AWS Compute Optimizer
ML-powered rightsizing service that analyzes utilization metrics and recommends optimal resource configurations.
Verified against AWS documentation · 2026-08-27
Key facts
- 01Opt-in service — analyzes CloudWatch metrics over a 14-day default lookback (32-day configurable, 93-day with paid enhanced infrastructure metrics)
- 02Covers EC2, Auto Scaling groups, EBS volumes, Lambda, ECS on Fargate, RDS/Aurora, and commercial licenses — plus idle-resource detection
- 03Findings labeled: under-provisioned, over-provisioned, optimized (EC2) — recommendations refresh daily
- 04Organization-wide view when opted in from the management account; preferences settable at org, account, or resource level
Exam triggers
Phrases and scenarios that point here.
- "Rightsizing recommendations", "identify over-provisioned instances", "optimal instance type based on utilization" → Compute Optimizer
- Uses machine learning on CloudWatch metrics → Compute Optimizer, not Trusted Advisor
- "Analyze memory utilization for rightsizing" → needs CloudWatch agent installed; CPU works out of the box
- Recommendations across an entire AWS Organization → opt in from the management account
Not this service
- Broad best-practice checks (security, limits, fault tolerance) → Trusted Advisor
- Visualizing and analyzing what you've already spent → Cost Explorer
- Reserved Instance / Savings Plans purchase recommendations → Cost Explorer, not Compute Optimizer
Use cases
- 01Quarterly rightsizing sweep: a team inherits 200 EC2 instances sized by guesswork years ago — Compute Optimizer flags the m5.2xlarges idling at 8% CPU and names the cheaper replacement
- 02Catching the opposite mistake: a t3.medium API server that's been throttling under load shows up as under-provisioned before customers start complaining
- 03Lambda memory tuning — finding functions where dropping (or raising) the memory setting cuts cost or duration without trial-and-error redeploys
- 04Idle-resource cleanup: surfacing unattached EBS volumes and forgotten dev instances that quietly bill every month
- 05Workloads with monthly patterns (payroll runs, month-end reporting) — 14 days of data misleads, so extend the lookback before trusting the recommendation
How it's charged
- 01The service itself is free — recommendations at the default lookback cost nothing
- 02Enhanced infrastructure metrics (93-day lookback) is a paid preference, billed per resource per hour
- 03Underlying CloudWatch metrics bill as normal — enabling the CloudWatch agent for memory metrics adds custom-metric charges
- 04Acting on recommendations is where the real money moves; the tool only points
