System Design: Personalized Recommendations for a Consumer App

Context

Assume you are building the home-feed recommendations for a large consumer app (choose one: video streaming, news, or e-commerce). For concreteness, you may anchor your design on a video streaming app with both subscription and ad-supported content. You must propose a production-ready system that balances user experience with engineering constraints.

Requirements

Design an end-to-end personalized recommendation system and address the following:

1. Objectives and Constraints
   • Clarify business objectives (e.g., engagement, retention, revenue) and trade-offs (engagement vs. monetization).
   • Define latency SLAs (p95/p99) and freshness requirements for new content and user signals.
2. Architecture (end-to-end)
   • Data ingestion and logging, streaming vs. batch.
   • Feature store (offline and online parity).
   • Candidate generation (recall) strategies and services.
   • Ranking model(s) and serving.
   • Re-ranking for diversity, deduplication, and business rules.
   • Online inference path and caching.
   • Feedback loops for model updates and real-time features.
3. Modeling and Algorithms
   • Specify models for candidate generation (e.g., two-tower, collaborative filtering, content-based, graph/sequence models) and ranking (e.g., GBDT/DNN/MTL).
   • Handle cold-start for users/items and exploration (e.g., contextual bandits).
   • Real-time updates, deduplication, diversity, and content safety.
4. Evaluation and Experimentation
   • Offline metrics and validation.
   • Online metrics, guardrails, and an A/B testing plan (power, duration, ramping, and guardrails).
5. Scale and Reliability
   • Traffic assumptions, QPS estimates, storage sizing, and sharding.
   • Caching strategy, failure modes, and graceful degradation.
   • Monitoring/alerting (SLOs), privacy, and bias/fairness considerations.

Deliverable

Provide a structured proposal with diagrams-as-text or clear component breakdowns, explicit assumptions, and concrete numbers where helpful (e.g., latency budgets, QPS, storage). Include alternatives and trade-offs where relevant.