What's changed: In-scope coverage: deployment/runtime/IaC services
3.2MLOps and SageMaker Pipelines
Understand MLOps: automating the ML lifecycle with SageMaker Pipelines, versioning/approval with the Model Registry, and CI/CD integration.
To continuously rebuild and deploy models, automate the ML lifecycle (MLOps). On AWS, SageMaker Pipelines is central.
3.2.1Pipelines and the Model Registry
- SageMaker Pipelines: automate prep→train→evaluate→register/deploy as a repeatable pipeline.
- Model Registry: version models and gate production deployment via approval status.
- Evaluation step: proceed to register/deploy only if metrics pass a threshold (a quality gate).
- Integrate with CI/CD (e.g., CodePipeline) to automate retraining/deploy from a commit.
Common on MLA: automate the ML lifecycle = SageMaker Pipelines, version/approve models = Model Registry, gate before prod on metrics = evaluation step.
MLOps runs "the ML lifecycle reproducibly, automatically, and under governance." SageMaker Pipelines defines process→train→evaluate→register/deploy as a DAG, tracking each step’s inputs/outputs (data, model, metrics) via lineage. An evaluation step + ConditionStep creates a quality gate that "proceeds to registration only if metrics pass a threshold." Trained models are versioned into the Model Registry as a model package group, with approval status (Approved/Rejected) governing production deployment—approval triggers automated deploy via CI/CD like CodePipeline. Use SageMaker Experiments for experiment tracking, ML Lineage Tracking for metadata, and Projects for templated setups. Automate retraining triggered by EventBridge (schedule) or Model Monitor drift detection (data/model quality, feature attribution, bias drift) to keep models fresh. The axes: "lifecycle automation = Pipelines," "version/approval = Model Registry," "pre-prod gate = evaluation step."
| Goal | Use |
|---|---|
| Automate the lifecycle | SageMaker Pipelines |
| Version/approve models | Model Registry |
| Pre-prod quality gate | Evaluation step + condition |
| Retrain on drift | Model Monitor + EventBridge |
Scenario: never ship sub-threshold models, and auto-deploy after approval. Build train → evaluation step + condition (e.g., F1 ≥ 0.85) in SageMaker Pipelines, registering to the Model Registry only when it passes. When an owner marks it Approved, CodePipeline auto-deploys. In production, Model Monitor watches drift and EventBridge triggers retraining on detection.
Q. Automate lifecycle? SageMaker Pipelines. Q. Version/approve? Model Registry. Q. Pre-prod gate? Evaluation step + condition. Q. Detect drift? Model Monitor. Q. Trigger retraining? EventBridge (schedule/event).
Watch the mix-ups: (1) Pipelines = automation / Model Registry = version-approval—don’t swap. (2) Without a quality gate, bad models reach prod—always add an evaluation step + condition. (3) Model Monitor detects but doesn’t auto-fix—design retraining/rollback separately. (4) Skipping the approval flow loses governance.
Automate retraining triggered by schedules (EventBridge) or data/model drift detection to keep models fresh.
3.2.2Section summary
- SageMaker Pipelines (automation) + Model Registry (version/approval)
- Quality gate via evaluation step; integrate with CI/CD
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. You want to automate the prep→train→evaluate→deploy ML lifecycle reproducibly. Which fits best?
Q2. You want to version models and deploy only approved ones to production. What do you use?
Q3. What lets a pipeline "proceed to registration only if the evaluation metric passes a threshold"?