Real-Time Last-Mile Delivery ETA Prediction System
Design a high-scale machine learning system to predict real-time Estimated Time of Arrival (ETA) for a last-mile logistics platform. The system must process 10M+ daily deliveries with a P99 latency of <100ms. Your design should specifically address: 1) Large-scale spatio-temporal feature engineering using spatial indexing, 2) Integration of real-time traffic and weather telemetry via streaming pipelines, 3) Modeling driver-specific behavior and heterogeneity, 4) Handling the 'last-100-meters' variance (parking/service time), and 5) An evaluation framework comparing predicted windows against actual delivery completions. Detail the data lifecycle from GPS ingestion to online inference and explain how you handle model drift in dynamic urban environments.
XGBoostH3Apache FlinkKafkaRedisQuantile RegressionTriton Inference ServerLightGBM
00