MLA-C01 AWS Certified Machine Learning Engineer - Associate Question and Answers

Amazon Kendra is an AI-powered search service designed for semantic search use cases. It allows ingestion of documents from an Amazon S3 bucket using the Amazon Kendra S3 connector. Once the documents are ingested, Kendra enables semantic searches with its built-in capabilities, removing the need to manually generate embeddings or manage a vector database. This approach is efficient, requires minimal operational effort, and meets the requirements for a Retrieval Augmented Generation (RAG) application.

To minimize communication overhead during distributed training:

1. Same VPC Subnet: Ensures low-latency communication between training instances by keeping the network traffic within a single subnet.

2. Same AWS Region and Availability Zone: Reduces network latency further because cross-AZ communication incurs additional latency and costs.

3. Data in the Same Region and AZ: Ensures that the training data is accessed with minimal latency, improving performance during training.

This configuration optimizes communication efficiency and minimizes overhead.

Amazon SageMaker supports direct file system access for training jobs through Amazon FSx. AWS documentation states that FSx for NetApp ONTAP file systems can be mounted directly to SageMaker training instances when they are located in the same VPC.

Mounting the FSx file system allows SageMaker to stream large datasets efficiently without copying data into Amazon S3. This is ideal for very large datasets such as 6 TB of image data and avoids unnecessary storage duplication.

Options involving SageMaker Data Wrangler are intended for data preparation and exploration, not large-scale training data access. Mountpoint for Amazon S3 is not required and introduces additional complexity.

Therefore, Option A is the correct and AWS-aligned solution.

Shadow testing allows you to send a copy of live production traffic to a shadow variant of the new model while keeping the existing production model unaffected. This enables you to evaluate the performance of the new model in real-time with live data without impacting end users. SageMaker endpoints support this setup by allowing traffic mirroring to the shadow variant, making it an ideal solution for assessing the new model's performance.

For real-time anomaly detection on streaming data, AWS recommends using Amazon Managed Service for Apache Flink, which provides serverless, scalable stream processing with built-in ML functions.

Apache Flink includes a native RANDOM_CUT_FOREST (RCF) function for unsupervised anomaly detection. This eliminates the need to deploy, manage, or scale custom ML endpoints. When combined with Amazon Kinesis Data Streams, the solution can process thousands of records per second with very low latency.

Option B introduces multiple services and custom logic, increasing operational overhead. Option C requires managing EC2 infrastructure and Kafka clusters. Option D is batch-oriented and does not meet real-time requirements.

AWS documentation explicitly highlights Kinesis + Managed Flink + RCF as the lowest-overhead solution for real-time anomaly detection.

Therefore, Option A is the correct and AWS-aligned choice.

Problem Description:

The dataset includes multiple data sources:

Transaction logs and customer profiles in Amazon S3.

Tables in an on-premises MySQL database.

There is a class imbalance in the dataset and interdependencies among features that need to be addressed.

The solution requires data aggregation from diverse sources for centralized processing.

Why AWS Lake Formation?

AWS Lake Formation is designed to simplify the process of aggregating, cataloging, and securing data from various sources, including S3, relational databases, and other on-premises systems.

It integrates with AWS Glue for data ingestion and ETL (Extract, Transform, Load) workflows, making it a robust choice for aggregating data from Amazon S3 and on-premises MySQL databases.

How It Solves the Problem:

Data Aggregation: Lake Formation collects data from diverse sources, such as S3 and MySQL, and consolidates it into a centralized data lake.

Cataloging and Discovery: Automatically crawls and catalogs the data into a searchable catalog, which the ML engineer can query for analysis or modeling.

Data Transformation: Prepares data using Glue jobs to handle preprocessing tasks such as addressing class imbalance (e.g., oversampling, undersampling) and handling interdependencies among features.

Security and Governance: Offers fine-grained access control, ensuring secure and compliant data management.

Steps to Implement Using AWS Lake Formation:

Step 1: Set up Lake Formation and register data sources, including the S3 bucket and on-premises MySQL database.

Step 2: Use AWS Glue to create ETL jobs to transform and prepare data for the ML pipeline.

Step 3: Query and access the consolidated data lake using services such as Athena or SageMaker for further ML processing.

Why Not Other Options?

Amazon EMR Spark jobs: While EMR can process large-scale data, it is better suited for complex big data analytics tasks and does not inherently support data aggregation across sources like Lake Formation.

Amazon Kinesis Data Streams: Kinesis is designed for real-time streaming data, not batch data aggregation across diverse sources.

Amazon DynamoDB: DynamoDB is a NoSQL database and is not suitable for aggregating data from multiple sources like S3 and MySQL.

Conclusion: AWS Lake Formation is the most suitable service for aggregating data from S3 and on-premises MySQL databases, preparing the data for downstream ML tasks, and addressing challenges like class imbalance and feature interdependencies.

AWS Lake Formation Documentation

AWS Glue for Data Preparation

The requirement that each parameter trial is built based on the performance of the previous trial is the defining characteristic of Bayesian optimization. In Amazon SageMaker Automatic Model Tuning, Bayesian optimization uses prior trial results to intelligently select the next set of hyperparameters, making it far more efficient than grid or random search—especially in large, mixed search spaces.

