Whenever you're ready, walk me through how you'd approach this.
How would you design a notification system that is both scalable and performant?
Discuss the end-to-end ML and serving architecture, including event ingestion, notification candidate generation, ranking, delivery, deduplication, and user controls. Explain how you would separate batch and online work, meet latency and availability goals, evaluate notification quality, and detect feature drift, training-serving skew, and delivery failures. State assumptions clearly and justify major tradeoffs.