Strategies

ofMachineLearningPlatformBuilding&Practicesin

eBayAgendaAIPlatformvision,designprinciplesandcore

capabilities1AI/MLuse

caseanalysis23Unified

datastrategiesAIUse

CasesOnlinedataservices–OTF

FEStreamingevents–NRTFEOfflinebatch/ETLdatasets–Batch

FEStructured

DataSemi/Unstructured

Data(image/video/text/3D/…)Data

Source- Contentgeneration/acquisitionNRT

pipelineUnifiedonline/offlinefeature

storeUnifiedonline/offlinecontent

storeStorageDataPiT

ParityOnline/offlinePiTdata

strategiesPiTdataparityisnot

requiredVendor/manual/auto

labellingDriver

set &trainingsetgeneration&management,catalog,datalineage,

etc.- GPUtrainingandinferencing

typicallyChallengesofBuildingEnterpriseML

PlatformTendstoinvestmoreonsolutionsinsteadof

platformLackofclearboundarybetween

solutionsand

platformLackofunifieddatastrategiesandself-servicesupportforMLPlatform

buildingTraditionally

focusmoreontraining,lackofenoughplatformsupportondata/featureand

inferencingLackofE2Eseamlessintegrationstrategies

crossfeature,trainingand

inferencingMLDevelopment

LifecycleAgendaAIPlatform

vision,designprinciplesandcorecapabilities23Unified

datastrategies1AI/MLuse

caseanalysisOur

VisionToempowereBayAIpractitionerstobuild,trainanddeploymachinelearningmodelswithfully-managed,efficientandself-serviceplatformat

scale.MLPlatformCoreCapability

MapMLPlatformArchitectural

PrinciplesEnableself-servicebasedoncentralizedconfigurationandmetadata-drivendesign,

withlifecyclemanagementandgovernancein

placeEnableunifiedmetadataanddefinitionscrossonlineandoffline,withenoughflexibility andextensibilitytosupportdomainlevel

customizationsProvideagroupofmanagementAPIs&servicesforMLPmanagedlifecycle,andenabletheE2Eseamlessintegrationbasedonthe

APIsProvideunifiedcatalogs(includingdata,storedvariables,features,models,solutions,etc.)topromotediscovery,reuseandbetter

governanceProvideE2EdatalineagesfortheAIPlatformdomain

entitiesApplyunifiedmonitoringcrossthewholeML

platformMLPlatformOnlineIntegration

ArchitectureEntityModelinginML

PlatformDependencyDAG&Execution

PlanUnifiedCPU/GPUInferencing

PlatformModelandFeature

MonitoringAgenda3Unified

datastrategies21AI/MLuse

caseanalysisAIPlatformvision,designprinciplesandcore

capabilitiesWhyDataStrategiesaresoImportantfor

AI/MLImagesource:Cognilytica,from

Batch

FeatureFeature

DSLNRTRoll-up

AbstractionNRTFeature

EngineeringNRT

FeatureSchemaDerivedputationOn-the-fly

FeatureparisonsofDifferentFeatures

TypesBatch

Feature NRT

FeatureOn-the-fly

FeaturePiTSimulation/FeatureSnapshottingFeatureSnapshotting

OnlyEasyto

reuseSolutionbysolution

supportTime-to-MarketFastFastexceptnewenrichedevent

acquisitionSlowMLP

ManagedSelf-servicebyEndUsers(DS)DelayofData

FreshnessData

SourceYesYesYesYesNoNo1Day+P99<5

secReal-timeETL/Batchdata/Snapshotted

DatasetEnriched

eventsRequestcontext

/Onlinedata

servicesEmbracingNRT

StrategyIntegratedData

StrategiesFeature

PlatformUnifiedFeature

StoreFeatureLifecyle

Mngt.FeaturePiT

SimulationTraining

PlatformTrainingSetGenerationDriver/training