Amazon Practice Questions, Discussions & Exam Topics by our Authors
A company wants to use a large language model (LLM) to generate concise, feature-specific descriptions for the company=E2=80=99s products....
Option Analysis:
A) Create one prompt that covers all products. Edit the responses to make the responses more specific, concise, and tailored to each product:
- Reasoning: This approach suggests using a single, general prompt for all products, then manually editing the generated responses. While this could save time upfront, it may require significant manual effort to ensure each response is accurate, concise, and tailored. This is not the most efficient approach, especially for a large range of products, as it doesn't leverage the model's potential for generating customized content for each product.
- Use case: This option might work if the company has only a few products and can afford the manual work, but it doesn't scale well and is time-consuming.
- Conclusion: Rejected.
B) Create prompts for each product category that highlight the key features. Include the desired output format and length for each prompt response:
- Reasoning: Creating specific prompts for each product category allows for more focused outputs tailored to the unique features of each product type. By highlighting key features, the model can generate more relevant and accurate descriptions. Additionally, specifying output length and format ensures consistency and aligns with the company’s needs for concise descriptions. This is an efficient approach because it combines specificity with flexibility.
- Use case: This method works well for a wide range of products where different categories have distinct features. It balances customization and efficiency, reducing the need for manual post-editing.
- Conclusion: Selected.
C) Include a diverse range of product features in each prompt to generate creative and unique descriptions:
- Reasoning: Including a broad range of features in the prompt may lead to descriptions that are overly complex and less con...
Author: Zain · Last updated Jul 19, 2026
A company is implementing intelligent agents to provide conversational search experiences for its customers. The company needs a database service that will support storage and queries of embeddings from a gene...
To meet the requirements of supporting storage and querying of embeddings from a generative AI model as vectors, the company needs a database service that is optimized for high-performance vector-based search, which is commonly required for storing and querying embeddings generated by AI models. Let’s evaluate the given options:
A) Amazon Athena
- Purpose: Amazon Athena is primarily a serverless query service that allows you to analyze data in Amazon S3 using SQL.
- Reason for rejection: While Athena can query structured data, it is not optimized for storing and querying vector embeddings (such as from AI models). It is primarily designed for ad-hoc queries on large datasets in S3, not for high-dimensional vector searches.
- Scenario where it can be used: If the company needs to perform SQL-based analysis on data stored in S3 (e.g., CSV, Parquet files), Athena could be helpful, but it doesn’t meet the need for vector search.
B) Amazon Aurora PostgreSQL
- Purpose: Amazon Aurora is a fully managed relational database that is compatible with PostgreSQL and MySQL.
- Reason for rejection: While PostgreSQL can store vectors using extensions like pgvector, Aurora is designed primarily for transactional relational workloads. It can support basic vector storage, but its performance for high-dimensional vector search may not be as optimized compared to specialized services.
- Scenario where it can be used: If the company has a mixed workload (both relational data and vector embeddings) and prefers a relational database, Aurora PostgreSQL could be a good fit, but it isn’t the best option for scalable, high-performance vector searches at scale.
C) Amazon Redshift
- Purpose: Amazon Redshift is a fully managed data warehouse service, optimized for OLAP (Online Analytical Processing) ...
Author: Victoria · Last updated Jul 19, 2026
A financial institution is building an AI solution to make loan approval decisions by using a foundation model (FM). For security and audit purposes, the company needs the AI solution's decisions to be...
To build an explainable AI solution for loan approval decisions, particularly for security and audit purposes, it's essential to understand what factors contribute to the explainability of the AI solution’s decisions. Let's evaluate each option in this context:
A) Model complexity
Model complexity refers to how complicated the underlying model is—such as the number of layers in a deep neural network, or the amount of interdependence between model components. More complex models (like large, deep neural networks) can be harder to explain because their decision-making processes are more difficult to trace or understand. In contrast, simpler models like decision trees are more transparent and their decisions can be explained in more straightforward terms.
Reason for selection: Model complexity is the most directly relevant factor when considering explainability. More complex models often result in harder-to-interpret decisions, which can create challenges for regulatory compliance, security, and audits. A balance must be struck between model accuracy and explainability, particularly in sensitive domains like financial decision-making.
B) Training time
Training time refers to how long it takes to train the AI model on data. While training time can be an important factor for model efficiency, it doesn't inherently affect the explainability of a model's decisions. A model can take a long time to train but still be either highly explainable or not explainable at all, depending on its structure.
Reason for rejection: Training time has no direct correlation with explainability. A model could be trained quickly or slowly and still have a similar level of explainability, so it is not a critical factor here.
C) Number of hyperparameters
Hyperparameters are the parameters set before training a model (like learning rate, number of layers, etc.). While the number of hyperparameters can influence the performance of a mode...
Author: ThunderBear · Last updated Jul 19, 2026
A pharmaceutical company wants to analyze user reviews of new medications and provide a concise overview for each m...
To analyze user reviews of new medications and provide a concise overview, we need a solution that focuses on natural language processing (NLP), particularly summarizing textual data such as user reviews.
Let’s break down the options:
A) Create a time-series forecasting model to analyze the medication reviews by using Amazon Personalize.
- Rejection Reasoning: Time-series forecasting is used for predicting trends and patterns over time, which is ideal for recommending products or predicting future behaviors. This doesn't align with the task of summarizing user reviews. User reviews are not inherently time-dependent and don't require forecasting but rather summarizing, which is why this is not the best solution.
B) Create medication review summaries by using Amazon Bedrock large language models (LLMs).
- Selection Reasoning: Amazon Bedrock provides access to powerful large language models (LLMs) like GPT or other NLP models, which are highly capable of processing and summarizing text. By leveraging these models, the company can generate concise and coherent summaries from user reviews. This directly addresses the need for creating review summaries and would provide a quick and accurate overview of feedback, minimizing manual effort while providing a scalable solution. The time and cost will be manageable cons...
Author: Isabella · Last updated Jul 19, 2026
A company wants to build a lead prioritization application for its employees to contact potential customers. The application must give employees the ability to view and adjust the weights assigned to different variables in...
To meet the company's requirement of building a lead prioritization application that allows employees to view and adjust the weights assigned to different variables based on their domain knowledge and expertise, we need to assess which machine learning (ML) model type fits best based on factors like interpretability, flexibility in adjusting weights, and ease of use.
A) Logistic regression model
- Benefits: Logistic regression is a linear model that is easy to understand and interpret. It allows users to directly view and adjust the weights (coefficients) associated with each input variable, which aligns perfectly with the requirement of allowing employees to adjust these weights based on their domain knowledge and expertise.
- Drawback: Logistic regression is limited to linear relationships between the features and the outcome. It may not capture complex interactions or non-linear relationships well, which could be a disadvantage depending on the complexity of the lead prioritization task.
- Use Case: This model is well-suited for tasks where the relationships between variables are relatively simple and interpretable. If the data can be reasonably approximated by a linear model, logistic regression is an ideal choice.
B) Deep learning model built on principal components
- Benefits: Deep learning models can capture highly complex patterns and interactions in data, especially when the data is high-dimensional. Principal component analysis (PCA) can be used to reduce dimensionality, which might improve training efficiency.
- Drawback: Deep learning models are typically more complex and harder to interpret. Employees would not be able to easily adjust the weights of the model’s features directly based on domain knowledge because the model operates in a non-linear, opaque manner. This contradicts the requirement for transparency and flexibility in adjusting weights.
- Use Case: Deep learning models are ideal for tasks where the data is very complex and contains intricate patterns that simpler models cannot capture. However, they are not suitable for this scenario because of their lack of interpretability and ease...
Author: Aarav2020 · Last updated Jul 19, 2026
A company needs to train an ML model to classify images of different types of animals. The company has a large dataset of labeled images and will not label more ...
To classify images of different types of animals, the company needs to select an appropriate type of machine learning (ML) based on several key factors such as the data they have, the effort required, the time needed for model training, and the associated cost.
Option Analysis:
A) Supervised Learning:
- Description: Supervised learning is the most appropriate method for situations where labeled data is available. Since the company has a large dataset of labeled images, supervised learning would allow the model to learn the relationship between the input images (features) and the labels (animal types).
- Reasoning: The company already has labeled data, meaning the effort of labeling new data is not required. Supervised learning is time-efficient because the labels are already available, and the model can learn directly from them. The cost is also low since no additional labeling effort is needed.
- Scenario Usage: This is the most straightforward and effective approach for the company’s goal of classifying images.
B) Unsupervised Learning:
- Description: Unsupervised learning is used when the data is not labeled, and the model needs to identify patterns or clusters without predefined labels.
- Reasoning: The company already has labeled data, so using unsupervised learning would be inefficient in this case. Unsupervised learning requires the model to find structures in the data on its own, which is not necessary when labels are already provided.
- Scenario Usage: Unsupervised learning is useful when you do not have labeled data and need to cluster or find patterns without specific labels, such as customer segmentation or anomaly detection.
...
Author: Sam · Last updated Jul 19, 2026
Which phase of the ML lifecycle determines compliance and regulatory requirements?
The phase of the ML lifecycle that determines compliance and regulatory requirements is Option D: Business goal identification.
Analysis of each option:
1. A) Feature engineering
- Feature engineering focuses on transforming raw data into useful features that can be fed into a machine learning model. While it’s a critical step in the ML lifecycle, it doesn't directly involve determining compliance or regulatory requirements. Compliance often involves ensuring that the data, the model, and the processes adhere to relevant laws and regulations, which is outside the scope of feature engineering.
- Effort: Medium, but unrelated to regulatory requirements.
- Time: Medium, but irrelevant to compliance determination.
- Cost: Medium, but it doesn't address regulatory needs.
2. B) Model training
- Model training involves optimizing the model's parameters using the data prepared in earlier steps. However, regulatory requirements aren't typically assessed during the training phase; rather, the focus is on improving model performance. Ensuring compliance is a preliminary task that should occur before training the model, during the goal-setting phase.
- Effort: High, but not related to compliance.
- Time: High, but compliance should be addressed earlier in the process.
