Best practices and solution design using data versioning for machine learning
summary
Claims a Git-first, branch-and-tag system that versions raw data, splits, models and monitoring so reproducibility is literal; demo uses Zab (Git-backed data store) or DVC style tooling, mlflow for experiment tracking, FastAPI for serving and Evidently for drift alerts. Specifics include data branches like raw/split/dev/prod, chunk deduplication to avoid storage bloat, commit-tagging models before training and persisting intermediate artifacts. Demo mostly runs end-to-end but shows a few realistic demo crashes.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 10:11Data versioning importance
- 22:16Branches and tags
- 31:10Demo setup raw data
- 40:10Developer branches model tracking
- 55:31Training models tagging
- 1:02:16Monitoring analysis branches
- 1:14:33Best practices recap
- 1:22:43Q&A
related talks
59mDevOps for Machine Learning Projects
32mReproducibility and versioning of ML systems
35mData versioning in machine learning projects
42mApplications in ML Drug Discovery pipelines
31mGetting ML Right in a Complex Data World
32mMachine Learning Models Versioning Using Open Source Tools
24mModel Versioning: Why, When, and How
37mOpen Source Tools for ML Experiments Management