AWS Compute Optimizer

TL;DR: Compute Optimizer watches how you actually use compute — CPU, memory, network, disk over 14 days — and hands you specific rightsizing moves for EC2, Auto Scaling Groups, EBS, Lambda, and Fargate, each with a dollar figure and a low/medium/high performance-risk rating. It's free, one-click to enable, and needs no infrastructure changes. The only catch: it can't see memory without the CloudWatch Agent, and you still have to act — too many teams enable it, glance once, and forget.

The numbers

• Findings: under-provisioned (upsize before an outage), over-provisioned (downsize — the common one), optimized, or none (insufficient data).
• Each recommendation shows current vs recommended config, estimated monthly savings, and performance risk.
• Coverage beyond EC2: ASG fleets, EBS (IOPS/throughput), Lambda memory (often raising memory cuts cost by finishing faster), and Fargate CPU/memory.
• Field examples: a startup downsized c5.4xlarge/r5.2xlarge → c5.xlarge/r5.large for $1,200/mo ($14,400/yr); an enterprise on legacy m4/c4 acted on its top 50 recs for $40,000/mo; a batch team combined "high-risk downsize" flags with Spot + ASG for a 60% cut.

Do this

1. Opt in (one click), install the CloudWatch Agent for memory metrics, and wait 14 days — without the agent it's CPU-blind and can't be trusted on memory-bound workloads.
2. Sort by estimated savings, then check performance risk — low-risk goes ahead; high-risk gets tested in staging first.
3. Read the utilization graphs before acting — sustained high usage means be cautious; a low 14-day average on a bursty workload can mislead.
4. Resize (stop → change type → start, or update the ASG launch template) and monitor CloudWatch for a few days after.
5. Make it a monthly habit — top 5 opportunities, test, deploy; recommendations age fast, and Enhanced Infrastructure Metrics (paid, ~$0.0003/instance-hr) extends the lookback to 93 days for irregular workloads.

Gotchas

• Memory needs the agent — the single most important setup step; missing it silently limits every memory recommendation.
• Savings assume on-demand list price — RI/SP-covered resources save less (still real).
• Bursty and batch workloads mislead it — an instance idle 99% of the time but pegged 5 min/day can be flagged for a downsize you'll regret; consider Spot + autoscaling instead.
• Custom bottlenecks are invisible — it only sees standard CloudWatch metrics, not app-level limits like DB connection pools.
• Recommendations lag reality — a workload that changed last week may not be reflected yet.

Skip this if

• Nothing — it's free and worth enabling even on a handful of instances. It's the analytical engine behind Trusted Advisor's and Cost Explorer's EC2 rightsizing. For the specialized cuts, see Compute Optimizer Memory Optimization and GPU Recommendations; rightsize before committing to Savings Plans, and pair high-risk downsizes with Spot Instances + Auto Scaling Groups.