A product API peaks at 8,000 requests/sec in the evening and must answer in under 120 ms at p99 with 99.95% availability and critical durability. The previous team left you:
- 8× m5.4xlarge on-demand in a single AZ (fixed, all day)
- no CloudFront in front, so every static asset hits the fleet
- no cache, so the database eats the full read load
- a single RDS instance — no Multi-AZ, no backups
The bill is ~$4,900/mo and it still misses the latency, availability, and durability targets.
Your goal: meet every SLA and bring the bill under $1,500/mo. Three stars if you get within ~10% of the leanest viable design.
Levers that matter: put CloudFront in front to offload static traffic, choose a cost-effective instance family (Graviton is cheaper per request), switch the fixed fleet to autoscaling across ≥2 AZs, pick the right purchasing model, add an ElastiCache tier to shed database load, and make RDS Multi-AZ with backups.