Building Resilient (ML) Pipelines for MLOps
summary
Claims a platform approach works where ad hoc boxes fail and shows a concrete architecture: feature training inference split, a metadata view abstraction with point‑in‑time joins layered over versioned state, and storage‑centric orchestration that stays tool‑agnostic (Python, SQL, Polars, Pandas, Flink). Presents a surprising operational scale — 2,000 models managed — describes an AI lakehouse feature store implemented at enterprise scale, offered serverless in cloud, and used successfully in production.
AI-written summary · about
chapters (13)
- 0:00MLOps introduction
- 1:51Production ML problems
- 3:18Data vs model performance
- 6:28MLOps definition
- 7:54Model lifecycle decay
- 9:38Modular ML pipelines
- 12:07Feature/training/inference
- 15:59Feature engineering
- 18:48Versioning & monitoring
- 20:57Temporal data joins
- 22:01Batch/real-time/LLM use
- 24:08Platform resources
- 25:39Q&A
related talks
25mBuilding Scalable MLOps Pipelines That Actually Work
53mExperience with an end-to-end MLOps pipeline: Practical insights and 3D architectural perspectives
13mMLOps at Scale
33mBuilding & Managing MLOps Pipelines in the Cloud
10mFrom Model to Production: Building Scalable MLOps Pipelines
41mTHE STATE OF MLOPS - MACHINE LEARNING IN PRODUCTION AT ENTERPRISE SCALE
13mMLOps: From AI Research to Production-Ready Systems
45mThe Culture, Architecture and Technology of MLOps