Real Time ML Pipelines in Multi-Tenant Environments
summary
Built a real‑time ML pipeline that keeps training and serving feature logic identical by using a common feature definition, a precomputed feature store and streaming scoring on Spark Streaming. Uses an in‑house Optimus Prime layer over Spark ML and Algebird‑style monoid aggregators to update sliding windows and avoid repeated joins. Targets Salesforce scale (100k+ tenants, planning for hundreds of thousands to millions of models); system runs now for thousands of tenants but model serving and CDC at true multi‑tenant scale remain challenges.
AI-written summary · about
chapters (6)
- 0:00Introduction and overview
- 5:15ML pipeline basics
- 11:25Real-time challenges
- 12:31Architecture and feature store
- 27:14Key takeaways and future work
- 31:40Q&A session
related talks
24mUnder Pressure: Applying ML in Real Time
49mImpact on business and society
44mBuilding Multi-Tenant Compute Systems in the Enterprise
36mReal-Time ML Pipelines: Feature Chaining with Chronon
41mInfrastructure for ML Applications
33mDeploying and Scaling Spark ML and Tensorflow AI Models
28mBuilding a Coud Native Data Pipeline for Your Real-Time ML & AI Apps
36mBuild prod ML systems with feature /training /inference pipelines