MLA-C01 Practice Question 8: Model Registry Approval Gate
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 1 · question 8 of 10
A regulated insurer retrains a pricing model weekly with a SageMaker AI pipeline that processes data, trains, and evaluates the candidate. Compliance requires that a named reviewer sign off on every model before it reaches the production endpoint, and that the sign-off be auditable. Once sign-off happens, deployment must proceed without any manual infrastructure work, and the pipeline itself must not sit idle waiting on a human.
How should the team implement this gate?
Show answer and explanation
Answer:
This is the standard SageMaker MLOps pattern and it satisfies all four constraints. The build pipeline finishes by registering a versioned model package in a model group with the status PendingManualApproval, so the training run terminates cleanly instead of blocking. A reviewer moves the version to Approved from Studio or with update_model_package, and that transition is recorded against the model version, giving the audit trail compliance asked for. SageMaker AI emits a model package state change event, so an EventBridge rule can trigger the deployment pipeline automatically; AWS documents that the PendingManualApproval to Approved transition is what initiates CI/CD deployment.
Option A inverts the control. Deploying first and reviewing afterwards means unapproved models serve production traffic, which is precisely what the regulator's sign-off requirement forbids.
Option B technically gates deployment, but it makes the pipeline execution hang on human latency; a reviewer who takes three days keeps a pipeline run and its polling Lambda alive for three days, risks hitting pipeline timeouts, and reimplements approval tracking that Model Registry already provides.
Option C removes the human entirely. A condition step is an automated quality gate on a metric, not a sign-off; registering as Approved up front means no person ever reviews the model. A condition step is a useful complement (reject weak candidates before a human ever sees them), but it is not a substitute for manual approval.
Reference: Update the Approval Status of a Model