Amazon Practice Questions, Discussions & Exam Topics by our Authors

An ML engineer is training a text generation model on Amazon SageMaker AI. After several epochs, the loss function does not converge, and the model's accuracy on the validation dataset starts to show oscillating results. The ML enginee...

The symptoms described are: Loss not converging Validation accuracy oscillating These strongly indicate unstable training dynamics, typically caused by: Learning rate being too high (overshooting minima) High gradient variance (small batch sizes) Poor optimization stability To achieve better generalization and stable convergence, we want: Smoother, more stable gradient updates Reduced step size in parameter updates Reduced noise in gradient estimation --- Correct Option: C) Decrease the learning rate and increase the mini-batch size Why C is correct Decreasing learning rate Reduces the step size of gradient updates Prevents overshooting the optimal minimum Helps loss converge smoothly instead of oscillating Increasing mini-batch size Reduces variance in gradient estimation Produces more stable and consistent updates Helps validation accuracy stabilize and improves generalization 👉 Together, these changes directly address both: Non-convergence (via smaller learning rate) Oscillating validation accuracy (via lower gradient noise) --- Why other options are incorrect A) Increase learning rate and decrease mini-batch size ❌ Increases instability: Higher learning rate → more overshooting Smaller batch size → noisier gradients This worsens oscillations and may cause divergence Suitable only for very early exploratory training (rare and risky), not conv...

Author: Ava · Last updated Jul 7, 2026

An ML engineer wants to use Amazon SageMaker Data Wrangler to perform preprocessing on a dataset. The ML engineer wants to use the processed dataset to train a classification model. During preprocessing, the ML engineer notices that a text feature has a range of thousands of values that differ only by spelling errors. The ML engineer needs to apply an ...

The key issue in this scenario is that the feature is a high-cardinality text/categorical field with many near-duplicate values caused by spelling errors. So the encoding method must not only handle many categories efficiently but also treat similar strings as related. --- Correct Answer: B) Perform similarity encoding to represent categories of the feature ✅ Why B is correct (Similarity Encoding) Similarity encoding is specifically designed for noisy categorical text data where values differ slightly due to: Spelling mistakes (e.g., “Color”, “Colour”, “Colr”) Minor variations in naming It works by computing string similarity (e.g., edit distance, Jaro-Winkler) and encoding categories based on how similar they are to each other. This allows the model to: Group near-duplicate values effectively Reduce sparsity caused by many unique spellings Improve generalization for classification tasks 👉 In Amazon SageMaker Data Wrangler, similarity encoding is recommended for dirty categorical text fields with spelling variations. --- ❌ Why other options are incorrect A) Ordinal encoding Assigns arbitrary integer values to categories Implies false ordering (e.g....

Author: Sofia · Last updated Jul 7, 2026

SNAPSHOT - A hospital wants to predict patient outcomes for the coming year. An ML engineer must improve several existing ML models that currently perform poorly. Select the correct regularization method from the following list to improve each model. Select each regularization meth...

Author: Carlos Garcia · Last updated Jul 7, 2026

A company deployed an Amazon SageMaker AI ML model to an endpoint by calling the CreateModel API operation. The network that was established with the API call includes two private subnets and one security group. The model must download data from an Amazon S3 bucket and must upload data to t...

Key requirement breakdown We need to ensure: SageMaker endpoint runs inside a VPC with private subnets The model must read/write to Amazon S3 No traffic to S3 must traverse the internet Must use correct AWS networking pattern for S3 access from VPC --- Correct AWS concept For Amazon S3 private access from a VPC without internet/NAT, AWS provides: Gateway VPC Endpoint for S3 This is the only option that: Keeps traffic on the AWS private network (AWS backbone) Does not require NAT Gateway or Internet Gateway Works by modifying route tables (prefix list routes for S3) --- Option analysis --- ❌ A) NAT Gateway Why it is wrong: NAT Gateway sends traffic out to the internet (even if private IP first) S3 access would go: VPC → NAT Gateway → Internet → S3 This violates requirement: “must not travel across the internet” When NAT Gateway is used: Private subnets need outbound internet access Example: downloading OS updates, calling external APIs --- ❌ C) Interface VPC Endpoint Why it is wrong: Interface endpoints are powered by AWS PrivateLink S3 does not use interface endpoint for general data access (S3 uses gateway endpoint for this use case) Also incorrect security group reasoning: Interface endpoints require SG inbound rules to ENI, but that’s not relevant for S3 bulk data path requirement ...