- Cost: High, but irrelevant for determining compliance.
3. C) Data collection
- Data collection involves gathering the data required to train the model. Although it’s essential to ensure that data used complies with legal and ethical guidelines, compliance and regulatory requirements are best addressed when defining the business goals. At this stage, the company identifi...
Author: Aarav · Last updated Jul 19, 2026
A food service company wants to develop an ML model to help decrease daily food waste and increase sales revenue. The company needs to continuously imp...
In this scenario, the food service company wants to develop a machine learning (ML) model to decrease food waste and increase sales revenue while continuously improving the model's accuracy. To meet these goals, the company needs a solution that facilitates the continuous improvement of the ML model based on new or evolving data.
Option Analysis:
A) Use Amazon SageMaker and iterate with newer data:
- Description: Amazon SageMaker is a fully managed service for building, training, and deploying ML models. It allows you to continuously retrain and improve the model by feeding it new data, thus improving accuracy over time.
- Reasoning: This is the most suitable option because it allows the company to develop and deploy ML models effectively. With SageMaker, they can iterate on their model by training it on newer data to improve its accuracy continuously. This is essential for addressing the dynamic nature of food waste patterns and sales trends. Moreover, SageMaker provides tools for model monitoring and retraining, which will enable ongoing optimization of the model.
- Scenario Usage: Ideal for developing and iterating ML models in an operational setting where continuous improvement is necessary, such as demand forecasting or inventory optimization.
B) Use Amazon Personalize and iterate with historical data:
- Description: Amazon Personalize is a machine learning service that helps businesses build recommendation systems. It is designed to personalize content, products, and services for users based on their behaviors.
- Reasoning: While Amazon Personalize is effective for recommendation systems (such as recommending products to customers), it’s not directly focused on solving the problem of reducing food waste or optimizing sales based on inventory and demand patterns. This makes it less suitable for the specific use case of decreasing food waste and improving sales revenue, which requires more control over demand forecasting and operational efficiency.
- Scenario Usage: Best used for personalization and recommendation tasks, like suggesting products o...
Author: MysticJaguar44 · Last updated Jul 19, 2026
A company has developed an ML model to predict real estate sale prices. The company wants to deploy the model to make predictions without managing se...
To meet the company's requirements of deploying an ML model to predict real estate sale prices without managing servers or infrastructure, we need to analyze the best solution considering key factors such as ease of deployment, cost, scalability, management overhead, and time efficiency.
A) Deploy the model on an Amazon EC2 instance
- Benefits: Amazon EC2 allows full control over the server environment and can be used to deploy models. It provides flexibility in terms of resources and configuration.
- Drawback: This solution requires manual management of the server, including setup, scaling, and maintenance of the infrastructure. It contradicts the requirement of not managing servers or infrastructure. Additionally, it might involve high overhead in terms of monitoring, patching, and scaling.
- Use Case: EC2 would be appropriate if the company requires full control over the infrastructure and has the resources to manage it. However, this is not ideal when the goal is to avoid server management.
B) Deploy the model on an Amazon Elastic Kubernetes Service (Amazon EKS) cluster
- Benefits: Amazon EKS provides a managed Kubernetes service that helps in deploying containerized applications. It allows for scaling and high availability of applications.
- Drawback: While EKS is a managed service, it still involves managing clusters, containers, and configurations. This might be overkill for a machine learning model deployment and would require more effort and expertise in Kubernetes. It is not a serverless solution.
- Use Case: EKS is best suited for large-scale, containerized applications that require sophisticated orchestration. It’s not ideal when you’re looking to deploy a model with minimal infrastructure management.
C) Deploy the model by using Amazon CloudFront with an Amazon S3 integration
- Benefits: Amazon CloudFront is a content delivery network (CDN)...
Author: Sophia Clark · Last updated Jul 19, 2026
A company wants to develop an AI application to help its employees check open customer claims, identify details for a specific claim, and access...
To address the requirements for the AI application, the solution must be able to help employees check open customer claims, identify details about specific claims, and access documents related to those claims. Let's analyze each option to determine which one meets the requirements with the least operational overhead and most effective use of services.
A) Use Agents for Amazon Bedrock with Amazon Fraud Detector to build the application
Amazon Fraud Detector is designed specifically for detecting fraudulent activity, such as identifying potentially fraudulent claims or transactions. However, this doesn't directly address the need to check open claims, identify details, and access documents. This service is more focused on fraud detection rather than claim management and information retrieval.
Reason for rejection: This solution is not a good fit because Amazon Fraud Detector focuses on fraud detection, not on managing or retrieving customer claims, details, and documents.
B) Use Agents for Amazon Bedrock with Amazon Bedrock knowledge bases to build the application
Amazon Bedrock is a service for creating and running AI agents. Using Bedrock with knowledge bases would allow you to create an AI-powered agent that could interact with data and answer queries related to customer claims. The knowledge base can store structured or unstructured information, making it ideal for identifying claim details and accessing related documents. This solution provides a natural language interface for employees to query claims and retrieve relevant information, reducing the need for manual searching.
Reason for selection: This option is the most suitable for the requirements because it allows you to build an AI-powered agent with integrated knowledge bases. It aligns well with the goal of checking open claims, identifying specific claim details, and accessing relevant documents, all while offering a low-effort solution for employees.
C) Use Amazon Personalize with Amazon Bedrock knowledge bases to build the application
Amazon Personalize is a service primarily used for creating personalized recommendations, such as product recommendations based on user preferences. While this service could potentially be used to tailor experiences, it doesn't directly address the need t...
Author: Leo · Last updated Jul 19, 2026
A manufacturing company uses AI to inspect products and find any damages or defects.Which type of AI...
To determine the correct type of AI application the manufacturing company is using, let's analyze the key information from the question: the company is using AI to inspect products and find damages or defects. This clearly indicates that the AI system is analyzing visual data (likely images or videos) to detect issues with the products.
A) Recommendation system
- Purpose: A recommendation system is designed to suggest items to users based on their preferences or behaviors, often used in e-commerce, streaming platforms, or content platforms.
- Reason for rejection: This option does not align with the use case described. The company is focused on inspecting products for damages, which is not related to suggesting items based on user behavior.
- Scenario where it can be used: A recommendation system would be relevant if the company were recommending products to customers, but not for detecting defects in products.
B) Natural language processing (NLP)
- Purpose: NLP is used for processing and understanding human language. It enables applications like chatbots, sentiment analysis, and text translation.
- Reason for rejection: NLP is not relevant to the task of inspecting products for damages or defects, as it deals with text and language, not images or visual data.
- Scenario where it can be used: NLP would be applicable if the company were processing customer feedback or support tickets regarding product issues, but not for the visual inspection of products.
C) Computer vision
- Purpose: Computer vision is a field of AI focused on enabling machines to interpret and understand visual d...
Author: Zara1234 · Last updated Jul 19, 2026
A company wants to create an ML model to predict customer satisfaction. The company needs fully automated model t...
To address the company’s need for fully automated model tuning for predicting customer satisfaction, let’s evaluate the given AWS services based on their capabilities, effort, time, cost, and suitability for the task.
A) Amazon Personalize
- Purpose: Amazon Personalize is a fully managed service that provides recommendations and personalization for customer-facing applications, using machine learning.
- Reason for rejection: While Personalize can be used for personalized recommendations, it is not specifically designed for predicting customer satisfaction through general ML modeling. It focuses on recommendation and personalization tasks, not broad machine learning use cases like customer satisfaction prediction.
- Scenario where it can be used: Amazon Personalize is great for creating personalized experiences, such as recommending products or content based on customer behavior, but it's not suitable for general-purpose ML model tuning or prediction tasks.
B) Amazon SageMaker
- Purpose: Amazon SageMaker is a fully managed service that covers the entire machine learning lifecycle, including data preparation, model building, training, deployment, and monitoring. It provides features like automated model tuning through automatic model tuning (Hyperparameter optimization or HPO) to help find the best model configuration.
- Reason for selection: SageMaker provides end-to-end capabilities for building and deploying ML models, and importantly, it offers automatic model tuning (Hyperparameter optimization), which directly addresses the requirement for fully automated model tuning. This makes SageMaker the best choice for creating a model to predict customer satisfaction while automating the tuning process.
- Scenario where it can be used: Amazon SageMaker is ideal for any type of custom machine learni...
Author: Harper · Last updated Jul 19, 2026
Which technique can a company use to lower bias and toxicity in generative AI applications during th...
In this scenario, a company wants to lower bias and toxicity in generative AI applications during the post-processing phase of the ML lifecycle. Post-processing generally refers to refining the output after the model generates results, and it's crucial to ensure that the output is fair, unbiased, and free of harmful content.
Option Analysis:
A) Human-in-the-loop:
- Description: Human-in-the-loop (HITL) refers to the practice of involving human judgment during various stages of the ML lifecycle, including post-processing. In this case, human reviewers can assess the outputs generated by the AI and apply corrections to remove biased or toxic content before it is finalized or deployed.
- Reasoning: Human-in-the-loop is highly effective for lowering bias and toxicity because it allows human experts to intervene in the decision-making process to identify and correct issues that may not have been caught by automated processes. This can include reviewing generative outputs for harmful stereotypes, offensive language, or biased assumptions. Given the complexity of detecting subtle biases or toxic content, human oversight is invaluable in ensuring high-quality outputs.
- Scenario Usage: Ideal for generative AI applications where human oversight is necessary to filter out problematic outputs, such as in content generation for social media, chatbots, or other platforms where accuracy and safety are crucial.
B) Data augmentation:
- Description: Data augmentation is the process of artificially increasing the size of the training dataset by applying transformations (like rotating, cropping, or altering images in computer vision) to the original data. It helps improve model generalization but doesn’t specifically address post-processing or issues like bias and toxicity in the generated output.
- Reasoning: While data augmentation is valuable in training models to generalize better, it doesn't directly affect the post-processing phase or mitigate bias and toxicity in the outputs that the model generates. It’s more focused on improving model performance and robustness durin...
