AWS MLA-C02: Machine Learning Engineer Associate Practice Exam

What the AWS MLA-C02 exam covers

• Data Preparation for ML and AI249 questions
• ML Model and Foundation Model (FM) Development214 questions
• Deployment and Orchestration of ML and AI Workflows214 questions
• Operating, Monitoring, and Securing ML and AI Solutions214 questions

Free AWS MLA-C02 sample questions

A sample of 10 questions with answers and explanations. Sign up free to practice all 891.

1. Question 1Data Preparation for ML and AI

   Which AWS service is commonly used as the data lake for storing raw ML training data?

   • AAmazon SQS
   • BAWS KMS
   • CAmazon S3Correct
   • DAmazon Route 53
   ✓ Correct answer: C

   Amazon S3 is the primary and most cost-effective service for storing large-scale raw ML training data in a data lake architecture. S3 provides virtually unlimited storage, high durability, integration with SageMaker for training workflows, and excellent performance when used with services like Athena and Glue for data transformation.

   Why the other options are wrong

   • AAmazon SQS is a message queue service for event processing, not data storage.
   • BAWS KMS is a key management service for encryption, not data storage.
   • DAmazon Route 53 is a DNS service with no relevance to ML data lakes.
2. Question 2ML Model and Foundation Model (FM) DevelopmentSelect all that apply

   A team automating model training wants reproducible experiments with tracked parameters, metrics, and lineage. Which TWO SageMaker features help track and govern this? (Choose TWO)

   • ASageMaker Model Registry to version and approve modelsCorrect
   • BAn S3 bucket made public for sharing the training data set
   • CSageMaker Experiments to track runs, parameters, and metricsCorrect
   • DAmazon Route 53 health checks
   ✓ Correct answer: A, C

   Reproducible ML experiments require both run-level tracking (hyperparameters, metrics, data versions) and model-level governance (versioning, approval workflows, lineage). SageMaker Experiments records every training run's parameters and metrics with full lineage, while SageMaker Model Registry provides versioned model packages with approval states and deployment history. Together they cover the full traceability requirement for MLOps without exposing data publicly or relying on ad hoc processes.

   Why the other options are wrong

   • BMaking an S3 bucket public for sharing exposes sensitive model artifacts and training data without access controls, which violates security best practices and does not provide experiment tracking or lineage.
   • DAmazon Route 53 health checks monitor endpoint availability for DNS failover purposes; they have no capability related to ML experiment tracking, parameter logging, or model versioning.
3. Question 3Operating, Monitoring, and Securing ML and AI Solutions

   A SageMaker endpoint is backed by a managed scaling policy. Which CloudWatch metric is the recommended target for an application auto scaling target-tracking policy on the endpoint?

   • ASageMakerVariantInvocationsPerInstanceCorrect
   • BCPUUtilization of the model artifact S3 bucket
   • CNumberOfMessagesVisible of an SQS queue
   • DNetworkPacketsIn of a NAT gateway
   ✓ Correct answer: A

   SageMakerVariantInvocationsPerInstance is the metric application auto scaling target tracking is designed around for a real-time endpoint: it expresses how much work each instance is handling, so a target value translates directly into a capacity decision and the policy adds or removes instances to hold it. Metrics belonging to other services describe queues, network interfaces or storage, none of which reflects the load on the model variant that scaling is meant to track.

   Why the other options are wrong

   • BCPUUtilization of the model artifact S3 bucket is incorrect because this option does not satisfy the requirement.
   • CNumberOfMessagesVisible of an SQS queue is incorrect because SQS is a queuing service; not applicable.
   • DNetworkPacketsIn of a NAT gateway is incorrect because NAT gateway does not provide the functionality.
4. Question 4Operating, Monitoring, and Securing ML and AI Solutions

   A company wants to grant a SageMaker notebook in a development account temporary, scoped access to read a dataset in an S3 bucket owned by a separate data account. Which approach follows least-privilege cross-account best practice?

