Priya had been the SRE lead at a mid-sized logistics software company for three years when the finance team sent her a short message: “Cloud bill up again. Any idea why?” She did not have a quick answer. Traffic was flat, no new products had launched, and nobody had asked for more capacity. So one Friday afternoon she blocked her calendar, opened the console and started what became the team’s quarterly cloud idle resource audit checklist. What she found, and the list she built from it, is worth copying.
Why idle resources pile up
Nobody sets out to waste money. Resources accumulate because creating them is easy and deleting them feels risky. An engineer spins up a test cluster for a proof of concept, the project pauses, and the cluster stays. A model endpoint is deployed for a demo and never retired. Each item is small. Together they turn the cloud account into a storeroom full of things the team once needed for a week.
Priya put it this way to her team: “We moved to the cloud so we would only pay for what we use. Somewhere along the way, we started owning things again.”
The cloud idle resource audit checklist
Here is the list she worked through that afternoon, refined over the following quarters. Each item comes with the moment she found it.
1. Compute instances with low average CPU
She sorted every instance by its thirty-day average utilisation. A cluster of eight machines sat below five percent. They belonged to a load test from the previous spring.
Action: stop or terminate anything persistently under-used; right-size the rest.
2. Development environments running around the clock
Three dev environments ran nights and weekends with nobody logged in.
Action: add schedules that shut them down outside working hours, or replace them with on-demand environments created per branch.
3. Unattached storage volumes
When instances are deleted, their disks often survive. Priya found dozens, some from engineers who had left the company.
Action: snapshot anything uncertain, then delete unattached volumes older than an agreed threshold.
4. Old snapshots and backups beyond retention
Daily snapshots had been kept indefinitely because nobody had written a retention policy.
Action: define retention by data class and automate cleanup.
5. Idle ML model endpoints
The data science team had four inference endpoints running, two of which had received no requests in weeks. Our guide on what every MLOps learner should watch once a model is live covers the request metrics that would have flagged this sooner.
Action: scale idle endpoints to zero or retire them; use serverless or batch inference where real-time is not needed.
6. Reserved IP addresses and load balancers with no targets
Small charges, but they added up across several regions.
Action: release anything not routing traffic.
7. Over-provisioned databases
A reporting database was sized for a migration that had finished months earlier.
Action: right-size, and consider serverless or auto-pausing tiers for intermittent workloads.
8. Untagged resources
About a fifth of resources had no owner tag. Nobody could say whether they were safe to remove.
Action: enforce owner and project tags at creation through policy; chase down untagged items each audit.
9. Software seats and licences
Paid monitoring and CI seats were assigned to people who had changed roles.
Action: reclaim inactive seats monthly.
The pushback, and how she handled it
Not everyone welcomed the clean-up. One developer worried that deleting an old environment would lose configuration he might need someday. A data scientist argued that keeping an endpoint warm saved time during demos. Priya did not overrule them. Instead she offered a deal: anything flagged would be snapshotted or captured as code first, and could be recreated within an hour if needed. Once people trusted that recreation was quick, the urge to keep things running “just in case” faded. In the following quarter, only one deleted resource was ever rebuilt, and it took twenty minutes. The lesson for her was that people hoard infrastructure for the same reason they hoard tools at home: getting it back feels harder than keeping it.
The other half: letting idle capacity serve others
Not everything idle needed deleting. Some of it simply needed a second customer. The team’s GPU nodes, bought for a training project, sat quiet between experiment cycles. Priya worked with the analytics group so their overnight batch jobs could run on those nodes. Utilisation rose, and the analytics team stopped paying for separate capacity.
The same thinking applied to the build farm. Instead of each squad owning a fixed pool of runners, they moved to a shared pool with autoscaling, described in our walkthrough of delivery pipelines for modern software teams. Fewer machines, busier machines.
A weekend thought that stuck
A few weeks after the first audit, Priya took a long weekend in Udaipur. Rather than relying on taxis, she rented a scooter through MotoShare, a marketplace where local owners and shops list two-wheelers and cars, collecting it near her guesthouse and paying at pickup. Riding around the lakes, she realised the owner was doing exactly what her team had just learned: putting a vehicle that would otherwise sit still to work for someone who needed it for three days. Access when needed, earnings when idle. She used the example in her next team retro.
Making the audit stick
The first audit cut a noticeable slice from the monthly bill. The bigger win was turning it into habit. Priya’s team now:
- Runs the checklist every quarter, rotating the engineer who leads it.
- Reviews a weekly report of the lowest-utilisation resources.
- Requires an expiry tag on any experimental resource, with automatic reminders.
- Shares one short summary with finance after each audit.
When finance asked again, months later, why the bill had changed, Priya had a better answer: “We stopped holding onto things we only needed for a while.”
Run your own audit this week
Pick a two-hour slot. Sort compute by utilisation, check for unattached storage, list endpoints with no traffic and find untagged resources. You do not need to fix everything in one sitting. Simply seeing how much sits idle is usually enough to change how the team provisions from that day on.

