Knative

Knative Serving brings serverless, scale-to-zero workloads to Kubernetes. Instead of paying for idle pods, your services sleep and wake up on demand. Like AWS Lambda but on your own infrastructure.

Apache 2.0. CNCF incubating project. Knative handles request-based autoscaling (including scale-to-zero), traffic splitting for canary deployments, automatic TLS, and revision management. You deploy a service, and Knative manages the lifecycle.

Fully free. No paid tier from the project. Google Cloud Run is built on Knative if you want managed.

The catch: Knative adds significant complexity to your cluster. It requires a networking layer (Istio, Kourier, or Contour), and the cold start latency when scaling from zero can be seconds, not milliseconds. If your services need instant response times, scale-to-zero defeats the purpose. And the resource overhead of Knative's own components (controller, activator, autoscaler, networking) means it only makes sense if you're running enough services that the savings from scale-to-zero outweigh the platform cost.