AWS cost optimization, 38% spend reduction without performance loss

A Series A data analytics platform, $42k monthly AWS, faced runway pressure. Engineering feared optimization meant throttling queries. Finance wanted savings inside one quarter without degrading customer-facing SLAs.

Assessment: 34% on-demand EC2 with flat utilization; oversized RDS instance class; CloudWatch log retention infinite; cross-AZ data transfer unmonitored; dev/staging running 24/7 at production scale; unused EBS snapshots from decommissioned experiments.

Problems: one legacy Java dependency blocked Graviton on two nodes, kept m5 reserved pair; CFO wanted showback per team, Cost Allocation Tags enforced via SCP with monthly FinOps PDF; engineering feared query regression, we held p95 within 3% margin throughout.

Actions: Compute Optimizer and Cost Explorer rightsizing review; one-year reserved capacity for baseline workers; m5 → m7g Graviton on stateless workers after ARM compatibility test suite; S3 Intelligent-Tiering on analytics exports; log retention 30 days with export to S3 Glacier for compliance; Instance Scheduler on non-prod with IST business-hours profile.

Technical implementation: Terraform modules for reservations and schedules; team showback tags enforced via SCP; Graviton migration required JVM heap tuning on two legacy workers; query p95 monitored in Datadog before/after each change.

Governance: monthly cost review with engineering leads; anomaly detection on daily spend; no production changes during month-end billing batch window.

Outcome: monthly bill $42k → $26k (-38%); query p95 unchanged within 3% margin. Platform anonymized, analytics SaaS, delivered from Bangalore with executive readout to founders.