Cloud FinOps Practice

FinOps: coined by J.R. Storment and Mike Fuller, formalized by the FinOps Foundation. Core principle: cloud spending decisions should be made by those closest to the trade-off (engineers), with financial visibility (real-time cost data), and organizational support (FinOps practitioners). Three lifecycle phases: Inform (visibility — who spends what), Optimize (identify and act on savings opportunities), Operate (continuous improvement, culture). Key shift from traditional IT: cloud is variable cost, not capital expense. FinOps enables treating cloud spend as a managed investment, not an uncontrolled liability.

Cost allocation: without it, cloud spend is a single company line item. With it, each team sees their contribution. Mechanisms: Resource tagging: AWS cost allocation tags (e.g., Team=payments, Env=production, Product=checkout). Tag everything: EC2, RDS, Lambda, S3, data transfer. Linked accounts: separate AWS accounts per team/product — natural cost boundary. Chargeback: cost allocated to the responsible team's P&L. Showback: cost reported to teams for visibility but not charged to their budget. Unit economics: cost per customer, cost per transaction, cost per API call — connecting cloud spend to business value. Shared costs (support, data transfer, shared services): allocate by usage proportion or fixed allocation rule.

Rightsizing: organizations typically provision 2-3x more than they use (fear of running out, easier to size up). Rightsizing process: (1) collect utilization metrics (CPU, memory, network) over 2-4 weeks, (2) identify instances using <5% CPU average, (3) downsize to a smaller instance type or family, (4) monitor after change. AWS Compute Optimizer, Azure Advisor, GCP Recommender: built-in rightsizing recommendations. Typical savings: 20-40% compute spend from rightsizing alone. Other cost levers: Reserved Instances / Savings Plans: 30-72% discount for 1-3 year commitments. Spot / Preemptible instances: 60-90% discount for interruptible workloads. Graviton/ARM instances: 20% better price/performance. Storage lifecycle: move cold data to S3 Glacier/Archive.

Cloud cost anomalies: often indicate: Misconfigured auto-scaling: scaling group without an upper bound, triggered by load test or DDoS. Runaway job: ETL job processing far more data than expected. Security incident: cryptomining using compromised credentials. Infrastructure leak: resources created but not destroyed (dev environments, test clusters). Data transfer spike: unexpected egress from data exfiltration or misconfigured CDN. Response process: (1) receive alert from cost anomaly detection (AWS Cost Anomaly Detection, GCP Budget Alerts), (2) identify which service/account/tag is spiking, (3) investigate root cause, (4) remediate (kill runaway job, fix auto-scaling, revoke compromised credentials), (5) postmortem and preventive controls. Budget alerts: set at 80% of expected spend, alert before the month ends.

Unit economics: the most mature FinOps practice. Instead of "we spent $500K on cloud this month", ask: "our cost-per-active-customer is $12, down from $15 last quarter". This frames cloud efficiency in business terms executives understand. Examples: SaaS: cost per active user per month (target: <10% of ARPU). E-commerce: cost per order processed. Data platform: cost per GB processed, cost per query. Building unit metrics: (1) tag resources by product/feature, (2) combine billing data with business metrics (orders from data warehouse), (3) compute the ratio, (4) track over time. Cloud efficiency formula: if unit cost is growing faster than revenue, you have a scaling problem. If unit cost decreases as you grow (economies of scale), architecture is efficient.