   • AAdd the dataset bucket's ARN to a shared broad allow-all policy
   • BThe execution role assumes a cross-account role allowed by the bucket policyCorrect
   • CCopy the dataset into a public bucket both accounts can read
   • DStore the data account's long-lived access keys in the notebook code
   ✓ Correct answer: B

   Cross-account access works by assuming a role in the target account, with the bucket policy trusting that role. The notebook's execution role assumes it and receives temporary credentials, so nothing long-lived is stored anywhere. Copying access keys into the notebook would work and is precisely the pattern this avoids.

   Why the other options are wrong

   • AA broad allow-all policy violates least privilege and is not scoped cross-account access.
   • CA public bucket exposes the data to everyone, not scoped access for one account.
   • DEmbedding long-lived keys is insecure and is not temporary, scoped cross-account access.
5. Question 5Deployment and Orchestration of ML and AI Workflows

   A SageMaker Pipeline has a ConditionStep that should branch to a deployment step only when validation accuracy exceeds 90%. The condition is not triggering correctly. Which resource type holds the metric value that the ConditionStep evaluates?

   • AA SageMaker Experiment trial component metric logged during training
   • BA JsonGet expression reading from a ProcessingStep or TrainingStep output property fileCorrect
   • CA CloudWatch custom metric published by the training container
   • DA SageMaker Model Registry approval status flag on the package
   ✓ Correct answer: B

   In SageMaker Pipelines, ConditionStep comparisons use JsonGet to extract a scalar value from a property file produced by a preceding ProcessingStep or TrainingStep. The upstream step writes evaluation metrics to a JSON file in S3, and the pipeline references it via a PropertyFile object. This is the native, in-pipeline mechanism for data-driven branching without requiring external services.

   Why the other options are wrong

   • ATrial component metrics track experiment lineage and are not what a ConditionStep evaluates.
   • CA ConditionStep reads pipeline property files, not CloudWatch custom metrics.
   • DAn approval status flag governs promotion, not the accuracy value the condition checks.
6. Question 6Data Preparation for ML and AI

   A team converts a free-text product description column into dense numeric vectors that capture semantic similarity for a downstream classifier. Which representation best captures semantic meaning rather than raw word counts?

   • AWord embeddingsCorrect
   • BOrdinal integer codes assigned alphabetically
   • CZ-score standardization of character lengths
   • DRemoving all whitespace from the text
   ✓ Correct answer: A

   Embeddings (such as those produced by Word2Vec, GloVe, or transformer models) place semantically similar words near each other in vector space, giving the classifier meaningful features. Simple counts or codes cannot express that two different words have similar meaning.

   Why the other options are wrong

   • BAlphabetical integer codes impose an arbitrary order and carry no semantic meaning.
   • CStandardizing character lengths describes string size, not semantic content.
   • DRemoving whitespace is a cleaning step that does not create numeric semantic features.
7. Question 7ML Model and Foundation Model (FM) DevelopmentSelect all that apply

   A SageMaker XGBoost training job overfits the training data. Which TWO hyperparameter adjustments would most directly reduce overfitting? (Choose TWO)

   • AIncrease the lambda (L2 regularization) termCorrect
   • BDecrease max_depth of the treesCorrect
   • CIncrease max_depth of the trees
   • DSet subsample to 1.0 (use all rows per tree)
   ✓ Correct answer: A, B

   A larger lambda penalizes large leaf weights, shrinking the model toward simpler solutions, while a smaller max_depth limits how many feature interactions each tree can capture. Both reduce variance and improve generalization on unseen data.

   Why the other options are wrong

   • CIncreasing max_depth makes trees more complex and typically worsens overfitting, the opposite of the goal.
   • DSetting subsample to 1.0 uses every row for each tree, removing the randomness that helps reduce overfitting; lowering subsample below 1.0 would help instead.
8. Question 8Deployment and Orchestration of ML and AI Workflows

   An MLOps team must guarantee that a SageMaker endpoint update performs zero-downtime by spinning up a full new fleet before terminating the old one, with no canary phase. Which traffic-shifting configuration should the deployment use?