Author: Leah · Last updated Jul 19, 2026
A bank has fine-tuned a large language model (LLM) to expedite the loan approval process. During an external audit of the model, the company discovered that the model was approving loans at a faster pace for a specific de...
To address the issue of bias in the loan approval process identified during the external audit, the most cost-effective solution would be Option A: Include more diverse training data and fine-tune the model again using the new data.
Analysis of each option:
1. A) Include more diverse training data. Fine-tune the model again by using the new data.
- This option directly addresses the issue by introducing more diverse data, which helps mitigate any biases present in the initial training data. Fine-tuning the model with more diverse and representative data is a cost-effective solution because it doesn't require starting from scratch. The bank can use its existing fine-tuned model, and adding more diverse data for further fine-tuning is often quicker and cheaper than re-training a model from the ground up.
- Effort: Low to medium, as it requires adding data and fine-tuning, but avoids the need to retrain the model entirely.
- Time: Medium, as fine-tuning will take time, but not as much as training a model from scratch.
- Cost: Low to medium, as it mainly involves adding data and retraining, which is a more affordable option than other solutions.
2. B) Use Retrieval Augmented Generation (RAG) with the fine-tuned model.
- RAG is a technique that combines a retrieval-based approach with generative models, which may not directly address the bias in the loan approval process. It is more useful when augmenting a model with external knowledge, not when correcting biases in model behavior. Implementing RAG could be more expensive and time-consuming without effectively solving the core issue of demographic bias in loan approvals.
- Effort: High, as integrating RAG would require significant engineering work.
- Time: High, s...
Author: MysticJaguar44 · Last updated Jul 19, 2026
A company is investigating actions that an IAM role performed. The company must find out when the role last accessed AWS Security Hub and when the role last used the DeleteInsi...
To determine when an IAM role last accessed AWS Security Hub and specifically when it last used the `DeleteInsight` action, we need detailed, time-stamped data about API activity. Let’s evaluate each option based on this requirement.
---
✅ Option B: Use the Access Advisor tab in AWS Identity and Access Management (IAM).
Strengths:
The Access Advisor tab shows the last accessed time for services by a role or user, including Security Hub.
This satisfies the first requirement: "when the role last accessed AWS Security Hub."
Limitations:
However, Access Advisor does not show individual API calls (like `DeleteInsight`)—it only reports at the service level.
---
❌ Option A: Use the checks for the security category in AWS Trusted Advisor.
Why it's rejected:
Trusted Advisor provides best practices and security checks, but does not track IAM role activity or API usage.
It cannot show when a specific action like `DeleteInsight` was used or when Security Hub was accessed.
Use case: Trusted Advisor is better for compliance, cost optimization, and identifying unused resources—not tracking API usage.
---
❌ Option C: Use AWS IAM to generate a credential...
Author: Carlos Garcia · Last updated Jul 19, 2026
A security engineer has created an Amazon GuardDuty detector in several AWS accounts. The accounts are in an organization in AWS Organizations. The security engineer needs centralized visibility of t...
To meet the requirement of centralized visibility of Amazon GuardDuty findings across multiple AWS accounts in an AWS Organization, we need a solution that:
Aggregates GuardDuty findings from multiple accounts
Provides centralized management and visibility
Is designed specifically for security findings, not just logs or metrics
Let’s evaluate the options one by one using key factors like centralization, integration, scalability, purpose-fit, and automation.
---
✅ Option C) Configure AWS Security Hub integrations
Why it's selected:
Purpose-built for security: AWS Security Hub is designed to centralize and aggregate security findings across AWS services like GuardDuty, Macie, and Inspector.
Organizational integration: Supports integration with AWS Organizations, so you can designate a Security Hub administrator account to collect findings from all member accounts.
Built-in GuardDuty integration: GuardDuty findings are automatically sent to Security Hub once integration is enabled.
Centralized visibility: Provides a unified dashboard and standardized findings (AWS Security Finding Format).
Automation and custom actions: Enables response workflows, insights, and integrations with ticketing and SOAR tools.
Scenario where this applies best:
You need security-focused visibility across all accounts in an AWS Organization.
You are using or plan to use multiple AWS security services (GuardDuty, Macie, Inspector).
You want to correlate findings and prioritize risks in one place.
---
❌ Option A) Configure Amazon CloudWatch Logs Insights
Why it’s rejected:
CloudWatch Logs Insights is used to query log data, not security findings.
GuardDuty findings are not automatically delivered to CloudWatch Logs unless you build a cu...
Author: James · Last updated Jul 19, 2026
A company uses Amazon Cognito for external user authentication for a web application. External users report that they can no longer log in to the application.
What is the F...
To determine the correct FIRST step in troubleshooting external users’ inability to log in via Amazon Cognito, we must consider the scope of the problem, likely root causes, and impact of the action. Let’s analyze each option:
---
🔍 Option A: Review AWS CloudTrail logs to identify authentication errors that relate to Cognito users.
Pros:
CloudTrail records AWS API calls and events, including those made by or affecting Cognito.
It helps identify what went wrong and when, e.g., failed login attempts, configuration changes, or access denials.
It provides an audit trail, which is essential in a security-focused approach.
Cons:
Might take longer to sift through logs without knowing what configuration changed.
Best used after narrowing down the issue (e.g., once we suspect a certain role, policy, or setting is broken).
Use when: You suspect an issue based on observed user behavior and want historical insight into what actions were taken recently.
---
🔐 Option B: Use AWS IAM Access Analyzer to delete all unused IAM roles and users.
Pros:
Helps identify unused roles and possible security risks.
Cons:
Not relevant to login failures.
Deleting IAM roles/users unrelated to the problem could exacerbate issues or cause new outages.
IAM Access Analyzer doesn’t help with Cognito-specific configuration or authentication issues directly.
Use when: You’re performing a security audit, not when investigating a live login issue.
---
⚙️ Option C: Review any recent changes in Cognito configura...
Author: Julian · Last updated Jul 19, 2026
A company is running its application on AWS. Malicious users exploited a recent promotion event and created many fake accounts.
The application currently uses Amazon CloudFront in front of an Amazon API Gateway API. AWS Lambda functions serve the different API endpoints. The GET registration endpoint is behind the path of /store/registration. The URI for submission of the new account details is at /store/newaccount.
A...
To prevent malicious users from exploiting promotions by creating fake accounts via your AWS-hosted application, we must focus on rate limiting, bot protection, and targeted mitigation at API endpoints. Let's analyze each option.
---
✅ A) Create an AWS WAF web ACL. Add the AWSManagedRulesACFPRuleSet rule group to the web ACL. Associate the web ACL with the CloudFront distribution.
Selected – This option is effective.
Why it works: The `AWSManagedRulesACFPRuleSet` (Account Creation Fraud Prevention) is specifically designed to detect and mitigate suspicious or fraudulent account creation attempts, such as bots or scripts registering fake accounts.
Key scenario: Protects endpoints like `/store/registration` and `/store/newaccount` from automated abuse.
Why it’s good: It is plug-and-play with best-practice rules, tuned for account fraud scenarios.
---
✅ B) Create an AWS WAF web ACL. Add a rate limit rule to the web ACL. Include a RateBasedStatement entry that has a SearchString value that points to /store/registration.
Selected – This is a strong complement to Option A.
Why it works: Rate-based rules in AWS WAF allow you to throttle excessive requests from the same IP, preventing rapid automated abuse (e.g., bots spamming the `/store/registration` or `/store/newaccount` endpoints).
SearchString use: You can scope the rule to trigger only when certain URI paths are involved (like `/store/registration`), so normal traffic elsewhere isn't blocked.
Key scenario: Useful during promotions, when high but abnormal traffic patterns appear.
---
❌ C) Specify /store/registration as the registration page path. Specify /store/newaccount as the account creation path....
Author: RadiantJaguar56 · Last updated Jul 19, 2026
A company is investigating an increase in its AWS monthly bill. The company discovers that bad actors compromised some Amazon EC2 instances and served webpages for a large email phishing campaign.
A security engineer must implement a solution to monitor for cost increases in the f...
To determine the earliest detection of cost increases due to malicious activity in AWS, we must evaluate each option based on timeliness, automation, relevance to billing anomalies, and operational overhead.
---
A) Amazon EventBridge + AWS Lambda for usage report analysis
Pros:
Fully customizable solution.
Can be scheduled to run hourly for reasonably fast detection.
You can tailor anomaly detection to business-specific needs.
Cons:
AWS usage reports (via Data Exports) can be delayed by several hours, so you’re not getting truly real-time data.
High operational complexity—you have to implement and maintain parsing logic, anomaly detection logic, notification mechanism, and data handling for every AWS service.
Cost anomalies may not be visible in usage reports until after the billing cycle has partially completed.
Use case: This is better for custom business logic or cost attribution analysis, not for earliest detection.
---
B) AWS Cost Anomaly Detection with alerts
Pros:
Purpose-built for early cost anomaly detection.
Uses machine learning models to analyze usage patterns and detect anomalies quickly after they occur.
Automatically notifies via SNS when thresholds are exceeded.
Requires minimal setup and no maintenance.
Cons:
...
Author: Andrew · Last updated Jul 19, 2026
A company's network security policy requires encryption for all data in transit. The company must encrypt data that is sent between Amazon EC2 instances and Amazon Elastic Blo...
To meet the requirement of encrypting data in transit between Amazon EC2 instances and Amazon EBS volumes, we need to examine what is possible and relevant within AWS.
Key Requirement:
> Encrypt data in transit between EC2 and EBS.
---
Option Analysis:
---
A) Configure Amazon EC2 to enable encryption in the EC2 network interface properties.
❌ Rejected.
Reason: There is no such setting in EC2 network interface properties to enable encryption for data in transit to EBS. EC2 network interfaces (ENIs) can be used to manage networking, but not encryption of data to EBS volumes.
When applicable: Not applicable in this context.
---
B) Configure Amazon EBS to enable volume encryption with AWS Key Management Service (AWS KMS) for data at rest.
❌ Rejected.
Reason: This enables encryption at rest, not in transit. It uses AWS KMS and protects stored data, but does not meet the requirement to encrypt data in transit between EC2 and EBS.
