Palantir ML System Design: Scalable Music Recommender

You are asked to design a scalable, low-latency music recommendation service for a large streaming platform (think Spotify or SoundCloud) that serves personalized playlists and song suggestions to millions of users.

Focus on a high-level architecture that separates candidate generation and ranking, supports real-time updates from listening events, and provides RESTful APIs such as GET /recommendations?user_id=123&limit=10. Describe storage for user profiles, song metadata, listening history, feature stores, and a fast embedding/nearest-neighbor store for real-time retrieval.

A typical interview flow: clarify constraints and SLAs (latency <100ms, 99.9% availability); propose an end-to-end architecture (event ingestion, feature pipeline, offline training, online feature service, ANN store, ranking service, caching, and API gateway); sketch data models and schema choices; and discuss operational concerns (scaling, monitoring, A/B testing, and cost trade-offs).

Skills and signals you should show: system design thinking for distributed services, knowledge of recommender architectures (candidate generation vs ranking), experience with streaming data (Kafka), feature stores and online features, approximate nearest neighbor (ANN) techniques (Faiss/HNSW), caching strategies (Redis/CDN), model deployment/versioning, and evaluation (offline metrics + online A/B tests). Also explain how you would handle cold-start, privacy, and evolving models without downtime.