Author: Noah · Last updated Jul 7, 2026

A company uses an Amazon SageMaker AI ML model to make real-time inferences. The company has configured auto scaling for the Amazon EC2 instances that SageMaker AI uses for the inferences. During times of peak usage, new instances launch before existing instances are fully ready. As a result, the ...

The core issue here is rapid scale-out during peak traffic, where SageMaker auto scaling triggers new EC2 instances before previously launched ones are fully initialized and ready to serve traffic. This leads to inefficient capacity usage and increased latency during warm-up periods. In AWS auto scaling behavior, this is typically caused by an overly aggressive scale-out cooldown period, which does not give newly launched instances enough time to become fully healthy before additional scaling decisions are made. --- Option A: Change to a multi-model endpoint configuration in SageMaker AI This is incorrect because multi-model endpoints are used to host multiple models on a single endpoint instance, optimizing cost and model deployment density. It does not address instance warm-up time or scaling behavior, so it won’t solve readiness delays. --- Option B: Integrate Amazon API Gateway and AWS Lambda This introduces additional orchestration layers. While it can help manage traffic or preprocessing, it: Adds extra latency Does not influence EC2 instance readiness or scaling mechanics So it is not suitable ...

Author: Elijah · Last updated Jul 7, 2026

An ML engineer is building an ML pipeline. The pipeline must process a dataset in two ways by using Amazon Athena. The pipeline must use batch processing to perform large-scale data transformations and for model training. The pipeline must also use near real-time processing to perform low-l...

The correct choice is: B) Apache Parquet Why Apache Parquet is selected Apache Parquet is a columnar storage format, which makes it the best option for both batch processing and low-latency queries in Amazon Athena. Key reasons: Columnar storage: Only relevant columns are scanned instead of entire rows, drastically reducing I/O. Predicate pushdown: Athena can skip irrelevant data blocks based on filters. Compression efficiency: High compression reduces storage and speeds up data transfer. Optimized for analytics: Designed specifically for large-scale analytical workloads (OLAP), which fits both training pipelines and inference/analytics queries. Use case fit: Batch processing: Efficient full-table scans and aggregations for ML training datasets. Near real-time queries: Fast query execution due to reduced data scanning and optimized reading patterns. --- Why other options are rejected A) CSV Row-based format; Athena must scan entire rows even if only a few columns are neede...

Author: Jack · Last updated Jul 7, 2026

A company is training a new ML model to replace a model that is deployed on an Amazon SageMaker AI real-time endpoint. An ML engineer needs to determine the latency and the accuracy of the new model. The ML engineer must evaluate the new model in a product...

The requirement is to evaluate a new model’s latency and accuracy in a real production environment without impacting existing users. This is a classic safe production validation problem where the new model must receive real traffic, but must not serve responses to end users. ✅ Correct Option: D) Perform shadow testing with a traffic sampling percentage of 100% Shadow testing in Amazon SageMaker allows you to: Send a copy of live production requests to the new model Run inference on the new model in parallel with the current production model Compare latency, accuracy, and performance metrics Ensure zero impact on users, since only the production model responses are returned Using 100% traffic sampling ensures the new model is evaluated against the full production workload, giving statistically reliable latency and accuracy measurements. --- Why other options are incorrect ❌ A) Blue/green deployment with linear traffic shifting Gradually shifts traffic from old to new model Users are directly affected during rollout Risky because the new model serves real user traffic before validation is complete ➡️ Not suitable when zero user impact is required --- ...