When applicable: Use this for data at rest encryption compliance.
---
C) Configure Amazon EBS to enable TLS encryption in the volume configuration properties.
❌ Rejected.
Reason: There is no such option in Amazon EBS configuration to enable TLS encryption. EBS volumes do not have TLS properties for encryption in transit configurable by the user.
When applicable: Not a valid option; AWS does not offer manual TLS ...
Author: Noah Williams · Last updated Jul 19, 2026
A company runs a web application on a fleet of Amazon EC2 instances that are in an Auto Scaling group. The EC2 instances are in the same VPC subnet as other workloads.
A security engineer deploys an Amazon GuardDuty detector in the same AWS Region as the EC2 instances. The security engineer also sets up an AWS Security Hub integration with GuardDuty.
The security engineer needs to implement an automated solution to detect and appropriately respond to anomalous traffic patterns for t...
To determine the best solution, we must align the options with AWS best practices for incident response, minimal disruption, and automated mitigation. Here's a breakdown of the options and key factors in choosing the correct one:
---
✅ Option B:
Create an Amazon EventBridge rule that invokes an AWS Lambda function when GuardDuty detects anomalous traffic. Program the Lambda function to disassociate the identified instance from the Auto Scaling group and to isolate the instance by using a new restricted security group.
Why it's selected:
Automated detection and response using GuardDuty + EventBridge + Lambda.
Follows AWS security incident response best practices:
Isolating the compromised instance (change to restricted security group).
Avoiding disruption to the rest of the application by not changing the entire subnet, Auto Scaling group, or VPC behavior.
Disassociation from Auto Scaling allows analysis without interference from scaling policies.
Minimizes disruption: Removes only the suspicious instance instead of modifying broader infrastructure like subnet NACLs.
---
❌ Option A:
Disable the EC2 instance profile access keys via Lambda.
Why it's rejected:
Only addresses IAM credential compromise, not network behavior or traffic anomalies.
GuardDuty "Behavior\:EC2/TrafficVolumeUnusual" is related to traffic patterns, not access keys mis...
Author: StarryEagle42 · Last updated Jul 19, 2026
A company has an application that needs to read objects from an Amazon S3 bucket. The company configures an IAM policy and attaches the policy to an IAM role that the application uses. When the application tries to read objects from the S3 bucket, the application receives AccessDenied errors.
A security engineer must...
To resolve the `AccessDenied` errors when an application tries to read from an Amazon S3 bucket without decreasing security, let's analyze the options step by step, focusing on key AWS factors such as IAM policy scope, role assumption, resource policy requirements, region independence, and access control mechanisms.
---
Option A: Attach a resource policy to the S3 bucket to grant read access to the role.
Use case: This is appropriate if the bucket policy is missing or incorrectly configured to allow access to the IAM role.
Strength: Explicitly grants access to a specific IAM role, preserving security boundaries.
Limitation: Might be unnecessary if the IAM policy already grants permission and the issue lies elsewhere (like incorrect role assumption).
Security consideration: Adding resource policies must be done carefully to avoid over-permissioning or accidental public access.
✅ Use this when: The IAM role has proper policies, but S3's bucket policy is blocking access.
---
Option B: Launch a new deployment of the application in a different AWS Region. Attach the role to the application.
Use case: This doesn't address the core access issue. S3 is a global service, but bucket data is region-specific. Changing regions won’t help unless the bucket is replicated.
Limitation: Introducing regional differences adds unnecessary complexity.
Security concern: Unnecessary redeployments could introduce operational and compliance risks.
❌ Reject this – it doesn't solve the permissions issue.
---
Option C: Review the IAM policy by using AWS IAM Access Analyzer to ensure that the policy grants the right permissions. Validate that the application is assuming the role correctly.
...
Author: Aarav · Last updated Jul 19, 2026
A security engineer is designing a solution that will provide end-to-end encryption between clients and Docker containers running in Amazon Elastic Container Service (Amazon ECS). This solution will also handle vol...
To determine the most scalable and lowest latency solution for end-to-end encryption between clients and ECS containers that also handles volatile traffic patterns, let's evaluate each option:
---
🔹 Option A: Configure a Network Load Balancer (NLB) to terminate the TLS traffic and then re-encrypt the traffic to the containers.
Pros:
NLBs are highly scalable and optimized for volatile traffic patterns (millions of requests per second).
Can handle sudden bursts and very high throughput with low latency.
Cons:
TLS termination at NLB breaks true end-to-end encryption.
Re-encryption adds complexity and potential latency.
Security concern: decrypted data exists temporarily on the load balancer.
> Use case: Suitable if inspection, logging, or WAF needs to occur at the LB layer, but not true E2E encryption.
---
🔹 Option B: Configure an Application Load Balancer (ALB) to terminate the TLS traffic and then re-encrypt the traffic to the containers.
Pros:
ALB supports HTTP/HTTPS and is suitable for application-layer routing.
Termination and re-encryption possible with ALB + service mesh (e.g., App Mesh).
Cons:
Higher latency than NLB due to application-layer processing.
Not as well-suited for volatile or high-volume traffic as NLB.
Still breaks end-to-end encryption unless a custom re-encryption system is in place.
> Use case: Good for HTTP routing and session-based applications, but not ideal for TLS passthrough or low-latency needs.
---
🔹 Option C: Configure a Network Load Balancer with a TCP listener to pass through TLS traffic to the containers.
Pros:
True end-to-end encryption: TLS is terminated in...
Author: Ella · Last updated Jul 19, 2026
A company has a large fleet of Amazon Linux 2 Amazon EC2 instances that run an application. The application processes sensitive data and has the following compliance requirements:
* No remote access management ports to the EC2 instances can be exposed internally or externally.
* All remote session activity must be recorded in an audit log.
* All remote access to the EC2 instances must be authenticated and authorized by AWS IAM Identity Center.
The company's ...
To determine the correct solution, let's break down the compliance requirements and analyze each option based on how well it meets them.
---
Compliance Requirements:
1. No remote access management ports (e.g., SSH/RDP) can be exposed internally or externally.
2. All remote session activity must be recorded in an audit log.
3. All remote access must be authenticated and authorized by AWS IAM Identity Center.
---
Option Analysis:
---
A) EC2 Serial Console with IAM Policy
Pros:
EC2 Serial Console access can be controlled via IAM.
No need to open SSH or RDP ports.
Cons:
The serial console provides low-level access (e.g., kernel boot issues), not standard shell access.
Limited to Nitro-based instances and certain troubleshooting scenarios, not ideal for general DevOps tasks.
Does not integrate with IAM Identity Center for authentication/authorization.
Audit logging is limited compared to Systems Manager Session Manager.
> ✅ Use case: Deep kernel-level or boot troubleshooting, not standard remote access.
---
B) EC2 Instance Connect with Security Group and Console Access
Pros:
Provides browser-based SSH access without exposing permanent SSH keys.
Cons:
Requires port 22 (SSH) to be open, at least to internal IPs, which violates the “no remote access port” requirement.
IAM policies can grant access, but does not natively support IAM Identity Center authorization.
No built-in session recording or logging unles...
Author: Sofia · Last updated Jul 19, 2026
A company is using Amazon SageMaker AI to create a classification model to categorize the company's sales performance for each month of the previous 20 years on a scale from 1 to 5. The dataset includes fields for month, sales region, regional aggregate sales, and the number of stores in each sales region. The company notices that during two months of every year, the aggregate sales values are unexpectedly high. The company performs one-hot encoding on all non-numerical features in the training and validation datasets. The company uses the training dataset to train the classif...
Let’s carefully analyze this AWS exam-style scenario:
---
✅ Correct Option
- D) Perform normalization on the aggregate sales feature for each sales region
- Why selected: The dataset shows unexpectedly high aggregate sales values during two months every year, which are outliers relative to normal sales patterns. If you normalize globally, regions with naturally higher sales will dominate the scale, skewing classification. By normalizing per region, you account for differences in scale across regions, ensuring fair comparison and improving model accuracy.
- Scenario use: A company with regions of vastly different store counts and sales volumes can normalize within each region to prevent bias toward larger regions.
---
❌ Incorrect Options
- A) Remove records with outliers
- Why rejected: Outliers here are systematic (two months every year), not random errors. Rem...
Author: Siddharth · Last updated Jul 7, 2026
A company is developing an ML model by using Amazon SageMaker AI. The company must monitor bias in the model and must display the results on a dashboard. An ML engineer creates a bias monitoring ...
The correct answer is:
✅ B) Capture Amazon CloudWatch metrics from SageMaker Clarify.
Step 1: Identify the key requirements
The question gives two important requirements:
1. Monitor bias in an ML model.
2. Display the results on a dashboard.
These keywords immediately point toward two AWS services:
Bias monitoring → Amazon SageMaker Clarify
Dashboard/metrics visualization → Amazon CloudWatch
---
Why Option B is Correct
Amazon SageMaker Clarify is the AWS service that detects and monitors:
Pre-training bias
Post-training bias
Feature attribution
Explainability
When a Clarify bias monitoring job runs (typically through SageMaker Model Monitor), it publishes metrics to Amazon CloudWatch.
CloudWatch metrics can then be visualized using:
CloudWatch Dashboards
CloudWatch Alarms
Amazon Managed Grafana (through CloudWatch)
This exactly satisfies:
> Capture bias metrics → Display on dashboard
Key reasoning
| Requirement | AWS Service |
| ------------- | -------------------- |
| Detect bias | SageMaker Clarify |
| Store metrics | CloudWatch |
| Dashboard | CloudWatch Dashboard |
This is a direct service integration.
---
Why the other options are wrong
A) Capture AWS CloudTrail metrics from SageMaker Clarify.
Why rejected
The keyword here is CloudTrail.
CloudTrail records:
API calls
User activity
Who invoked an operation
When an operation happened
Example:
CreateTrainingJob
CreateModel
StartMonitoringSchedule
CloudTrail does NOT collect ML bias metrics.
It only logs management events.
Key factor
CloudTrail answers:
> Who did what?
