Autoscaling & scale-to-zero

Some resources run at a fixed instance count, or hold a minimum number of replicas warm around the clock, even though their load is bursty or drops to nothing for long stretches. Letting capacity follow demand — autoscale, or a scale-to-zero floor — means you stop paying for idle replicas without losing the ability to handle peaks.

What we look for

• A workload running a fixed instance count (no autoscale configured) whose utilization chart shows clear, repeated headroom.
• A minimum replica/instance floor above zero on a service that tolerates a cold start — the floor keeps paying through every idle window.
• A cluster or job with no idle-timeout / auto-termination, so transient compute lingers after the work is done.

We only recommend scale-to-zero where the platform supports it and the workload can absorb the first-request cold start (typically a second or two), and we call that trade-off out explicitly.

How to act

Enabling autoscale or dropping the floor is reversible — restore the original minimum and the resource is back to fixed capacity immediately. leancosts drafts the autoscale rule (or the --min-replicas 0 flip), what to watch for cold-start impact, and the rollback. Follow Act on a cost finding.