Model decommissioning and archiving
Instruction and application
In ProgressWhen a model has run its course, retiring it responsibly is just as important as deploying it in the first place. Ignoring this phase can lead to wasted resources, compliance risks and even unexpected system failures.
This section explores why models need to be decommissioned, the protocols for doing it safely and how to archive them for future reference and regulatory needs.
Why decommission ML models?
Models don’t last forever. Without a clear retirement strategy, organisations risk cluttered environments, rising costs and compliance gaps.
Common drivers for decommissioning include:
- Obsolescence: Accuracy declines as data drifts, business needs evolve or better alternatives emerge. For instance, a retail forecasting model may fail to reflect post-pandemic shopping habits.
- Cost: Even unused models consume compute, storage and monitoring resources that could be better applied elsewhere.
- Complexity: Too many active models create governance challenges, version confusion and conflicting insights.
- Compliance: Regulations such as GDPR’s 'right to be forgotten’ may require retiring models trained on sensitive personal data.
- Security: Older models often run on outdated infrastructure, increasing the attack surface and breach risk.
Protocols for safe decommissioning
Retiring a model isn’t a delete action. It’s a structured process to remove it without disrupting operations, breaking dependencies or leaving hidden risks. Poorly managed retirements can cause outages, data loss and compliance gaps.
Use the following step-by-step approach to decommission safely:
Assess the impact
Identify all downstream dependencies: Dashboards, APIs or applications that rely on the model. For example, an HR attrition model feeding a workforce planning dashboard.
Communicate with stakeholders
Notify stakeholders (e.g. data scientists, engineers, business users) well in advance to prevent surprises.
Divert traffic
Gradually shift requests to a replacement model, or cease serving if none is needed. Canary-style rollbacks can help detect gaps.
De-provision resources
Shut down endpoints, remove compute resources (VMs, containers), clean up storage (unless archived) and disable related configurations (API gateways, load balancers).
Monitor during decommissioning
Watch for failures or outages to ensure dependent systems aren’t disrupted.
Verify
Confirm that all resources are safely de-provisioned and that traffic has been fully redirected.
Archiving procedures and compliance
Retiring a model doesn’t always mean erasing it. Many must be preserved for compliance, reproducibility or institutional learning. Archiving ensures models are stored securely for audits, investigations and future reference, while protecting organisations from compliance gaps or loss of historical insight. Why archive?:Reproducibility: Archived models let teams revisit past predictions to validate outcomes or debug issues. For example, if a medical diagnosis is challenged years later, the archived model and data provide defensible evidence.
- Compliance: Industries such as finance and health care often require retaining models tied to regulated decisions for a set period to meet laws such as HIPAA, SOX or the EU AI Act.
- Knowledge preservation: Archived models capture the organisation’s ML journey, helping teams spot trends, avoid past mistakes and improve future models.What to archive?:Model artefacts: The trained model files.
- Model metadata: Hyperparameters, training metrics, lineage and version history.
- Training data snapshots: Or secure references to versioned datasets used during training.
- Relevant code: Training scripts, inference code and feature engineering logic, all properly versioned.Deployment logs: Serving history, performance records and system events.Archiving strategies: Usemodel registrieswith an ‘archived’ life cycle stage to distinguish active from retired models.
- Store artefacts in long-term,cost-efficient storagesuch as AWS S3 Glacier, Azure Archive or GCP Archive.
- Apply retention rules that define how long models must be kept, who can access them and where they are stored.Document the archive with clear metadata to ensure future teams can easily interpret what was preserved, and why.
Special consideration
Regulations such as GDPR’s ‘right to be forgotten’ may require removing personal data used in training. In practice, this can mean retraining the model without those records or invalidating older models entirely. Handling these requests effectively is a crucial part of compliant archiving.
Action item: Retiring a model safely
This activity will challenge you to apply decommissioning principles in practice, linking technical steps with compliance and business continuity.
Scenario
Your company’s HR department has decided to replace an old employee attrition prediction model with a newer, more accurate and compliant solution. While the old model has only a few active users left, it is still running in production and feeding into a workforce planning dashboard. Leadership has asked you to design a decommissioning plan that ensures business continuity, avoids hidden risks and meets compliance obligations. Fill out the form below as you work through each task in the activity.