Amazon EKS Costs: Node Groups, Namespaces and Workloads

The two layers of an EKS bill

AWS charges every EKS cluster a per-cluster, per-hour fee based on its Kubernetes version: $0.10 per hour in standard support and $0.60 per hour in extended support. Worker capacity, the EC2 instances and EBS volumes that run your pods, is billed separately. The control plane fee continues as long as the cluster exists, so a schedule saves on nodes, not on the cluster itself. A cluster that has drifted into extended support is paying six times the standard fee for staying on an old version.

Two ways ZopNight can quiet a cluster

• Node group scaling. EKS is on the Start/Stop supported list through node pool scaling: stop scales a node group down, start brings it back.
• Namespace stop. For shared clusters, a namespace can be stopped on its own. Deployments and StatefulSets are scaled to zero and CronJobs suspended; HPAs and PodDisruptionBudgets are removed on stop and restored on start. System namespaces such as kube-system and default are never touched.

Emptying a namespace frees node capacity only if the cluster autoscaler then removes the idle nodes. Cluster start and stop run synchronously today; there is no background watcher that retries a long-running cluster operation.

ZopNight’s EKS rules

• EKS Underutilized Node needs both CPU and memory rollups from Container Insights before it fires.
• Running EKS Node Group Idle prices a node group whose nodes sit under 3% CPU.
• EKS Abandoned Node Group finds degraded node groups and prices their orphaned attachments.
• EKS Spot Node Pool for Dev/Test uses live Spot rates.
• EKS Cluster Autoscaling Not Configured, EKS Control Plane Logging Disabled and EKS Cluster Network Policy Not Enforced cover configuration.

Inside the cluster, the Kubernetes workload rules add reliability, security, idle, orphan and rightsizing checks such as Missing CPU/Memory Requests, Single Replica Deployment, Privileged Container, Unbound PVC and CPU Over-Provisioned. A single-replica fix is guided only; ZopNight never scales a workload you chose to run as a singleton without confirmation.

See the EKS cluster fact sheet, node group scheduling and namespace-level scheduling.