This scenario includes both categorical (20) and continuous (7) hyperparameters and allows up to 1,000 training jobs, which is well within the supported limits of SageMaker AMT. Bayesian optimization natively supports mixed parameter types and is explicitly recommended by AWS for large, high-dimensional search spaces where exhaustive grid search is impractical.

Option A and D (grid search) do not meet the requirement because grid search evaluates combinations independently and does not learn from previous trials. Additionally, grid search becomes computationally infeasible as dimensionality increases.

Option B (random search) also evaluates trials independently and does not leverage previous results, violating the core requirement.

Therefore, defining both categorical and continuous parameters and using Bayesian optimization with a maximum of 1,000 jobs is the correct and AWS-recommended solution.

The SageMaker Canvas user needs permissions to access the Amazon S3 bucket where the model artifacts are stored to retrieve the model for use in Canvas.

Registering the model in the SageMaker Model Registry allows the model to be tracked and managed within the SageMaker ecosystem. This makes it accessible for tuning and deployment through SageMaker Canvas.

This combination ensures proper access control and integration within SageMaker, enabling the Canvas user to work with the model.

The ETL process has two critical characteristics: it is long-running (6 hours) and written in R. AWS Lambda is unsuitable because it has a maximum execution time of 15 minutes. AWS Glue primarily supports Spark-based ETL and does not natively support custom R-based workloads.

AWS documentation recommends using Amazon SageMaker Processing Jobs for long-running, custom data processing workloads. Processing jobs allow users to run arbitrary code in custom Docker containers, making them ideal for migrating on-premises ETL jobs written in R.

By building a custom Docker image that includes the R runtime and required libraries and storing it in Amazon ECR, the company can run the ETL job at scale on managed infrastructure without rewriting the code.

SageMaker script mode is intended for training, not ETL. Therefore, SageMaker processing jobs with a custom container are the correct solution.

Logistic regression is a suitable modeling approach for this requirement because it is designed for binary classification problems, such as predicting whether a customer will require long-term support ("yes" or "no"). It calculates the probability of a particular class and is widely used for tasks like this where the outcome is categorical.

The company lacks programming expertise, so a no-code/low-code solution is required. Amazon SageMaker Canvas includes Data Wrangler’s visual interface, which enables users to import, transform, and prepare data using a graphical workflow.

Data Wrangler in Canvas supports common preprocessing tasks—such as handling missing values, normalization, outlier detection, and feature engineering—without writing code. It integrates directly with Amazon S3 and is suitable for large tabular datasets.

Options A, C, and D require coding skills (Spark, SQL, or Python), which violates the stated constraint.

Therefore, using the visual interface of SageMaker Data Wrangler in Canvas is the correct solution.

For asynchronous inference on large datasets, AWS recommends SageMaker batch transform, which is designed for offline and large-scale inference workloads. Batch transform jobs do not require real-time endpoints and can process large volumes of data efficiently.

For monitoring data quality, AWS provides Amazon SageMaker Model Monitor, which supports scheduled monitoring of data quality, schema drift, and feature distribution drift. Model Monitor can publish metrics to Amazon CloudWatch and trigger alerts when thresholds are breached.

The other options lack native ML-specific monitoring capabilities. CloudTrail audits API activity, not data quality. Glue, Batch, and ECS would require extensive custom monitoring logic.

Therefore, SageMaker batch transform combined with Model Monitor is the correct solution.

Amazon SageMaker Pipelines is designed for creating, automating, and managing end-to-end ML workflows, including complex data preprocessing tasks. It supports handling large datasets and can integrate with custom steps, such as NLP transformations. By combining SageMaker Pipelines with Amazon EventBridge, the entire workflow can be triggered and automated efficiently, meeting the requirements for scalability, automation, and processing complexity.

SageMaker script mode allows you to bring existing custom Python scripts and run them on AWS with minimal changes. SageMaker provides prebuilt containers for ML frameworks like PyTorch, simplifying the migration process. This approach enables the company to leverage their existing Python scripts and domain knowledge while benefiting from the scalability and managed environment of SageMaker. It requires the least effort compared to building custom containers or retraining models from scratch.

AWS Secrets Manager is designed for securely storing, managing, and automatically rotating secrets, including API tokens. By configuring a Lambda function for custom rotation logic, the solution can automatically rotate the API tokens every 3 months as required. Secrets Manager simplifies secret management and integrates seamlessly with other AWS services, making it the ideal choice for this use case.

Logging the metrics to Amazon CloudWatch allows the metrics to be tracked and monitored effectively.

CloudWatch Alarms can be configured to trigger when metrics breach a predefined threshold.

The alarm can be set to notify through Amazon Simple Notification Service (SNS), which can send email messages to the configured recipients.

This is the standard and most efficient way to achieve the desired functionality.

AWS provides Amazon SageMaker Pipelines as the native CI/CD solution for ML. Pipelines can automatically trigger retraining when new data arrives in Amazon S3, handle large datasets (such as 10 GB files), and orchestrate preprocessing, training, evaluation, and deployment steps.

For governance and auditing, Amazon SageMaker Model Registry integrates directly with Pipelines to manage model versions, approval status, and metadata such as training metrics. This combination eliminates the need for custom tracking logic and ensures traceability across the ML lifecycle.

Option A lacks native ML lineage and version governance. Option C is unsuitable for large data and complex workflows. Option D is manual and not scalable.

Therefore, using SageMaker Pipelines with the Model Registry is the correct and AWS-recommended solution.

