I'm seeing a lot of teams ditching the management of worker nodes entirely. With GKE Autopilot and Fargate, it feels like the "Kubernetes Administrator" role is evolving into more of a "Cloud Architect" role. Is anyone running a full production environment on serverless K8s? What are the limitations you’ve found regarding custom CNI configurations or complex sidecar patterns?
3 answers
We migrated our entire retail frontend to GKE Autopilot six months ago. The biggest benefit is that we stopped worrying about node patching and bin-packing. However, the limitation is definitely the "opinionated" nature of the platform. If you need a custom DaemonSet for a niche security agent, or if you need to tune the sysctl settings on the host, you’re out of luck. It’s perfect for 90% of web apps, but for our specialized data-scraping jobs that need high-performance networking, we still maintain a "standard" cluster where we have full control.
Kimberly, have you noticed a price premium with Autopilot? Usually, "managed" means a higher cost per vCPU compared to running your own spot instances.
We are using Knative for event-driven functions. It’s the best of both worlds—serverless scaling but with the ability to use any Docker image we want.
Knative is great for those who want that Lambda-like experience without the vendor lock-in of AWS. It really makes event-driven architecture on K8s manageable.
Jason, on paper the unit price is higher, but our total bill actually went down. Because Autopilot only bills for the resources your pods actually request (and not the idle space on a node), we stopped paying for the "empty gaps" between pods. For a spiky traffic pattern like ours, the efficiency gains outweighed the higher per-unit cost.