Opscraft
Cloud cost control (FinOps) · Lesson 6 of 6 · 3 min read

Interview questions and model answers

"Our AWS bill doubled last month. What do you do?"

Open Cost Explorer and group by service, then by account, tag and usage type, to find exactly which line grew and when. Correlate the date with deploys and infrastructure changes. Common culprits are a new NAT gateway data path, logs or metrics exploding, a forgotten load test environment, autoscaling stuck high, or a change that added cross-AZ traffic. Fix the cause, then add budgets and anomaly detection per account and team, and tagging so attribution is faster next time.

"Spot, Savings Plans or Reserved Instances?"

They combine. Savings Plans cover the steady baseline I'm confident about for one to three years, with Compute Savings Plans for flexibility across families, Fargate and Lambda. Reserved Instances cover services Savings Plans don't, like RDS and ElastiCache. Spot runs interruptible, flexible workloads such as CI, batch and stateless tiers at up to 90% off, diversified across instance types with interruption handling. Peaks stay on-demand.

"How do you split EKS costs between teams?"

Namespaces per team with labels, then Kubecost or OpenCost, or AWS split cost allocation data, to allocate node, storage and network costs by requests and usage. Shared and idle capacity is reported separately and owned by the platform team. Showing teams their requested-versus-used gap is usually the biggest lever, because nodes are bought for requests.

"How do you reduce Kubernetes costs without hurting reliability?"

Set requests from real usage with VPA recommendations, let Karpenter bin-pack and consolidate with disruption budgets, use Spot for stateless workloads with PDBs and spread across AZs while keeping critical workloads on on-demand, scale non-production to zero at night, and use Graviton. I'd keep headroom and anti-affinity for critical services and watch SLOs as changes roll out.

"NAT gateway costs are high. Why, and what do you do?"

NAT gateways charge per GB processed, so private workloads pulling container images, talking to S3 or calling AWS APIs through NAT add up. I'd check VPC Flow Logs or the NAT metrics to find the top destinations, then add free gateway endpoints for S3 and DynamoDB and interface endpoints for ECR, STS, CloudWatch and others where traffic justifies them, and use one NAT per AZ only where needed.

"How do you get engineers to care about cost?"

Make it visible to them and attributable: accounts and tags per team, showback dashboards and budgets alerting their own channels, cost estimates on Terraform pull requests, and unit-cost metrics they can relate to. Then build it into habits: a regular review with owners, cost as part of design reviews, and recognising savings. Guardrails like SCPs stop the expensive mistakes.

"What are the risks of buying Savings Plans?"

Over-committing: if usage drops or moves to services the plan doesn't cover, you pay for unused commitment. I'd commit only to the steady baseline from recent usage, start with one-year terms and Compute Savings Plans for flexibility, buy in stages, and track utilisation and coverage monthly.