Thoughts on Load Testing

We are Funkload fans at Mozilla Services. Writing a load test against one of our web service is dead easy using that tool.

It boils down to writing a functional test in Python, that calls the service, then ask Funkload to run it in parallel, accross many threads and boxes.

We've written a web app on the top of Funkload called Marteau - initial blog post about it that now allows us to provision slaves on demand on Amazon WS and run distributed load tests.

You can watch this hangout to better understand what Marteau & Funkload are https://plus.google.com/106436370949746015255/posts/Lq7t4jkiwNR

Marteau is already useful to us but I am now facing some difficulties to improve the tool and add some features we want, like:

  • realtime JS charts of the ongoing load test
  • web socket load testing
  • the ability to run load tests written in other languages or frameworks, like node.js

Funkload is a project that was started 10 years ago, and is very robust. But it has evolved from a tool that runs on a single box to a distributed load testing tool by using SSH. In other words, a distributed load test in Funkload is just several load tests running in parallel, then a merge of all the results files that are copied back to a single box through SSH.

It means that there is no specific client/server protocol and no way to interact live with the slaves that are running in the distributed mode. You can kill them of course, or just wait for them to return.

I started to add some communication channel in the tool but the core itself was not built with this in mind.

I have found the locust.io project, which has this approach and which looks very promising, but does not exactly provide what I am looking for.

For example, I don't really want developers to have to write load tests using yet another set of APIs. The concept of writing Python unit tests is fabulous and I don't want to lose it.

Starting Loads

I am experimenting on something new, based on what we've learned from our experience with Funkload and what we need in Marteau.

It's called Loads and it's a client/server architecture based on ZMQ that will use a very simple protocol based on Message Pack or maybe BJson - we will see.

It's quite similar to locust.io in the principles, but instead of introducing new APIS, it's going to rely on a set of API people know & like : Requests.

So, how will a load test with Loads look like ?

import unittest
from loads import Session


class TestWebSite(unittest.TestCase):

    def setUp(self):
        self.session = Session()

    def test_something(self):
        res = self.session.get('http://blog.ziade.org')
        self.assertTrue('Tarek' in res.content)

That's it. A unittest class that uses a Session class. The Session object is the same one you find in Requests.

I am not sure yet how I will extend the API so it can work with web sockets. That'll come later.

This test can then be executed using the loads command. Example for 10 concurrent users and 10 runs each:

$ bin/loads test_blog.TestWebSite.test_something -c 10 -n 10
[====================================================================================================]  100%

Like locust.io, Loads uses greenlets here, so you can already push quite some load from a single box.

Everytime a request is done, the status code returned by the server and the time it took are pushed in a stream. The stream can be the standard output or a ZMQ stream.

And I will be using the ZMQ stream to actually drive distributed tests.

Each agent will connect to a master through ZMQ. The master will be able to drive them through a dedicated ROUTE socket and will ask agents to run some load tests.

Results will be sent back via ZMQ in real-time using a dedicated channel.

The master will then publish all results in a merged stream - a ZMQ pub socket.

From there, the Marteau web app will be able to register to that stream of result to display some real time charts and allow any kind of interaction. Or any app that whishes to do something with the results.

And since there are some ZMQ bindings in most languages, it's possible to implement a node.js client for example, so the system can have agents able to run Javascript-based tests and report back results to the master like the built-in Python agent.

That's the plan. I have started the prototype here: https://github.com/tarekziade/loads and I am very excited about this project.