Autoscaling Policy - CI/CD Glossary Definition
An autoscaling policy is the rule set that decides when to add or remove runners based on a signal such as queue depth, CPU utilization, or a custom metric. It defines scale-up and scale-down thresholds, cooldown periods, and minimum and maximum instance counts.
Queue-depth scaling
For CI runner fleets, the most useful signal is the number of queued (pending) jobs. A policy might add one runner per N queued jobs and scale down after a runner sits idle for a cooldown window. Scaling on CPU alone is a poor proxy because a runner waiting for a job is idle even when work is queued.
Avoiding thrash
Cooldown periods and hysteresis (different up and down thresholds) prevent rapid add/remove cycles, which waste boot time and money. Setting a minimum of zero enables scale-to-zero during quiet hours.