The configuration of my home bare metal kubernetes cluster

NOTE There are some security implications to be aware of if you are copying this configuration! See below.

Master Node (galactica): Dell R610, Dual Xeon X5670 @ 2.93 GHz, 120 GB RAM

Worker Nodes (starbuck, viper): Dell R610, Dual Xeon X5670 @ 2.93 GHz, 72 GB RAM

Labgrid Coordinator Edge Node (coordinator-labgrid): Dell Optiplex 7040, i7-6700 @ 3.7 GHz, 16 GB RAM

Labgrid Exporter Edge Node (pegasus-labgrid): Rapsberry Pi 4, 8 GB RAM

The labgrid nodes are attached to my PiRATE test boards

The primary purpose of this cluster is to demonstrate a Kubernetes based CI setup for building and testing embedded products with the Yocto Project

The idea is that you can describe and entire CI pipeline that will run anywhere (Cloud Provider, Bare Metal, etc.) with minimal changes. The CI pipeline should be able to build everything required, and also run CI tests on remote devices that do not need to be co-located with the cluster (e.g. imagine you wanted to do your builds in the cloud, but have it test on hardware located next to your desk)

To test out these concepts, this cluster is setup to build a Yocto DOOM Demo, but could be configured to build other projects

The CI pipeline is setup using Tekton. Builds are fairly straight forward, with typical clone/build tasks. Testing of remote devices is using Labgrid. The Tekton pipeline runs tests using pytest with the Labgrid plugin, and it connects to a remote Labgrid coordinator over ssh. The coordinator is also the SSH proxy for the labgrid exporter(s). The Labgrid nodes are attached to the cluster as edge nodes using KubeEdge, which allows orchestrating them using Kubernetes. Finally, testing of x86 emulated devices is done using KubeVirt to allow the testing image to run QEMU using KVM acceleration.

Mostly everything works now, but I have a few things I'd like to do yet:

• Get webhooks working with GitHub/Tekton so it will build from GitHub events (see the security implications)
• Figure out a way to automatically cleanup old builds. My cluster has pretty limited disk space. Right now, I have to manually delete all the PipelineRuns before I trigger more or I run our of PersistantVolumeClaims. Ideally, it would only keep around one of each build type and delete them when a new on appears
• Try building on CEPH volumes using ROOK. This would give a more "cloud" like expirence, since it would allow a pipeline's tasks to be run on different nodes (currently, a pipeline is locked to a specific node because it's using a local volume). This would optimize build scheduling, but I'm curious how the I/O will affect build times. I'm waiting on this until I get a LCAP-enabled switch so I can gang 4x1GbE links between build nodes
• Figure out a better way of managing the Kubernetes deployment. I'm a complete K8s noob, but I couldn't find anything that easily let me describe my entire clusters configuration and manage it (I've already rebuilt the entire thing 2 or 3 times) sanely, so I wrote the deploy script. I don't like it... someone please point me to something better.

• The build pipeline is setup to use Tekton secrets for the labgrid SSH private key. However, since it is executing code from a GitHub repo, it is trivial to exfiltrate those keys outside of the cluster which would give anyone access to your Labgrid cluster. I would not recommend running these pipelines on any untrusted repos (e.g. pull requests). It's on my TODO list to resolve this, probably by moving the keys to a sidecar running SSH agent

I needed a naming scheme, so I named my cluster nodes after Battlestar Galactica characters/ships