Migrating to GKE: Preemptible nodes and making space for the Chaos Monkeys

Expel migrated its legacy tasking infrastructure from statically provisioned VMs to Google Kubernetes Engine to achieve high availability, automatic scaling, and major cost reductions. The post explains using regional GKE clusters, cluster autoscaler, vertical-pod-autoscaler, preemptible node pools with non-preemptible fallbacks, node termination handling, and chaos-engineering principles to run largely stateless task-executor workloads reliably and cheaply, outlines trade-offs and future improvements, and reports roughly 70% compute cost savings.