Strategies of
Machine Learning Platform
Building & Practices in eBay
Agenda
AI Platform
vision, design
principles and
core capabilities
1
AI/ML use case
analysis
2
3
Unified data
strategies
AI Use Cases
- Online data services – OTF FE
- Streaming events – NRT FE
- Offline batch/ETL datasets – Batch FE
Structured Data Semi/Unstructured Data
(image/video/text/3D/…)
Data Source
- Content generation/acquisition NRT pipeline
Unified online/offline feature store Unified online/offline content storeStorage
Data PiT Parity Online/offline PiT data strategies PiT data parity is not required
Feedback Loop - Short: Continuous online training
- Long: Offline PiT feature simulation Vendor/manual/auto labelling
mon Driver set & training set generation & management, catalog, data
lineage, etc.
CPU/GPU - CPU training and inferencing typically - GPU training and inferencing typically
Challenges of Building Enterprise ML Platform
Tends to invest
more on solutions
instead of platform
Lack of clear
boundary
between solutions
and platform
Lack of unified
data strategies
and self-service
support for ML
Platform building
Traditionally focus
more on training,
lack of enough
platform support
on data/feature
and inferencing
Lack of E2E
seamless
integration
strategies cross
feature, training
and inferencing
ML Development Lifecycle
Agenda
AI Platform vision,
design principles
and core
capabilities
2
3
Unified data
strategies
1
AI/ML use case
analysis
Our Vision
To empower eBay AI practitioners to build, train and
deploy machine learning models with fully-managed,
efficient and self-service platform at scale.
ML Platform Core Capability Map
ML Platform Architectural Principles
Enable self-service based on centralized configuration and metadata-driven design, with
lifecycle management and governance in place
Enable unified metadata and definitions cross online and offline, with enough
flexibility and extensibility to support domain level customizations
Provide a group of management APIs & services for MLP managed lifecycle, and enable
the E2E seamless integration based on the APIs
Provide unified catalogs (including data, stored variables, features, models, solutions,
etc.) to promote discovery, reuse and better governance
Provide E2E data lineages for the AI Platform domain entities
Apply unified monitoring cross the whole ML platform
ML Platform Online Integration Architecture
Entity Modeling in ML Platform
Dependency DAG & Execution Plan
Unified CPU/GPU Inferencing Platform
Model and Feature Monitoring
Agenda
3
Unified data
strategies
2
1
AI/ML use case
analysis
AI Platform
vision, design
principles and
core capabilities
Why Data Strategies are so Important for AI/ML
Image source: Cognilytica, from
Batch Feature
Feature DSL
NRT Roll-up Abstraction
NRT Feature Engineering
NRT Feature
Schema
Event processing
Derivedputation
On-the-fly Feature
parisons of Different Features Types
Batch Feature NRT Feature On-the-fly Feature
Online/offline PiT Strategy
PiT Simulation / Feature
Snapshotting
PiT Simulation / Feature
Snapshotting Feature Snapshotting Only
Reusability Easy to reuse Easy to reuse Solution by solution support
Time-to-Market Fast
Fast except new enriched
event acquisition
Slow
MLP Managed
Self-service by End Users
(DS)
Delay of Data Freshness
Data Source
Yes
Yes
Yes
Yes
No
No
1 Day+ P99 < 5 sec Real-time
ETL/Batch
data/Snapshotted Dataset
Enriched events
Request context /
Online data services
Embracing NRT Strategy
Integrated Data Strategies
Feature Platform
Unified Feature Store
Feature Lifecyle Mngt.
Feature PiT Simulation
Training Platform
Training Set Generation
Driver/training Set Mngt.
High-throughput Data Access
Inferencing Platform
Feature/Model Snapshotting
Unified Model Spec
API Spec Auto-Gen
么 ．ArchSumniit
全 球 架 构 师 歧 会
lnfoQ* 极 客 传焊