1️⃣ Simultaneous calls to the endpoint must reach models with the same configuration

Correct strategy: All at once traffic shifting

Why:

“All at once” replaces the old model with the new model immediately. After the switch, all concurrent requests hit only the new model configuration, guaranteeing configuration consistency across simultaneous calls.

2️⃣ The new model must receive only a fraction of the requests for validation before receiving all the traffic

Correct strategy: Canary traffic shifting

Why:

Canary deployments route a small percentage of traffic (for example, 5% or 10%) to the new model first. This allows validation of correctness and performance before shifting 100% of traffic.

3️⃣ Traffic to the new model must increase gradually to ensure that pipelines that rely on the endpoint do not fail because of changes in latency

Correct strategy: Linear traffic shifting guardrail

Why:

Linear traffic shifting gradually increases traffic in equal increments over time and includes guardrails (such as CloudWatch alarms) to automatically roll back if latency or errors exceed thresholds.

AWS recommends shadow testing to evaluate a new model against a production model with minimal operational overhead. Using production variants on a single SageMaker endpoint allows traffic to be routed to multiple models without managing additional endpoints.

With a shadow variant, the new model receives a copy of live traffic but does not affect production responses. Performance metrics such as latency, accuracy, and error rates can be compared directly against the current model using Amazon CloudWatch metrics. This approach is natively supported by Amazon SageMaker Endpoints.

Options A, B, and D introduce unnecessary complexity by requiring additional endpoints, traffic routing infrastructure, or custom code.

Therefore, deploying the new model as a shadow variant on the same endpoint is the most efficient solution.

Amazon Macie is a fully managed data security and privacy service that uses machine learning to discover and classify sensitive data in Amazon S3. It is purpose-built to identify sensitive data with minimal operational overhead. After identifying the sensitive data, you can use AWS Lambda functions to automate the process of removing or redacting the sensitive data, ensuring efficiency and integration with the hybrid cloud environment. This solution requires the least development effort and aligns with the requirement to handle sensitive data effectively.

The requirement specifies real-time streaming ingestion and column-level transformation before sharing data with a vendor. Amazon Kinesis Data Streams is designed for low-latency, real-time data ingestion and delivery.

To extract only required columns from the stream, AWS recommends using Amazon Managed Service for Apache Flink as a stream consumer. Flink enables real-time transformations such as filtering, projection, and enrichment on streaming data before delivering it downstream.

Option A and D are batch-oriented and not suitable for real-time streaming. Option C is incorrect because S3 bucket policies cannot enforce column-level access controls.

Therefore, Kinesis Data Streams combined with Apache Flink meets all requirements.

Large Parquet files can cause out-of-memory (OOM) issues during training if individual files exceed the memory capacity of the training instances. AWS documentation recommends repartitioning large datasets into smaller files to enable efficient streaming and parallel loading.

By using Apache Spark on Amazon EMR to repartition the Parquet files, the ML engineer can split the dataset into multiple smaller files that can be read incrementally during training. This approach avoids loading a single large file into memory and improves data parallelism.

Attaching larger EBS volumes increases storage capacity but does not solve memory constraints. Switching to memory-optimized instances increases cost and is not necessary when the dataset can be restructured. SageMaker distributed data parallelism focuses on model parameter synchronization across GPUs, not dataset file size.

AWS best practices explicitly recommend partitioning large Parquet datasets to improve memory efficiency during training.

Therefore, Option B is the correct and AWS-aligned solution.

The primary requirement in this scenario is achieving sub-second latency for a real-time analytics dashboard powered by Amazon OpenSearch Service. The current architecture uses Amazon Data Firehose, which buffers incoming records based on time or size before delivering them to the destination. A buffer interval of 60 seconds introduces unavoidable latency, making it unsuitable for near-real-time or sub-second use cases.

According to AWS documentation, reducing or eliminating buffering in Firehose is the correct approach when low-latency ingestion is required. Setting the Firehose buffer interval to zero seconds forces Firehose to deliver records as soon as they are received. Additionally, tuning the PutRecordBatch batch size allows efficient ingestion while minimizing delivery delay. This configuration is explicitly recommended for latency-sensitive analytics pipelines.

Option B is incorrect because AWS DataSync is designed for batch-oriented data transfers between storage systems, not real-time streaming. Enhanced fan-out consumers are a feature of Amazon Kinesis Data Streams, not DataSync, making this option invalid.

Option C directly contradicts the requirement. Increasing the buffer interval from 60 seconds to 120 seconds would further increase latency and degrade real-time performance.

Option D is also incorrect because Amazon SQS is a message queueing service, not a streaming ingestion service optimized for indexing data into OpenSearch with minimal latency. Using SQS would add additional processing layers and would not inherently provide sub-second ingestion into OpenSearch.

Therefore, using zero buffering in the Firehose stream and tuning the PutRecordBatch batch size is the only change that aligns with AWS best practices for achieving sub-second latency in real-time analytics pipelines.

Amazon SageMaker Automated Model Tuning supports several hyperparameter search strategies. Bayesian optimization is explicitly designed to model the relationship between hyperparameters and objective metrics using regression techniques. Based on results from previous training jobs, Bayesian optimization predicts which hyperparameter combinations are most likely to improve model performance and evaluates those next.

