restriction means that the put operation will block until there is space on the queue. Conversely the get operation will block if there are no elements on the queue to fetch, it also takes a timeout argument to allow for the queue to exit with the exception gevent.queue.Empty if no work can found within the time frame of the Timeout.

[[[cog import gevent from gevent.queue import Queue, Empty

tasks = Queue(maxsize=3)

def worker(n): try: while True: task = tasks.get(timeout=1) # decrements queue size by 1 print('Worker %s got task %s' % (n, task)) gevent.sleep(0) except Empty: print('Quitting time!')

def boss(): """ Boss will wait to hand out work until a individual worker is free since the maxsize of the task queue is 3. """

for i in xrange(1,10):
print('Assigned all work in iteration 1')

for i in xrange(10,20):
print('Assigned all work in iteration 2')

gevent.joinall([ gevent.spawn(boss), gevent.spawn(worker, 'steve'), gevent.spawn(worker, 'john'), gevent.spawn(worker, 'bob'), ]) ]]] [[[end]]]

Groups and Pools

A group is a collection of running greenlets which are managed and scheduled together as group. It also doubles as parallel dispatcher that mirrors the Python multiprocessing library.

[[[cog import gevent from gevent.pool import Group

def talk(msg): for i in xrange(3): print(msg)

g1 = gevent.spawn(talk, 'bar') g2 = gevent.spawn(talk, 'foo') g3 = gevent.spawn(talk, 'fizz')

group = Group() group.add(g1) group.add(g2) group.join()

group.add(g3) group.join() ]]] [[[end]]]

This is very usefull for managing groups of asynchronous tasks that.

As mentioned above Group also provides an API for dispatching jobs to grouped greenlets and collecting their results in various ways.

[[[cog import gevent from gevent import getcurrent from gevent.pool import Group

group = Group()

def hello_from(n): print('Size of group', len(group)) print('Hello from Greenlet %s' % id(getcurrent())), xrange(3))

def intensive(n): gevent.sleep(3 - n) return 'task', n


ogroup = Group() for i in ogroup.imap(intensive, xrange(3)): print(i)


igroup = Group() for i in igroup.imap_unordered(intensive, xrange(3)): print(i)

]]] [[[end]]]

A pool is a structure designed for handling dynamic numbers of greenlets which need to be concurrency-limited. This is often desirable in cases where one wants to do many network or IO bound tasks in parallel.

[[[cog import gevent from gevent import getcurrent from gevent.pool import Pool

pool = Pool(2)

def hello_from(n): print('Size of pool', len(pool)), xrange(3)) ]]] [[[end]]]

Often when building gevent driven services one will center the entire service around a pool structure. An example might be a class which polls on various sockets.

from gevent.pool import Pool

class SocketPool(object):

    def __init__(self):
        self.pool = Pool(1000)

    def listen(self, socket):
        while True:

    def add_handler(self, socket):
        if self.pool.full():
            raise Exception("At maximum pool size")
            self.pool.spawn(self.listen, socket)

    def shutdown(self):

Locks and Semaphores

A semaphore is a low level synchronization primitive that allows greenlets to coordinate and limit concurrent access or execution. A semaphore exposes two methods, acquire and release The difference between the number of times and a semaphore has been acquired and released is called the bound of the semaphore. If a semaphore bound reaches 0 it will block until another greenlet releases its acquisition.

[[[cog from gevent import sleep from gevent.pool import Pool from gevent.coros import BoundedSemaphore

sem = BoundedSemaphore(2)

def worker1(n): sem.acquire() print('Worker %i acquired semaphore' % n) sleep(0) sem.release() print('Worker %i released semaphore' % n)

def worker2(n): with sem: print('Worker %i acquired semaphore' % n) sleep(0) print('Worker %i released semaphore' % n)

pool = Pool(), xrange(0,2)), xrange(3,6)) ]]] [[[end]]]

A semaphore with bound of 1 is known as a Lock. it provides exclusive execution to one greenlet. They are often used to ensure that resources are only in use at one time in the context of a program.

