| react to one sample | stabilization window |
| scale on total lag only | age and lag derivative |
| rapid scale cycles | hysteresis and cooldown |
| scale down mid-work | graceful drain and commit |
| more replicas than partitions | hard useful ceiling |
| Monitor oldest-event age, arrival/service rate, rebalance duration, idle consumers, processing p99, and replica oscillation. Test 10x spikes and dependency slowdown separately. | |
| ## Test yourself | |
| 1. What causes autoscaling flapping? | |
| 2. Why can adding replicas temporarily reduce throughput? | |
| 3. Which signal distinguishes spike from slow sink? | |
Continue to professional.md. | |