Kubernetes autoscaling is often treated as a solved problem. Enable HPA, add a node autoscaler such as Cluster Autoscaler or Karpenter, and expect the system to balance performance, reliability, and cost automatically. In real production environments, this assumption often breaks down.

This talk explores why autoscaling fails at scale, not because individual tools are broken, but because scaling decisions are made in isolation. Node autoscaling, workload autoscaling, bin packing, and cost optimization may work well independently, but their interactions introduce failure modes when combined in real clusters.

Based on production experience operating large Kubernetes environments, the session highlights pitfalls such as misleading utilization signals, disruption during scale-down events, and cost optimizations that negatively impact reliability. The talk reframes autoscaling as an orchestration problem and offers a practical mental model for evaluating trade-offs in production systems.