Thread Locals


The actor model is a higher level concurrency model popularized by the language Erlang. In short the main idea is that you have a collection of independent Actors which have an inbox from which they receive messages from other Actors. The main loop inside the Actor iterates through its messages and takes action according to its desired behavior.

Gevent does not have a primitive Actor type, but we can define one very simply using a Queue inside of a subclassed Greenlet.

import gevent
from gevent.queue import Queue

class Actor(gevent.Greenlet):

    def __init__(self):
        self.inbox = Queue()

    def receive(self, message):
        Define in your subclass.
        raise NotImplemented()

    def _run(self):
        self.running = True

        while self.running:
            message = self.inbox.get()

In a use case:

import gevent
from gevent.queue import Queue
from gevent import Greenlet

class Pinger(Actor):
    def receive(self, message):
        print message

class Ponger(Actor):
    def receive(self, message):
        print message

ping = Pinger()
pong = Ponger()


gevent.joinall([ping, pong])

Real World Applications

Gevent ZeroMQ

ZeroMQ is described by its authors as "a socket library that acts as a concurrency framework". It is a very powerful messaging layer for building concurrent and distributed applications.

ZeroMQ provides a variety of socket primitives, the simplest of which being a Request-Response socket pair. A socket has two methods of interest send and recv, both of which are normally blocking operations. But this is remedied by a briliant library by Travis Cline which uses gevent.socket to poll ZeroMQ sockets in a non-blocking manner. You can install gevent-zeromq from PyPi via: pip install gevent-zeromq


Note: Remember to pip install pyzmq gevent_zeromq

import gevent from gevent_zeromq import zmq

Global Context

context = zmq.Context()

def server(): server_socket = context.socket(zmq.REQ) server_socket.bind("tcp://")

for request in range(1,10):
    print('Switched to Server for ', request)
    # Implicit context switch occurs here

def client(): client_socket = context.socket(zmq.REP) client_socket.connect("tcp://")

for request in range(1,10):

    print('Switched to Client for ', request)
    # Implicit context switch occurs here

publisher = gevent.spawn(server) client = gevent.spawn(client)

gevent.joinall([publisher, client])

]]] [[[end]]]

Simple Telnet Servers

# On Unix: Access with ``$ nc 5000`` 
# On Window: Access with ``$ telnet 5000`` 

from gevent.server import StreamServer

def handle(socket, address):
    socket.send("Hello from a telnet!\n")
    for i in range(5):
        socket.send(str(i) + '\n')

server = StreamServer(('', 5000), handle)

WSGI Servers

Gevent provides two WSGI servers for serving content over HTTP. Henceforth called wsgi and pywsgi:

  • gevent.wsgi.WSGIServer
  • gevent.pywsgi.WSGIServer

In earlier versions of gevent before 1.0.x, gevent used libevent instead of libev. Libevent included a fast HTTP server which was used by gevent's wsgi server.

In gevent 1.0.x there is no http server included. Instead gevent.wsgi is now an alias for the pure Python server in gevent.pywsgi.

Streaming Servers

If you are using gevent 1.0.x, this section does not apply

For those familiar with streaming HTTP services, the core idea is that in the headers we do not specify a length of the content. We instead hold the connection open and flush chunks down the pipe, prefixing each with a hex digit indicating the length of the chunk. The stream is closed when a size zero chunk is sent.

HTTP/1.1 200 OK
Content-Type: text/plain
Transfer-Encoding: chunked




The above HTTP connection could not be created in wsgi because streaming is not supported. It would instead have to buffered.

from gevent.wsgi import WSGIServer

def application(environ, start_response):
    status = '200 OK'
    body = '<p>Hello World</p>'

    headers = [
        ('Content-Type', 'text/html')

    start_response(status, headers)
    return [body]

WSGIServer(('', 8000), application).serve_forever()

Using pywsgi we can however write our handler as a generator and yield the result chunk by chunk.

from gevent.pywsgi import WSGIServer

def application(environ, start_response):
    status = '200 OK'

    headers = [
        ('Content-Type', 'text/html')

    start_response(status, headers)
    yield "<p>Hello"
    yield "World</p>"

WSGIServer(('', 8000), application).serve_forever()

But regardless, performance on Gevent servers is phenomenal compared to other Python servers. libev is a very vetted technology and its derivative servers are known to perform well at scale.

