Yes, it works, but will it scale?
I’ve gotten so used to the reliability of Amazon autoscaling groups and load balancers that, the one time I left their comfort zone, I didn’t have a good tool to answer that question. Nor does there seem to be a simple, cheap, industry-standard solution.
You have a website or web API. You want to see how it behaves when a few thousand users hit it with simultaneous HTTP requests - specifically, how many of these requests fail or are severely delayed. You want those results displayed as pretty graphs, both for your own convenience and to impress your clients.
In addition, you don’t want to run that test from your local machine. Your data would hopelessly confuse website performance issues with your office’s faulty internet connection and the twelve tabs of YouTube videos you forgot you were streaming in the background. This is clearly a job for cloud computing.
Finally, you don’t want all these users to show up at the same time. AWS autoscaling groups are designed to handle a gradually increasing load, not a sudden 10,000 users spike. The same is true of the load balancer itself (which quietly autoscales under heavy load). Amazon recommends that you increase the load "at a rate of no more than 50 percent every five minutes". (The entire article is well worth reading.) You might also be interested in ramping down, to test scale-down behaviour.
For this basic use case, RedLine13's "simple test" proved almost sufficient.
It’s very easy to set up, and while the paid subscription is expensive, the free tier already provides most everything you need. It doesn’t have ramp-down, but that can be achieved through its Gatling integration. I wasn’t familiar with Gatling at the time though, so I looked elsewhere.
Clojure runs through the soul of JUXT, so I experimented with clj-gatling, a Clojure wrapper around Gatling.
I quickly ran into its limitations. It is not a full wrapper; in fact it covers very few of Gatling’s many functionalities. In particular, clj-gatling currently does not support time-based ramp-up. The completion-based ramp-up it offers is a poor substitute and behaves unreliably. As for clojider, its AWS integration, I did not manage to get it to work at all.
It ultimately convinced me to use Gatling directly.
Even with no knowledge of Scala, it isn’t too hard to figure out how to write simple simulation scripts. It also comes with a tool that detects your browser activity and turns it into a simulation script - though this does not play nice with HTTPS.
What it doesn’t have is a convenient way to run it remotely on AWS. RedLine13 would work, but since I didn’t realise that at the time, I ended up writing my own tool.
Ramp is an open-source shell script + AWS CloudFormation template to simplify remote load testing. In one command, it spins up an EC2 instance (or reuses the existing one if it hasn’t been terminated), sets it up for Gatling use, runs the Gatling simulation script provided, and uploads the report to an S3 bucket.
Since that report is a nicely formatted HTML file, and since S3 buckets can be used as static websites, you can see the report on your browser, with all its glorious multicoloured graphs. Or share it with a colleague or client.
Ramp comes with a Scala simulation script that does a basic load test (spamming a URL with GET requests, with ramp-up and ramp-down). I have since used it to simulate more complex behaviours, including mass registration and log-in. Writing more sample scripts is on my to-do list, but Gatling has decent documentation on its own website.
(I didn’t realise before just how easy it is to make a website unusable via DDOS attack. Poor caching, poor autoscaling, poor ephemeral port management, there are any number of potential weakest links. Implementing load testing in all JUXT projects as a matter of course is, well, a load off our minds.)
Raffi Krikorian (of Twitter and Uber fame) pointed out to us that it’s quite hard to create a simulated user that’s as, shall we say, unpredictable, as the real thing. Even if your simulation script is quite complex and makes full use of your website’s various subsystems. Ultimately, it’s still written by an engineer who’s making assumptions about how the website is supposed to be used.
Real users click a button five times in a row because they think the page is too slow to load. Or go back a page right in the middle of execution. Or refresh a resource-intensive page every second as they impatiently wait for news.
Gatling is a valuable tool to find out how your website handles load, but isn’t enough to be certain it’ll handle users. For that, you should ideally capture the behaviour of multiple real users, and replay that thousands of times on your staging site.
At JUXT, we’ve yet to tackle a project that has to scale as much as a Twitter or an Uber - but we’re slowly moving in that direction and are starting to feel a need for this sort of testing. We don’t have the tools for it but we’re investigating options.