Design a real-time game recommendation system
Overview
This question evaluates proficiency in designing scalable, low-latency ML-powered recommendation system architectures, covering real-time data ingestion and streaming feature computation, real-time feature stores, service contracts for candidate generation and ranking, caching and freshness, personalization state, scalability, experimentation, failure recovery, and operational concerns like monitoring, privacy, and security. It is commonly asked to gauge an engineer's ability to reason about trade-offs between latency, freshness, consistency, and cost for global high-throughput platforms; category: ML System Design (recommender systems), domain: systems architecture and operations, level of abstraction: practical application and architectural design rather than model internals.
System Design: Real‑Time Game Recommendation System (Architecture Focus)
You are designing a real‑time recommendation system for a large gaming platform with millions of games and tens of millions of users. The focus is on system architecture and operations, not the internals of ML models.
Requirements
Describe the following aspects clearly:
- End‑to‑end request flow and latency SLOs.
- Online/offline data ingestion (user events, game metadata) and schemas.
- Streaming feature computation and a real‑time feature store.
- Candidate generation and ranking service interfaces.
- Caching strategy and freshness/TTL.
- User personalization state and session handling.
- Scalability plans (QPS estimates, partitioning, replication, storage).
- Experimentation and rollout (A/B, feature flags).
- Cold‑start strategies.
- Failure modes, consistency, and backfill/lag recovery.
- Privacy, security, and abuse prevention.
- Monitoring, alerting, and cost controls.
Assume:
- Tens of millions of MAU/DAU, mobile and desktop clients.
- Real‑time freshness for recent interactions (sub‑second to low seconds).
- Global user base with multi‑region traffic.
- The ML models exist but their internals are out of scope; you must define their service contracts and integration points.
Solution
Login required