GKE Cost Control: Node Pools, Namespaces, Fee
The cluster fee is fixed; the nodes underneath it are where schedules and rightsizing pay off.
Three parts of a GKE bill
The GKE pricing page separates the charges:
- Cluster management fee: a flat $0.10 per cluster per hour, charged in 1-second increments, for every GKE cluster regardless of mode, size or topology.
- Free tier: $74.40 in monthly credits per billing account, equivalent to one zonal Standard or Autopilot cluster, applied only to zonal and Autopilot clusters and not rolled over.
- Compute: in Standard node pools you pay for the underlying Compute Engine instances until the nodes are deleted. Autopilot workloads are priced from the compute resources the Pods request.
The practical consequence: a Standard cluster with nodes scaled down still pays its management fee, and the saving from a quiet night lives almost entirely in the node VMs.
Where ZopNight acts
GKE is on ZopNight’s start/stop list. The finer-grained option is namespace scheduling: a namespace Stop scales every Deployment and StatefulSet in it to 0 and suspends its CronJobs, saving prior replica counts, HPA and PodDisruptionBudget settings so Start restores them. System namespaces such as kube-* and default are refused before any API call. That lets a shared dev cluster keep running for one team while another team’s namespace sleeps.
On the recommendation side, GKE gets two families. Infrastructure rules from the GCP catalog include GKE Spot Node Pool Opportunity, GKE Node Pool Underutilized and GKE Node Pool VM-Size Rightsizing, plus posture checks like GKE Node Auto-Upgrade Not Configured, GKE Cluster Network Policy Not Enforced and GKE Cluster Not Private. Workload rules come from the shared Kubernetes framework, instantiated for GKE: Missing CPU/Memory Requests, Single Replica Deployment, HPA At Max Capacity, Unbound PVC, and CPU or memory over-provisioning read from HPA ScalingLimited signals.
The single-replica fix is worth knowing about: it scales 1 to 2 through a live manifest edit, and ZopNight treats it as guided-only, never automatic, because a singleton may be deliberate.
Cluster start and stop run synchronously today; there is no background watcher for long-running cluster operations yet, so a cluster that takes a while to come up is reported when the call returns.
Fact sheets: GKE clusters and node pools. Schedules: GKE cluster, node pool. Going deeper: namespace-level scheduling and Kubernetes cost management.
Questions we get a lot.
Does scaling a GKE Standard cluster to zero nodes remove all charges?
No. Google charges the cluster management fee for every cluster regardless of size; scaling nodes down removes the Compute Engine node charges.
Can ZopNight schedule one namespace without stopping the whole cluster?
Yes. Namespace Stop scales that namespace's Deployments and StatefulSets to zero and suspends its CronJobs, then Start restores the saved replica counts, HPA and PDB settings.