FinOps: Turning Off What Nobody Uses

Overview

Cost dashboards were full of "potential savings", but many of them were stale or wrong. Resources were already compliant, already stopped or still needed. I ran an estate-wide review and only reported what I'd checked against live resource state. The result was a prioritised plan worth a five-figure sum every month, including quick wins that could be actioned straight away with no risk.

Findings

Method

1. Inventory everything with Resource Graph: power state, SKU, attachment state, tags and shutdown schedules.
2. Validate each candidate. Several VMs already had shutdown schedules, and one GPU VM was already deallocated, so I excluded them. One VM had a schedule that was disabled.
3. Trace orphans to their cause. The snapshots came from a past database migration, so the recommendation was "confirm with that change's owner", not "delete on sight".
4. Flag compliance as well as cost. A developer-edition licence turned up in production during the licensing review.

resources
| where type =~ 'microsoft.compute/disks'
| where properties.diskState =~ 'Unattached'
| project name, resourceGroup, sku = sku.name,
          sizeGB = toint(properties.diskSizeGB), created = properties.timeCreated
| order by sizeGB desc

The AKS Lever

AKS clusters can't stop themselves, and an in-cluster CronJob can't start the cluster it stopped. So I recommended an external schedule, with opt-in by tag so teams stay in control:

schedules:
  - cron: "0 20 * * 1-5"
    branches: { include: [ main ] }
    always: true
steps:
  - task: AzureCLI@2
    inputs:
      azureSubscription: nonprod-automation
      scriptType: bash
      inlineScript: |
        az aks list --query "[?tags.autoshutdown=='true'].[name,resourceGroup]" -o tsv |
        while read -r name rg; do az aks stop -n "$name" -g "$rg" --no-wait; done

Key Outcomes

• A five-figure monthly saving identified, with more than half coming from scheduling non-production AKS.
• Zero-risk quick wins separated out so they could be actioned immediately.
• Every item verified against live state, so leadership could act without second-guessing.
• Guardrails proposed to stop waste returning: schedules, shutdown policy, tagging and budgets.

Lessons Learned

• Validate before you report. One wrong recommendation costs more credibility than ten right ones earn.
• Orphaned resources always have a story. Find the change that created them.
• Show cost when infrastructure is created, not just after the bill arrives.

Conclusion

The biggest saving wasn't exotic. It was switching things off when nobody uses them. The value was in the evidence behind each line.

Technologies Used