Cloud7 min read

Kubernetes cost optimisation: a FinOps checklist for growing teams

Kubernetes makes it easy to ship and just as easy to overspend. This checklist covers visibility, right-sizing, autoscaling, pricing models and the habits that keep cluster costs under control.

By INFIDEC Cloud & DevOps Team

Cover: Kubernetes cost optimisation

Why Kubernetes bills drift upward

Kubernetes separates the people who request resources from the people who pay for them. A developer sets CPU and memory requests in a manifest, the scheduler reserves that capacity on a node, and the cloud provider bills for the node whether the pod uses its reservation or not. Multiply that by dozens of services and several environments and the gap between what you pay for and what you use can become large without anyone making a single bad decision.

The fix is not a one-off clean-up. It is a set of practices that make cost visible, give teams sensible defaults, and let the platform scale capacity to real demand. The checklist below is ordered roughly by how quickly each item tends to pay back.

Start with visibility and allocation

You cannot optimise what you cannot attribute. Cloud bills show nodes, disks and load balancers, not teams or services, so the first job is to map spend back to the workloads that cause it.

  • Label every workload consistently with team, service and environment, and enforce the labels with an admission policy such as Kyverno or OPA Gatekeeper.
  • Install a cost allocation tool such as OpenCost or Kubecost to split node costs across namespaces and labels based on requests and usage.
  • Report idle cost separately: capacity that is provisioned but not requested by any pod is a platform problem, not a team problem.
  • Share a weekly cost view per team, alongside their own reliability metrics, so cost is discussed in the same place as performance.

Right-size requests and limits

Requests drive scheduling and therefore drive node count. Most clusters we review have requests set once at launch, often copied from another service, and never revisited. Compare requested CPU and memory with observed usage over a representative period, including peaks, and adjust.

The Vertical Pod Autoscaler in recommendation mode is a low-risk way to get suggested values without letting it change pods automatically. Be more careful with memory than CPU: CPU over its limit is throttled, while memory over its limit gets the container killed. Many teams set memory limits equal to requests for predictability and leave CPU limits unset or generous to avoid unnecessary throttling, while keeping CPU requests accurate.

  • Set default requests and limits per namespace with LimitRange so new workloads do not start unbounded.
  • Use ResourceQuota to cap what a namespace can reserve.
  • Review the largest twenty workloads by requested cost first; that is usually where most of the waste sits.

Scale pods and nodes to demand

Right-sized pods only save money if nodes are removed when they are no longer needed. Pair workload autoscaling with node autoscaling so both layers follow demand.

  • Use the Horizontal Pod Autoscaler on CPU, memory or custom metrics such as queue depth; KEDA is useful for event-driven scaling, including scaling to zero.
  • Use Cluster Autoscaler or, on AWS, Karpenter to add and remove nodes. Karpenter can also choose instance types per workload and consolidate underused nodes.
  • Set PodDisruptionBudgets so consolidation can move pods safely without taking a service below its minimum.
  • Scale non-production environments down outside working hours. Development and staging clusters often run around the clock for no reason.

Use the right pricing models

Once usage is efficient, reduce the unit price. Spot or preemptible instances are substantially cheaper than on-demand capacity but can be reclaimed at short notice, so they suit stateless, fault-tolerant workloads such as batch jobs, CI runners and horizontally scaled web services with several replicas.

Keep a stable on-demand baseline for stateful and critical components, and run spot capacity across several instance types and availability zones to reduce the chance of losing it all at once. For the steady baseline, commitment discounts such as AWS Savings Plans, Azure reservations or Google Cloud committed use discounts are worth evaluating once usage has stabilised. Arm-based instances, such as AWS Graviton, often offer better price-performance if your images build for arm64.

Do not forget storage, network and observability

Compute is the largest line item for most clusters, but the remaining costs are often easier to cut because nobody is watching them.

  • Delete orphaned persistent volumes and old snapshots, and choose storage classes that match the performance actually needed.
  • Watch cross-zone data transfer. Chatty services spread across zones can generate meaningful charges; topology-aware routing can keep traffic local where it is safe to do so.
  • Consolidate load balancers behind a shared ingress or gateway instead of one per service.
  • Review log and metric volume. Debug-level logging in production and high-cardinality metrics can make your observability bill grow faster than your cluster.

Make it a habit, not a project

Savings erode when the attention moves elsewhere. Put a monthly cost review on the calendar with engineering and finance, track a unit metric such as cost per thousand requests or cost per active customer, and set budget alerts so surprises arrive in days rather than at month end.

Most importantly, make the efficient path the default. Platform templates with sensible requests, autoscaling and labels already in place will do more for your bill than any one-off optimisation sprint.

Strategy

How much does custom software development cost in 2026?

There is no single price for custom software, but there is a reliable way to estimate it. Here are the factors that drive cost, how engagement models differ, and how to reduce budget risk.

7 min read

Have a product to build or a system to fix?

Tell us where you are and what you need. A senior engineer replies within 24 hours with next steps, not a sales script.