《Lyft 市场中的实时 ML.pdf》由会员分享,可在线阅读,更多相关《Lyft 市场中的实时 ML.pdf(25页珍藏版)》请在三个皮匠报告上搜索。
1、Real Time ML in Marketplace LyftBy RakeshMarketplace Org1Agenda Introduction Use cases Overall architecture Dev Ex Lesson Learned Data Points QA2Introduction3Dynamic PricingSupply/Demand curveETAPricingNotificationsDetect DelaysCouponsUser DelightFraudBehavior FingerprintingMonetary ImpactImperative
2、 to act fastTop DestinationsCore Experience4Use Cases5Use Case 1:Dynamic PricingDynamically adjust prices to ensure service reliability and optimal throughput,considering imbalances in marketplace demand and supply6Use Case 2:Supply ManagementManage Supply(drivers)in a given region by creating real-
3、time incentive and directing them to under supplied areas to improve service reliability7Use Case 3:Short-Term ForecastingPredict diverse elements of a fluid marketplace to enable subsystems to adapt efficiently,thereby enhancing overall performance and maximizing throughput.8Use Case 4:Fraud Detect
4、ionThe fraud detection team requires access to real-time data points to promptly identify and prevent any fraudulent activities occurring on the Lyft platform.9High-Level Architecture10Feature Generation PipelineTBD Enriches raw data into high-quality tagged data Shared between aggregation pipelines
5、 Aggregation pipelines generate real time ML features Features are indexed in feature store and published to Kafka topics11Model Training PipelinesTBD Models training is triggered by smart triggerModel training frequency=30 seconds Trained and serialized model is stored in artifactory for model serv
6、ing12Model Inference Pipelines Consists of Model DAG Triggered by smart trigger Stores computed value in a service or Kafka topic13Entire Ecosystem10,000 feet view14Dev Ex15Dev ExperienceYaml Based Pipelines16Dev ExperienceModel Debugging/Data replayability17Lessons Learned18Lessons LearnedSeparatio
7、n of Concerns19Lessons LearnedSeparation of Concerns20Lessons LearnedSeparation of Concerns21Lessons Learned Standardization Guidelines Observability Better abstraction for usability22Data Points23Data Points Powering Business critical products at Marketplace Dynamic Pricing,driver incentive,real time supply management,etc.Computing dynamic pricing for 10s million geohashes Forecasting marketplace parameters for 350+regions Computing driver incentive in real time for 1 million hotspot24Q&A25