FinOps Certified Engineer cheat sheet

Container cost allocation starts with the pod: request-based allocation uses sum of container CPU/memory requests times pod runtime; usage-based allocation uses measured consumption and is preferred when requests are heavily over-provisioned.

Kubernetes scheduling is driven by requests, not limits: a pod's cost floor is its requests, so rightsizing requests to P95 usage is the primary lever for reducing reserved-but-unused spend.

Idle and unallocated capacity equals node cost minus the sum of allocated pod costs; it must be attributed explicitly (to a shared/unallocated cost center) rather than silently spread across workloads.

Namespace and label conventions (team, app, environment, cost-center) are the allocation key; missing or inconsistent labels produce untaggable spend, so label policy enforcement is a cost control.

Cluster autoscaler and Karpenter reduce node count when pods are unschedulable or underutilized; scale-down thresholds and pod disruption budgets determine how fast savings are realized.

Bin packing efficiency = sum of pod requests / sum of node allocatable capacity; low efficiency means fragmentation, which is fixed by instance-type diversity, topology spread constraints, and rightsizing.

Horizontal Pod Autoscaler scales on metrics such as CPU utilization relative to requests, so lowering requests without adjusting HPA targets changes scaling behavior and can raise cost.

Shared cluster overhead (kube-system, monitoring, service mesh sidecars, ingress) is real cost: measure it per cluster and allocate by namespace share or treat as a platform chargeback line.

Spot and preemptible node pools trade discount for interruption risk; only interruption-tolerant workloads (stateless, checkpointed, queue-driven) should be placed there.

FinOps lifecycle: Inform (visibility, allocation, benchmarking), Optimize (rightsizing, commitment, rate and usage optimization), Operate (governance, automation, budgets, anomaly response) - engineering work maps to each phase.

Amortization rule: for a fully upfront or partial upfront commitment, spread the upfront fee across the term; unblended cost shows the charge on the purchase date, amortized cost shows the daily share.

Effective savings rate (ESR) = (on-demand equivalent cost − actual cost) / on-demand equivalent cost; report it per commitment family and per scope, not as a single blended number.

Commitment coverage = spend covered by commitments / total eligible spend; utilization = commitment used / commitment purchased. Low coverage means rate risk; low utilization means waste.

Unit economics anchor decisions: cost per request, per tenant, per build minute, or per transaction normalizes spend against business volume and reveals whether growth is efficient.

Showback vs chargeback: showback reports cost to owners without financial consequence; chargeback posts cost to budgets. Both require the same allocation and tagging integrity.

Anomaly detection baselines expected daily spend by account/service with seasonality; thresholds set too tight create alert fatigue, so pair detection with an owner and a runbook.

Tagging enforcement patterns: mandatory tags at provisioning (IaC policy, cloud policy-as-code), automated remediation for untagged resources, and a periodic untagged-spend report as the exception queue.

Rightsizing decisions use percentile usage over a lookback window plus headroom for peaks; the tradeoff is performance risk versus savings, and each recommendation needs an owner and a validation step.