AWS documentation highlights Bayesian optimization as the preferred strategy when the hyperparameter search space is small to medium and when training jobs are expensive. Because the algorithm learns from prior runs, it avoids wasting resources on unpromising configurations and converges efficiently.

Grid search exhaustively evaluates all combinations and becomes inefficient even with moderately sized search spaces. Random search does not use information from prior runs and is less efficient. Hyperband focuses on aggressive early stopping and resource allocation, not regression-based sequential selection.

Therefore, Option C is the correct and AWS-verified solution.

The problem described is a patient segmentation use case, which is a classic example of unsupervised learning. The objective is to group patients with similar characteristics without predefined labels. AWS documentation clearly states that Amazon SageMaker k-means is designed specifically for clustering and segmentation tasks.

The SageMaker k-means algorithm groups data points into clusters based on feature similarity and requires the user to define the number of clusters using the k hyperparameter. This directly satisfies the requirement to “determine the number of groups by using hyperparameters.” AWS recommends k-means for applications such as customer segmentation, risk grouping, and pattern discovery in healthcare data.

Option A (XGBoost) is a supervised learning algorithm used for classification and regression. The max_depth hyperparameter controls tree complexity, not the number of groups, making it unsuitable for this task.

Option C (DeepAR) is a time-series forecasting algorithm optimized for predicting future values, not clustering patients.

Option D (Random Cut Forest) is an anomaly detection algorithm. While useful for identifying outliers or unusual patient behavior, it does not perform clustering or group segmentation.

AWS SageMaker documentation explicitly identifies k-means as the correct choice when the goal is to partition data into a predefined number of clusters using a tunable hyperparameter.

Therefore, Option B is the correct and AWS-verified answer.

Use case 1:

Summarize the text of a research paper

➡️ Sequence-to-Sequence (seq2seq) algorithm

Why:

Seq2seq models are designed for natural language generation tasks such as text summarization, translation, and paraphrasing. AWS documentation explicitly lists text summarization as a primary use case for the SageMaker seq2seq algorithm.

Use case 2:

Scan every pixel of an image to help self-driving cars identify objects in their path

➡️ Semantic segmentation algorithm

Why:

Semantic segmentation performs pixel-level classification, assigning a class label to every pixel in an image. This is exactly what is required for applications such as autonomous driving, road scene understanding, and object boundary detection.

Use case 3:

Identify abnormal data points in a dataset

➡️ Random Cut Forest (RCF) algorithm

Why:

Random Cut Forest is an unsupervised anomaly detection algorithm. AWS SageMaker RCF is purpose-built to identify outliers, unusual patterns, and anomalies in numerical datasets, making it ideal for fraud detection, monitoring, and abnormal data point detection.

This is a classic class imbalance problem, where fraudulent transactions (minority class) are severely underrepresented. AWS documentation for SageMaker Data Wrangler recommends SMOTE (Synthetic Minority Oversampling Technique) as an effective approach for improving model performance in such scenarios.

SMOTE generates synthetic minority samples by interpolating between existing minority class examples. This improves the model’s ability to learn decision boundaries without simply duplicating data, which can cause overfitting.

Random undersampling removes valuable majority class data, reducing overall model robustness. Random oversampling duplicates data and increases overfitting risk. Changing algorithms does not address the root cause.

AWS best practices highlight SMOTE as the preferred technique for fraud detection and other highly imbalanced datasets.

Therefore, Option C is the correct and AWS-verified answer.

This scenario describes a severely imbalanced classification problem. In such cases, accuracy is a misleading metric, because the model can achieve high accuracy by predicting only the majority class.

AWS ML best practices recommend using F1 score (or precision/recall) when evaluating imbalanced datasets. The F1 score balances false positives and false negatives, making it ideal for assessing minority-class performance.

For image data, image augmentation (rotations, flips, crops, color jitter) is the preferred technique to increase minority-class representation. SMOTE is designed for tabular data and is not suitable for image pixel data.

Therefore, the correct solution is to optimize for F1 score and apply image augmentation.

Thus, Option B is the correct and AWS-aligned answer.

AWS Lake Formation provides fine-grained access control and simplifies data governance for data lakes. By configuring Lake Formation tags to map ML engineers to their specific campaigns, you can restrict access to both structured and unstructured data in the data lake. This method is operationally efficient, as it centralizes access control management within Lake Formation and ensures consistency across Amazon Athena and S3 bucket access without requiring manual updates to policies or DynamoDB-based custom logic.

AWS ML best practices state that data preprocessing applied during training must be applied identically during inference. For min-max normalization, this requires reusing the minimum and maximum values calculated from the training dataset.

If production data is normalized using different statistics, the feature distributions will differ from what the model learned, leading to degraded prediction accuracy. AWS documentation explicitly warns against recomputing normalization parameters on inference data.

Options A, C, and D introduce data leakage or inconsistent feature scaling. Option B ensures consistency between training and inference pipelines and preserves model integrity.

Therefore, Option B is the correct and AWS-aligned solution.

For near real-time anomaly detection on streaming IoT data, AWS recommends Amazon Managed Service for Apache Flink. Flink is designed for stateful, low-latency stream processing and is well suited for time-series sensor analytics.

A key requirement is application resilience during deployments. Managed Flink supports system rollback, which automatically reverts to the last stable application version if a new deployment fails. This capability ensures uninterrupted processing even if application bugs are introduced, meeting the company’s reliability requirement.

