AWS MLA-C01: Machine Learning Engineer Associate Practice Exam

What the AWS MLA-C01 exam covers

• Data Preparation for ML202 questions
• ML Model Development164 questions
• Deployment and Orchestration188 questions
• Monitoring and Security151 questions

Free AWS MLA-C01 sample questions

A sample of 10 questions with answers and explanations. Sign up free to practice all 705.

1. Question 1Data Preparation for ML

   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 Development

   A deep learning training job repeatedly fails near the end after many hours due to Spot interruptions, wasting compute. Which design change reduces wasted work while keeping costs low?

   • AEnable managed Spot training with periodic checkpointing to S3 so jobs resume from the last checkpointCorrect
   • BTurn off checkpointing so each epoch runs marginally faster before interruption
   • CShrink the dataset until the job happens to finish before interruptions occur
   • DSwitch entirely to on-demand instances and remove checkpointing altogether
   ✓ Correct answer: A

   SageMaker Managed Spot Training uses Amazon EC2 Spot Instances at up to 90% cost savings compared to on-demand. By configuring periodic checkpointing to S3, the job saves its progress at regular intervals so that when a Spot interruption occurs and the job is restarted, it resumes from the last checkpoint rather than restarting from scratch - combining cost efficiency with resilience to interruption.

   Why the other options are wrong

   • BTurning off checkpointing means every Spot interruption discards all progress, increasing wasted compute.
   • CShrinking the dataset degrades model quality and is an unreliable way to dodge unpredictable interruptions.
   • DMoving fully to on-demand removes Spot savings and, without checkpointing, still loses work on any failure.
3. Question 3Deployment and Orchestration

   A team wants a serverless event-driven integration: when a model file is uploaded to S3, automatically run a lightweight transformation and register the model. Which service is the typical glue for the event-triggered logic?

   • AAWS Lambda triggered by an S3 event notificationCorrect
   • BAmazon EventBridge Scheduler on a fixed cron interval
   • CAWS Step Functions Activities polled by a worker
   • DAmazon SQS with a long-polling consumer you operate
   ✓ Correct answer: A

   AWS Lambda is the standard serverless compute for event-driven AWS integrations. When a model artifact is uploaded to S3, an S3 event notification invokes a Lambda function that can run lightweight transformation logic and call the SageMaker Model Registry API to register the model - all without provisioning servers and with sub-second trigger latency.

   Why the other options are wrong

   • BEventBridge Scheduler fires on a time schedule, not in direct response to a model file being uploaded to S3.
   • CStep Functions Activities require an external worker to poll for tasks and are not the direct S3-upload glue described.
   • DAn SQS queue with a self-run consumer needs you to operate the poller and is not the serverless S3-triggered function pattern.
4. Question 4ML Model Development

   A team runs many fault-tolerant SageMaker training jobs that checkpoint frequently and can resume after interruption. They want to cut training compute cost as much as possible. Which option should they enable?

   • AProvisioned Concurrency enabled on the training job for warm capacity
   • BOn-Demand instances with frequent checkpointing turned off entirely
   • CA SageMaker Savings Plan commitment applied to one-off experimental jobs
   • DManaged Spot Training, which uses spare capacity at a steep discount and resumes from checkpointsCorrect
   ✓ Correct answer: D

   This correctly addresses the requirement following AWS best practices.

   Why the other options are wrong

   • AProvisioned Concurrency is a serverless inference feature and does not apply to training jobs.
   • BOn-Demand instances pay full price and gain no discount from the job's checkpointing tolerance.
   • CSavings Plans commit to steady usage and do not discount sporadic one-off experimental jobs.
5. Question 5Monitoring and SecuritySelect all that apply

   An organization is implementing defense-in-depth for a SageMaker real-time endpoint exposed to internal applications. Which TWO controls reduce the network attack surface? (Choose TWO)

   • APlace the endpoint behind a VPC interface endpoint and restrict it with security groupsCorrect
   • BDisable CloudTrail to avoid logging overhead
   • CUse an endpoint resource policy / IAM to limit which principals can invoke itCorrect
   • DOpen the security group to 0.0.0.0/0 on all ports for flexibility
   ✓ Correct answer: A, C

   Use an endpoint resource policy / IAM to limit which principals can invoke it This correctly addresses the requirement following AWS best practices.

   Why the other options are wrong

   • BDisable CloudTrail to avoid logging overhead is incorrect because this option does not satisfy the requirement.
   • Dopen the security group to 0.0.0.0/0 on all ports for flexibility is incorrect because this option does not satisfy the requirement.
6. Question 6ML Model DevelopmentSelect all that apply

   A team is designing a multi-label text classification model where each document can belong to zero or more of five categories. Select TWO design choices that differ from standard single-label multi-class classification.

