从构建 LinkedIn 人工智能数据平台中汲取的经验教训.pdf

1、Lessons Learned From Building LinkedIns AI Data Platform QCon London 2024Flix GVFlix GVVenice CommitterData Infra since ConfigurationData CollectionFeature ExtractionData VerificationProcess Management ToolsMachine Resource ManagementAnalysis ToolsML CodeServing InfrastructureMonitoring“Only a small

2、 fraction of real-world ML systems is composed of the ML code.The required surrounding infrastructure is vast and complex.”Adapted from Hidden Technical Debt in Machine Learning SystemsAdvances in Neural Information Processing Systems 28(NIPS 2015)1AI LinkedIn2AI Ecosystem3Data Infra4Venice5Challeng

3、es&Lessons LearnedAgendaAI LinkedInAI LinkedInPYMKAI LinkedInPYMK(People You May Know)See the People You May Know:Fast Recommendations over Massive Data talk for more info.AI LinkedInThe FeedAI EcosystemAI PlatformProductive Machine LearningDouble our productivity developing ML and AI solutions.End-

4、to-end,unified,and opinionated platform.Easily develop,train,retrain,experiment with,deploy,and continuously maintain,monitor and debug a combination of state-of-the-art ML models.Cater to both AI researchers and AI engineers.AI PlatformFrameOpen ConnectModel DeploymentModel CloudHealth AssuranceFed

5、ExKing KongAI Metadata StoreReMixTritonFeature MarketplaceTensorFlow/PyTorchAI StudioQuasarML Agility DashboardFrameOpen ConnectModel DeploymentModel CloudHealth AssuranceFedExKing KongAI Metadata StoreReMixTritonFeature MarketplaceTensorFlow/PyTorchAI StudioQuasarML Agility DashboardModel Deploymen

6、tFrame Offline AnchorFrame Offline Anchor(Very)Brief OverviewFedExFrame Offline AnchorFrame Offline Feature DefinitionVenice Push JobVeniceFeature Join/PrepModel CloudFrame Online Feature DefinitionVenice ClientModelTop KOpen ConnectTrainingFeature Join/PrepModelRequestResponseFeature GenerationInfe