Running the Runners: How CNCF Powers GitHub Actions at Scale
summary
CNCF runs a shared hosted runner platform on donated Oracle Cloud credits using Argo CD, actions-runner-controller and packer to produce packer-generated VM images, supporting 230+ projects and ~200 million GitHub Actions minutes in 2025. A custom pod-to-VM launcher boots VMs on demand (startup ~5 min, now ~2 min with a KubeVirt POC), cut projected costs from roughly $1,000/day to ~$1,000/month, and mostly works but still faces networking, nested-virt and cleanup edge cases.
AI-written summary · about
chapters (23)
- 0:00CNCF runner intro
- 1:07CNCF projects' needs
- 3:36Donated cloud challenges
- 5:46CI infrastructure needs
- 6:29GitHub Actions as CNCF CI
- 8:37GitHub policy changes
- 11:03GitHub Actions Runner Controller
- 13:32Argo CD automation
- 13:55Container runner limits
- 15:42Transitioning from containers to virtual machines
- 18:13Custom VM images
- 20:01GHA Runner VM builds
- 21:46VM runner performance
- 23:32Exploring KubeVirt for faster VM provisioning
- 24:17Cluster ops challenges
- 25:42CI monitoring and alerts
- 27:04Component failures
- 28:10Cgroups compatibility
- 29:12Success and improvements
- 30:14Key decisions timeline
- 31:40Current and future infra
- 32:43Agnostic cloud strategy
- 33:04Closing and Q&A
related talks
36mRunning the Runners: How CNCF Powers GitHub Actions at Scale
41mScaling GitHub Actions To the CNCF
34mMigrating GitHub Actions with Nested Virtualization to a Cloud Native Ecosystem
50mGitHub Actions Self-Hosting Runners in k8s
43mRewriting an Old CI/CD on GitHub Actions for 280 Contributors?
30mPragmatic, scalable, and ergonomic CI/CD using Github Actions and K8s.
30mSpend More Time Coding & Automate the Boring Stuff
29mDeploying CI/CD Infrastructure with Terraform and Packer