1. Feature Stores Building ML Infrastructure on Apache Pulsar 

2. Simba Khadder Co-Founder & CEO StreamSQL.io Using Apache Pulsar to power our feature store for >100m MAU 

3.Agenda ● The ML process ● Moving our ML Pipelines w/ Pulsar ● Building a Feature Store on top of Pulsar ● Q&A 

4. Machine Learning :: Model(Features) = Output Input Features Model Output Last 5 articles read Current article Top Genre Recommend Next Article Average Content Length Diversity of reading tastes .... 

5.Feature Engineering > Model Research* 

6. Behind every great model is a set of great features Input Features Model Output Last 5 articles read Current article Top Category Recommend Next Article Total time spent reading Diversity of reading tastes .... 

7.

8.Credit: Microsoft Azure Sales deck 

9.Our ML teams spent >80% of their time building and maintaining ML pipelines for feature generation and feature engineering. 

10. The Feature Engineering Cycle Hypothesis Deploy Feature to New Feature Production Validate New Feature Generate Training Dataset Increases Performance with new Feature 

11.Online Features Serving User ID Feature Set Training Data Arr([FeatureSet, Actual]) Train 

12. Input Features Model Output Last 5 articles read Current article Top Category Recommend Next Article Total time spent reading Diversity of reading tastes .... 

13. Generating Point-in-Time Correct Training Data Events Storage Training Data Features at Features at Features at timestamp timestamp timestamp read read read Time 

14. Generating Features for Serving (in a perfect world) Processor Online Features Event Stream Features at Features at Features at timestamp timestamp timestamp read read read Time 

15. Most Features are Stateful Input Features Model Output Last 5 articles read Current article Top Category Recommend Next Article Total time spent reading Diversity of reading tastes .... 

16. Stateful Features must be Bootstrapped Input Features Model Output Total time spent reading 

17. Bootstrapping Stateful Features with Historical Data in S3 SELECT user, SUM(readtime) FROM read_events GROUP BY user; 

18. Time Persisted in S3 Not in S3, but in Kafka retention period MsgID 

19. Finish bootstrapping & start stream processing from Kafka SELECT user, SUM(readtime) FROM read_events GROUP BY user; 

20.Full Feature Deployment Process 

21.

22. Combine Batch & Stream Processing with an Immutable Ledger ● Each new event appends to the end of the ledger ● Cut at an arbitrary point, and the ledger looks like a batch problem ● Only read from the head of the ledger and it looks like streaming problem 

23.Pulsar Based Architecture with Infinite Retention 

24.Pulsar’s offloading makes Event-Sourcing achievable 

25.Pulsar’s Tiered Architecture enhances Processing on Infinite Retention 

26.Feature are the building blocks of ML models; however, they are developed and maintained in ad-hoc ways. They lack a dedicated system of management. 

27. ML Pipelines < Feature Stores ● No concrete feature definitions, feature logic is split across Flink jobs. ● No feature versioning and rollback. ● No feature sharing, re-use, and discovery. ● No integrations into Tensorflow, Jupyter, etc. 

28.A Platform for features allows for teams to work together. Features are easily defined, shared, and re-used. There exists a single source of truth for features. 

29. Models across an organization may benefit from some of these features. Input Features Model Output Last 5 articles read Current article Top Category Recommend Next Article Total time spent reading Diversity of reading tastes ....