Cloud cost control (FinOps) · Lesson 3 of 6 · 5 min read
Optimising compute
Rightsizing and modernising
- Rightsize: compare CPU, memory and network use against instance size over a few weeks (Compute Optimizer needs the CloudWatch agent for memory data). A server averaging 8% CPU on an
m5.2xlargecan often run on anm7g.large. - Move to current generations and Graviton: newer instance families give more performance per dollar, and Graviton (arm64) is typically around 20% cheaper for similar or better performance. Containers and interpreted languages usually move easily; build multi-architecture images.
- Turn things off: development and test environments running nights and weekends burn about 70% of their hours idle. Schedule them down (Instance Scheduler, scaling Auto Scaling groups or node pools to zero, stopping RDS) and spin up preview environments on demand.
- Delete the forgotten: stopped instances still pay for EBS; unattached volumes, idle load balancers, unused Elastic IPs and old AMIs all cost money every month.
Paying less for what you use
| Option | Discount vs on-demand | Commitment | Good for |
|---|---|---|---|
| Compute Savings Plans | Up to about 66% | $/hour for 1 or 3 years, any instance family, region, OS; also Fargate and Lambda | Steady baseline that may change shape |
| EC2 Instance Savings Plans | Up to about 72% | $/hour for one instance family in one region | Very stable fleets |
| Reserved Instances | Similar | A specific instance type | RDS, ElastiCache, OpenSearch, Redshift (where Savings Plans don't apply) |
| Spot Instances | Up to about 90% | None, but AWS can reclaim with a 2-minute warning | Stateless, fault-tolerant, flexible workloads |
How to use them together:
Enlarge diagram
Swipe sideways to see all of it, or tap to open it full screen.
- Cover the steady baseline (the lowest usage level you're confident about for the term) with Savings Plans; leave peaks on on-demand or Spot. Over-committing wastes money just like over-provisioning.
- Track coverage (what share of eligible usage is discounted) and utilisation (what share of the commitment you actually use). Aim for high utilisation first.
- Spot: diversify across many instance types and AZs, use capacity-optimised allocation, handle the interruption notice (drain the node, checkpoint work), and keep a slice of on-demand for critical pods. Good fits: CI runners, batch jobs, stateless web tiers behind load balancers, dev clusters.
Enlarge diagram
Swipe sideways to see all of it, or tap to open it full screen.
Kubernetes-specific savings
- Requests drive cost: nodes are bought to fit requested resources. Set requests from observed usage (VPA in recommendation mode, Goldilocks, or Kubecost's suggestions), and review the largest gaps between requested and used.
- Bin packing: Karpenter chooses instance types to fit pending pods and consolidates by moving pods off under-used nodes and removing them. Set
consolidationPolicy: WhenEmptyOrUnderutilizedwith disruption budgets. - Mix capacity: a Karpenter NodePool allowing Spot and on-demand, multiple instance families and both architectures gives the scheduler room to pick the cheapest option.
- Scale to demand: HPA or KEDA scale pods down overnight; Karpenter or the Cluster Autoscaler removes the empty nodes.
- Avoid tiny clusters per team: each EKS cluster has a control-plane fee and its own overhead (system pods, spare capacity). Shared multi-tenant clusters with namespaces are usually cheaper, as long as isolation needs allow it.
- Clusters on old Kubernetes versions past standard support pay a much higher per-hour extended support fee; keep upgrades current.
Enlarge diagram
Swipe sideways to see all of it, or tap to open it full screen.