Question

A morning traffic ramp at 09:00 triggers a cascading collapse instead of a smooth scale-up. Dashboards: as load rose, autoscaler added pods; new pods take ~90s to warm caches/JIT before serving fast; during warm-up they're slow, so the LB's per-pod success rate drops, the LB marks them unhealthy and removes them, which dumps their share of load back onto the already-hot old pods, which then also start failing health checks — a removal spiral. Effective capacity is DROPPING as you add pods. No deploy. How do you triage and mitigate?