cloud rightsizing: reduce waste without breaking workloads
Reviewed: 20 October 2026Get a quote for my projectCloud rightsizing matches resource capacity to the workload it must serve. Use performance evidence, availability requirements and a controlled change process before reducing an instance size, altering a schedule or deleting idle resources.

On this page
Cloud rightsizing: What each tool looks at, and for how long
| Tool | Observation window | Points to know |
|---|---|---|
| AWS Compute Optimizer | 14 days by default; up to 93 days with the paid enhanced infrastructure metrics; lookback configurable in rightsizing preferences | The memory utilization graph requires the unified CloudWatch agent; confirm which metrics are available for the recommendation. Findings are Under-provisioned, Over-provisioned or Optimized, each with a performance risk from very low to very high. Refreshed daily. |
| Azure Advisor | Stated per recommendation: 7 days for scale sets and App Service plans, 30 days of no activity for idle Cosmos DB containers | Shutdown and right-size recommendations for VMs, disks not attached to a VM, stopped Data Explorer clusters after at least 60 days. |
| Google Compute Engine recommender | 8 days, from 60-second averages | Short spikes can be missed. Ops Agent improves memory data. No recommendation for instances with GPUs, TPUs or ephemeral disks, or in the memory-optimized family. |
A window of one to two weeks says nothing about month-end processing, quarterly batch jobs or a seasonal campaign. Ask the workload owner for the calendar and extend the window where a peak is missing.
Stop, resize or delete: three different risks
- Stop or schedule off: reversible. Disks, reserved IP addresses and licences usually keep billing, and any commitment covering the instance keeps billing too.
- Resize: needs a restart on most platforms. AWS lists platform differences to check for a recommended type: architecture, hypervisor, instance store, network interface and storage interface. A move to Arm-based Graviton is rated from low to high migration effort depending on the workload, and may need a recompiled application or another machine image.
- Delete: irreversible for disks. Azure Advisor says to snapshot first and confirm the data is no longer needed. Applying a Compute Engine machine type change requires permission to stop and start the instance.
Protocol for each change
- The resource owner confirms the purpose, the business calendar and any recovery dependency, in writing.
- Check memory, disk and network, not CPU alone. A database with low CPU can hold a large working set in memory.
- Take a snapshot or image and note the rollback steps.
- Schedule a change window and name who watches the workload afterwards.
- Observe for a full cycle of the workload, then compare the tool's predicted utilization with reality.
- Close the ticket with the billing effect from the invoice, not from the tool.

Why the invoice may not move
- A Savings Plan or Reserved Instance was covering the instance. Usage falls, the commitment does not, and utilization drops instead of the bill.
- Compute Optimizer prices its estimate at On-Demand rates unless savings estimation mode or Cost Optimization Hub is on. Against a discounted instance, that overstates the change.
- Attached storage, snapshots, public addresses and licences remain.
- An autoscaling group or deployment pipeline recreates the old size at the next release.
- Demand grew elsewhere at the same time.
Measure the net effect across every service touched by the change, for at least one full billing month after it.
Evidence pack per item
- Resource ID, owner, environment and business purpose.
- Tool finding with its date and observation window, plus the extra metrics the provider checked.
- Proposed action, expected risk and approver.
- Change date, rollback used or not, and incidents in the following cycle.
- Billing effect one month later and the commitment position at that date.
Setting the roles
Put the owner of each environment on the approval line and keep the provider on the analysis side. The provider can rank findings and prepare tickets, your engineers apply them under your change process, and finance validates the billing effect. Non-production environments can be a first batch when the owner confirms a suitable schedule, recovery path and acceptable business impact. Shared test services or release dependencies can still be critical.
Want these steps applied to your own accounts? Get my quote.
Official sources
Official documentation used for this page.
- docs.aws.amazon.com/compute-optimizer/latest/ug/what-is-compute-optimizer.html
- docs.aws.amazon.com/compute-optimizer/latest/ug/view-ec2-recommendations.html
- learn.microsoft.com/en-us/azure/advisor/advisor-reference-cost-recommendations
- docs.cloud.google.com/compute/docs/instances/apply-machine-type-recommendations-for-instances
- docs.cloud.google.com/recommender/docs/recommenders
- www.finops.org/framework/capabilities/usage-optimization/
Rightsizing questions from engineering and finance
Is low CPU proof that a resource is unnecessary?
No. Check memory, storage, network, the workload calendar and recovery needs. The resource owner must validate the purpose and the proposed action.
Will stopping compute remove every related charge?
Not necessarily. Storage, networking, licences and commitments may keep billing. Measure the net effect across the affected services.
Which environments should go first?
Start with owner-approved resources whose interruption and recovery risks have been tested. A non-production label alone does not establish low risk. Apply the validated protocol before changing more critical workloads.
Continue your review
- FinOps consulting: cloud cost review and implementation scope
- Request a scoped FinOps quote
- How your request is handled: from enquiry to proposal
- Cloud forecasting: explain budget variance to finance
- Cloud egress cost optimization: audit data movement
- Kubernetes cost optimization consulting for EKS, AKS and GKE
Get a quote for my cloud costs
A specialist in cloud cost control reads your request and proposes a quote.
Get my quoteWhat is rightsizing in cloud cost optimization? It is the controlled adjustment of resource capacity to the demand and service behavior the workload must support. For AWS rightsizing, use the Compute Optimizer EC2 recommendation evidence as an input to application validation, with the observation period and metric limits recorded.