Not
> How biased is my model?
When CloudTrail is used
Use CloudTrail when you need:
Security auditing
Compliance
API tracking
Investigating who deleted a model
Governance
Never use CloudTrail for monitoring model performance or bias.
---
C) Capture SageMaker Model Monitor metrics from Amazon EventBridge.
Why rejected
EventBridge is an event routing service.
It receives events like:
Monitoring job completed
Training job failed
Endpoint...
Author: Emma · Last updated Jul 7, 2026
An ML engineer is building a model to predict house and apartment prices. The model uses three features: Square Meters, Price, and Age of Building. The dataset has 10,000 data rows. The data includes data points for one large mansion and one extremely small apartment.
The ML engineer must perform preprocessing on the datas...
The correct answer is:
Step 1: Understand the problem
The model predicts house/apartment prices using:
Square Meters (Numerical)
Price (Target)
Age of Building (Numerical)
Dataset:
10,000 rows
One very large mansion
One extremely small apartment
Requirement:
> Ensure accurate predictions for the typical house or apartment.
The key phrase is typical house or apartment. The mansion and tiny apartment are outliers.
---
Analyze each option
A) Remove the outliers and perform a log transformation on the Square Meters variable. ✅
Why it is correct
There are two separate preprocessing steps here.
1. Remove outliers
The dataset has only two extreme observations out of 10,000.
These values are not representative of normal houses.
Outliers can:
pull regression lines
distort feature distributions
increase prediction error for normal observations
Since the question explicitly wants good predictions for typical houses, removing these outliers is appropriate.
---
2. Log transformation
Square Meters is a numerical feature.
House sizes are often right-skewed.
Example:
```
40
55
70
85
110
150
220
450
1000
```
Most houses are small, with a few very large ones.
A log transformation compresses large values.
Example:
```
log(40)=3.69
log(1000)=6.91
```
Instead of:
```
40 → 1000 (difference = 960)
```
It becomes
```
3.69 → 6.91 (difference = 3.22)
```
Benefits:
reduces skewness
makes relationships more linear
improves many regression models
This is a common preprocessing step for numerical variables with long tails.
Key factors
✔ Numerical feature
✔ Right-skewed distribution
✔ Few extreme values
✔ Improve regression performance
Therefore this is the best answer.
---
B) Keep the outliers and perform normalization on the Square Meters variable.
Why it is wrong
Normalization scales values into a fixed range, such as:
```
0 to 1
```
Formula:
```
(x-min)/(max-min)
```
Suppose
Normal houses:
```
60–250 m²
```
One mansion:
```
5000 m²
```
After normalization:
```
60 → 0.00
120 → 0.012
180 → 0.024
250 → 0.038
5000→1.00
```
Almost all normal houses get compressed into a tiny interval.
The outlier still dominates the scaling.
Normalization does not remove the effect of outliers.
Since the requirement is to improve predictions for typical houses, keeping outliers is a poor choice.
---
When normalization is useful
Use normalization when:
features have different scales
distance-based algorithms are used
KNN
K-Me...
Author: SilverBear · Last updated Jul 7, 2026
A company is preparing data to train a new ML model on Amazon SageMaker AI. The data has not been used before for ML training. The data includes duplicates and is missing some values.
The company needs to increase the data qu...
The correct answer is B) Use SageMaker Data Wrangler to create data quality rules. Use SageMaker Clarify to detect bias.
Let's analyze the requirements first.
Requirements from the question
The dataset:
✅ Has never been used before (preprocessing stage)
✅ Contains duplicates
✅ Has missing values
✅ Needs higher data quality
✅ Needs statistical bias detection
There are two independent tasks:
1. Improve data quality (clean the dataset)
2. Detect bias in the dataset
---
Understanding each AWS service
1. SageMaker Data Wrangler
Purpose: Data preparation and feature engineering before training.
It is used to:
Remove duplicates
Handle missing values
Detect anomalies
Transform columns
Encode categorical features
Normalize data
Create data quality checks
Think of it as the ETL/Data Cleaning service inside SageMaker.
Key factor
> If the question talks about cleaning raw datasets before ML training, think Data Wrangler.
Typical exam keywords:
Missing values
Duplicate rows
Feature engineering
Data preprocessing
Data transformations
Data quality
---
2. SageMaker Clarify
Purpose:
Detect bias
Explain model predictions
Clarify works:
Before training → detects dataset bias
After training → detects model bias
Explains feature importance (SHAP values)
The question specifically asks:
> detect any statistical bias in the data
This is exactly Clarify.
Typical keywords:
Bias detection
Fairness
Explainability
SHAP
Feature importance
---
Option Analysis
A) Use SageMaker Clarify to create data quality rules. Use SageMaker Model Monitor to detect bias.
First statement
> SageMaker Clarify to create data quality rules
❌ Incorrect.
Clarify is not a data cleaning service.
It does not:
Remove duplicates
Fill missing values
Improve data quality
Its job is:
Bias detection
Explainability
---
Second statement
> Model Monitor to detect bias
Mostly incorrect.
Model Monitor monitors:
Data drift
Model quality
Feature drift
Prediction quality
Bias detection is primarily done by Clarify.
Therefore option A is rejected.
---
B) Use SageMaker Data Wrangler to create data quality rules. Use SageMaker Clarify to detect bias.
First statement
Data Wrangler
✔ Correct.
It handles:
Missing values
Duplicate records
Invalid data
Feature engineering
Data quality
---
Second statement
Clarify
✔ Correct.
Detects:
Dataset bias
Model bias
Feature importance
This satisfies both requirements perfectly.
---
C) Use SageMaker Debugger to create data quality rules. Use SageMaker Model Monitor t...
Author: Ming88 · Last updated Jul 7, 2026
A bank needs to use Amazon SageMaker AI to create an ML model to determine which customers qualify for a new product. The bank must use algorithms that SageMaker AI directly supports. The model must be expla...
The key requirements in this question are:
1. Use Amazon SageMaker AI built-in algorithms (directly supported by SageMaker).
2. Determine which customers qualify for a new product → This is a classification problem.
3. The model must be explainable to regulators → Interpretability is a major requirement.
Let's evaluate each option.
---
A) Train the model by using the Object2Vec algorithm.
What is Object2Vec?
Object2Vec is a SageMaker built-in algorithm that learns embeddings (vector representations) of objects.
Typical use cases:
Recommendation systems
Similarity search
Document similarity
Product matching
Why it is rejected
Although it is a SageMaker built-in algorithm, it is not designed for an explainable classification model.
It creates dense vector embeddings, which are difficult to interpret.
Regulators usually require understanding why a customer was accepted or rejected.
When should you use Object2Vec?
Use it when you need:
Similar item recommendations
Similar customer matching
Search ranking
Semantic similarity
Key factor: Good for embeddings, not explainable classification.
❌ Rejected
---
B) Train the model by using the linear learner algorithm.
What is Linear Learner?
Linear Learner is a SageMaker built-in supervised learning algorithm.
It supports:
Binary classification
Multiclass classification
Regression
For this question:
Customer qualifies or does not qualify
Binary classification
Why it is selected
This satisfies every requirement.
Requirement 1: Built-in SageMaker algorithm
✔ Yes
Requirement 2: Classification
✔ Yes
Requirement 3: Explainability
✔ Excellent
Linear models are among the easiest ML models to explain.
You can explain:
Feature coefficients
Positive influence
Negative influence
Relative importance of variables
Example:
Income coefficient = +2.3
Debt coefficient = -1.1
This clearly explains why a customer qualified.
Banks, insurance companies, healthcare, and governments often prefer linear models because regulators can audit them.
When should you use Linear Learner?
Use it for:
Loan approval
Credit scoring
Fraud detection
Customer churn
Risk prediction
Any explainable classification/regression problem
Key factor: Explainable + Classification + Built-in SageMaker.
✅ Correct
---
C) Train a neural network.
What is a Neural Network?
Neural networks are deep learning models.
Examples:
TensorFlow
PyTorch
MXNet
They are excellent for:...
Author: Deepak · Last updated Jul 7, 2026
A company has built, trained, and tuned two new ML models:
* Model A detects if a transaction is fraudulent based on the IP address, location, and user credentials. This model will be accessed every time a transaction occurs.
* Model B forecasts sales totals for the next month based on historical sales data. This model will be accessed one time each month.
The compa...
The correct answer is:
> ✅ C) Host Model A with a real-time endpoint. Use batch transform for Model B.
Let's analyze the requirements first, then evaluate every option.
---
Step 1: Identify the workload characteristics
Model A
Detects fraud during every transaction.
Invoked every time a customer makes a transaction.
Users expect an immediate response before allowing or rejecting the payment.
Low latency is critical.
Key factor:
Online inference
Millisecond/second response required
High request frequency
This is exactly what Real-Time Endpoints are designed for.
---
Model B
Forecasts next month's sales.
Runs only once every month.
No immediate response required.
Works on historical data.
Key factor:
Offline inference
Large dataset
Infrequent execution
Latency doesn't matter
This is exactly what Batch Transform is designed for.
---
Understanding SageMaker hosting options
1. Real-Time Endpoint
Best for
Immediate predictions
Low latency
Online applications
High request volume
Examples
Fraud detection
Recommendation systems
Spam detection
Credit card approval
Chatbot inference
Think:
> User is waiting for the answer.
---
2. Batch Transform
Batch Transform is for offline inference.
Instead of sending one request at a time,
you submit
S3 input data
Model
SageMaker processes everything
Output saved back to S3
Best for
Daily predictions
Weekly predictions
Monthly forecasting
Large datasets
One-time processing
Examples
Sales forecasting
Customer segmentation
Predicting churn for all customers
Scoring millions of records
Think:
> Nobody is waiting for the prediction.
---
3. Asynchronous Endpoint
This is often confusing.
It still exposes an endpoint, but
request accepted immediately
processing happens later
result stored in S3
Designed for
Long-running inference
Large payloads
Processing can take minutes
Examples
Image generation
Video processing
Large NLP models
Huge document analysis
Think:
> Online request, but long processing time.
