mhausenblas / kboom

The Kubernetes scale & soak load tester

A simple Kubernetes load testing tool

NOTE: this is WIP and also this is not an official AWS tool. Provided as is and use at your own risk.

Think of kboom as the Kubernetes equivalent of boom, allowing you to create short-term load for scale testing and long-term load for soak testing. Supported load out of the box for scale testing are pods and custom resources via CRDs for soak testing is planned.

Check out the interactive demo.

Why bother?

I didn't find a usable tool to do Kubernetes-native load testing, for scalability and/or soak purposes. Here's where I can imagine kboom might be useful for you:

• You are a cluster admin and want to test how much "fits" in the cluster. You use kboom for a scale test and see how many pods can be placed and how long it takes.
• You are a cluster or namespace admin and want to test how long it takes to launch a set number of pods in a new cluster, comparing it with what you already know from an existing cluster.
• You are developer and want to test your custom controller or operator. You use kboom for a long-term soak test of your controller.

Install

Before you begin, you will need kubectl client version v1.12.0 or higher for kubectl plugin support.

To install kboom, do the following:

$ curl https://raw.githubusercontent.com/mhausenblas/kboom/master/kboom -o kubectl-kboom
$ chmod +x kubectl-kboom
$ sudo mv ./kubectl-kboom /usr/local/bin

From this point on you can use it as a kubectl plugin as in kubectl kboom. However, in order for you to generate the load, you'll have to also give it the necessary permissions (note: you only need to do this once, per cluster):

$ kubectl create ns kboom
$ kubectl apply -f https://raw.githubusercontent.com/mhausenblas/kboom/master/permissions.yaml

Now you're set up and good to go, next up, learn how to use kboom.

Use

Here's how you'd use kboom to do some scale testing. The load test is run in-cluster as a Kubernetes job so you do multiple runs and compare outcomes in a straight-forward manner. Note that by default kboom assumes there's a namespace kboom available and it will run in this namespace. If this namespace doesn't exist, create it with kubectl create ns kboom or otherwise use the --namespace parameter to overwrite it.

So, first we use the generate command to generate the load, launching 10 pods (that is, using busybox containers that just sleep) with a timeout of 14 seconds (that is, if a pod is not running within that time, it's considered a failure):

$ kubectl kboom generate --mode=scale:14 --load=pods:10
job.batch/kboom created

From now on you can execute the results command as often as you like, you can see the live progress there:

$ kubectl kboom results
Server Version: v1.12.6-eks-d69f1b
Running a scale test, launching 10 pod(s) with a 14s timeout ...

-------- Results --------
Overall pods successful: 6 out of 10
Total runtime: 14.061988653s
Fastest pod: 9.003997546s
Slowest pod: 13.003831951s
p50 pods: 12.003529448s
p95 pods: 13.003831951s

When you're done, and don't need the results anymore, use kubectl kboom cleanup to get rid of the run. Note: should you execute the cleanup command too soon for kboom to terminate all its test pods, you can use kubectl delete po -l=generator=kboom to get rid of all orphaned pods.

Known issues and plans

• Need to come up with stricter permissions, currently too wide and not following the least privileges principle.
• Add support for custom resources and soak testing (running for many hours).
• Add support for other core resources, such as services or deployments.