Unified Feature Store

1. Unified Feature Store @ eBay Yucai Yu eBay AI Architect

2.

3. Agenda • Feature Pain Points in Production AI • Unified Feature Store @ eBay • Demo • Landing Use Cases

4. AI Project Life Cycle Applied machine learning is basically feature engineering - Andrew Ng

5. Major Pain Points in Feature Long TTM (Time-to-Market) - Lack of point-in-time online/offline feature parity - Communication cost cross teams is not cheap - Lack of self-service and automation tools for feature life-cycle management and integration Feature discovery & reuse - Need feature catalog for discovery and exploration - Need feature data lineage information - Lack of unified loading key management cross all checkpoints Data quality monitoring & alerting - Need active monitoring & alerting on feature data validation and data shift

6. Agenda • Feature Pain Points in Production AI • Unified Feature Store @ eBay • Demo • Landing Use Cases

7. Architecture Visions & Capabilities • Enable self service for feature engineering with governance • Leverage unified DSL to define features for both online and offline • Manage the whole feature life cycle • Support seamless fast feature deployment & release • Enable online/offline data parity, support feature point-in-time & backfill • Provide feature catalog for discovery and reuse cross eBay

8. Taichi Feature Store Feature versioning Online/offline feature parity Feature lifecycle management Fast feature backfill Feature consistency Feature access control Feature discovery and reuse Feature documentation & analysis

9. Feature Store in Industry • Uber: Michelangelo Palette • Airbnb: zipline A Declarative Feature Engineering Framework • Google: Vertex Feature Store • AWS: SageMaker Feature Store • Databricks: https://www.databricks.com/product/feature-store • Feast/Tecton: https://github.com/feast-dev/feast • OpenMLDB: https://github.com/4paradigm/OpenMLDB • Feathr: https://github.com/linkedin/feathr

10. Unified Feature Store @ eBay • Architecture • Feature Management • Feature Engineering & Parity • Feature Serving & Reuse

11. Architecture

12. Unified Feature Store @ eBay • Architecture • Feature Management • Feature Engineering & Parity • Feature Serving & Reuse

13. Feature Organization eBay-speci c curated and crowd-sourced feature Organized as {domain}:{feature type}:{feature group name}#{feature template name}{^args (optional)}:{key name(s)} "risk:nrt:fg1#page_view_duration_userid:Buyer.Id" "ads:batch:PromotedList#page_view_duration_userid:Buyer.Id"

14. Key Management Manage all the loading keys Each key should have its unique name, belong to a key dimension Provide catalog for key discovery and life-cycle management

15. Feature Definition • Strong data types • UDF extension • Interoperable with Java • High execution efficiency • Static validation and dependency extraction

16. Feature Definition

17. Feature Life Cycle Management Centralized con guration and metadata-driven design for feature lifecycle management

18. Unified Feature Store @ eBay • Architecture • Feature Management • Feature Engineering & Parity • Feature Serving & Reuse

19. Batch Feature

20. NRT(Near-real-time) Feature

21. NRT(Near-real-time) Feature Generic pipeline with built-in roll-up variable types support • Sliding Window • LastK • Time Decay • Event Sequencing

22. NRT Feature - Write & Read

23. MDC Strategy

24. On-the-fly Derived Feature

25. Unified Feature Store @ eBay • Architecture • Feature Management • Feature Engineering & Parity • Feature Serving & Reuse

26. Online Serving • Feature in production: reliability, scale, low latency • Complex data type support • Schema backword compatibility

27. Feature Reuse & Isolation

28. Offline Training Set Features Driver Set • “risk:otf:fg1#connected_bad_slr:Buyer.Id” • “risk:nrt:fg1#sum_lstg_price_by_slr_sw_6h:Seller.Id” • “risk:batch:fg2#trust_mgid_txn_amt:Buyer.Id”

29. Offline Training with Feature Store

30. Agenda • Feature Pain Points in Production AI • Unified Feature Store @ eBay • Demo • Landing Use Cases

31. Feature Catalog

32. Feature Onboarding

33. Feature Point-in-time Simulation

34. Feature Release & Online Serving

35. Agenda • Feature Pain Points in Production AI • Unified Feature Store @ eBay • Demo • Landing Use Cases C

36. Landing Use Cases Domains: Use Cases: • Ads, Buyer Experience, Marketing • Knowledge Hub, Search • Risk, Shipping, Customer Support • • • • • • • Promotion Display Guidance Merch Home Page Personalization Sneaker Size Image Embedding …….

37.

38. Offline Feature Store ○ Point-in-time features is not the CURRENT features

39. Batch Feature Point-in-time • Build on top of HBase/HDFS • Assuming transaction happened in 20211213 00:00:59, look up key "ship_addr_cb_txn_var:1.0#7653|20211212 01:10:10".

40.

41. NRT Feature Point-in-time