---
Evaluate each option
Option A
> Host both models in one container behind one real-time endpoint.
Why Model A fits
✔ Fraud detection requires real-time.
---
Why Model B does NOT fit
Sales forecasting
runs once per month
doesn't need endpoint
keeping endpoint running 24/7 wastes money
Real-time endpoints incur ongoing infrastructure costs.
Using them for monthly jobs is inefficient.
❌ Reject
---
Option B
> Host Model A with an asynchronous endpoint.
> Host Model B with a real-time endpoint.
Model A
Fraud detection needs response immed...
Author: Arjun · Last updated Jul 7, 2026
An ML engineer uses A/B testing to dynamically select recommendation models. The models are deployed on Amazon SageMaker AI endpoints. The ML engineer needs to monitor system metrics such as latency, call volume, and HTTP status codes when the ...
The correct answer is:
✅ B) Configure Amazon CloudWatch dashboards with AWS Lambda log processors.
Step 1: Identify the key requirements
The question asks for monitoring:
System metrics
Latency
Call volume
HTTP status codes
Amazon SageMaker endpoints
Least operational overhead
The phrase "system metrics" is the biggest clue.
---
Evaluate each option
A) Use AWS X-Ray tracing to monitor SageMaker AI endpoints.
Why rejected
AWS X-Ray is designed for:
Distributed application tracing
End-to-end request flow
Finding bottlenecks across microservices
It is not primarily a monitoring service for endpoint metrics like request count or HTTP status codes.
Key factors
✅ Good for request tracing
✅ Root cause analysis
❌ Not intended as primary monitoring dashboard
❌ Doesn't directly provide endpoint operational metrics
When would you use X-Ray?
Use X-Ray when you need to answer questions like:
Which microservice is slow?
Where is latency introduced?
How does one request travel across services?
Example:
API Gateway → Lambda → DynamoDB → SageMaker
You want distributed traces.
---
B) Configure Amazon CloudWatch dashboards with AWS Lambda log processors.
Why selected
Amazon SageMaker automatically publishes many endpoint metrics to Amazon CloudWatch, including:
Invocation count
Model latency
Overhead latency
Error metrics
HTTP status-related metrics
CloudWatch dashboards provide visualization.
If additional processing or custom aggregation of logs is needed, Lambda can process CloudWatch Logs automatically.
Key factors
✅ Native AWS monitoring service
✅ Supports latency
✅ Supports request volume
✅ Supports error metrics
✅ Low operational overhead
✅ Integrates directly with SageMaker
This is exactly what AWS recommends.
---
Why "Lambda log processors"?
Even though CloudWatch already provides many metrics, Lambda can:
Process logs
Create custom metrics
Generate alerts
Since it's serverless, operational overhead remains low.
---
When should you use this?
Whenever you need:
Monitoring dashboards
Metrics
Alarms
Endpoint health
Operational monitoring
Examples:
Lambda
SageMaker
ECS
EC2
API Gateway
CloudWatch is almost always the first choice.
---
C) Enable IAM Access Analyzer to track SageMaker AI endpoint metrics.
Why rejected
This option is completely unrelated.
IAM Access Analyzer checks:
Resource access
External access
Security permissions
Cross-account sharing
It has nothing to do with:
Latency
Requests
HTTP status codes
Key factors
✅ Security analysis
✅ IAM policy validation
❌ No monitori...
Author: Maya · Last updated Jul 7, 2026
A company is using Amazon SageMaker AI to build an ML model to predict customer behavior. The company needs to explain the bias in the model to an auditor. The explanation must focus on de...
The correct answer is A) Use SageMaker Clarify to generate a bias report. Send the report to the auditor.
Let's analyze the question like you would in an AWS certification exam.
---
Step 1: Identify the key requirements
The question contains important keywords.
Amazon SageMaker AI
Explain the bias
ML model
Auditor
Focus on demographics data
The biggest keyword is:
> Explain the bias
AWS has a dedicated service for bias detection and explainability:
> Amazon SageMaker Clarify
Whenever you see:
Bias detection
Fairness
Sensitive attributes (gender, race, age, demographics)
Explain predictions
Feature importance
Think:
> SageMaker Clarify
---
Option A
> Use SageMaker Clarify to generate a bias report. Send the report to the auditor.
Why this is correct
SageMaker Clarify is specifically built for:
Detecting bias before training
Detecting bias after training
Measuring fairness
Evaluating demographic bias
Producing explainability reports
Feature attribution (SHAP values)
Since the question specifically mentions:
> focus on demographics data
Clarify can analyze bias against protected demographic groups like:
Gender
Age
Race
Income
Geographic region
Other sensitive attributes
It produces reports that are intended for governance, compliance, and auditing.
Exactly what the auditor wants.
Key exam clue
Whenever you see:
bias
fairness
discrimination
explainability
sensitive attributes
demographics
→ Choose SageMaker Clarify.
✅ Correct
---
Option B
> Use AWS Glue DataBrew to create a job to detect drift in the model's data quality.
Why this is incorrect
Notice the keywords:
Data quality
Drift
These are completely different concepts.
The question is asking about:
> Bias
Data drift means:
Incoming data distribution changes over time.
Example:
Training data:
```
Average age = 35
```
Production data:
```
Average age = 52
```
That is drift.
Bias means:
```
Model predicts loans unfairly against women.
```
These are unrelated.
AWS Glue DataBrew is used for:
Cleaning data
Preparing datasets
Profiling data
Finding missing values
Transformations
It is not an ML fairness tool.
When DataBrew is used
Use DataBrew when you need to:
Clean CSV files
Remove duplicates
Normalize columns
Prepare data before ML
Never for model bias.
❌ Rejected
---
Option C
> Use Amazon Quick Suite (QuickSight) integration with SageMaker AI to generate a bias report.
Why this is incorrect
QuickSight (now Amazon Quick Suite) is a BI and visualization service.
It creates:
Dashboards
Charts
Reports
Business analytics
It does not detect ML bias.
Even if it visualizes Clarify outputs, it is no...
Author: Matthew · Last updated Jul 7, 2026
An ML engineering team is spread across multiple locations. When the lead ML engineer opens an Amazon SageMaker Al notebook, the ML engineer does not see the latest merged notebook made by other team members from a Git repository.
The lead ML...
The correct answer is:
> ✅ A) Run the `!git pull origin master` command.
Key factors in the question
Look for the important clues:
Team is spread across multiple locations → They are collaborating using Git.
Latest merged notebook made by other team members → Changes have already been merged into the remote repository.
Lead engineer does not see the latest notebook → The local SageMaker notebook is outdated.
Must see the latest updates → Need to download the latest changes from the remote repository.
The keyword is "see the latest merged notebook".
---
Option A) Run the `!git pull origin master` command. ✅ (Correct)
Why?
`git pull` fetches the latest commits from the remote repository and merges them into the local working copy.
```
Remote Git Repository
↓
git pull
↓
Local SageMaker Notebook
```
Since other team members have already merged their notebook changes into the remote Git repository, the lead engineer simply needs to pull those updates.
Key reasoning
Downloads latest commits.
Updates local notebook.
Exactly matches the requirement of viewing the newest merged notebook.
When is this used?
Use git pull whenever:
Someone else has pushed changes.
You want your local repository to be up to date.
You begin work for the day.
Before making your own changes (to reduce merge conflicts).
---
Option B) Run the `!git commit` command. ❌
Why rejected?
`git commit` only saves your local changes into your local Git repository.
It does not:
contact the remote repository,
download changes,
show teammates' updates.
Key factor
The question is about receiving others' changes, not saving your own.
When is this used?
Use git commit when:
You have modified files.
You want to record your changes locally before pushing.
Workflow:
```
Edit files
↓
git add
↓
git commit
```
---
Option C) Run the `!git pu...
Author: Michael · Last updated Jul 7, 2026
A recommendation model uses ML and calls an Amazon SageMaker AI endpoint to get recommendations. An ML engineer must ensure that the model stays available during an expecte...
The requirement here is:
A SageMaker endpoint is already in use
There will be an expected increase in traffic
The goal is to ensure availability and performance under load
No mention of changing model type or architecture—only handling scaling demand
So the key concept is scaling inference capacity dynamically to handle traffic spikes.
---
Option A: Configure auto scaling on the SageMaker endpoint
Amazon SageMaker Endpoint Auto Scaling dynamically adjusts the number of instances behind a real-time endpoint based on metrics like:
Invocations per instance
CPU/GPU utilization
Request count per target
This ensures:
The endpoint can scale out during traffic spikes
It maintains availability and low latency
It avoids manual intervention
This is the standard and correct solution for handling variable or increased traffic on SageMaker endpoints.
When it is used:
Traffic spikes or seasonal load increases
Maintaining SLA for real-time inference
Preventing throttling or overloaded instances
✅ Best fit
---
Option B: Create a new SageMaker endpoint
Creating a new endpoint:
Does not inherently solve scaling
Introduces unnecessary duplication
Increases cost and operational overhead
Still requires auto scaling to handle t...
Author: MoonlitPantherX · Last updated Jul 7, 2026
A company wants to use large language models (LLMs) that are supported by Amazon Bedrock to develop a chat interface for the company's internal technical documentation. The company stores the documentation as dozens of text files that are several megabytes in total si...
This is a classic RAG (Retrieval-Augmented Generation) vs fine-tuning vs training-from-scratch decision question, and the key constraints are:
Internal technical documentation (private data)
Stored as many text files (multi-MB total)
Frequently updated
Need a chat interface using Amazon Bedrock models
Most cost-effective solution required
---
Key evaluation factors
1. Cost
Training or fine-tuning models is expensive.
Maintaining retraining pipelines for frequent updates increases cost further.
