CI/CD, Automation, and ML Governance
Domain 7 covers CI/CD for code, data, and models, plus governance evidence: approvals, lineage, model cards, and audit trails.
Pipeline rules
- CI runs unit tests, linting, data validation, and model behavior tests on every change.
- CD promotes immutable artifacts through registry stages; production changes require an approval gate and a rollback path.
- Tests should cover schema, feature logic, prediction invariants, bias checks, and performance budgets.
- Express release policy as code-thresholds, required approvals, and evidence checks-so gates are repeatable and auditable.
- Governance artifacts include the model card, intended use, limitations, training data lineage, evaluation results, and approval records.
Worked example
A pull request changes a feature transformation. CI runs unit tests, a data-contract check, a retraining job, and an evaluation gate: accuracy must be at least the current baseline, p95 latency at most 50 ms, and subgroup recall gap at most 5 points. The run passes accuracy and latency but fails the gap at 7 points. Reasoning: the gate blocks promotion even though the headline metric improved, because fairness is a release criterion. The team adds reweighted training data, reruns the gate, registers the model, and requests manual approval for a canary.
Traps
- CI that tests only application code, not data or model behavior.
- Gates with no baseline or no owner.
- Mutable artifacts that cannot be reproduced.
- Governance performed after deployment.
- Approvals with no audit record.