Tools for the Marketplace server

https://marketplace-cdn.addons.mozilla.net/media/img/ecosystem/open_icon.png

We had an Apps work week in Mountain View this week and worked with the fine folks from the WebDev team on the web app called Marketplace.

We're joining the Marketplace team with Alexis and a couple of other folks from Services to try to help them polishing the app before it moves from its preview stage to full speed.

Marketplace is the Django-based application WebDev wrote to allow the community to publish Open Web Apps that will be installable on your Firefox OS phone.

This app is a derivation of the addons web app, using the same basis: Zamboni -- except that add-ons are Open Web Apps in this case.

On the code side, most things are done and we'll mostly help with the bug fixing.

Although, coming from Services, the real value we can bring in this project is trying to apply on Marketplace some good practices that were successful on Services projects.

The two main things we started to work on with Alexis are load testing and connectors testing.

Load testing => Marteau

If you follow this blog you know I am a Funkload fan. Funkload is a load testing tool we're using at Services, written by a co-worker I had at Nuxeo back in the old days.

Funkload is cool for many reasons:

  • writing load tests is done by writing vanilla unit tests in Python -- no Jython involved like in the Grinder.
  • the load runner is a simple command line tool that can distribute the load across many nodes.
  • Funkload creates nice HTML reports that are most of the time sufficient to detect any issue an app has.

Alexis has started to write a couple of load tests we'll be running on Marketplace, but we thought it would be a good idea to take it to the next stage.

Several people at Mozilla have expressed interest on being able to run a load test on a web app without having to worry about finding free VMs and managing a cluster.

So we've started a very simple tool called Marteau, that will let people enqueue a load test through a web interface, and execute it for them whenever there are free resources.

We've created for this a YAML file you can simply stick into your project repository, in which you define where is your load test and how you want it to run.

Like how Travis does, any project that has this YAML file is eligible for being used by Marteau.

The Token Server for instance has one now here

The web interface will simply display the queue of pending load tests, and pick free VMs to run them, then ping you when it's done, with a link to the Funkload report.

I'll blog about it more when we have a more advanced tool.

Connections Testing => Vaurien

What happens to your web app when the SQL database is down for a couple of minutes ?

Are your connectors properly handling the case ? are you able to get back on track once the database is back online ? can you survive when memcache is down ? if not, are you sending back the proper 50x to the end user ?

To answer to all of these questions, you have to manually test each of of those scenario, which can be painful.

Or you could set up a Chaos Monkey but I suspect this is a bit overkill in most cases.

Another way to automate these tests is to run a proxy between the web app and the back end server, that breaks things on purpose.

The proxy can:

  • add delays
  • drop requests
  • send back errors
  • simulate any bad behavior we've already seen in a backend

Having this proxy up and running during a load test is the best way to detect if your web application is robust enough to handle those situations.

The tool we've started is called Vaurien and is a TCP proxy we can run in front of any backend server the web app uses.

It pings statsd everytime it does a bad thing, so we can corellate the application errors with the proxy actions.

It's also meant to be completely pluggable, so if you have a specific back end behavior you want to simulate, you can write a Python function for it and hook it via configuration.

The tool is currently using a configuration file where you define a list of behaviors -- from "normal" to "total blackout" --, and for each one of them their % of occurrences.

More on this as soon as we have a decent v1.

What's next

These two tools we're building can probably be useful to more people around the Mozilla & Python community, so if you are interested in helping building them, please ping us.

Comments !