2. Freshness of data
Documentation is updated often → solution must support easy updates without retraining models
3. Best AWS-native pattern
For knowledge bases + LLM chat: AWS strongly recommends Bedrock Knowledge Bases (RAG)
4. Scalability
Must handle growing documentation without model changes
---
Option analysis
---
❌ A) Create a new LLM and train it on company data
Why it is wrong:
Training a new LLM from scratch is extremely expensive and unnecessary
Requires massive compute, data, and ML expertise
Not supported as a typical Bedrock use case (Bedrock provides foundation models, not custom LLM training from scratch)
Updates to documents would require retraining the entire model
When this is used:
Only large AI companies building foundation models (like Claude, Llama-scale training)
Not for enterprise documentation chat systems
---
❌ B) Use Bedrock guardrails with documentation
Why it is wrong:
Amazon Bedrock Guardrails are for:
Content filtering
Safety policies
PII redaction
They do not store or retrieve knowledge
They cannot ingest documents or provide context-aware answers
When this is used:
To enforce safety rules like:
No toxic content...
Author: Nathan · Last updated Jul 7, 2026
An ML engineer at an insurance company trains a regression model to predict the number of insurance policy sales each month. After training the model, the ML engineer uses Amazon SageMaker AI to deploy the model for inference. The ML engineer wants to monitor the model predictions to detect whether the production data distribut...
The requirement is to detect whether the production data distribution differs from the training data distribution due to changing customer behavior. In Amazon SageMaker Model Monitor, this specifically corresponds to data drift (data quality drift / feature distribution drift).
✅ Correct Option: A) Determine whether there is drift in the data quality
---
Why Option A is correct (Key reasoning)
In Amazon SageMaker Model Monitor, Data Quality monitoring compares:
Baseline training dataset statistics
Real-time inference data statistics
It detects:
Feature distribution shifts (data drift)
Schema violations (missing values, type changes)
This directly matches the requirement: “production data distribution differs from training data distribution.”
No labels are required, making it ideal for detecting behavioral changes early.
---
Why other options are incorrect
❌ B) Drift in model quality
Requires ground truth labels
Measures model performance d...
Author: GlowingTiger · Last updated Jul 7, 2026
A company is performing A/B testing on a model that provides product recommendations. The company has deployed two versions of the model and is showing each version of the model to 50% of users randomly.
...
The goal of the experiment is to determine whether users actually act on recommendations generated by two different versions of a model. This is an online A/B testing scenario, so the correct evaluation metric must directly reflect user behavior in production, not offline model properties or system performance.
Correct option: A) The conversion rates between the two versions of the model
Why A is correct:
Conversion rate measures the percentage of users who take a desired action (e.g., click, purchase, add-to-cart) after seeing a recommendation.
In an A/B test, the key objective is to compare user behavior between control (Model A) and treatment (Model B).
It directly captures the causal impact of each model version on real user engagement and business outcomes.
This is the standard North Star metric for recommendation effectiveness in production experiments.
---
Why other options are rejected:
B) The number of recommendations provided to each user
This measures system output volume, not user response.
A model could show many recommendations but still have poor en...
Author: Aditya · Last updated Jul 7, 2026
A company is developing an application that reads animal descriptions from user prompts and generates images based on the information from the prompts. The application reads a message from an Amazon Simple Queue Service (Amazon SQS) queue. Then the application uses Amazon Titan Image Generator on Amazon Bedrock to generate an image based on the information in the mess...
The application has two distinct responsibilities:
1. Consume messages from Amazon SQS
2. Call Amazon Bedrock (Titan Image Generator) to generate images
3. Delete processed messages from the queue
So the IAM role must include permissions for both SQS message processing and Bedrock model invocation.
---
Correct Options
A) Allow the `bedrock:InvokeModel` action for the Amazon Titan Image Generator resource.
This is required because the application is calling the generative AI model hosted on Amazon Web Services via Amazon Bedrock.
Key factor: Generating images requires model invocation, not data retrieval.
`bedrock:InvokeModel` is the core permission used to run inference on foundation models like Titan Image Generator.
Without this, the application cannot generate images at all.
Used when: Any application calls Bedrock foundation models for inference (text, image, embeddings, etc.).
---
C) Allow the `sqs:ReceiveMessage` action and the `sqs:DeleteMessage` action for the SQS queue resource.
These permissions are required for proper SQS message consumption using Amazon Simple Queue Service.
Key factors:
`sqs:ReceiveMessage` → reads messages from the queue
`sqs:DeleteMessage` → removes processed messages to prevent reprocessing
Why both matter:
Without `ReceiveMessage`, the app cannot fetch jobs.
Without `DeleteMessage`, messages would be processed repeatedly (at-least-on...
Author: Daniel · Last updated Jul 7, 2026
A company is creating an ML model to identify defects in a product. The company has gathered a dataset and has stored the dataset in TIFF format in Amazon S3. The dataset contains 200 images in which the most common defects are visible. The dataset also contains 1,800 images in which there is no defect visible.
An ML engineer trains the model and notices poor...
The key issue here is class imbalance: 200 defect images (minority class) vs 1,800 non-defect images (majority class). This causes the model to bias toward predicting the majority class, leading to poor performance on defect detection.
---
Correct approach: C) Oversample the 200 images in which the most common defects are visible
Why C is correct
Oversampling the minority class is a standard technique to fix class imbalance. In this case, the 200 defect images should be:
Duplicated or synthetically augmented (e.g., rotation, flipping, brightness changes)
Balanced against the 1,800 non-defect images
Optionally combined with class weighting during training
In AWS workflows, this is commonly used when training models with services like Amazon Rekognition Custom Labels, where balanced datasets improve precision/recall for minority classes.
---
Why other options are incorrect
A) Use a few hundred images and Amazon Rekognition Custom Labels to train a new model
❌ Incorrect because it does not address class imbalance
Reducing dataset size can actually worsen model performance
Amazon Rekognition Custom Labels works best when the dataset is already balanced or properly augmented
Also vague (“few hundred images”) and risks losing valuable majority-class information
When this would be used:
When you are prototyping quickly...
Author: Aria · Last updated Jul 7, 2026
An ML engineer needs to develop an AI assistant by using Retrieval Augmented Generation (RAG) in Amazon Bedrock. The company stores a collection of PDF text files in an Amazon S3 bucket. The ML engineer must create an Amazon Bedrock knowledge base to process t...
To build an Amazon Bedrock Knowledge Base for RAG, the key requirement is to convert PDF text chunks into vector embeddings and store them in a vector database (vector store) for semantic retrieval.
So the critical capability needed here is:
> Document embedding model (not instruction tuning or text generation model)
---
✅ Correct Option: D) Use Cohere Embed Multilingual to perform document embedding
Why this is correct
Amazon Bedrock knowledge bases require an embedding model to:
Convert PDF content into dense vector representations
Store those embeddings in a vector store (e.g., OpenSearch, Aurora PostgreSQL, Pinecone via integration)
Enable semantic search for RAG retrieval
Cohere Embed Multilingual is specifically designed for:
Generating high-quality embeddings
Supporting multilingual text, useful for diverse document corpora
Powering retrieval use cases in RAG systems
This makes it the only valid option for building the knowledge base vector index.
---
❌ Why the other options are incorrect
A) Amazon Titan Text Premier for embedding
Amazon Titan Text Premier is a text generation model, not an embedding model
Cannot produce vector embeddings required for retrieval
Used for: summarization, Q&A generation, content creation
👉 When it is used:
Final response generation in RAG pipelines
...
Author: Chloe · Last updated Jul 7, 2026
An ML engineer is tuning an image classification model that shows poor performance on one of two available classes during prediction. Analysis reveals that the images whose class the model performed poorly on represent an extremely small fraction of the whole traini...
This is a classic class imbalance problem in image classification, where one class is severely underrepresented, leading to poor model performance on that minority class.
Key observations from the scenario
One class has very small representation → class imbalance.
Model performs poorly on that class → likely bias toward majority class.
This is image data, not tabular data.
Goal is to improve performance on the minority class, not just overall correctness.
---
Option Analysis
❌ C and D (SMOTE-based options)
SMOTE (Synthetic Minority Oversampling Technique) is mainly designed for structured/tabular datasets, where feature vectors are well-defined.
It works by interpolating between feature vectors of minority samples.
For raw images, SMOTE is generally not suitable, because:
Pixel-space interpolation creates unrealistic images.
It does not preserve spatial/semantic structure.
In AWS ML practice, SMOTE is rarely used for vision problems.
👉 Therefore:
C (accuracy + SMOTE) → wrong metric + wrong technique
D (F1 + SMOTE) → correct metric idea but still wrong technique for images
---
⚠️ A (Accuracy + image augmentation)
Image augmentation is correct and highly relevant for image imbalance:
rotations, flips, crops, color jitter, etc.
increases diversity of minority class samples.
However, optimizing for accuracy i...
Author: Jack · Last updated Jul 7, 2026
An ML engineer at an entertainment company is refining an ML model to predict audience preferences. As the model learns from data, the ML engineer notices that the model's performance peaks early and then begins to gradually decline. The ML engineer m...
The situation described is a classic case of overfitting during training.
The model improves initially (learning useful patterns), reaches peak performance, and then validation performance starts degrading while training continues to improve. This typically happens when the model begins memorizing noise in the training data instead of learning generalizable patterns.
Correct Option: B) Implement early stopping
Early stopping is a regularization technique where training is halted as soon as performance on a validation set stops improving (or begins to degrade). This directly addresses the issue:
Monitors validation loss/metric during training
Stops training at the point of best generalization
Prevents overfitting from continued training epochs
Commonly used in deep learning workflows on AWS services like SageMaker
---
Why other options are incorrect:
A) Increase the number of layers
Increases model complexity
Makes the model more expressive but more prone ...
Author: NightmareDragon2025 · Last updated Jul 7, 2026
An ML engineer decides to use Amazon SageMaker AI automated model tuning (AMT) for hyperparameter optimization (HPO). The ML engineer requires a tuning strategy that uses regression to slowly and sequentially select the next set of hyperparameters based on previou...
The correct choice is C) Bayesian optimization.
