Question bank
AWS Certified Machine Learning Engineer – Associate
43 AWS Certified Machine Learning Engineer – Associate questions with the correct answer and a written explanation, organised across 5 exam domains.
- 43 questions
- 5 domains
- Foundation level
certilumQuestion index
Jump straight to any question. Answers and explanations are inline.
Data Preparation for Machine Learning
8 questions in this domain
- Question 1Single Choice
Which AWS service is commonly used as a central data lake for ML data ingestion?
- AAmazon S3Correct answer
- BAmazon Kinesis
- CAmazon Redshift
- DAWS Glue
Why this is the answer
Amazon S3 is the most common target for ML data ingestion due to its durability, scalability, and low cost.
- Question 2Multiple Select
Which two AWS services are used for real-time streaming data ingestion? (Select TWO)
- AAmazon Kinesis Data StreamsCorrect answer
- BAmazon Kinesis Data FirehoseCorrect answer
- CAmazon S3
- DAmazon Redshift COPY
Why this is the answer
Kinesis Data Streams and Kinesis Data Firehose are primary services for ingesting streaming data in real-time.
- Question 3Interactive
Examine the code snippet below. Which step is missing before this normalization can be applied to test data?
- ASaving the fitted scaler to S3Correct answer
- BShuffling the test data
- COne-hot encoding categories
- DDropping duplicates first
Why this is the answer
The fitted scaler must be saved to S3 so it can be loaded and applied to test data during inference.
- Question 4Single Choice
What is the purpose of using a validation set during model training?
- ATo evaluate final model performance
- BTo tune hyperparameters and early stopCorrect answer
- CTo train the model parameters
- DTo generate synthetic data
Why this is the answer
The validation set guides hyperparameter tuning and early stopping without influencing the final test evaluation.
- Question 5Multiple Select
Which techniques can be used to handle missing values in a dataset? (Select TWO)
- AMean imputationCorrect answer
- BDropping rowsCorrect answer
- CAdding random noise
- DIncreasing sample size
Why this is the answer
Mean imputation and dropping rows are basic but effective ways to handle missing values, depending on the pattern.
- Question 6Interactive
Review the code. What type of scaling is being implemented?
- AMin-Max scaling
- BStandardization (Z-score)Correct answer
- CRobust scaling
- DLog transformation
Why this is the answer
The code subtracts the mean and divides by standard deviation, which is Standardization (Z-score normalization).
- Question 7Single Choice
Which scaling method is most robust to outliers?
- AMin-Max scaling
- BRobust scalingCorrect answer
- CStandardization
- DUnit vector scaling
Why this is the answer
Robust scaling uses median and IQR, which are not influenced by extreme outliers.
- Question 8Multiple Select
Which two actions are considered data quality issues in ML pipelines? (Select TWO)
- AMissing valuesCorrect answer
- BDuplicate recordsCorrect answer
- CHigh dimensionality
- DLow model accuracy
Why this is the answer
Missing values and duplicates are classic data quality issues that degrade model performance.
Deployment and Orchestration
12 questions in this domain
- Question 9Single Choice
What is the primary purpose of Amazon SageMaker real-time endpoints?
- ATo process large datasets asynchronously
- BTo serve predictions with low latency on a per-request basisCorrect answer
- CTo train models using batch data
- DTo store model artifacts in a versioned repository
Why this is the answer
Real-time endpoints provide a persistent HTTPS endpoint for low-latency, synchronous predictions, ideal for customer-facing applications.
- Question 10Multiple Select
Which two statements are true about SageMaker Serverless Inference? (Choose two.)
- AIt supports GPU instances for deep learning
- BIt automatically scales to zero when idleCorrect answer
- CIt requires you to specify an instance type
- DIt charges per inference request and compute timeCorrect answer
Why this is the answer
Serverless Inference auto-scales to zero, has no idle cost, and charges per request + compute. It does not support GPU or require instance type selection.
- Question 11Interactive
Refer to the code snippet. What SageMaker resource is being created?
- AA SageMaker Model
- BAn Endpoint ConfigurationCorrect answer
- CA Transform Job
- DA Monitoring Schedule
Why this is the answer
The code creates an EndpointConfiguration resource with a production variant, including instance type and model ARN.
- Question 12Single Choice
What is the purpose of the BatchStrategy parameter in a SageMaker Batch Transform job?
- AIt defines the maximum payload size
- BIt determines whether records are sent singly or in multi-record batchesCorrect answer
- CIt sets the number of concurrent transform instances
- DIt specifies the output S3 bucket path
Why this is the answer
BatchStrategy controls how records are grouped: SingleRecord sends each record individually, MultiRecord sends multiple records together.
- Question 13Multiple Select
Which two AWS services can be used to trigger a SageMaker Batch Transform job? (Choose two.)
- AAWS LambdaCorrect answer
- BAmazon CloudWatch Logs
- CAmazon EventBridgeCorrect answer
- DAWS CloudFormation
Why this is the answer
Lambda and EventBridge can invoke Batch Transform via SDK or scheduled rules. CloudWatch Logs does not directly trigger jobs; CloudFormation provisions infrastructure.
- Question 14Interactive
Examine the code. What does this AWS Lambda function do?
- AIt creates a SageMaker endpoint
- BIt starts a SageMaker training job
- CIt initiates a SageMaker Batch Transform jobCorrect answer
- DIt updates an endpoint configuration
Why this is the answer
The Lambda calls create_transform_job with parameters for batch inference, indicating it is starting a transform job.
- Question 15Single Choice
In AWS MLOps, which service is designed specifically to orchestrate machine learning workflows?
- AAWS CodePipeline
- BAmazon SageMaker PipelinesCorrect answer
- CAWS Step Functions
- DAWS CodeBuild
Why this is the answer
SageMaker Pipelines is purpose-built for ML workflows, handling training, evaluation, and registration steps via a DAG.
- Question 16Multiple Select
Which two components are part of a SageMaker Model Registry workflow? (Choose two.)
- ARegister a model version with metadataCorrect answer
- BDeploy a model directly from the registry without a Model object
- CSet an approval status for a model versionCorrect answer
- DRun a Batch Transform job
Why this is the answer
The Model Registry supports registering model versions with metadata and setting approval statuses. Deploying requires a SageMaker Model object, not direct registry deployment.
- Question 17Interactive
Look at the pipeline definition snippet. Which step type is missing a dependency declaration?
- AProcessingStep
- BTrainingStep
- CEvaluationStepCorrect answer
- DRegisterModelStep
Why this is the answer
The EvaluationStep in the code does not specify a DependsOn, meaning it may start before TrainingStep completes, causing a failure.
- Question 18Single Choice
What is the purpose of enabling Data Capture on a SageMaker endpoint?
- ATo automatically retrain the model on new data
- BTo store inference requests and responses in S3 for monitoringCorrect answer
- CTo increase the endpoint's throughput
- DTo enable GPU acceleration
Why this is the answer
Data Capture logs input and output payloads to an S3 bucket, which is essential for later analysis like drift detection.
- Question 19Multiple Select
Which two types of drift can be detected by SageMaker Model Monitor? (Choose two.)
- AData quality driftCorrect answer
- BInfrastructure drift (e.g., instance type changes)
- CModel quality drift (accuracy drop)Correct answer
- DNetwork latency drift
Why this is the answer
SageMaker Model Monitor can detect data quality drift (input feature changes) and model quality drift (performance changes with ground truth).
- Question 20Interactive
Examine the monitoring schedule creation. What is the purpose of the 'baseline' parameter?
- AIt defines the maximum number of instances for the monitoring job
- BIt provides the expected statistical ranges for features to compare againstCorrect answer
- CIt sets the frequency of the monitoring runs
- DIt specifies the endpoint to monitor
Why this is the answer
The baseline (constraints and statistics) provides the reference for comparing live data to detect drift.
ML Operations (MLOps)
7 questions in this domain
- Question 21Single Choice
Which AWS service is purpose-built for creating ML retraining pipelines with native steps like training and conditional evaluation?
- AAmazon SageMaker PipelinesCorrect answer
- BAWS Step Functions
- CAWS Glue
- DAmazon EventBridge
Why this is the answer
Amazon SageMaker Pipelines is a fully managed CI/CD service for ML that includes native steps such as TrainingStep, ConditionStep, and RegisterModelStep, making it purpose-built for retraining pipelines.
- Question 22Multiple Select
Which two statements about Amazon SageMaker Model Registry are true? (Choose two.)
- AIt stores the actual model artifacts.
- BIt provides versioning and approval status tracking.Correct answer
- CIt can automatically deploy approved models to endpoints.
- DIt integrates with SageMaker Pipelines for automated registration.Correct answer
Why this is the answer
The Model Registry stores references to model artifacts in S3, not the artifacts themselves. It tracks version and approval status but does not auto-deploy; deployment requires a separate process. It does integrate with SageMaker Pipelines via RegisterModelStep.
- Question 23Interactive
Refer to the code block. Which AWS service's logs are being queried?
- AAWS CloudTrail
- BAmazon CloudWatch LogsCorrect answer
- CAmazon S3 access logs
- DAWS CloudTrail Insights
Why this is the answer
The query uses CloudWatch Logs Insights syntax (fields, filter, sort, limit) and filters for error messages, indicating the log source is Amazon CloudWatch Logs.
- Question 24Single Choice
Which instance purchasing option offers the largest discount but can be interrupted by AWS?
- AOn-Demand
- BSpot InstancesCorrect answer
- CReserved Instances
- DSavings Plans
Why this is the answer
Spot Instances offer the deepest discounts (up to 90%) but can be interrupted with a two-minute warning. They are ideal for fault-tolerant batch training that uses checkpointing.
- Question 25Multiple Select
Which two event types can trigger a SageMaker Pipeline execution for retraining? (Choose two.)
- ANew file upload to an S3 bucketCorrect answer
- BScheduled cron expression via EventBridgeCorrect answer
- CChange in an IAM policy
- DNew version registration in Model Registry
Why this is the answer
SageMaker Pipelines can be triggered by S3 events (e.g., new data) or scheduled cron rules. IAM policy changes do not trigger pipelines, and Model Registry version creation can be an output, not a trigger.
- Question 26Interactive
Refer to the code block. What is the effect of setting `ModelApprovalStatus` to `'Approved'`?
- AIt deploys the model to the endpoint immediately.
- BIt marks the model version as ready for deployment.Correct answer
- CIt deletes all other model versions in the group.
- DIt triggers a new retraining pipeline.
Why this is the answer
The `update_model_package` API with `ModelApprovalStatus='Approved'` changes the status of that model version to approved, indicating it is ready for deployment. It does not automatically deploy, delete other versions, or trigger retraining.
- Question 27Single Choice
Which AWS service records all API calls made to ML services for auditing purposes?
- AAmazon CloudWatch Logs
- BAWS CloudTrailCorrect answer
- CAWS Config
- DAWS X-Ray
Why this is the answer
AWS CloudTrail captures all API calls (control plane) for AWS services, including SageMaker, logging identity, source IP, and request parameters. CloudWatch Logs stores application logs; Config records resource configurations; X-Ray traces requests.
ML Security and Compliance
5 questions in this domain
- Question 28Single Choice
Which S3 server-side encryption mode provides a CloudTrail audit trail for each decrypt request?
- ASSE-S3
- BSSE-KMSCorrect answer
- CSSE-C
- DClient-side encryption
Why this is the answer
SSE-KMS logs each KMS Decrypt API call in AWS CloudTrail, providing an audit trail. SSE-S3 does not log individual decrypt requests, and SSE-C does not use AWS for key management.
- Question 29Multiple Select
Which two IAM actions must a SageMaker training job execution role have to read training data from an S3 bucket encrypted with SSE-KMS?
- As3:GetObjectCorrect answer
- Bkms:DecryptCorrect answer
- Cs3:ListBucket
- Dkms:GenerateDataKey
Why this is the answer
The training job must perform s3:GetObject to retrieve the object and kms:Decrypt to decrypt it. kms:GenerateDataKey is used for encryption, not decryption. s3:ListBucket is not required for reading a single object.
- Question 30Interactive
An ML engineer creates the following VPC endpoint policy for S3. What is the effect of this policy?
- AAllows only GetObject on the bucket named my-bucket through the endpointCorrect answer
- BAllows all S3 actions on my-bucket through the endpoint
- CAllows GetObject on all S3 buckets
- DDenies access to S3 through the endpoint
Why this is the answer
The policy allows the s3:GetObject action only on the resource arn:aws:s3:::my-bucket/*. It does not allow other actions or other buckets. The endpoint policy is evaluated in addition to IAM policies.
- Question 31Single Choice
Under the AWS Shared Responsibility Model, who is responsible for enabling encryption on training data stored in Amazon S3?
- AAWS
- BThe customerCorrect answer
- CBoth AWS and the customer
- DNeither (encryption is optional)
Why this is the answer
AWS is responsible for the security of the cloud (infrastructure), but the customer is responsible for security in the cloud, including data encryption. The customer must enable and manage encryption settings for S3 (e.g., SSE-KMS).
- Question 32Multiple Select
Which two statements accurately describe envelope encryption with AWS KMS?
- AData keys are stored encrypted with the customer master key (CMK)Correct answer
- BData keys are stored in plaintext alongside the encrypted data
- CEnvelope encryption reduces the number of KMS API calls needed for large objectsCorrect answer
- DEnvelope encryption requires client-side encryption libraries
Why this is the answer
In envelope encryption, KMS generates a data key that is encrypted under a CMK. The encrypted data key is stored with the ciphertext, reducing KMS API calls because only the data key needs to be decrypted. Client-side libraries are optional, not required.
Model Development
11 questions in this domain
- Question 33Single Choice
What is the primary difference between a built-in algorithm and a custom architecture in Amazon SageMaker?
- ABuilt-in algorithms abstract infrastructure concernsCorrect answer
- BCustom architectures are always faster
- CBuilt-in algorithms require custom Docker containers
- DCustom architectures have no hyperparameters
Why this is the answer
Built-in algorithms abstract infrastructure, while custom architectures require more manual setup.
- Question 34Multiple Select
Which business constraints act as primary filters when selecting an algorithm in SageMaker?
- AInference latencyCorrect answer
- BThroughputCorrect answer
- CCost of hardwareCorrect answer
- DNumber of hidden layers
Why this is the answer
Latency, throughput, and cost are primary constraints; hidden layers are a model architecture detail.
- Question 35Interactive
In the code snippet, what does the highlighted line within the run context log?
- AA metric named 'accuracy' with value 0.95Correct answer
- BA parameter named 'accuracy' with value 0.95
- CAn artifact file named 'accuracy'
- DA tag for the run
Why this is the answer
run.log_metric logs a metric (accuracy) and its value (0.95).
- Question 36Single Choice
Which hyperparameter optimization strategy is known to be sample-efficient?
- ABayesianCorrect answer
- BRandom
- CHyperband
- DGrid
Why this is the answer
Bayesian optimization is sample-efficient, using prior results to guide search.
- Question 37Multiple Select
Which metrics should be used when evaluating a classifier on an imbalanced dataset?
- APrecisionCorrect answer
- BRecallCorrect answer
- CF1 scoreCorrect answer
- DAccuracy
Why this is the answer
Precision, recall, and F1 give insight into minority class performance; accuracy can mislead.
- Question 38Interactive
In the SageMaker Clarify snippet, what does the 'facet_name' parameter identify?
- AThe sensitive attribute column (e.g., gender)Correct answer
- BThe target label column
- CThe name of the model
- DThe S3 output path
Why this is the answer
facet_name defines the sensitive group attribute for bias analysis.
- Question 39Single Choice
What does demographic parity measure in bias detection?
- ADifference in positive prediction rates between groupsCorrect answer
- BDifference in true positive rates between groups
- CDifference in false positive rates between groups
- DDifference in accuracy between groups
Why this is the answer
Demographic parity looks at prediction rates (positive outcomes) across groups.
- Question 40Multiple Select
Which AWS services can be used to perform error analysis on ML models?
- ASageMaker ClarifyCorrect answer
- BSageMaker Model MonitorCorrect answer
- CSageMaker DebuggerCorrect answer
- DAmazon Rekognition
Why this is the answer
Clarify, Model Monitor, and Debugger are for custom model analysis; Rekognition is a pre-built service.
- Question 41Interactive
According to the MLflow snippet, where should the artifact store be located for production use?
- AAn S3 bucketCorrect answer
- BLocal filesystem
- CA mounted NFS volume
- DAn EFS file system
Why this is the answer
For production, MLflow artifact store should be an S3 bucket to persist artifacts.
- Question 42Single Choice
What is the correct hierarchy for SageMaker Experiments?
- AExperiment > Trial > TrialComponentCorrect answer
- BTrial > Experiment > TrialComponent
- CTrialComponent > Trial > Experiment
- DExperiment > TrialComponent > Trial
Why this is the answer
SageMaker Experiments organize as Experiment (group), Trial (run), TrialComponent (step).
- Question 43Multiple Select
Which capabilities are provided by SageMaker Model Registry?
- AModel versioningCorrect answer
- BStage transitions (Pending, Approved, Rejected)Correct answer
- CAutomatic deployment to endpoints
- DLinking to experiment run metadataCorrect answer
Why this is the answer
Model Registry supports versioning, stage transitions, and lineage; deployment requires additional steps.
Frequently asked questions
Everything about the AWS Certified Machine Learning Engineer – Associate question bank and how to use it.