Kubernetes Cost Optimization: Cut Cloud Spending 50%

The Kubernetes Cost Problem

Kubernetes makes it easy to deploy workloads. Too easy, perhaps - resources get requested and forgotten. Clusters grow without scrutiny. Before you know it, cloud bills are enormous and nobody knows why.

The typical Kubernetes cluster runs at 20-40% utilization. That means 60-80% of spending is waste. With proper optimization, 50% cost reduction is achievable without impacting performance.

Understanding Resource Requests and Limits

Requests are what the scheduler uses for placement decisions. If you request 1 CPU, the scheduler finds a node with 1 CPU available. Requests also determine your guaranteed resources.

Limits cap resource usage. Exceeding memory limits causes OOM kills. Exceeding CPU limits causes throttling.

The common mistake: setting requests equal to limits (guaranteed QoS) wastes resources. Set requests based on typical usage, limits based on peak usage. This enables bin packing - fitting more pods on fewer nodes.

Right-Sizing Strategies

Most teams guess at resource requirements, usually too high. Use actual metrics to right-size:

Vertical Pod Autoscaler (VPA): Analyzes historical usage and recommends (or automatically applies) right-sized requests. Start in recommendation mode to understand your workloads before enabling auto-updates.

Metrics analysis: Query Prometheus for actual CPU and memory usage over time. p95 usage is often 10-20% of requested resources.

Goldilocks: Open-source tool that runs VPA in recommendation mode for all deployments and provides a dashboard showing recommendations.

Spot Instances for Non-Critical Workloads

Spot instances cost 60-90% less than on-demand. The tradeoff: they can be terminated with little warning.

Good candidates for spot:

Stateless services with multiple replicas

Batch processing jobs

Development and testing environments

Any workload that handles interruption gracefully

Use node affinity to prefer spot nodes while allowing fallback to on-demand. Configure Pod Disruption Budgets to maintain availability during spot terminations.

Autoscaling: Scale Down Aggressively

Cluster Autoscaler removes underutilized nodes. Default settings are conservative - tune for cost optimization:

Reduce scale-down delay (how long a node must be underutilized)

Lower utilization threshold (what counts as underutilized)

Enable scale-down for nodes with local storage if safe

Horizontal Pod Autoscaler scales pods based on metrics. Scale down quickly when load decreases. Consider KEDA for event-driven scaling.

Namespace Quotas and Limits

Without quotas, any team can consume unlimited resources. Implement:

Resource quotas: Limit total CPU/memory per namespace

Limit ranges: Default and maximum resources per pod

Cost allocation: Tag resources by team for chargeback

Visibility creates accountability. When teams see their costs, they optimize.

Best Practices

Always set requests: Enables bin packing and scheduling

Use VPA for right-sizing: Data-driven resource allocation

Embrace spot instances: 60-90% savings for tolerant workloads

Autoscale aggressively: Scale down quickly when load drops

Implement quotas: Prevent runaway resource consumption

Monitor and alert: Track cost per namespace, per team