Design a ranking and recommendation system

Design a ranking + recommendation system for a personalized feed (e.g., short videos, posts, or products).

Requirements

Core goals

• Given a user U , return a ranked list of items for their feed.
• Support both:
  • Home feed (personalized recommendations)
  • Optionally: Following feed (content only from followed creators)

Functional

• Candidate generation (retrieve a large pool of potentially relevant items).
• Ranking (order items by predicted utility).
• Handle freshness/recency, diversity, and content policies (blocks, safety filters).
• Support exploration vs. exploitation.

Non-functional

• Low latency per request (e.g., p95 < 200 ms) with high QPS.
• High availability, graceful degradation.
• Near real-time adaptation (user actions should influence feed quickly).
• Observability and experiment platform (A/B tests).

Deliverables

Explain:

• High-level architecture and data flow (online serving + offline training).
• Storage choices (feature store, embeddings, logs).
• Model/feature strategy and how you ensure freshness.
• How you evaluate: offline metrics + online metrics/guardrails.
• Key failure modes and mitigations (cold start, feedback loops, abuse).