Version control tracks and manages successive versions of the code, models, features, data artifacts, and dependencies that make up a machine-learning system. It supports reproducible delivery by making changes identifiable, testable, and recoverable. section 1.3 section 5.5.2
Treat versioning as delivery infrastructure rather than an administrative afterthought. Establish it while building pipelines and data tests, before exploratory analysis and model development depend on shared artifacts. section 5.0 section 5.5.2
Use a compact verification process for every important artifact:
- Identify and register the artifact and its dependencies. section 5.5.2
- Use checksums to confirm that the expected information is present. section 5.5.2
- Use signed binaries to verify ownership, source, and trust. section 5.5.2
- Sample known values from binaries and check that they remain correct. section 5.5.2
Production delivery checkpoint
Before a CI/CD build runs, verify that the pipeline, data, model, feature, and dependency versions match the registered configuration. Build systems such as Jenkins, GCP Cloud Build, and AWS Code Build can then deliver that configuration consistently instead of relying on manual assembly. section 5.5.2 section 1.3
