Whenever you're ready, walk me through how you'd approach this.
How would you serve a large model with low latency requirements to millions of users?
Explain the end-to-end serving pipeline, including request ingestion, preprocessing, model execution, response delivery, deployment orchestration, autoscaling, caching, and observability. Address how you would handle traffic spikes, model versioning, partial failures, capacity limits, and safe rollbacks without compromising latency or availability.