KubeCon + CloudNativeCon North America 2024

Databricks manages over a thousand k8s clusters across three major cloud providers which run critical workloads in cloud regions around the world. This talk describes the system we built to upgrade nodes’ operating system, k8s version, and other configs monthly, supporting EKS, AKS, GKE, and self-managed k8s. Our system is built on k8s operators and performs zero-downtime blue-green rolling updates, respects contracts with services with features like PDBs, maintenance windows, deferred node draining, and custom workload handling plugins. It enables easy rollbacks, has good observability, and incurs minimal human operational cost. This has allowed us to patch vulnerabilities and release infrastructure changes quickly and reliably across the fleet. We will also share our lessons learned on building several operators that work together using the controller-runtime framework, designing the declarative interfaces between them, and achieving consistent behavior across three clouds.