AWS Rightsizing Recommendations: How to Act on Them Safely

How to read AWS rightsizing recommendations from Compute Optimizer and Cost Explorer, check the data gaps and apply changes safely.

Rightsizing recommendations are suggestions from AWS tools to move an over-provisioned resource, usually an EC2 instance, to a smaller or cheaper type. They are a good starting point for cutting cost, but they are not safe to apply blindly. This guide shows how to read them and act without hurting performance.

Where the recommendations come from

AWS Compute Optimizer analyses usage and suggests changes for EC2 instances, Auto Scaling groups, EBS volumes, Lambda functions and more. AWS Cost Explorer also shows EC2 rightsizing recommendations in the billing console. Both rely on utilisation metrics, so the quality of the advice depends on the quality of the data.

Know what the data misses

  • Memory: EC2 does not report memory by default. Without the CloudWatch agent, a recommendation may look only at CPU, network and disk.
  • Short lookback: Compute Optimizer uses a limited period of recent data by default. A monthly report, a quarter-end job or a seasonal peak may not appear. Longer lookback options exist, so check the current settings.
  • Averages hide peaks: a server at 10% average CPU can still hit 90% for ten minutes every hour.

A safe way to apply them

  • Rank by saving and risk. Start with large, steady, non-critical workloads, not the busiest production database.
  • Check the peak, not only the average. Look at maximum CPU, memory and network over a full business cycle.
  • Check compatibility. Moving to a different instance family, or to Graviton (Arm), can require software changes and testing. Also check licensing that depends on cores.
  • Change one step at a time. Test a scoped candidate, observe a representative business cycle, and then decide on the next step.
  • Test in non-production first, and keep a rollback plan.
  • Record the result. Save the before and after cost and performance so the next review is easier.

Our rightsizing playbook goes deeper on how to do this without causing incidents.

What a good result looks like

An accepted rightsizing change should improve the agreed cost measure while latency, errors and queue time remain within the service objectives. If an objective is breached, use the agreed rollback procedure and investigate what the evidence missed.

Rightsize before you commit

Savings Plans involve a one- or three-year spending commitment; Reserved Instances apply discounts under their purchased configuration and term. If you buy them first and rightsize later, you may commit to capacity you no longer need. Rightsize first, then buy commitments for what is left. Also check that the inputs behind a commitment recommendation still hold after a change, as we explain in Is this Savings Plans recommendation still valid?

Make it a routine

Rightsizing is not a one-off task. Workloads change, so review recommendations monthly and watch for drift with our cloud cost drift tool.

If you want a second pair of eyes before you change production, our cloud cost optimization team can review your recommendations and rank them by saving and risk. Contact us to start.

Worked example: a recommendation with missing memory

Assume a fictional reporting worker runs on a general-purpose EC2 instance. An optimization dashboard recommends a smaller instance because recent CPU demand is low. The workload owner knows that the worker also processes a large month-end export, but memory measurements for that job are missing. The recommendation is a candidate for investigation, not authorization to change production. The immediate output should be an evidence gap assigned to the owner.

Record the instance identifier, Region, current configuration, recommended configuration, observation window and recommendation timestamp. Preserve the projected saving as a modeled result from that recommendation. Then record the known business cycle and missing signals. The Compute Optimizer resource requirements make clear that recommendation generation depends on resource-specific metric requirements. A displayed recommendation does not prove that every important application condition was observed.

Ask whether the observed window includes the month-end export, a backup overlap and the busiest scheduled batch. If it does not, either observe a representative cycle or run an authorized representative test with explicit limitations. Do not purchase enhanced telemetry or enable broad data collection merely to make a dashboard look complete. Confirm the access, cost and retention decision with the responsible owner first.

For an illustrative test, suppose a representative export on the candidate instance completes in 46 minutes against an agreed 60-minute deadline. The current instance takes 38 minutes on the same input. That is eight minutes slower but still within this one acceptance boundary. It does not establish readiness for larger inputs, concurrent workloads or a zone failure. Retain the input size, concurrency and software version alongside the timing so another engineer can reproduce the comparison.

Compare the whole change, not just the rate

Assume invented monthly compute estimates of $240 before the change and $180 afterward, both on the same billing basis. The modeled compute reduction is $60, or 25% of that compute line. If unchanged storage and networking add $80 on each side, the modeled workload total changes from $320 to $260. The same $60 is 18.75% of the original workload total. Neither percentage is a verified AWS saving, and the rates are teaching inputs rather than current prices.

Now consider an existing commitment that remains payable after the resize. A smaller instance may reduce consumed capacity without reducing that month's cash payment. Separate changes in resource usage, amortized allocation and actual invoice impact. The Savings Plans overview describes the spending commitment that accompanies discounted usage. Check which eligible workload can use remaining commitment before reporting avoided compute consumption as money saved.

Changing architecture or instance family also changes the compatibility question. A move to Arm is not simply a resize. Check binary dependencies, agents, base images, native extensions and licensing, then test the deployed artifact. A smaller instance within the existing family can still reduce network or storage throughput allowances. Record each changed resource characteristic rather than assuming that memory and virtual CPU are the only differences that matter.

Create an acceptance and rollback record

FieldWhat to recordStop condition
Business boundaryDeadline, latency or queue-age objectiveThe objective has no agreed owner
Representative demandInput size, concurrency and peak windowsThe sample excludes a known critical event
Dependency behaviorDatabase, storage and network observationsA new bottleneck moves outside the instance
Financial basisRate assumptions and existing commitmentsThe estimated cash reduction is unsupported
Recovery pathPrior configuration and restoration procedureThe prior configuration cannot be restored in time

Write rollback triggers before the change window. They should refer to the application objective, such as missed completion deadlines or sustained queue growth, rather than only to host CPU. A worker can have comfortable CPU while waiting on storage. Conversely, higher CPU after a resize can be acceptable if the service remains within its objectives and has tested headroom. A single utilization threshold cannot decide every workload's readiness.

Recovery has dependencies too. Restoring a previous instance type requires access, deployment tooling and available capacity. Keep the previous launch configuration and application artifact identifiable. For stateful workloads, specify how in-flight work and writes are handled; replacing compute is not a data rollback. Do not make a production database the first experiment merely because its recommendation shows the largest estimated saving.

Limitations, ownership and next action

A smaller instance can be unsuitable when bursts, single-thread performance, memory pressure, specialized hardware or contractual licensing dominate the workload. Recommendations can also become stale after an application release, a schedule change or a new customer launch. Tie the review to a recent deployment and demand snapshot. Missing memory is not zero memory use, and a quiet metric period is not proof of unused business capacity.

The workload engineer owns performance acceptance, the service owner approves the operating risk and finance validates the cost basis. These roles can be held by the same person in a small team, but the decisions remain different. Keep a short result record that says accepted, rejected or insufficient evidence, with the reason and the next observation needed. Do not present a rejected recommendation as a failed optimization program; rejecting an unsafe change is a useful outcome.

The next action is to choose one reversible recommendation and complete the acceptance record before scheduling it. If a known peak is absent from the evidence, collect or test that peak first. After release, compare the same workload and accounting basis over a representative period. Preserve any adverse effect and recovery action alongside the saving estimate so future reviewers can judge the change honestly.

Related services