Why C is correct (Bayesian optimization)
Amazon SageMaker AI Automated Model Tuning (AMT) supports Bayesian optimization, which is specifically designed to:
Build a probabilistic regression model (surrogate model) of the objective function (e.g., validation loss vs hyperparameters)
Use results from previous training jobs
Sequentially choose the next hyperparameter set by balancing exploration and exploitation
Perform especially well in small to medium hyperparameter search spaces
Key alignment with the question:
“uses regression” → Bayesian optimization uses a surrogate regression model
“slowly and sequentially select next set” → it is inherently sequential
“based on previous runs” → it learns from prior trials
“small hyperparameter ranges” → it performs best when search space is not extremely large or purely categorical
---
Why other options are incorrect
A) Grid search
Tries all combinations exhaustively
No learning from previous runs
Not sequential or adaptive
Becomes computationally expensive quickly
When it is used:
Very small search spaces
When reproducibility and comp...
Author: VioletCheetah55 · Last updated Jul 7, 2026
An ML engineer is developing a linear regression ML model. The model shows high accuracy on the training dataset but performs poorly on unseen new data.
...
The problem described is a classic case of overfitting: the model performs very well on training data but poorly on unseen data. This means the model has learned noise and patterns specific to the training dataset rather than generalizable patterns.
Key objective
Improve generalization by reducing overfitting.
---
✅ Correct Option: B
B) Apply ML techniques such as cross-validation and regularization. Use Amazon SageMaker Experiments to track and compare different model versions and their performance metrics.
Why this is correct
This option directly addresses the root cause:
Regularization (L1/L2) penalizes overly complex models and reduces coefficient magnitude in linear regression, preventing overfitting.
Cross-validation ensures the model is evaluated on multiple data splits, giving a more reliable estimate of generalization performance.
Amazon SageMaker Experiments helps compare different runs (e.g., with/without regularization, different hyperparameters) and select the best-performing generalizable model.
When this is used
High training accuracy + low validation/test accuracy
Overfitting in regression or classification models
Hyperparameter tuning and model selection phases
---
...
Author: Vikram · Last updated Jul 7, 2026
An ecommerce company trains an ML model to forecast demand for near real-time inventory management based on historical customer activity. The company successfully deploys the trained model to a production Amazon SageMaker AI endpoint. However, the company notices that the model's forecast performance degrades over time. Th...
The key issue in this scenario is model performance degradation over time in production, which typically indicates data drift, concept drift, or changes in customer behavior patterns. The requirement is for a long-term, automated way to detect and respond to this degradation so the model can be retrained appropriately.
Let’s evaluate each option:
---
A) Amazon SageMaker Debugger
Amazon SageMaker Debugger is designed to monitor training jobs, not production inference.
What it does well:
Detects training issues such as exploding gradients, vanishing gradients, and tensor anomalies during model training.
Why it is wrong here:
The problem is happening after deployment, not during training. Debugger does not monitor live endpoint performance or data drift.
Use case:
Debugging and optimizing training phase behavior, not production monitoring.
---
B) AWS X-Ray
AWS X-Ray focuses on request tracing and application performance monitoring.
What it does well:
Tracks latency, service calls, and debugging distributed applications.
Why it is wrong here:
It does not evaluate model accuracy, prediction quality, or data drift. It cannot detect that forecasts are becoming inaccurate.
Use case:
Troubleshooting microservice latency or API performance issues, not ML model degradation.
---
C) Amazon SageMaker Ground Truth
Amazon SageMaker Ground Truth is used to build labeled datasets.
What it does well:
Helps create high-quality labeled training data for supervised learning.
Why it is partially relevant but incorrect as a solution:
It does not provide automated monitoring or detection of degr...
Author: SolarFalcon11 · Last updated Jul 7, 2026
A company is using an ML model to classify motion in videos. The data is stored in MP4 format in Amazon S3. When the company created the model, the company needed 4 months to label all the video frames.
The company needs to retrain the model with an existing training workflow in Amazon SageMake...
The key requirement is to reduce labeling time for video frames while still integrating with an existing Amazon SageMaker AI training workflow. The dataset is already in Amazon S3 (MP4 videos), so the best solution should provide efficient video annotation + automation (pre-labeling or assisted labeling).
---
✅ Correct Answer: A) Use Amazon SageMaker Ground Truth to annotate the video frames
✔ Why A is correct
Amazon SageMaker Ground Truth is AWS’s purpose-built data labeling service, and it is specifically designed to:
Label images, videos, text, and 3D data
Support video frame annotation tools (object tracking, frame-by-frame labeling)
Use labeling workflows with humans + automation
Enable model-assisted labeling (MAL) → pre-labels data using an ML model, reducing human effort
Integrate directly with S3 and SageMaker training pipelines
🔑 Key reasoning factors
The problem is time-consuming manual labeling (4 months)
We need efficiency improvements in labeling, not just model improvement
Ground Truth provides:
Automatic pre-labeling via ML models
Frame interpolation in videos (label propagation across frames)
Active learning (label only most informative samples)
👉 This directly reduces labeling time while keeping SageMaker workflow intact.
---
❌ Why other options are incorrect
B) SageMaker JumpStart with pre-trained models
Amazon SageMaker JumpStart
JumpStart provides pre-trained models and solutions
It is used for:
Fast model prototyping
Transfer learning
...
Author: Ahmed97 · Last updated Jul 7, 2026
A company uses an ML model to recommend videos to users. The model is deployed on Amazon SageMaker AI. The model performed well initially after deployment, but the model's performance has degraded o...
The correct answer is B) Create a baseline from the training dataset. Then create a monitoring job in SageMaker Model Monitor.
Why B is correct
In Amazon SageMaker, model drift detection using SageMaker Model Monitor follows a specific workflow:
1. Create a baseline first
You generate statistics and constraints from the training dataset (or a validation dataset).
This baseline represents the “expected” distribution of data and model behavior at deployment time.
2. Then create a monitoring job (monitoring schedule)
The monitoring job continuously compares incoming real-time inference data against the baseline.
It detects:
Data drift (feature distribution changes)
Concept drift (performance degradation patterns, indirectly)
It produces reports and can trigger alerts when thresholds are violated.
This order is essential because the monitoring job needs a reference baseline to compare against.
Why other options are incorrect
A) Create monitoring job first, then baseline
Incorrect sequence.
A monitoring job cannot function pr...
Author: Akash · Last updated Jul 7, 2026
An ML engineer used Amazon SageMaker Studio to train a neural network. The neural network logs its information into TensorBoard and uses stochastic gradient descent (SGD) as the optimizer.
The ML engineer reviewed training graphs and discovered that the accuracy was not increasing and the loss was decreasing very slowly. The M...
The key observation is: accuracy is not improving and loss is decreasing very slowly. In SGD-based training, this pattern most commonly indicates that the model is learning too slowly, which is typically caused by a learning rate that is too low.
---
Option A: Increase the initial learning rate (Correct)
Increasing the learning rate allows SGD to make larger parameter updates per step, helping the model:
Converge faster
Escape slow plateaus in the loss landscape
Improve accuracy within the same number of training steps (thus not increasing training time)
In AWS SageMaker training scenarios, this is a common fix when:
Loss decreases very slowly
Model appears to underfit or learn inefficiently
Training is stable but not progressing
However, it must be tuned carefully to avoid divergence.
---
Option B: Decrease the initial learning rate (Incorrect)
Reducing the learning rate makes updates even smaller, which would:
Further slow convergence
Potentially worsen the already slow d...
Author: CrystalWolfX · Last updated Jul 7, 2026
An ML engineer is importing a custom model from the Hugging Face Hub into Amazon Bedrock. The ML engineer wants to use the model with the Amazon Bedrock API in an agentic AI applicat...
Correct Answer Logic (AWS Bedrock Model Import from Hugging Face Hub)
We are importing a custom model from the Hugging Face Hub into Amazon Bedrock and then using it via the Bedrock API in an agentic AI application.
This requires understanding two key stages:
1. Model preparation (format + storage requirements)
2. Model import into Bedrock via S3
3. Runtime usage through Bedrock APIs (not SageMaker)
---
✅ Correct Options
✅ B) Convert the model to a quantized format. Import the quantized model after conversion.
Why this is correct (exam intent):
Amazon Bedrock model import requires models to be in a compatible and optimized format.
Quantization is commonly used to:
Reduce model size
Meet inference efficiency constraints
Ensure compatibility with Bedrock-supported hosting limits
This is part of the model preparation step before import, especially for large Hugging Face transformer models.
When this applies:
Large LLMs (e.g., LLaMA-style models) that exceed memory/latency constraints
When optimizing for inference cost/performance before Bedrock ingestion
---
✅ E) In the import job, provide the complete Amazon S3 URI that points to the Hugging Face model files. Ensure that the S3 URI is in the same AWS account as the Amazon Bedrock import job.
Why this is correct:
Bedrock Model Import requires model artifacts to be stored in Amazon S3
You must provide:
Full S3 URI (e.g., s3://bucket/model-path/)
Same AWS acc...
Author: NebulaEagle11 · Last updated Jul 7, 2026
A digital media entertainment company needs real-time video content moderation to ensure compliance during live streaming events.
Which solution w...
We need a solution for real-time video content moderation during live streaming with least operational overhead.
Key requirements breakdown
Real-time processing of video frames
Content moderation (detect unsafe/inappropriate visual content)
Minimal infrastructure management
AWS-native, managed services preferred
---
Option A: Amazon Rekognition + AWS Lambda
Why this is correct
Amazon Rekognition provides fully managed computer vision APIs for:
Detecting explicit content
Detecting unsafe imagery
Label detection (objects/scenes)
It works directly on image frames extracted from video streams
AWS Lambda adds serverless orchestration:
Automatically triggers processing
No server management needed
Scales automatically for live streaming bursts
Key advantages
Fully serverless architecture
No ML model training required
No infrastructure management
Built specifically for image/video moderation use cases
Very common AWS exam pattern for “lowest operational overhead”
When to use this
Real-time or near-real-time video/image moderation
Compliance filtering (nudity, violence, unsafe content)
Event-driven pipelines
---
Option B: Rekognition + LLM on Amazon Bedrock
Why this is NOT ideal
Bedrock LLMs are designed for text reasoning, not image frame analysis directly
You would still need Rekognition or another vision...
