Free MLA-C02 sample questions with answers
Try these before reading the answers. One from each domain, plus an extra from data preparation to match the weighting.
Question 1. A model trained on transaction data performs far better in testing than in production. Investigation shows a column derived from the final transaction status was included as a feature. What has happened?
- A. Overfitting caused by too many training epochs
- B. Data leakage — a feature encodes information unavailable at prediction time
- C. Model drift since deployment
- D. Insufficient endpoint compute capacity
Show answer
Answer: B
A feature derived from information only available after the outcome is data leakage: the model learned from something it will not have at prediction time, inflating test performance. Overfitting would show as memorising the training set generally, drift develops over time after deployment, and insufficient capacity would depress test performance too.
Question 2. An ML pipeline must score ten million records once each night. Latency per record is irrelevant; cost matters. Which inference approach is most appropriate?
- A. A real-time endpoint with auto scaling
- B. A real-time endpoint with provisioned concurrency
- C. Batch inference on a nightly schedule
- D. An always-on GPU endpoint fleet
Show answer
Answer: C
Batch inference processes a large set on a schedule without maintaining a persistent endpoint, which is the cheapest fit when per-record latency does not matter. A real-time endpoint bills continuously for capacity nobody needs overnight, provisioned concurrency adds cost for fast cold starts, and an always-on GPU fleet is the most expensive option.
Question 3. A fraud detection dataset contains 0.3% fraudulent transactions. A model reports 99.7% accuracy. What should the engineer conclude?
- A. Accuracy is misleading here; evaluate precision and recall on the minority class
- B. The model is performing well and is ready for production
- C. The model needs more training epochs to exceed 99.9%
- D. The endpoint should be scaled up before evaluation
Show answer
Answer: A
With a class this imbalanced, predicting the majority class every time yields 99.7% accuracy while detecting no fraud at all, so accuracy is the wrong metric and recall or precision on the minority class is needed. The model is not necessarily production ready, more epochs would not fix the metric choice, and the endpoint size is unrelated.
Question 4. An application calls a deployed model from AWS Lambda. Security policy prohibits storing long-lived credentials anywhere in the deployment. What should the engineer configure?
- A. An access key in a Lambda environment variable
- B. An access key in a configuration file inside the deployment package
- C. An access key stored in an encrypted S3 object read at runtime
- D. An IAM execution role attached to the Lambda function
Show answer
Answer: D
An IAM execution role lets the Lambda function obtain temporary credentials automatically, removing any stored secret, which is what the policy requires. Environment variables, a config file in the deployment package and an encrypted S3 object all still store a credential somewhere.
Question 5. A model deployed nine months ago is producing steadily less accurate predictions. No code, configuration or infrastructure has changed. What is the most likely cause and the appropriate response?
- A. Endpoint under-provisioning; increase the instance size
- B. Data drift; implement drift monitoring and a retraining schedule
- C. A transient fault; restart the endpoint
- D. Excessive logging; reduce the log retention period
Show answer
Answer: B
Gradually declining accuracy with no change to the system is the signature of drift, as live data diverges from the training distribution, and the response is monitoring for drift with scheduled retraining. Scaling the endpoint addresses throughput, restarting addresses transient faults, and reducing logging addresses cost.
How did you do?
Notice that every question describes a constraint — cost matters, latency does not, no stored credentials, nothing changed — and the constraint decides the answer. That is how MLA-C02 is written throughout.
If the data preparation questions felt hardest, that matches the weighting: it is the largest domain at 28%. See weeks 1 and 2 of the study plan.