Manual rollback (Option B) introduces operational risk and delays. Amazon Data Firehose (Options C and D) is a delivery service and does not support complex anomaly detection logic or application version rollback.

Therefore, using Managed Flink with system rollback enabled is the correct solution.

This scenario clearly describes an overfitting problem. The neural network performs well on training data but fails to generalize to evaluation data. According to AWS Machine Learning and deep learning best practices, overfitting occurs when a model learns noise or overly specific patterns in the training data instead of generalizable relationships.

L1 and L2 regularization are well-established techniques to combat overfitting. They work by penalizing large weight values during optimization, effectively constraining the model’s complexity. L1 regularization encourages sparsity, while L2 regularization (weight decay) smooths the weight distribution. AWS documentation recommends regularization as a primary method for improving generalization when a model shows a large gap between training and evaluation performance.

Option B is incorrect because removing dropout layers would increase overfitting, not reduce it. Dropout is itself a regularization technique.

Option C would likely worsen overfitting, as training for more epochs allows the model to memorize the training data even further.

Option D increases model complexity, which again exacerbates overfitting rather than resolving it.

Therefore, penalizing large weights using L1 or L2 regularization is the correct solution.

Scenario: The company wants to perform online validation of a new ML model on 10% of the traffic before fully deploying the model in production. The setup must have minimal operational overhead.

Why Use SageMaker Production Variants?

Built-In Traffic Splitting: Amazon SageMaker endpoints support production variants, allowing multiple models to run on a single endpoint. You can direct a percentage of incoming traffic to each variant by adjusting the variant weights.

Ease of Management: Using production variants eliminates the need for additional infrastructure like separate endpoints or custom ALB configurations.

Monitoring with CloudWatch: SageMaker automatically integrates with CloudWatch, enabling real-time monitoring of model performance and invocation metrics.

Steps to Implement:

Deploy the New Model as a Production Variant:

Update the existing SageMaker endpoint to include the new model as a production variant. This can be done via the SageMaker console, CLI, or SDK.

Example SDK Code:

import boto3

sm_client = boto3.client('sagemaker')

response = sm_client.update_endpoint_weights_and_capacities(

EndpointName='existing-endpoint-name',

DesiredWeightsAndCapacities=[

{'VariantName': 'current-model', 'DesiredWeight': 0.9},

{'VariantName': 'new-model', 'DesiredWeight': 0.1}

]

)

Set the Variant Weight:

Assign a weight of 0.1 to the new model and 0.9 to the existing model. This ensures 10% of traffic goes to the new model while the remaining 90% continues to use the current model.

Monitor the Performance:

Use Amazon CloudWatch metrics, such as InvocationCount and ModelLatency, to monitor the traffic and performance of each variant.

Validate the Results:

Analyze the performance of the new model based on metrics like accuracy, latency, and failure rates.

Why Not the Other Options?

Option B: Setting the weight to 1 directs all traffic to the new model, which does not meet the requirement of splitting traffic for validation.

Option C: Creating a new endpoint introduces additional operational overhead for traffic routing and monitoring, which is unnecessary given SageMaker's built-in production variant capability.

Option D: Configuring the ALB to route traffic requires manual setup and lacks SageMaker's seamless variant monitoring and traffic splitting features.

Conclusion:

Using production variants with a weight of 0.1 for the new model on the existing SageMaker endpoint provides the required traffic split for online validation with minimal operational overhead.

Problem Description:

The F1 score, which is a balance of precision and recall, has decreased significantly. This indicates the model's predictions are no longer aligned with the real-world data distribution.

Why Concept Drift?

Concept drift occurs when the statistical properties of the target variable or features change over time. For example, customer behaviors or subscription cancellation patterns may have shifted, leading to reduced model accuracy.

Signs of Concept Drift:

Deviation in performance metrics (e.g., F1 score) over time.

Declining prediction accuracy for certain groups or scenarios.

Solution:

Monitor for drift using tools like SageMaker Model Monitor.

Regularly retrain the model with updated data to account for the drift.

Why Not Other Options?:

B: Model complexity is unrelated if the model initially performed well.

C: Data quality issues would have been detected during baseline analysis.

D: Incorrect ground truth labels would have resulted in a consistently poor baseline.

Conclusion: The decrease in F1 score is most likely due to concept drift in the customer data, requiring retraining of the model with new data.

AWS documentation identifies Amazon SageMaker Clarify as the primary service for detecting, measuring, and explaining bias in ML models, particularly across demographic and sensitive attributes such as age, gender, and location. Clarify can analyze bias before training, after training, and during inference, making it suitable for audit and compliance requirements.

SageMaker Clarify generates bias reports using established fairness metrics such as difference in positive proportions, disparate impact, and conditional demographic disparity. These reports are exportable and auditor-friendly, directly meeting the requirement to explain bias to an external party.

AWS Glue DataBrew focuses on data preparation and quality, not bias detection. Amazon QuickSight does not provide ML fairness metrics. Amazon CloudWatch captures operational metrics, not demographic bias indicators.

AWS best practices explicitly recommend SageMaker Clarify for model transparency, fairness evaluation, and regulatory reporting.

Therefore, Option A is the correct and AWS-verified solution.

AWS recommends Amazon SageMaker Model Monitor for detecting data drift and model drift in deployed models. Model Monitor works by comparing live inference data against a baseline, which must first be created from the training dataset.

