Whenever you're ready, walk me through how you'd approach this.
How do you approach designing a system for high availability and low latency in a cloud-native environment?
Explain how you would clarify requirements, decompose the ML serving path, and choose between online and batch components. Cover deployment topology, traffic management, caching, autoscaling, observability, graceful degradation, and disaster recovery. Discuss how you would validate latency and availability before production, and how you would detect feature drift and training-serving skew.