Does this Karpenter and Argo CD setup make sense for EKS?

November 16, 2026 4 answers
Asked by m.593~buttonkestrel

I'm moving workloads from a managed node group to Karpenter on EKS. The plan is to keep a small managed node group for the Karpenter controller, Argo CD and critical cluster components, with Karpenter deployed as an app through Argo CD.

Does that sound like a reasonable setup? What should I watch out for during bootstrap and recovery?

Answers

v.408~drumladle

Yes, that's a reasonable split. Keep enough managed node group capacity for Karpenter, Argo CD and the cluster components they need, then let Karpenter provide capacity for ordinary workloads. The controller needs somewhere to run even when its NodePools have no nodes.

Think through the bootstrap order too. Argo CD needs compute before it can deploy Karpenter, and Karpenter needs its configuration before it can provision workload nodes. Managing the controller as an Argo CD app is fine, but Argo CD isn't required for the controller to keep running once it's deployed.

9r+foldbasket_354:

I'd also keep a manual way to scale the managed node group without relying on Karpenter. That gives you a recovery path if the controller is down.

Reply to this answer
q.369+bracketfaucet

Keep the static group small and protect it from becoming a landing spot for every workload. Taints and appropriate tolerations can reserve it for the components that actually need it. Check placement for Argo CD and the controller rather than assuming they'll land there.

For disruption, consider a PodDisruptionBudget for components such as the Argo CD repo server, especially if they're running on nodes subject to consolidation. A budget needs to fit the replica count and availability you want, and it won't protect against every kind of disruption.

Reply to this answer
4u!ferrulewinder_608

Fargate is another option for the controller and components such as CoreDNS if you want them separate from node management. Check the requirements of each component first, though. Fargate doesn't run DaemonSets, so a managed node group may be the simpler choice for system components that need them.

Reply to this answer
n.472!buscrank

Test the failure path before calling the migration done. Let the Karpenter-managed capacity reach zero, then test losing a static node and see whether the controller and Argo CD still have somewhere to run and can restore workload capacity. Steady state alone won't tell you much about bootstrap dependencies.

Reply to this answer

Add your answer

Answers are reviewed before publication.

Related Discussions