About these questions
These five questions are original. We wrote them in the style of the MLA-C02 skills AWS lists, one or more per domain. They are not real exam questions, and the real exam will word things differently.
How to use these
Try each question without looking at your notes, and pick an answer before you open the explanation. Then read the explanation even when you were right, because the reason the closest wrong option fails is where the learning is. Note any domain where you slipped and revisit it on the domains explained page.
Question 1 (Data Preparation for ML and AI)
A furniture retailer already stores its product catalog and order history in an Amazon Aurora PostgreSQL cluster. The team wants semantic search over product descriptions and needs to filter the results by columns such as price and stock level in a single SQL statement. Which approach fits best?
- A. Store the embeddings in an Amazon DynamoDB table with a global secondary index
- B. Store the embeddings in an Amazon S3 Vectors index and merge the results in the application
- C. Add the pgvector extension to the existing Aurora PostgreSQL cluster
- D. Create an Amazon OpenSearch Serverless collection and copy the catalog into it
Show the answer and explanation
Correct answer: C. The pgvector extension adds a vector column type and distance operators to the database the team already runs. Similarity search and filters on ordinary columns then happen in one query. OpenSearch Serverless, the closest wrong option, is a capable vector store, but it keeps the data in a second system, so the filter on price and stock would need a separate query or a copy of the catalog. S3 Vectors has the same split problem. A DynamoDB secondary index supports key-based lookups and has no vector distance search.
Question 2 (ML Model and Foundation Model Development)
Two payment-fraud models both show an AUC above 0.98 on a test set where three transactions in a thousand are fraudulent. The team must choose the model that finds fraud better at the one threshold it will deploy. Which comparison is most informative?
- A. The mean squared error of the scores
- B. The accuracy of each model on the test set
- C. The AUC of each model on a test set ten times larger
- D. The precision-recall curve of each model
Show the answer and explanation
Correct answer: D. A precision-recall curve looks only at the rare positive class, so it shows how many flagged transactions are truly fraud (precision) against how much fraud is caught (recall) at each threshold. AUC, the closest wrong option, summarizes performance across all thresholds and stays high when negatives dominate, so a bigger test set will not separate the models. Accuracy is almost perfect for a model that never flags anything, and mean squared error is a regression metric.
Question 3 (Deployment and Orchestration of ML and AI Workflows)
A SageMaker AI endpoint runs a large model on 30 GPU instances. The account is close to its quota for that instance family, so the team cannot start a second full fleet during an update. They still want traffic shifted in steps with automatic rollback. Which two actions meet the requirement? Choose two.
- A. Use a blue/green deployment with linear traffic shifting
- B. Update the endpoint with a rolling deployment
- C. Move the endpoint to serverless inference
- D. Request a quota increase and keep the current update process
- E. Set the rolling batch size to a small share of the fleet
Show the answer and explanation
Correct answers: B and E. A rolling deployment shifts traffic in batches and removes old instances after each shift, so only one batch of extra capacity is needed. A small batch size keeps that extra capacity low, at the cost of a slower update. Blue/green with linear shifting, the closest wrong option, still starts a complete new fleet first, which the quota will not allow. Serverless inference does not support GPUs. A quota increase is a request that may not be granted, and it does not change how the update works.
Question 4 (Operating, Monitoring, and Securing ML and AI Solutions)
An endpoint hosts two production variants. The team wants to move traffic from 80/20 to 50/50 between them without creating a new endpoint configuration and without redeploying either model. Which API call should they use?
- A. UpdateEndpoint with a new endpoint configuration
- B. CreateEndpointConfig with the new weights
- C. UpdateEndpointWeightsAndCapacities
- D. RegisterScalableTarget
Show the answer and explanation
Correct answer: C. UpdateEndpointWeightsAndCapacities changes the weights of existing variants, or the capacity of one variant, on a running endpoint. That is how you ramp an A/B test. UpdateEndpoint with a new configuration, the closest wrong option, replaces the deployment, which is what the team wants to avoid. CreateEndpointConfig only defines a configuration and does nothing until the endpoint is updated to use it. RegisterScalableTarget registers a variant with auto scaling and never moves traffic between variants.
Question 5 (Operating, Monitoring, and Securing ML and AI Solutions)
A compliance reviewer asks for a record of who called InvokeEndpoint on a SageMaker AI endpoint last week. The account has no trails or extra logging configured. The reviewer opens CloudTrail Event history and searches for those calls. What will the reviewer find?
- A. The calls, recorded as management events
- B. The calls, recorded as CloudTrail Insights events
- C. The calls, with full request bodies
- D. No such records, because they are data events that are not logged by default
Show the answer and explanation
Correct answer: D. InvokeEndpoint is a data event, and CloudTrail does not log data events unless a trail with data event selectors is set up. Event history does not show them. Control plane calls such as CreateEndpoint are management events, which is the closest wrong option, but invocation is a runtime call and falls on the data side. Insights is a separate feature for unusual activity. Even with data event logging on, InvokeEndpoint does not record the request parameters.
More guides for MLA-C02
- MLA-C02 Exam Format, Scoring and Retakes
- MLA-C02 Study Plan: 4-Week and 8-Week
- Is the MLA-C02 Exam Hard?
- MLA-C02 Prerequisites and Skills Check
- MLA-C02 Renewal and Retakes Explained
- Is MLA-C02 Worth It? An Honest Look
- MLA-C02 Cheat Sheet: Key Facts to Review
- AIF-C01 vs MLA-C02: Which First?
Next step
If you missed a question, find its domain on the domains explained page and revisit that section. For more practice, our MLA-C02 practice tests have five original sets across the four domains.