Self-managed k8s sucks. EKS isn't that.

EKS used to be an instant no from us. Running Kubernetes yourself means babysitting a control plane at 3am, and we'll take Lambda and a queue over that any day. But at a certain scale the math flips, and we finally get why teams reach for it.

The compute ladder

We pick compute in a fixed order. Lambda first. If the work doesn't fit a function, drop down to containers on Fargate. Small team, a handful of services? That's the whole story. Stop there, you're done.

Where it breaks

Then you scale, and the ladder snaps. A JVM service that needs 90 seconds to warm up. A model server that wants a GPU. A vendor tool that only ships as a Helm chart. None of that fits a function. Now hand 30 teams their own setup and watch one security fix turn into a pull request in 30 repos. Run hundreds of services flat out and the Fargate premium stops feeling convenient and starts feeling like money you're lighting on fire.

EKS is not self-managed Kubernetes

Here's what changed our minds. Every nightmare up top, the control plane, the etcd backups, the 3am upgrades, comes from running Kubernetes yourself. EKS hands all of it to AWS. They run the control plane across AZs and charge you ten cents an hour for it. That's 73 bucks a month. For an org with hundreds of services that's a rounding error, and the thing that used to need a whole team is now a line on the bill.

That's the part that flipped us. The rest of the post is the actual blueprint: Flux syncing everything from Git, Karpenter picking your nodes and parking most of them on Spot to gut the bill, and the trick that lets a brand-new cluster boot itself. Plus the spots where it bites.

🚀 Read the full blueprint →