A platform wants to eliminate infrastructure waste. What are the highest-return categories, and which apparent waste should be left alone?

The highest-return categories

1. Orphaned resources. Unattached storage volumes, unused load balancers, idle addresses, snapshots of deleted systems, environments for finished projects. These deliver zero value and are the easiest to remove safely. Automated detection and reclamation, not tickets.

2. Non-production environments running continuously. Development and test environments used during working hours and billed around the clock. Scheduled shutdown is usually the single largest and simplest saving available.

3. Over-provisioned storage tiers. Data on hot storage that has not been accessed in months, and retention set to the maximum because nobody chose a number.

4. Duplicate telemetry. The same information emitted as a metric, a log line and a trace attribute, all retained.

5. Forgotten scheduled jobs. Batch processes for reports nobody reads, and pipelines producing datasets with no consumers.

6. Over-generous default configurations — a log level, a retention period, a sampling rate set once at a generous value and inherited by every service since.

What should be left alone

1. Headroom. Capacity above current utilisation is the buffer that keeps latency bounded and failures survivable, because queueing delay rises non-linearly as utilisation increases. Removing it is not eliminating waste; it is spending reliability.

2. Redundancy. Idle standby capacity is doing its job by being idle. Cutting it converts a survivable failure into an outage.

3. Failure-domain capacity. Running at two-thirds utilisation so that losing one zone is survivable looks like 33% waste and is the price of the redundancy the architecture claims.

4. Observability that has not been queried recently but covers rare events. Telemetry for a failure mode that occurs annually is not waste; it is insurance. The distinction from genuinely unused telemetry is whether anyone can name what it would be used for.

5. Test and staging environments that are realistic. Reducing them to save money means testing stops predicting production, and the cost surfaces as incidents.

The framing that avoids the mistake

Waste is spend that produces no value under any circumstance. Headroom, redundancy and insurance produce value in circumstances that have not occurred yet.

Conflating them is how cost programmes cause incidents, and it is common because both look identical on a utilisation dashboard. The test is to ask, for each candidate: under what circumstance does this become valuable, and how likely is that circumstance? If there is no answer, it is waste.