Milvus 实战系列 #2：元宇宙到家，那些“聪明”的设计工具

1 .那些“聪明”的设计工具 

2 .Contents 1 Business Background 2 Technology Selection 3 System Design 4 Search in Action 5 Summary 

3 .Business Background 

4 .Business Background Business unit Scenario Pain point Home furnishing enterprise System furnishing selling Difficult to do MCC pre-build 1 

5 .Real World 2 

6 .Real World 3 

7 .Overview Design tool Business logic Data generation • Item attributes • Rule base • User journey • Prior knowledge • Extra operations • Item describe • …… • …… Data smart offering platform 4 

8 .Technology Selection 

9 .Data flow overview The data flow is similar with the traditional search engine / recommendation engine, which include offline data preparing, online serving and post processes Business logic Data fulfillment Extra operation Serving Recall Ranking Reorder BFF Data preparing Definition Generation Deployment Online Offline 5 

10 .Requirements 01. Scenario 02. Serving 03. Storage 04.Support H5 Recall Stable Evolution App Ranking Threshold Self-repair Mini-program Response Compatibility Downgrade & fusing …… …… …… …… Multi-platform High-performance Block-free Robustness 6 

11 .Online Components Ranking • Multiple roles Recall • Model base • Role base • Vector distance • Multiple dataset • Sample sorting VS ES BE Re-order • BU support • Business insight • Diversity 7 

12 .Vector Search Descriptive Vector search Differ from the traditional search solutions, vector search engine afford the store & search ability on vector data. Evolution With the improvements on AI technology, especially the ML/DL ones, we use unstructured data (vector base) to describe the items, trying to finger out the attributes, behavers, interests for our machine understanding, and do our best to serving our customers Fusion 8 

13 .Solutions Comparison 9 

14 .System Design 

15 .System Architecture 10 

16 .Dataflow offline Definition • Items definition against business insight Data Generation • Data calculation • Scoring & quality assurance Feature Engineering • Label recognition Recall Data Preparing • Feature encoding • Data convert & searchable data generation • Dataset management Ranking Data Preparing • Build / upgrade model dataset • Dataset management 11 

17 .Dataflow Online BE QU SP RA BE Receive Recall Post-process Receive & prepare query params Prepare query dsl Data convert & fulfillment + + + User profiles Multiple dataset Other operations for downstream Understand Rank Finger out the key params Do data scoring & ranking + + Prepare policy & sort params Other biz operations 12 

18 .Search in Action 

19 .Real Data TagsGroup1: 2360 * 1000 * 58 TagsGroup2: 居家型衣柜 TagsGroup3: 挂衣、叠衣、被褥放置、裤子收纳 TagsGroup4: 舒适型价格区间 TagsGroup5: T型分割件 TagsGroup6: …… 13 

20 .Index Design (ES) 14 

21 .Real Data Vector1: [1, 0, 0, 1, 0, 1, 1, 0 …] Vector2: [1.134246, -0.000498, 0.176506, 0.971405, -1.875313, …] Vector3: …… 15 

22 .Index Design (Milvus) 16 

23 .Resource Estimation ES PostgreSQL v7.13 v11 4C8G20G M *3 1C2G * 1 4C8G20G IC * 3 8C32G500G ID * 8 OSS Milvus 1.5 ~ 2 TB V1.0 8C16G500G * 1 17 

24 .Recall Dsl (ES) 18 

25 .Recall Dsl (Milvus) 19 

26 .Summary 

27 .Issues & Solutions Complex dataflow Dataset Separate Node optimization Too much calc nodes Separate into sub-sets Hyperparameter tuning Performance Parallel calculation Algorithm tuning Quality DAG optimization Expert system Issue Solution 1 Solution 2 20 

28 .Issues & Solutions Performance More candidate Customized build Approximate calculation Multiple dataset recall More algorithms involve Vector representation Reflow data Manually blind test Comprehensibility MCC BU base Issue Solution 1 Solution 2 21 

29 .Metric comparison 01 Speed up Average design time has been decreased from 180+ mins to around 30 mins 03 Serving number Global rollout which serving X BU, Y costumers in total 02 Recall rate Wardrobe proofing from 3.2 to 20 04 Sales raise Help to settle over X orders, ATV raises from Y to Z 22