Bringing CI/CD Practices to Machine Learning with MLOps
summary
Claims that MLops is 80% operations and recommends treating data processing as versioned, unit‑tested libraries instead of notebooks, shipping wheels via CI/CD and using CalVer for dataset versions. Advises to bake weights into containers (reduces cold start by several minutes at one company), manage GPU images with Packer/Terraform, and use inference stacks like Triton + pre‑queries and MLPerf for performance. Emphasizes shadow deployments for real traffic testing and continuous data observability for drift detection.
AI-written summary · about
chapters (12)
- 0:00MLOps introduction
- 1:46MLOps is 80% software ops
- 3:33Model development cycle
- 5:17Data prep as code
- 7:02Continuous data delivery
- 8:47Model training challenges
- 10:29Embedding model weights
- 11:34Serving models with GPUs
- 13:20Performance tuning
- 14:43Shadow deployments
- 16:48Observability and drift
- 18:56Closing remarks
related talks
39mContinuous Delivery for Machine Learning — The Future of MLOps
01mMLOps: Applying DevOps to ML
43mNavigating the MLOps Journey: Key Considerations for Successful Implementations
13mMLOps at Scale
41mTHE STATE OF MLOPS - MACHINE LEARNING IN PRODUCTION AT ENTERPRISE SCALE
25mBuilding Scalable MLOps Pipelines That Actually Work
13mMLOps: From AI Research to Production-Ready Systems
30mMLOps for Continuous Learning