AWS documentation clearly specifies the required order:

Create a baseline using training data statistics

Create a monitoring schedule to compare incoming data against the baseline

Option A reverses this order and is therefore incorrect. Option C is incorrect because SageMaker Clarify focuses on bias and explainability, not ongoing drift detection. Option D is reactive and does not provide continuous monitoring.

Model Monitor integrates with Amazon CloudWatch, enabling automated alerts and downstream retraining workflows. This proactive approach allows companies to detect degradation early and maintain model quality.

Therefore, Option B is the correct and AWS-verified answer.

S3 Gateway Endpoint: Allows private access to S3 from within a VPC without requiring a public IPv4 address, ensuring that data transfer between the primary and secondary accounts is secure and private.

Bucket Policy Update: The S3 bucket policy in the secondary account must explicitly allow access from the primary account's IAM principals to provide the necessary permissions.

Interface VPC Endpoints: Required for private communication between the VPC and Amazon SageMaker and Amazon Redshift services, ensuring the solution operates without public internet access.

This configuration meets the requirement to avoid public IPv4 addresses and allows secure and private communication between the accounts.

Recall measures the ability of a model to correctly identify all positive cases (true positives) out of all actual positives, minimizing false negatives. Since the cost of false negatives is much higher than false positives in this scenario, the ML engineer should prioritize models with high recall to reduce the likelihood of missing positive cases.

This dataset shows severe class imbalance, with only 2% of records representing patients with the disease. AWS ML best practices recommend correcting imbalance only in the training dataset, while keeping validation and test sets representative of real-world distributions.

Synthetic Minority Oversampling Technique (SMOTE) generates synthetic samples of the minority class by interpolating between existing minority examples. This improves the model’s ability to learn disease-related patterns without discarding data.

PCA is a dimensionality reduction method, not an oversampling technique. Oversampling the majority class worsens imbalance. Altering the test dataset would invalidate evaluation results.

Therefore, applying SMOTE to the training dataset is the correct approach.

Step 1:

The company commits code changes or new training datasets to a Git repository.

This is the trigger point. A source code or data change initiates the CI/CD pipeline.

Step 2:

CodePipeline detects code changes and starts to build automatically.

CodePipeline monitors the Git repository (for example, AWS CodeCommit, GitHub, or Bitbucket) and automatically triggers the pipeline when changes are detected.

Step 3:

The company builds and deploys ML models and applications to staging servers for testing.

The pipeline runs build, training, and test stages (often using AWS CodeBuild and SageMaker) and deploys artifacts to a staging or test environment for validation.

Step 4:

Human approval is provided after testing is successful.

A manual approval action is a best practice for ML workflows to ensure governance, compliance, and quality checks before production deployment.

Step 5:

CodePipeline deploys ML models and applications to production.

After approval, the pipeline automatically deploys the validated model or application to the production environment.

An oscillating loss pattern during training with stochastic gradient descent (SGD) is a strong indicator that the learning rate is too high. When the learning rate is excessive, the optimizer takes overly large steps during gradient updates, causing the model to repeatedly overshoot the optimal minimum of the loss function. This results in unstable convergence behavior, where training and validation loss decrease briefly and then increase again in a repeating cycle.

AWS Machine Learning documentation and general deep learning best practices recommend reducing the learning rate when training loss and validation loss both remain high and fluctuate rather than steadily decreasing. Lowering the learning rate allows the optimizer to take smaller, more precise steps toward the minimum, leading to smoother convergence and improved generalization on the test dataset.

Option A, early stopping, is used primarily to prevent overfitting when validation loss increases while training loss continues to decrease. In this scenario, both losses remain high and unstable, indicating an optimization issue rather than overfitting.

Option B is incorrect because increasing the test set size does not affect the training dynamics or convergence behavior of the model.

Option C would worsen the problem, as increasing the learning rate would further amplify oscillations and instability.

Therefore, decreasing the learning rate is the correct corrective action to stabilize SGD training and improve model performance.

Step 1:

Create a Shapley Additive Explanations (SHAP) baseline for the model by using Amazon SageMaker Clarify.

Step 2:

Schedule an hourly model explainability monitor.

Step 3:

Check the analysis results on the SageMaker Studio console.

When monitoring a deployed model for data drift and explainability, AWS prescribes a specific workflow using SageMaker Clarify and SageMaker Model Monitor:

Create a SHAP baseline (Step 1)Before any monitoring can occur, SageMaker Clarify must establish a baseline explainability configuration. This baseline captures the reference SHAP values for feature importance using training or baseline data. Model Monitor uses this baseline to compare future inferences and detect drift in feature attributions.

Schedule the model explainability monitor (Step 2)After the baseline is created, an explainability monitoring schedule must be configured (hourly in this case). The monitor periodically analyzes inference data, compares it against the SHAP baseline, and generates reports that highlight drift or anomalies in feature contributions.

Inspect results in SageMaker Studio (Step 3)Once monitoring jobs run, SageMaker stores the analysis results in Amazon S3 and surfaces them in the SageMaker Studio console, where engineers can review metrics, violations, and visual reports.

This sequence is mandatory because:

A monitor cannot run without a baseline

Results cannot be reviewed until the monitor executes

The task described is unsupervised topic modeling, where topics are unknown in advance. Latent Dirichlet Allocation (LDA) is a probabilistic generative model specifically designed to discover latent topics from a corpus of documents without labeled data. AWS provides LDA as a built-in algorithm in Amazon SageMaker, making it well suited for this requirement.