   • ABlue/green with an ALL_AT_ONCE traffic-shifting policyCorrect
   • BBlue/green with a CANARY traffic-shifting policy
   • CBlue/green with a LINEAR traffic-shifting policy
   • DIn-place update that mutates the existing fleet
   ✓ Correct answer: A

   SageMaker managed deployment guardrails support blue/green with ALL_AT_ONCE, CANARY, and LINEAR policies. ALL_AT_ONCE provisions the complete new fleet, verifies it healthy, then routes 100% of traffic to it and decommissions the old fleet, avoiding downtime while skipping incremental traffic phases.

   Why the other options are wrong

   • BCANARY routes a small slice of traffic first, which the requirement explicitly excludes.
   • CLINEAR shifts traffic in equal increments over time, again introducing a phased rollout.
   • DAn in-place update modifies the running fleet directly and risks downtime, not a blue/green swap.
9. Question 9Deployment and Orchestration of ML and AI Workflows

   A team must be able to reproduce which data, code and parameters produced a model that was deployed four months ago. What provides this?

   • ARecorded pipeline lineage and the model versionCorrect
   • BThe current contents of the training bucket
   • CThe team's notes from that quarter
   • DThe latest notebook in the shared workspace
   ✓ Correct answer: A

   Reproducing an older model needs two facts, which artefact was deployed and what produced it, and the correct answer supplies both: the registry identifies the exact model version, and the pipeline's recorded lineage links it to the dataset, parameters and code used. Everything else describes present state rather than recorded history. The bucket holds today's data rather than what existed four months ago, notes are not evidence and rarely survive an auditor's questions, and the latest notebook has been edited many times since. Designing for this before it is needed is the only way it is available when it is.

   Why the other options are wrong

   • BThe bucket holds current data rather than what it contained four months ago.
   • CNotes are not evidence and are unlikely to be complete or accurate.
   • DThe notebook has been edited since and no longer reflects that run.
10. Question 10ML Model and Foundation Model (FM) Development

    Why should an evaluation set be held fixed while prompts and retrieval settings are tuned?

    • AOtherwise a score change cannot be attributed to the change madeCorrect
    • BOtherwise the evaluation cannot be run automatically
    • COtherwise the foundation model has to be retrained
    • DOtherwise the retrieved passages cannot be cited
    ✓ Correct answer: A

    If the prompts change alongside the configuration, an improved score might reflect an easier set rather than a better system, and the whole tuning exercise loses its evidence - a fixed set is what makes each comparison meaningful. Automation is unaffected by whether the set is fixed. No retraining is implied, and citation depends on retrieval rather than on the evaluation set.

    Why the other options are wrong

    • BThe evaluation being unrunnable automatically is unaffected by whether the prompt set is held fixed.
    • CThe foundation model having to be retrained is not implied by changing an evaluation set.
    • DThe retrieved passages being uncitable depends on retrieval rather than on how evaluation is run.

Who this AWS MLA-C02 practice exam is for

This practice set is for anyone preparing for the AWS MLA-C02: Machine Learning Engineer Associate exam at the intermediate level - from first-time candidates building a foundation to experienced AWS practitioners doing a final review before test day. If you learn best by working through realistic questions and reading why each answer is right or wrong, it is built for you.

How to use this AWS MLA-C02 practice exam

1. Start with the free sample questions above to gauge your current baseline.
2. Read the full explanation on every question, including why each wrong option is wrong.
3. Track your weak domains and focus your study where you are losing the most marks.
4. Once you are scoring consistently well, take a timed, full-length mock exam.
5. Use your readiness score to decide when you are ready to book the real AWS MLA-C02 exam.

Related AWS resources

• AWS MLA-C02 study guideKey concepts
• AWS practice examsAll AWS
• Certification pathWhere this fits
• Certification exam guides & tipsBlog
• Plans & pricingFree & paid
• AWS SCS-C03 practice examRelated
• AWS Certified AI Practitioner (AIF-C01) practice examRelated
• AWS CLF-C02 practice examRelated

AWS MLA-C02 practice exam FAQ