ML Model Traceability and Reproducibility by Design
summary
Claims a reproducible deployment pipeline that ties code, runs, artifacts and serving together using GitHub Actions to pass git SHA, MLflow tags with Databricks run ID, and Docker image with embedded model and env vars. The pipeline uses Databricks workflows for training, MLflow model registry, Azure Container Registry and AKS for serving, returns git SHA and run ID with each prediction, and enables rollbacks via workflow dispatch. It works end to end in production and next work is data versioning and registry cleaning.
AI-written summary · about
chapters (11)
- 0:00ML traceability challenges
- 3:21Defining traceability
- 5:53Traceability design
- 6:37Five steps to traceability
- 17:26Traceability workflow
- 18:44Docker metadata Q&A
- 21:44Data versioning future
- 23:23Reusable pipelines
- 25:03Data science empowerment
- 27:17Shared infrastructure
- 28:26Closing remarks
related talks
32mReproducibility and versioning of ML systems
24mFrom Model to Production: Automating the ML Lifecycle with CI/CD
57mManaging Machine Learning Lifecycle with MLflow
17mContinuous deployment for ML: The new software development lifecycle
33mOperationalizing Model Lifecycle with MLFlow
57mDevOps for Machine Learning
33mBuilding Reproducible ML Processes with an Open Source Stack
27mThe Model Review: improving transparency, reproducibility, & knowledge sharing...