Author: VioletCheetah55 · Last updated Jul 7, 2026

A company is building an Amazon SageMaker AI pipeline for an ML model. The pipeline uses distributed processing and training. An ML engineer needs to encrypt network communication between instances that run distributed jobs. The ML engineer configures the ...

We need to ensure encryption of network communication between instances in a distributed SageMaker training/processing job running inside a private VPC. Let’s evaluate each option with AWS SageMaker behavior in mind. --- ✅ Correct Answer: C) Enable inter-container traffic encryption In Amazon SageMaker distributed training/processing, multiple instances (or containers) communicate with each other over the network (e.g., parameter servers, data parallel workers). To encrypt this inter-node communication, SageMaker provides a specific feature: Inter-container traffic encryption It uses TLS encryption between containers/instances It is designed specifically for distributed ML workloads in SageMaker This directly satisfies the requirement: > encrypt network communication between instances in distributed jobs --- ❌ Why other options are incorrect A) Enable network isolation This isolates containers from external internet access entirely. It ensures no outbound internet traffic and no inbound access, except what is explicitly allowed inside the job. ❌ It does NOT encrypt inter-node communication It only improves security boundary iso...

Author: Michael · Last updated Jul 7, 2026

SNAPSHOT - An airline company deploys ML models to one dozen Amazon SageMaker AI inference endpoints. The inference endpoints must be able to handle different types of workloads in a cost-effective way. Select the correct inference option from the following list to handle each type of workload. Select each ...

Author: Suresh · Last updated Jul 7, 2026

SNAPSHOT - A company develops an ML model to classify products. The model uses textual data and visual data to classify the products into a hierarchical taxonomy. An ML engineer must use specific strategies to enhance the model's accuracy and handle class imbalances. Select the correct strategy from the following list for each use case. Select each strategy one time. (Choos...

Author: Emma · Last updated Jul 7, 2026

An ML engineer is using Amazon QuickSight anomaly detection to detect very high or very low machine operating temperatures compared to normal. The ML engineer sets the Severity parameter to Low and above. The ML engineer sets the Direction parameter to All. What effect will the ML e...

In Amazon QuickSight anomaly detection, the Direction parameter controls whether the model flags: All deviations (both higher and lower than expected) Only higher than expected values Only lower than expected values Given setup: Severity = Low and above → keeps sensitivity fairly broad (does not restrict detection much) Direction changes from All → Lower than expected → model now ignores high temperature anomalies and only looks for low temperature anomalies --- Effect of changing Direction to “Lower than expected” By restricting detection to only one side of deviation: 1. Decreased anomaly identification frequency Previously: both high + low anomalies were detected Now: only low anomalies are detected ➡️ Fewer total anomalies flagged 2. Decreased recall High temperature anomalies are no longer detected These were previously true positives under “All” ➡️ Some actu...

Author: Alexander · Last updated Jul 7, 2026

A company needs to analyze a large dataset that is stored in Amazon S3 in Apache Parquet format. The company wants to use one-hot encoding for some of the columns. The company needs a no-code solution to transform the data. The solution must store the t...

Key requirement breakdown Data in Amazon S3 (Parquet format) Need one-hot encoding transformation Must be a no-code solution Output must be written back to the same S3 bucket Used for model training (data prep pipeline) --- ✅ Correct Answer: A) AWS Glue DataBrew Why A is correct AWS Glue DataBrew is a fully managed, visual (no-code/low-code) data preparation service designed exactly for this type of requirement. Key reasons: No-code requirement satisfied Provides a visual interface (UI-driven) recipe builder Supports data transformations like one-hot encoding Built-in transformations for encoding categorical variables Directly reads from and writes to Amazon S3 Works natively with Parquet datasets via S3 import Ideal for data preparation for ML pipelines When A is used Data cleaning / transformation for ML No-code or low-code ETL needs Business analysts or data engineers who prefer UI over coding --- ❌ Why other options are incorrect --- B) AWS Glue Data Catalog + Amazon Athena SQL Why incorrect: Amazon Athena is SQL-based, not no-code One-hot encoding requires: Complex SQL (CASE WHEN or pivot logic) Manual feature engineering → not “no-code” Athena is query engine only, not a transformation workflow tool Limited orchestration for reusable ML preprocessing pipelines When B is used: Ad-hoc querying of S3 data Simple transformations using SQL Quick analytics...

