- Over-provisioning reduces latency
- No immediate risk of capacity exhaustion
- Scaling down may reintroduce errors
