COPING WITH THE CHALLENGE OF SORTING LARGE PRODUCT CATALOGS

1. COPING WITH THE CHALLENGE OF SORTING LARGE PRODUCT CATALOGS ONLINE - SHOP WINDOW ARRANGEMENT CAGDAS SENOL 19-06-2019

2. TABLE OF CONTENTS ● Challenges ● Data Driven Sorting ● Three Improvements for faster iteration ● One Open Source Contribution ● Status Quo ● Q&A 2

3. ZALANDO AT A GLANCE ~ 5.4 billion EUR > 300 million revenue 2018 3 > 15,500 > 80% employees in Europe of visits via mobile devices visits per month > 400,000 > 27 product choices million ~ 2,000 17 brands countries active customers

4. DISCLAIMER 4

5. Window Dressing 5

6. “ Window dresser Window dressers arrange displays of goods in shop windows or within a shop itself. Such displays are themselves known as "window dressing". They may work for design companies contracted to work for clients or for department stores, independent retailers, airport or hotel shops. “ 6

7. DATA DRIVEN SORTING 7

8. CHALLENGES 400k 15 CHALLENGES 4k ups 17 8

9. 9

10. Fail fast, Iterate faster 10

11. ITERATE FAST DESIGN A/B TESTS ANALYSIS 11

12. Three Improvements ● Steering ● Fast Index Updates ● Sorting with functions 12

13. First Improvement: Sort Steering 13

14. Sort Steering - SQL Analogy Id Bucket Popularity sku1 1 0.2332332 sku2 2 0.123233 sku3 1 0.4533 Sku2 SELECT * FROM articles ORDER BY Bucket Desc, Popularity DESC Sku3 Sku1 14

15. Sort Steering - SQL Analogy Id Bucket Popularity Popularity_male sku1 1 0.2332332 0.4 sku2 2 0.123233 0.6 sku3 1 0.4533 0.1 If category_gender == “men” SELECT * FROM articles ORDER BY Bucket DESC, Populartiy_male DESC Else SELECT * FROM articles ORDER BY Bucket DESC, Popularity DESC 15

16. Pre-Sort Steering Architecture 16

17. Sort Steering Added 17

18. Sort Steering - SQL Analogy Id Bucket Popularity Popularity_male sku1 1 0.2332332 0.4 sku2 2 0.123233 0.6 sku3 1 0.4533 0.1 If category_gender == “men” SELECT * FROM articles ORDER BY Bucket DESC, Populartiy_male DESC Else SELECT * FROM articles ORDER BY Bucket DESC, Popularity DESC 18

19. 19

20. 2nd Improvement: Decoupled Data Ingestion 20

21. Indexing - SQL Analogy Id Price Stock Size Partner Performance Performance_new_formula sku1 9.99 100 32 false 0.5 0.4 sku1 9.99 100 32 false 0.3 0.6 INSERT INTO articles VALUES(9.99, 100, 32, false, 0.5, 0.4) INSERT INTO articles VALUES(9.99, 100, 32, false, 0.3, 0.6) 21

22. Intake Architecture 22

23. Indexing - SQL Analogy Id Price Stock Size Partner Id Performance Performance_new_formula sku1 9.99 100 32 false sku1 0.5 0.4 sku1 0.3 0.6 JOINS => Elasticsearch 23 Id Price Stock Size Partner Performance Performance_new_formula sku1 9.99 100 32 false 0.3 0.6

24. Intake Architecture Now 24

25. 3rd Improvement: Sorting with Functions 25

26. Painless Scripts 26

27. Sorting with Functions - Eliminate Reindexing Id Price Stock Size Partner clicks sales sku1 9.99 100 32 false 10000 300 SELECT * FROM articles ORDER BY popularity(sales,clicks) popularity (sales, 27 clicks) = sales/clicks

28. Sorting with Functions - Personalization Id Price Stock Size Partner popularity article_features sku1 9.99 100 32 false 1.2 [9.99, 100, 32] If known_customer : SELECT * FROM articles ORDER BY dot_product(article_feature, customer_features) Else SELECT * FROM articles ORDER BY popularity 28

29. Sorting with Functions - Fulltext Search Id Price Stock Size Partner clicks sales sku1 9.99 100 32 false 10000 300 If fulltext_search : SELECT * FROM articles ORDER BY f(relevance_score, clicks, sales, customer_features, article_features) Else SELECT * FROM articles ORDER BY g(clicks, bucket, sales, customer_features) 29

30. EXAMPLE SORTING RULES 30

31. Personalization 31

32. Query Relevance 32

33. Inline Popularity Calculation 33

34. Open Source Contribution 34

35. CODE IN CONFIG ???????? 35

36. TGYHT - Thanks God You Have Tests 36

37. Make Painless Script Development Painless • • • 37 Painless Lacks Tooling Elasticsearch Painless Execute API https://www.elastic.co/guide/en/elasticsearch/painless/cur rent/painless-execute-api.html

38. Painless Scripts Development Tool • • 38 https://github.com/csenol/plsd Integrated with CI/CD Pipelines

39. Painless Script Development Environment 39

40. Painless Script Performance Tests 40

41. TESTING PAINLESS SCRIPTS/ CI-CD Integration 41

42. Sum up ● Sort Steering => A/B tests ● Decoupled Indexing => Data Enrichment ● Sorting With Functions => Faster Implementation + Personalization 42

43. Notable window dressers [edit] ● Giorgio Armani, the fashion designer once worked as a window dresser. [1] 43

44. Cagdas Senol cagdassenol@gmail.com

45. Q&A 45