Model and prompt versioning: MLA-C02 study notes and practice questions
What to know about Model and prompt versioning for the MLA-C02 exam, from the Deployment and Orchestration of ML and AI Workflows domain (24% of the exam guide), then 1 free question to test it. Part of the MLA-C02 study guide.
These questions ask how to track which model version is in production, who approved it, and how to roll back.
SageMaker Model Registry
A model package group holds the versions of one model; each version is a model package.
Each version records the image, artifact location, metrics, data and lineage.
Approval status:PendingManualApproval, Approved or Rejected. Status changes emit EventBridge events that can trigger deployment.
Pipelines deploy only approved versions, giving an audit trail for compliance.
Model cards can be attached to document intended use and risks.
Groups can be shared across accounts (resource policies or RAM) for a central registry with dev, staging and prod accounts.
Rolling back means deploying the previous approved version.
Other versioning
Item
Tool
Prompt templates
Bedrock Prompt Management versions
Agents
Bedrock Agent versions and aliases; point the alias at a tested version
Experiment runs
MLflow model registry, which can sync with SageMaker
Foundation models
Pin a specific model ID and version in code, not “latest”
Data
S3 versioning, Feature Store, dataset snapshots
Code and containers
Git commits and image digests in ECR
Why it matters on the exam
When a question mentions compliance, approvals, audits or “which model is deployed”, the answer usually includes the Model Registry with approval status, not tags, file names or spreadsheets.
Common traps
Deploying the newest artifact in S3 without approval.
Tags or spreadsheets to track approvals.
Changing an agent alias before testing the new version.
Keeping only the latest artifact in S3, so there is no earlier approved version to roll back to.
A regional bank deploys its credit models through a CI/CD pipeline. Compliance requires a named reviewer to approve each model version before it can reach the production endpoint, and the deployment must always use the version that was approved last.
The team wants the approval record and the model versions to stay in one managed place, so auditors can see which version was approved and when.
Which approach meets these requirements?
Choose one answer.
AA. Store each model artifact in a versioned S3 prefix, and have the deployment pipeline deploy whichever artifact in the folder is the newest.
BB. Register each model version in a model package group with approval status pending, and deploy only versions whose status is approved.
CC. Record the approval in a spreadsheet, and have the deployment pipeline read the version number from the spreadsheet each time.
DD. Give the production endpoint a tag with the approved version, and have the reviewer edit the tag after each approval.
Show the answer and explanation
Answer: B
A. Wrong: The newest file is not necessarily the approved one, and there is no approval record. The stem requires an approval step tied to the version.
B. Right: Correct. The Model Registry keeps model versions in a model package group with an approval status, so the pipeline can deploy only approved versions and auditors can see the approval history. The registry is the managed source of approved models. Learn more: Model Registration Deployment with Model Registry (opens in a new tab)
C. Wrong: A spreadsheet is outside the managed service and has no versioned link to the models. The option does not give the audit trail that the stem requires.
D. Wrong: Tags describe resources and do not record model approval status or deploy only approved versions. The option does not enforce the approval.
Explanation
The SageMaker AI Model Registry holds versions in a model package group with an approval status, and a pipeline can deploy only approved versions. This gives the audit trail compliance needs. Newest files, spreadsheets and tags do not enforce or record approval.
Want more on Model and prompt versioning? The 5 full-length practice sets have 7 more Model and prompt versioning questions, 325 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets →