What the exam covers
Domain 1Data Preparation for Machine Learning28%
- 1.1Ingest and store data
- 1.2Transform data and perform feature engineering
- 1.3Ensure data integrity and prepare for modeling
Domain 2ML Model Development24%
- 2.1Choose a modeling approach
- 2.2Train and refine models
- 2.3Analyze model performance
Domain 3Deployment and Orchestration of ML Workflows24%
- 3.1Select deployment infrastructure
- 3.2Create and script infrastructure
- 3.3Automated orchestration and CI/CD
Domain 4ML Solution Monitoring, Maintenance, and Security24%
- 4.1Monitor model inference
- 4.2Monitor and optimize infrastructure and costs
- 4.3Secure AWS resources
Sample questions
A company's documents use a custom structure in which every section starts with a special marker line. The team wants Bedrock Knowledge Bases to split documents at these markers, which none of the built-in chunking strategies do. What should the ML engineer do?
- Choose semantic chunking and set a high similarity threshold
- Configure a custom transformation that calls an AWS Lambda function to chunk the documents during ingestion
- Choose hierarchical chunking with very small child chunks
- Rename the files so each section name appears in the file name
Answer
B. Bedrock Knowledge Bases can invoke a Lambda function during ingestion to apply custom chunking logic, such as splitting on specific markers. Semantic and hierarchical chunking cannot be told to use a custom marker, and renaming files does not split their content.
A team runs its feature pipelines as Apache Spark jobs on Amazon EMR. They want to add unit-test style checks, such as uniqueness and completeness constraints, directly in the existing Spark code using an open-source library. Which tool should they use?
- Amazon Macie
- AWS X-Ray
- SageMaker Debugger
- Deequ
Answer
D. Deequ is an open-source library built on Spark for defining and running data quality constraints, so it fits directly into the existing EMR jobs. Debugger inspects model training, Macie finds sensitive data, and X-Ray traces application requests.
A SageMaker AI training job fails with an AccessDenied error when reading training data from an S3 bucket encrypted with SSE-KMS using a customer managed key. The execution role already has s3:GetObject on the bucket. What is the MOST likely fix?
- Add s3:PutObject to the execution role
- Disable default encryption on the bucket and re-upload the data
- Enable inter-container traffic encryption on the training job
- Allow the execution role to use the KMS key, for example kms:Decrypt, in the key policy or the role's IAM policy
Answer
D. Reading SSE-KMS objects requires permission to use the KMS key in addition to S3 read access. Disabling encryption weakens security, PutObject is for writing, and inter-container encryption is unrelated to reading from S3.
A data analyst is preparing a tabular dataset in SageMaker Data Wrangler in SageMaker Canvas. Before handing it to the ML team, the analyst wants a quick, low-code check for duplicate rows, class imbalance, and features that predict the target suspiciously well. What should the analyst use?
- Create a SageMaker Model Monitor data quality baseline from the dataset
- Run a SageMaker Debugger profiling report on a sample training job
- Generate a Data Quality and Insights Report for the dataset in Data Wrangler
- Run an AWS Glue crawler and review the inferred schema in the Data Catalog
Answer
C. The Data Quality and Insights Report in Data Wrangler flags duplicates, imbalance, and possible target leakage, and gives a quick model score without code. Debugger analyzes training jobs, and a Model Monitor baseline is meant for monitoring a deployed endpoint, not for leakage checks.
A retailer trained a demand model on sales data from its urban stores only. The model will now be used for all stores, including rural stores, which have different customer behavior. Before retraining, what should the ML engineer do about the data?
- Keep the urban-only data and apply stronger L1 regularization
- Remove the store location feature so the model cannot tell store types apart
- Collect and include representative data from rural stores, and check how each store type is represented in the training set
- Oversample the urban records so the model is more confident
Answer
C. The training data must represent the population the model will serve, so rural stores need to be included and their representation checked. Regularization and removing the location feature do not fix missing data, and oversampling urban records increases the skew.
A payments company must compute rolling 5-minute aggregates over a transaction stream in Amazon Kinesis Data Streams and pass them to a fraud model within seconds. The aggregation logic is stateful and uses event-time windows. Which service is BEST suited for this processing?
- AWS Glue crawlers
- AWS DataSync
- Amazon Athena
- Amazon Managed Service for Apache Flink
Answer
D. Managed Service for Apache Flink runs stateful stream processing with event-time windows on Kinesis Data Streams and produces results continuously with low latency. Athena runs ad hoc queries on data at rest, crawlers catalog schemas, and DataSync transfers files, so none of them process a live stream.
A data science team must choose between two binary classifiers before the business has decided on the decision threshold to use in production. The classes are reasonably balanced. The team wants a metric that summarizes how well each model ranks positives above negatives across all thresholds. Which metric should the team use?
- Precision at a threshold of 0.5
- Area under the ROC curve (ROC-AUC)
- Accuracy at a threshold of 0.5
- Mean absolute error
Answer
B. ROC-AUC summarizes true positive rate versus false positive rate across every possible threshold, so it compares ranking quality without committing to a threshold. Accuracy or precision at 0.5 depends on one arbitrary threshold that may not be used in production.
An ML engineer must tune six hyperparameters for a deep learning model whose training jobs run for many epochs. The engineer wants SageMaker AI Automatic Model Tuning to stop under-performing configurations early and shift resources to promising ones, using a strategy that is built around this behavior. Which tuning strategy should the engineer choose?
- Grid search
- Random search
- Bayesian optimization
- Hyperband
Answer
D. Hyperband is a multi-fidelity strategy that evaluates many configurations with small budgets, stops weak ones early and reallocates resources to the best ones. Bayesian optimization chooses promising values from earlier results, but it is not designed around stopping and reallocating resources within training.
A manufacturer trains a model to detect defective parts, and defects make up about 0.2% of the data. Two models have nearly identical ROC-AUC scores of 0.97, but the team cares mainly about how well each model identifies the rare defective parts across thresholds. Which metric BEST distinguishes the models for this goal?
- R-squared of the predicted scores
- Overall accuracy at the default threshold
- Area under the precision-recall curve (PR-AUC)
- Specificity at the default threshold
Answer
C. With a very rare positive class, ROC-AUC can look high because the huge number of true negatives keeps the false positive rate small. The precision-recall curve focuses on performance on the positive class, so PR-AUC exposes differences that ROC-AUC hides. Accuracy and specificity are dominated by the majority class.
A furniture maker wants to classify photos of its products into 30 custom style categories. It has only about 150 labeled images per category, and a model trained from scratch performed poorly. The team wants better accuracy with limited data and minimal setup. Which approach should the team use?
- Use Amazon Textract to read style information from the product photos
- Use the SageMaker AI K-Means algorithm to group the photos into 30 clusters
- Fine-tune a pre-trained image classification model from SageMaker JumpStart on the labeled images
- Train a new deep network from scratch for more epochs on the same images
Answer
C. Transfer learning starts from a model already trained on a large image dataset, so fine-tuning it needs far fewer labeled images, and JumpStart provides such models ready to fine-tune. Training from scratch for longer on a small dataset tends to overfit, K-Means ignores the labels, and Textract extracts text from documents rather than classifying visual style.
An engineer trains a recurrent neural network on long sequences. Training loss suddenly jumps to very large values and the weights become NaN, and the logged gradient norms are extremely large just before each jump. Which technique directly addresses this problem?
- Removing dropout layers
- Training for more epochs
- Switching from L2 to L1 regularization
- Gradient clipping
Answer
D. Very large gradient norms just before divergence are exploding gradients, a common problem in recurrent networks. Gradient clipping caps the gradient norm at a threshold so that no single update can blow up the weights. More epochs or a different regularization type does not limit the size of each update.
Several applications call Amazon Bedrock with prompt templates that are hard-coded in their source code. Prompt engineers want to change and test prompts, keep a version history, and release new prompts without redeploying the applications. Which solution should the ML engineer use?
- Add the templates to a Bedrock guardrail so that they are applied to every model invocation.
- Store the templates as text files in Amazon S3 and track versions with a custom naming scheme.
- Move the templates into Lambda environment variables and update the variables for each prompt change.
- Store the templates in Bedrock Prompt Management with variables, and have the apps reference a prompt version.
Answer
D. Prompt Management stores prompts with variables and immutable versions, and applications can call a specific prompt version, so prompts change without code redeployment. Environment variables and S3 files need custom versioning and testing work. Guardrails filter content; they do not hold prompt templates.
A company is building an agent on Amazon Bedrock AgentCore. The agent must use existing Lambda functions and internal REST APIs as Model Context Protocol (MCP) tools, and it must remember each user's preferences across separate sessions. Which TWO AgentCore services should the ML engineer use? (Choose two.)
- AgentCore Browser.
- AgentCore Observability.
- AgentCore Memory.
- AgentCore Gateway.
- AgentCore Code Interpreter.
Answer
D, C. AgentCore Gateway turns APIs and Lambda functions into MCP-compatible tools that agents can call, and AgentCore Memory provides long-term memory that can keep user preferences across sessions. Code Interpreter runs code in a sandbox, Browser lets an agent use websites, and Observability provides tracing and metrics; none of them expose tools or store preferences.
A SageMaker AI serverless endpoint serves a dashboard that sees a sharp, predictable rise in traffic at 9 AM on weekdays. Users complain about slow first responses each morning caused by cold starts. Which change addresses this with the LEAST operational overhead?
- Configure provisioned concurrency on the serverless endpoint
- Replace the endpoint with a nightly batch transform job
- Convert the endpoint to asynchronous inference
- Increase the maximum concurrency setting of the endpoint
Answer
A. Provisioned concurrency keeps a set number of serverless workers initialized and ready, removing cold-start delays for predictable traffic, and it can be scheduled through auto scaling. Raising maximum concurrency only allows more simultaneous invocations but does not keep workers warm, and batch or asynchronous options do not serve an interactive dashboard.
A production SageMaker AI endpoint runs in a VPC and must keep serving requests if one Availability Zone fails. The team wants to avoid running and routing between extra endpoints. Which configuration meets these requirements with the LEAST operational overhead?
- Enable the multi-AZ option in the endpoint configuration and keep a single instance in the variant.
- Run one larger instance in the production variant so that it can absorb traffic from a failed zone.
- Create one endpoint per Availability Zone and use Amazon Route 53 failover records to route between them.
- Run at least two instances in the production variant and give the model subnets in at least two Availability Zones.
Answer
D. SageMaker AI spreads an endpoint's instances across Availability Zones when there are several instances and the VPC configuration includes subnets in several zones. Separate endpoints with Route 53 failover would work but add significant overhead. There is no multi-AZ option for a single instance, and one instance is a single point of failure.
In a SageMaker AI pipeline, a data governance system owned by another team must review each new dataset before training starts. The review can take several hours, and the governance system can call AWS APIs when it finishes. How should the ML engineer pause the pipeline until the review is done?
- Add a callback step that sends a message to an Amazon SQS queue, and have the governance system call SendPipelineExecutionStepSuccess when the review is done.
- Add a fail step before training and restart the pipeline manually after the review has finished.
- Add a condition step that checks a review flag and retries every few minutes until it becomes true.
- Add a Lambda step that polls the governance system in a loop until the review is marked complete.
Answer
A. A callback step publishes a token to SQS and waits until an external system reports success or failure for that token, which fits a long external review. A Lambda step is limited by the Lambda timeout and would waste compute polling. Condition steps evaluate once and do not retry, and a fail step plus manual restart defeats automation.
A training job fails with an AccessDenied error when reading its input data. The data is in an S3 bucket that uses SSE-KMS with a customer managed key, and the execution role already has s3:GetObject on the bucket. What is the MOST likely fix?
- Add s3:PutObject on the input bucket to the execution role.
- Disable network isolation on the training job.
- Enable inter-container traffic encryption on the training job.
- Allow the execution role to use kms:Decrypt on the customer managed key.
Answer
D. Reading an SSE-KMS object requires permission to decrypt with the key as well as permission to get the object, so the role needs kms:Decrypt through its policy and the key policy. Write access and network settings don't affect decrypting the input data.
A team wants to deploy an open-weight model downloaded from a public model hub. The weights are distributed as Python pickle files, and the security team warns that loading a pickle file can run arbitrary code. Which actions reduce this risk? (Choose two.)
- Load the pickle files on an instance type that has more memory
- Encrypt the pickle files with an AWS KMS key before they are loaded
- Use weights published in the safetensors format, which stores only tensor data
- Verify the publisher and checksums and scan the artifacts before they enter the deployment pipeline
- Rename the pickle files with a .bin extension before they are loaded
Answer
C, D. Pickle deserialization can execute code embedded in the file, so a malicious model artifact is a supply chain risk. Safetensors cannot carry executable code, and verifying and scanning artifacts before use catches tampered files. Encryption protects files at rest but the decrypted pickle still runs the same code when loaded.
A retail company that already uses SageMaker Model Monitor hosts an XGBoost demand model on an Amazon SageMaker AI real-time endpoint. The ML engineer must be alerted when the statistical distribution of incoming feature values moves away from the distribution of the training data. Ground truth labels are not available in production. Which solution meets these requirements with the LEAST operational overhead?
- Create a model quality monitoring schedule that compares captured predictions with the training labels and alerts on accuracy changes.
- Attach SageMaker Debugger built-in rules to the endpoint so that tensor statistics are compared with the training job on every request.
- Turn on AWS CloudTrail logging for InvokeEndpoint calls and run scheduled Amazon Athena queries that compare request fields with the training data.
- Enable data capture on the endpoint, run a SageMaker Model Monitor baseline job on the training dataset, and create a data quality monitoring schedule.
Answer
D. A data quality monitor compares statistics of captured inference inputs with a baseline computed from the training data and needs no labels, which matches the requirement. A model quality monitor needs production ground truth, which is not available, and CloudTrail does not record request payloads. Note: SageMaker Model Monitor has been closed to new AWS customers since July 2026; existing customers can keep using it.
A security team is hardening the SageMaker AI notebook instances used by contractors. Contractors must not install software as the root user, and notebook traffic must not reach the internet directly. Which settings should be configured? (Choose two.)
- Disable direct internet access and launch the notebook instances in a VPC subnet.
- Disable root access on the notebook instances.
- Encrypt the notebook instance volumes with a customer managed KMS key.
- Enable network isolation in the notebook instance lifecycle configuration.
- Enable inter-container traffic encryption on the notebook instances.
Answer
B, A. Notebook instances let you turn off root access and turn off direct internet access, which forces traffic through the chosen VPC. Volume encryption protects data at rest but changes neither requirement, and inter-container encryption and network isolation are training and hosting settings, not notebook settings.
A company runs SageMaker AI training, processing and inference continuously in two AWS Regions. Usage is steady, but the team changes instance families every few months as new hardware becomes available. Which purchasing option will reduce cost while keeping this flexibility?
- Amazon EC2 Reserved Instances in each Region
- Spot Instances for all real-time endpoints
- A SageMaker AI Savings Plan
- An EC2 Instance Savings Plan for the current instance family
Answer
C. SageMaker AI Savings Plans apply to eligible SageMaker AI usage regardless of instance family, size or Region, so they keep discounting usage as the team changes hardware. EC2 Instance Savings Plans and Reserved Instances do not apply to SageMaker AI usage, and real-time endpoints cannot run on Spot.