   • AThe output layer should use sigmoid activations (one per label) rather than a single softmax over all classes.Correct
   • BBinary cross-entropy loss should be applied independently to each output unit rather than categorical cross-entropy over the full output vector.Correct
   • CThe model must use a separate encoder for each label to avoid label interference during feature extraction.
   • DThe number of training epochs must equal the number of labels to ensure each label is learned separately.
   ✓ Correct answer: A, B

   Multi-label classification requires that each label be predicted independently because a document can belong to multiple classes simultaneously. Softmax enforces a probability distribution that sums to 1, which is wrong for multi-label problems where multiple labels can all be true. Instead, a sigmoid on each output unit produces independent probabilities in [0,1] for each label. Correspondingly, the loss is binary cross-entropy computed per label and summed, treating each label as a separate binary classification problem.

   Why the other options are wrong

   • CSeparate encoders per label are not required and would be computationally wasteful. A shared encoder extracts general features, and separate output heads or threshold adjustments handle per-label predictions. Shared encoders benefit from all labeled data during feature learning.
   • DTraining epochs are a training duration hyperparameter unrelated to the number of labels. Setting epochs equal to the number of labels has no theoretical or practical basis and would result in undertrained models for most datasets.
7. Question 7Data Preparation for MLSelect all that apply

   A team needs to securely mask PII in a dataset before it is used for ML training. Which TWO AWS services provide built-in capabilities to detect and redact or mask PII from unstructured text or structured data? (Select TWO)

   • AAmazon ComprehendCorrect
   • BAWS Glue sensitive data detectionCorrect
   • CAmazon Rekognition
   • DAmazon Translate
   ✓ Correct answer: A, B

   Amazon Comprehend's DetectPiiEntities and ContainsPiiEntities APIs identify PII entities (names, SSNs, credit card numbers, etc.) in free-form text and support redaction. AWS Glue's sensitive data detection feature uses machine learning to detect PII in structured data during ETL jobs and can automatically apply hashing or redaction transformations.

   Why the other options are wrong

   • CAmazon Rekognition analyzes images and video for object detection, facial analysis, and content moderation; it does not detect textual PII in tabular or document datasets.
   • DAmazon Translate provides language translation between human languages; it has no PII detection or redaction capabilities.
8. Question 8Data Preparation for MLSelect all that apply

   A data scientist is preparing a dataset and wants to reduce the impact of outliers on a numeric feature while keeping all rows. Which TWO techniques accomplish this? (Choose TWO)

   • AWinsorize (clip) the feature at the 1st and 99th percentilesCorrect
   • BApply robust scaling using the median and interquartile rangeCorrect
   • CDelete every row whose feature value exceeds the mean
   • DOne-hot encode the numeric feature
   ✓ Correct answer: A, B

   Winsorization caps extreme values at chosen percentiles, limiting their leverage while keeping records. Robust scaling centers on the median and divides by the IQR, which is far less sensitive to outliers than mean/standard-deviation scaling. Both retain all rows.

   Why the other options are wrong

   • CDeleting rows above the mean discards roughly half the data and does not keep all rows.
   • DOne-hot encoding is for categorical variables and is inappropriate for a continuous numeric feature.
9. Question 9ML Model Development

   A team uses SageMaker Automatic Model Tuning and wants tuning to learn from past trials and intelligently propose promising hyperparameter values rather than sampling blindly. Which tuning strategy provides this adaptive behavior?

   • ABayesian optimizationCorrect
   • BGrid search over every combination
   • CPure random search with no feedback
   • DA single fixed hyperparameter set
   ✓ Correct answer: A

   Bayesian search treats tuning as guided optimization: it fits a probabilistic model of how hyperparameters affect the objective and selects new candidates that balance exploration and exploitation. This typically finds good configurations in fewer jobs than uninformed methods.

   Why the other options are wrong

   • BGrid search exhaustively tries combinations without learning from prior results and scales poorly.
   • CRandom search samples independently and does not use feedback from earlier trials.
   • DA single fixed set performs no search at all.
10. Question 10Deployment and Orchestration

    A SageMaker Batch Transform job scoring many small files is underutilizing instances. Which two parameters should the team tune together to increase throughput by batching records per request?

    • AMaxConcurrentTransforms and BatchStrategy/MaxPayloadInMBCorrect
    • BInitialInstanceCount and InstanceType only
    • CVolumeKmsKeyId and ModelDataUrl
    • DVariantName and InitialVariantWeight
    ✓ Correct answer: A

    Batch Transform throughput depends on how many records are sent per request (BatchStrategy=MultiRecord with a suitable MaxPayloadInMB) and how many concurrent requests each instance issues (MaxConcurrentTransforms). Tuning both fills the instances and avoids the low utilization seen when sending one tiny record at a time.

    Why the other options are wrong

    • BAdding instances without batching still sends small underutilizing requests; instance settings alone do not raise per-request efficiency.
    • CVolumeKmsKeyId and ModelDataUrl govern encryption and artifact location, not throughput batching.
    • DVariantName and InitialVariantWeight apply to real-time endpoint variants, not Batch Transform.

Related AWS resources

• AWS MLA-C01 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-C01 practice exam FAQ