Cutting Kubernetes Costs by 60% Without Sacrificing Performance

The Kubernetes Cost Dilemma

Scaling workloads on Kubernetes (K8s) is remarkably easy, but without proper controls, cloud hosting bills can spiral out of control. Over-provisioning resources is the number one cause of waste in modern cloud budgets. Here is how we auditing and optimising K8s resource distribution.

1. Pod Resource Rightsizing

Developers often allocate arbitrarily high CPU and memory requests to prevent out-of-memory (OOM) kills. By implementing monitoring tools like Goldilocks or Kubecost, we analyze actual resource usage over time and adjust limits to realistic baselines.

resources:
  requests:
    memory: "256Mi"
    cpu: "200m"
  limits:
    memory: "512Mi"
    cpu: "500m"

2. Strategic Spot Instance Pools

Using spot instances (AWS Spot or GCP Preemptible VMs) can cut VM costs by up to 90%. However, spot nodes can be reclaimed at any time. We structure node groups to use mixed instances, reserving standard on-demand nodes for critical components (like API endpoints) while running queue workers and background jobs on spot nodes.

• Use Karpenter or Cluster Autoscaler with multi-family node templates.
• Add node-affinity rules to route fault-tolerant workloads to spot pools.
• Set up termination handlers to gracefully drain nodes upon shutdown signals.