Scalable Competitive Matchmaking System Design
Design a machine learning system to facilitate competitive matchmaking for a global online gaming platform with 50 million monthly active users. The system must balance players into two teams of five to ensure a fair competitive environment (win probability near 50%) while keeping wait times under 30 seconds. Discuss the end-to-end lifecycle, including regional data ingestion via streaming, feature engineering for player skill and behavior, the choice of a ranking model vs. a heuristic baseline, and how to handle low-latency inference at a scale of 1M+ concurrent users. Additionally, address how you would evaluate match quality offline and manage the trade-off between skill gap and queue latency in real-time.
LightGBMXGBoostKafkaFlinkSparkRedisSimulated AnnealingTrueSkillFeature Store
00