Exam-style
A team has fine-tuned an open-weight model on SageMaker AI. Every training run must be kept alongside its evaluation metrics so an auditor can later see which artefact served which quarter, and a human has to approve a candidate before any traffic moves to it. Which mechanism records the runs and gates the promotion?
Reveal the answer
C. Register a model package version in a SageMaker Model Registry model package group, carrying the metrics and an approval status a pipeline reads
A model package version records the artefact URI, the inference container image and the evaluation metrics, and the registry integrates with SageMaker Model Cards for the audit trail. Group those versions and a fine-tuning line reads as an ordered set of candidates. The approval status gates promotion: a version is registered as PendingManualApproval, a person moves it to Approved or Rejected, and a versioned model has to be Approved before it deploys. Bedrock prompt versions snapshot the draft prompt and its configuration. An agent version snapshots an agent, while the alias is a movable pointer at a version and holds no record of its own. Bedrock Agents has also been Agents Classic since 30 July 2026, closed to accounts with no prior use. A custom model id names imported weights. None of these carries training metrics or an approval state, and S3 tags hold the same strings with nothing enforcing them.
Q. Two versioning vocabularies sit next to each other on this track. What does SageMaker Model Registry version, and what does Bedrock version?
A. SageMaker Model Registry versions model packages inside a model package group: artefact URI, inference container image, evaluation metrics and model card. Each version carries an approval status a CI/CD pipeline reads before promotion. Bedrock versions other things: prompts, guardrails and agents. An agent alias points at a version rather than holding one, and Bedrock Agents has been Agents Classic, closed to accounts with no prior use, since 30 July 2026. None of that describes a training artefact. See importing custom weights for where the two meet, and versioning the Bedrock side for the other half.
Why? Rollback differs with the vocabulary. A SageMaker AI endpoint update rolls back through deployment guardrails: a canary takes a slice of traffic, and a CloudWatch alarm tripping during the baking period returns it to the old fleet. A model imported into Bedrock runs on demand and rolls back by repointing the invocation at the earlier model ARN. A model customised inside Bedrock serves on demand through a custom model deployment, rolling back to the earlier deployment ARN, or behind Provisioned Throughput, where the rollback moves that commitment.