To benchmark, try Apache Benchmark ab or see this Benchmark of Python WSGI Servers for comparison with other servers.

$ ab -n 10000 -c 100

Long Polling

import gevent
from gevent.queue import Queue, Empty
from gevent.pywsgi import WSGIServer
import simplejson as json

data_source = Queue()

def producer():
    while True:
        data_source.put_nowait('Hello World')

def ajax_endpoint(environ, start_response):
    status = '200 OK'
    headers = [
        ('Content-Type', 'application/json')

    start_response(status, headers)

    while True:
            datum = data_source.get(timeout=5)
            yield json.dumps(datum) + '\n'
        except Empty:


WSGIServer(('', 8000), ajax_endpoint).serve_forever()


Websocket example which requires gevent-websocket.

# Simple gevent-websocket server
import json
import random

from gevent import pywsgi, sleep
from geventwebsocket.handler import WebSocketHandler

class WebSocketApp(object):
    '''Send random data to the websocket'''

    def __call__(self, environ, start_response):
        ws = environ['wsgi.websocket']
        x = 0
        while True:
            data = json.dumps({'x': x, 'y': random.randint(1, 5)})
            x += 1

server = pywsgi.WSGIServer(("", 10000), WebSocketApp(),

HTML Page:

        <title>Minimal websocket application</title>
        <script type="text/javascript" src="jquery.min.js"></script>
        <script type="text/javascript">
        $(function() {
            // Open up a connection to our server
            var ws = new WebSocket("ws://localhost:10000/");

            // What do we do when we get a message?
            ws.onmessage = function(evt) {
                $("#placeholder").append('<p>' + + '</p>')
            // Just update our conn_status field with the connection status
            ws.onopen = function(evt) {
            ws.onerror = function(evt) {
            ws.onclose = function(evt) {
        <h1>WebSocket Example</h1>
        <div id="conn_status">Not Connected</div>
        <div id="placeholder" style="width:600px;height:300px;"></div>

Chat Server

The final motivating example, a realtime chat room. This example requires Flask ( but not neccesarily so, you could use Django, Pyramid, etc ). The corresponding Javascript and HTML files can be found here.

# Micro gevent chatroom.
# ----------------------

from flask import Flask, render_template, request

from gevent import queue
from gevent.pywsgi import WSGIServer

import simplejson as json

app = Flask(__name__)
app.debug = True

rooms = {
    'topic1': Room(),
    'topic2': Room(),

users = {}

class Room(object):

    def __init__(self):
        self.users = set()
        self.messages = []

    def backlog(self, size=25):
        return self.messages[-size:]

    def subscribe(self, user):

    def add(self, message):
        for user in self.users:
            print user

class User(object):

    def __init__(self):
        self.queue = queue.Queue()

def choose_name():
    return render_template('choose.html')

def main(uid):
    return render_template('main.html',

def join(room, uid):
    user = users.get(uid, None)

    if not user:
        users[uid] = user = User()

    active_room = rooms[room]
    print 'subscribe', active_room, user

    messages = active_room.backlog()

    return render_template('room.html',
        room=room, uid=uid, messages=messages)

@app.route("/put/<room>/<uid>", methods=["POST"])
def put(room, uid):
    user = users[uid]
    room = rooms[room]

    message = request.form['message']
    room.add(':'.join([uid, message]))

    return ''

@app.route("/poll/<uid>", methods=["POST"])
def poll(uid):
        msg = users[uid].queue.get(timeout=10)
    except queue.Empty:
        msg = []
    return json.dumps(msg)

if __name__ == "__main__":
    http = WSGIServer(('', 5000), app)

Jump to Line
Something went wrong with that request. Please try again.