MLOps Architecture for a Recommendation Platform
summary
A presentation describes an MLOps architecture used to deploy a recommendation platform across Sber’s ecosystem, covering services from streaming to e‑commerce. It explains the platform’s universal approaches for building personalized recommendations and the operational components needed to run those models in production.
AI-written summary · about
chapters (17)
- 0:00MLOps intro
- 1:50Recommendation platform
- 5:47Data processing
- 10:26Offline recommendations
- 13:15Online recommendation systems
- 17:34Near-real-time recs
- 20:02Platform tools
- 21:05Q&A: Data spikes
- 23:11Q&A: Experiment reproducibility
- 24:38Q&A: Measuring rec quality
- 26:00Q&A: Artifact storage
- 26:44Q&A: Spark & Kubernetes
- 28:54Q&A: Feature store workflow
- 30:43Q&A: Inference orchestrator
- 31:47Q&A: Vector DB scaling
- 33:16Q&A: GPU resources
- 35:48Q&A wrap-up
related talks
41mTHE STATE OF MLOPS - MACHINE LEARNING IN PRODUCTION AT ENTERPRISE SCALE
47mCentralized MLOps Platform: From R&D to Inference in a Single Loop
53mExperience with an end-to-end MLOps pipeline: Practical insights and 3D architectural perspectives
45mThe Culture, Architecture and Technology of MLOps
13mMLOps at Scale
33mModel Development and Deployment Tools. An Introduction to MLOps.
52mA MLOps Platform for All Data Personas
41mMicroservices for Machine Learning