Real-Time Hyper-Local Ad Recommendation System
Design a low-latency machine learning system for a mobile advertising platform that serves hyper-local ads based on real-time GPS coordinates and user movement trajectories. The system must handle 50M DAU and 100k active ads with a P99 latency under 100ms. Detail the end-to-end lifecycle: from streaming GPS ingestion and spatial indexing (e.g., H3/S2) to a two-stage retrieval and ranking architecture. Address specific challenges including high-cardinality spatial features, movement pattern encoding, privacy-compliant data handling, and the management of delayed feedback loops for physical store visits.
DeepFMH3KafkaFlinkRedisSparkONNXFeastTritonMLflow
00