Why this matters
As an Applied Scientist, your models influence users, revenue, and compliance. Governance ensures your data and models are discoverable, documented, approvable, and auditable—so launches are safe, repeatable, and fast.
- Ship models with clear owners, approvals, and rollback plans.
- Answer audit questions quickly: what data was used, how it was processed, and why.
- Prevent avoidable incidents: data leakage, privacy violations, or biased outputs.
- Enable trustworthy collaboration with Legal, Security, and Product.
Real tasks you’ll face
- Register a new dataset with purpose, access, retention, and quality checks.
- Create a model card and get sign-off for a launch.
- Set risk-based gates (e.g., extra testing for high-impact decisions).
- Monitor drift and fairness; file an incident report if metrics degrade.
Concept explained simply
Data governance is about how data is collected, labeled, stored, accessed, and retired. Model governance is about how models are built, evaluated, approved, deployed, monitored, and retired.
- Data governance artifacts: data inventory, lineage, consent/purpose, retention, access controls, quality checks.
- Model governance artifacts: model registry, versioning, model cards, evaluation reports, approvals, monitoring, incident logs.
Mental model: a governed factory
Imagine a factory. Data are the raw materials with labels, sources, and safety notes. Models are finished products leaving with manuals and quality certificates. Governance tracks every step so you can recall, improve, or prove safety anytime.
Core components and roles
Artifacts
- Data Inventory entry: owner, purpose, legal basis/justification, sensitivity, retention, access, quality checks.
- Lineage: where data comes from and where it flows.
- Model Registry entry: owner, version, training data references, code snapshot, metrics, risk rating, approvals.
- Model Card: intended use, limitations, data summary, performance, fairness, robustness, safety mitigations.
- Approval record: who signed, when, and under which conditions.
- Monitoring plan: metrics, thresholds, frequency, alert routes.
- Incident log: what happened, impact, mitigation, lessons.
Typical roles
- Data Owner / Steward: accountable for dataset quality and access.
- Model Owner: accountable for model lifecycle and outcomes.
- Reviewer(s): Privacy/Security/Compliance provide guidance and approvals.
- MLOps/Platform: ensures versioning, deployment, and monitoring are reliable.
- Risk Committee (as needed): adjudicates higher-risk launches.
Minimal workflow (risk-based)
- Register data and model (inventory + registry).
- Document: data sheet and model card.
- Evaluate and sign-off: metrics meet pre-defined thresholds; reviewers approve.
- Deploy with monitoring and rollback plan.
- Operate: monitor, log incidents, review periodically, retire responsibly.
Risk-based gating tips
- Higher risk (e.g., impacts eligibility, safety, or vulnerable groups) → require deeper testing, fairness review, and multi-party approval.
- Lower risk (e.g., internal analytics) → lighter but still documented controls.
Worked examples
Example 1: Adding a new labeled dataset
Scenario: You receive a new batch of user feedback with labels.
- Create a Data Inventory entry: purpose (improve classification), sensitivity (may contain personal data), retention (90 days), access (project group only), legal/justification (user-submitted feedback).
- Lineage: source (feedback system), transformations (PII redaction pipeline), storage location, downstream models that consume it.
- Quality checks: label agreement ≥ 0.8, missing rate < 5%.
- Controls: role-based access, audit logging.
Example 2: Launching a moderation classifier
Scenario: A content moderation model will influence visibility decisions.
- Model Card: intended use (flag content), not for use (legal decisions), data summary, thresholding policy, performance and fairness metrics, known failure modes, fallback (human review).
- Approval: privacy review (no raw PII outputs), policy review (appeals workflow), risk committee sign-off (high impact).
- Deployment: canary release, monitoring (precision/recall by segment, false positive appeals rate), rollback plan (switch to previous model).
Example 3: Scheduled retraining with drift
Scenario: Quarterly retraining detects distribution shift.
- Registry: new model version with training snapshot and data references.
- Evaluation: compare against champion; require lift ≥ defined threshold and stable fairness metrics.
- Approval: standard sign-off if within risk bounds; escalate if fairness degraded.
- Operations: configure drift alerts; file a decision log entry justifying promotion.
Exercises
Try these before viewing the solutions. They mirror the exercises section where you can submit your answers.
Exercise 1: Map the governance for a dataset journey
Scenario: A clickstream dataset will be used to personalize recommendations. It includes user IDs and timestamps.
- List the governance artifacts you must create or update.
- Specify at least three controls to reduce risk.
- Name the accountable role for each artifact (e.g., Data Owner, Model Owner).
Checklist to guide you
- Inventory has purpose, sensitivity, retention, access, owner.
- Lineage shows sources, transforms, destinations.
- Quality checks defined and automated.
- Access is least-privilege; logs enabled.
Show a sample solution
See the Exercises section below for the full solution and hints.
Exercise 2: Draft a minimal Model Card
Scenario: You trained a staging toxicity classifier.
- Write a concise model card with: intended use, limitations, data summary, core metrics, fairness notes, safety mitigations, owner, approval status.
- Add a clear fallback if confidence is low.
Checklist to guide you
- States where not to use the model.
- Includes segment-level performance if available.
- Names the dataset versions used.
- Has monitoring and rollback plan.
Common mistakes and how to self-check
- Missing purpose/retention in the data inventory. Self-check: can you state why you hold the data and for how long?
- Leaking test data into training. Self-check: are data lineage and splits documented and reproducible?
- Skipping segment analysis. Self-check: do you track performance for key user groups?
- No approval evidence. Self-check: is there a signed record tied to the model version?
- Unclear rollback. Self-check: can you revert within minutes with a known process?
- Silent drift. Self-check: alerts configured for drift, performance, and fairness thresholds?
Quick self-audit mini-list
- Owner named for every dataset/model.
- Inventory and registry entries exist and are current.
- Model card links to data versions and metrics.
- Monitoring dashboard and on-call route defined.
- Last review date is recorded.
Practical projects
- Create a dataset entry and lineage diagram for one project dataset; include quality checks and access rules.
- Write a model card for your latest model; add approval notes and a rollback plan.
- Set up a monitoring plan: pick 5 metrics (e.g., accuracy, calibration, drift, fairness by segment, incident rate) and define alert thresholds.
Acceptance criteria
- All artifacts identify owners and versions.
- Risk level is stated with matching controls.
- Decisions are logged with rationale.
Who this is for
- Applied Scientists shipping or maintaining ML models.
- MLOps/Engineers who enable safe deployments.
- Data Scientists needing clear approval paths.
Prerequisites
- Basic ML lifecycle knowledge (data prep, training, evaluation, deployment).
- Familiarity with version control and experiment tracking.
- Awareness of privacy and fairness concepts.
Learning path
- Map your current assets: list datasets and models with owners.
- Create or update data inventory entries and lineage notes.
- Register models and draft model cards; define metrics and thresholds.
- Implement risk-based approvals and document sign-offs.
- Configure monitoring and an incident response template.
- Schedule periodic reviews and retirement criteria.
Next steps
- Deepen fairness and robustness testing for high-impact models.
- Add privacy-enhancing techniques where needed (e.g., minimization, pseudonymization).
- Automate documentation generation from your pipelines to reduce drift.
Mini challenge
Pick one production model. In 30 minutes, gather its owner, version, data sources, last approval, and rollback plan. If any item is missing, create the artifact or schedule the needed review.
Quick Test
Take the quick test to check your understanding. It’s available to everyone; if you sign in, your progress will be saved.