LDA models documents as mixtures of topics and topics as mixtures of words, enabling interpretable topic discovery at scale. This aligns precisely with the need to organize documents into topics when the topics are not predefined.

BlazingText is optimized for word embeddings and supervised text classification, not topic modeling. Sequence-to-sequence models are used for translation or summarization. Object2Vec creates embeddings but does not itself perform topic discovery without additional clustering steps.

Therefore, LDA is the correct and purpose-built solution.

To generate images, the application must invoke a foundation model hosted on Amazon Bedrock. Invoking the Amazon Titan Image Generator requires the bedrock:InvokeModel permission. Read-only permissions such as bedrock:Get* do not allow inference execution and are insufficient.

The application also interacts with Amazon Simple Queue Service. To process messages, the application must be able to receive messages from the queue and delete them after successful processing. These actions require sqs:ReceiveMessage and sqs:DeleteMessage.

sqs:GetQueueAttributes alone does not allow message consumption. SageMaker permissions are unrelated because Bedrock is a separate service.

Therefore, permissions to invoke the model and receive/delete SQS messages are required.

SageMaker Data Wrangler provides a no-code/low-code interface for preparing and transforming data, including dropping unnecessary columns. By creating a data flow and configuring a transform step, the ML engineer can easily remove correlated or unneeded columns from the Parquet file with minimal effort. This approach avoids the need for custom coding or managing additional infrastructure.

When predicting continuous variables, such as apartment prices, it's essential to evaluate the model's performance using appropriate regression metrics. The Mean Absolute Error (MAE) is a widely used metric for this purpose.

Understanding Mean Absolute Error (MAE):

MAE measures the average magnitude of errors in a set of predictions, without considering their direction. It calculates the average absolute difference between predicted values and actual values, providing a straightforward interpretation of prediction accuracy.

Advantages of MAE:

Interpretability: MAE is expressed in the same units as the target variable, making it easy to understand.

Robustness to Outliers: Unlike metrics that square the errors (e.g., Mean Squared Error), MAE does not disproportionately penalize larger errors, making it more robust to outliers.

Comparison with Other Metrics:

Accuracy, AUC, F1 Score: These metrics are designed for classification tasks, where the goal is to predict discrete labels. They are not suitable for regression problems involving continuous target variables.

Mean Squared Error (MSE): While MSE also measures prediction errors, it squares the differences, giving more weight to larger errors. This can be useful in certain contexts but may be sensitive to outliers.

Conclusion:

For evaluating the performance of a model predicting apartment prices—a continuous variable—MAE is an appropriate and effective metric. It provides a clear indication of the average prediction error in the same units as the target variable, facilitating straightforward interpretation and comparison.

AWS Bedrock Knowledge Bases support multiple chunking strategies to optimize retrieval quality. Hierarchical chunking is specifically designed for structured documents such as PDFs with headings, sections, and subsections.

Hierarchical chunking allows fine-grained retrieval at the subsection level while automatically preserving parent section context. This ensures that when a small portion is retrieved, the surrounding section is also provided to the foundation model for better understanding.

Fixed-size and maximum-token chunking can split content arbitrarily, breaking semantic and structural boundaries. Semantic chunking focuses on meaning but does not guarantee structured context preservation without additional logic.

AWS documentation highlights hierarchical chunking as the preferred strategy when documents are structured and contextual integrity is required.

Therefore, Option A is the correct and AWS-aligned solution.

Amazon SageMaker Data Wrangler supports both direct and cataloged connections. To ensure data is always up to date, AWS recommends using cataloged connections backed by AWS Glue Data Catalog.

Cataloged connections allow Data Wrangler to reference the source systems dynamically, ensuring that each import reflects the latest data changes without manual reconfiguration. This approach supports Amazon S3, Amazon Redshift, and Snowflake and integrates securely using managed credentials.

Direct connections are point-in-time imports and do not automatically reflect schema or data updates. Glue and Lambda-based solutions introduce unnecessary complexity and operational overhead.

Therefore, using cataloged connections in Data Wrangler is the correct solution.

Change the number of samples used in each iteration of training

Correct selection:

Change the batch size

Why:

Increasing the batch size reduces the number of iterations per epoch, which can significantly shorten training time while maintaining model quality when tuned appropriately. AWS explicitly recommends batch size tuning as a primary performance optimization.

Increase the number of instances used during training

Correct selection:

Use the SageMaker AI distributed data parallelism (SMDDP) library

Why:

SMDDP is designed to efficiently distribute training data across multiple GPU instances with optimized gradient synchronization. This accelerates training without affecting model convergence or accuracy, unlike naive scaling approaches.

Stop training before the maximum number of epochs are reached if performance is sufficient and not improving

Correct selection:

Use early stopping on the training job

Why:

Early stopping automatically terminates training when validation metrics stop improving. AWS recommends this to reduce wasted compute time while preserving optimal model performance.

In the Amazon SageMaker Python SDK, the fit() method is used to start a training job after an estimator has been configured. AWS documentation explicitly states that once an estimator (such as PyTorch or TensorFlow) is defined with parameters like instance type, framework version, and hyperparameters, the fit() method is responsible for launching the training process.

The fit() method accepts the training data location (commonly an Amazon S3 URI) and initiates the managed training job on SageMaker infrastructure. SageMaker then provisions the required compute resources, stages the data, executes the training script, and stores model artifacts in Amazon S3.

The create_model() method is used after training to create a SageMaker model object from trained artifacts. The deploy() method deploys a trained model to an endpoint for inference. The predict() method is used only after deployment to request predictions from an endpoint.

AWS documentation clearly separates these lifecycle steps and identifies fit() as the correct method to initiate training.

Therefore, Option A is the correct and AWS-verified answer.

In regression problems such as house price prediction, extreme values can significantly distort model learning. In this dataset, the presence of a large mansion and an extremely small apartment represents clear outliers in the Square Meters feature. According to AWS Machine Learning best practices, outliers can disproportionately influence loss functions (such as mean squared error), leading to poor predictions for the majority of typical data points.

Removing these outliers helps the model focus on learning patterns that apply to the majority of houses and apartments, which aligns with the requirement to produce accurate predictions for typical properties. After removing outliers, applying a log transformation to the Square Meters feature further improves model performance by reducing skewness and stabilizing variance. Log transformations are commonly recommended in AWS and general ML documentation when numerical features span multiple orders of magnitude.

Option B is incorrect because normalization alone does not address the undue influence of extreme outliers. Option C and D are incorrect because one-hot encoding is intended for categorical variables, not continuous numerical features such as square meters.

Therefore, removing outliers and applying a log transformation is the most statistically sound preprocessing approach.

Class imbalance occurs when one class significantly outnumbers another, causing models to bias predictions toward the majority class. In this case, images without defects (1,800) vastly outnumber images with defects (200). AWS ML best practices recommend oversampling the minority class to improve class representation without discarding valuable data.

Oversampling techniques—such as duplicating minority samples or applying data augmentation—help the model better learn defect-related features. This approach preserves all available data and improves recall and precision for underrepresented defect classes.

Option B is incorrect because undersampling the minority class would further worsen imbalance. Option A unnecessarily reduces dataset size. Option D does not address the imbalance problem.

Thus, oversampling defect images is the correct solution.

AWS provides Amazon SageMaker Multi-Model Endpoints to host multiple trained models behind a single inference endpoint. This feature dynamically loads models into memory based on request traffic, significantly reducing endpoint sprawl and operational overhead.

Multi-model endpoints are ideal when an organization has hundreds or thousands of related models, such as geographic or regional models, that are not all invoked simultaneously. AWS documentation highlights this approach as a best practice for cost optimization and simplified deployment management.

Option B (multi-container endpoints) is intended for chaining a small number of models, not hundreds. Option C does not address endpoint consolidation. Option D changes the modeling strategy and is not required.

Therefore, using a single SageMaker multi-model endpoint is the correct solution.

Callback steps in Amazon SageMaker Pipelines allow you to integrate external processes, such as AWS Glue jobs, into the pipeline workflow. By using a Callback step, the SageMaker pipeline can trigger the AWS Glue workflow and pause execution until the Glue jobs complete. This approach provides seamless integration with minimal operational overhead, as it directly ties the pipeline's execution flow to the completion of the AWS Glue jobs without requiring additional orchestration tools or complex setups.

Amazon SageMaker Automatic Model Tuning extracts objective metrics directly from training logs. For custom containers, AWS requires the ML engineer to define a metric definition that specifies a regular expression to parse the desired metric value from standard output.

The ObjectiveMetricName in the tuning job must match the metric captured by the regex. This is the only supported method for integrating custom metrics with SageMaker AMT.

TrainingInputMode does not define metrics. CloudWatch metrics cannot be used as tuning objectives. AMT cannot read metrics from S3 artifacts.

AWS documentation explicitly states that regex-based metric definitions are required for custom metrics in hyperparameter tuning jobs.

Therefore, Option B is the correct and AWS-verified solution.

For near real-time predictions, AWS documentation recommends Amazon SageMaker real-time inference endpoints. These endpoints support automatic scaling based on metrics such as ConcurrentInvocationsPerInstance, ensuring low latency and high availability during traffic spikes.

For long-running historical analysis, SageMaker Processing jobs are the appropriate solution. Processing jobs are designed for batch workloads, can run for hours, and integrate cleanly with SageMaker pipelines. Scheduling them with Amazon EventBridge provides a fully managed, scalable, and serverless solution.

AWS Lambda is unsuitable for multi-hour workloads. EC2 Auto Scaling adds unnecessary infrastructure management overhead.

Therefore, Options A and C together meet all requirements and align with AWS best practices.

Apache Parquet is a columnar storage file format optimized for complex and large datasets. It provides efficient reading and processing by accessing only the required columns, which reduces I/O and speeds up data handling. This makes it ideal for use with Amazon SageMaker Canvas, where minimizing processing time is important for training ML models. Parquet is also compatible with S3 and widely supported in data analytics and ML workflows.

SageMaker Pipelines provides a directed acyclic graph (DAG) view for managing and visualizing ML workflows with fine-grained control. It integrates seamlessly with SageMaker Studio, offering an intuitive interface for workflow orchestration.

SageMaker ML Lineage Tracking keeps a running history of experiments and tracks the lineage of datasets, models, and training jobs. This feature supports model governance, auditing, and compliance verification requirements.