Author: FrostFalcon88 · Last updated Jul 7, 2026

An ML engineer has trained an ML model by using Amazon SageMaker AI. The ML engineer determines that the model is overfitting and that the training data contains unnecessary features. The ML engineer must reduce the overfit...

The requirement is to reduce overfitting and reduce the impact of unnecessary features in a model trained using Amazon SageMaker AI. Key idea for solving this Overfitting happens when the model learns noise and irrelevant patterns from training data. “Unnecessary features” suggest feature selection or feature shrinkage is needed. The solution must both regularize the model and reduce feature impact automatically. --- ✅ Correct Option: B) Apply L1 regularization to the training data. Retrain the model. Why B is correct L1 regularization (Lasso) adds a penalty proportional to the absolute value of feature weights. This causes: Some feature weights to become exactly zero Automatic feature selection Reduced model complexity → less overfitting Directly addresses both requirements: ✔ Reduces overfitting ✔ Eliminates or reduces impact of unnecessary features --- ❌ Why other options are incorrect A) Increase the number of training iterations. Retrain the model. Increasing iterations typically: Improves fit to training data Increases overfitting risk Does not remove or reduce feature impact Useful only when ...

Author: Vivaan · Last updated Jul 7, 2026

A company is using an Amazon SageMaker AI ML model to predict traffic accidents that potholes cause. An ML engineer has configured SageMaker Model Monitor to run as part of a SageMaker AI pipeline. In the MonitoringExecution output, the ML engineer observes several bas...

In Amazon SageMaker Model Monitor, a `baseline_drift_check` violation indicates that the incoming production data distribution differs significantly from the baseline statistics/constraints that were computed during model setup (typically from training data or a validation dataset). In a SageMaker pipeline, this is a quality gate, so the pipeline fails when drift exceeds thresholds. --- Key idea to solve the issue You must ensure consistency between: The model (trained on current data patterns) The baseline statistics/constraints used for monitoring When drift is consistently detected, the correct remediation is to update both the model and its baseline using fresh representative data. --- Correct Option Analysis ✅ C) Retrain the model with new training data. Use the new baseline in Model Monitor. This is the correct approach. Why it works: New training data reflects the current real-world distribution You generate a new baseline (constraints + statistics) from this updated dataset Model Monitor now compares production traffic against a relevant reference distribution This removes false failures caused by an outdated baseline When to use: Data distribution has genuinely changed (real concept/data drift) Model is being retrained as part of periodic refresh Monitoring system nee...

Author: Daniel · Last updated Jul 7, 2026

An ML engineer is using Amazon SageMaker Canvas to build a custom ML model from an imported dataset. The ML engineer wants the model to make continuous numeric predictions based on 10 years of d...

This is a regression problem in Amazon SageMaker Canvas, because the model is required to make continuous numeric predictions based on historical (10 years) data. Let’s evaluate each option using key ML problem–metric alignment: --- A) Accuracy Accuracy measures the proportion of correct classifications in classification problems (e.g., spam vs not spam). ❌ Not suitable here because: The target is continuous numeric values, not discrete classes. Accuracy cannot measure how “far off” a numeric prediction is. ✅ Used in: Binary or multi-class classification models. --- B) Inference Latency Inference latency measures the time taken for the model to return a prediction. ❌ Not a performance quality metric. It does not measure prediction correctness. ✅ Used in: Real-time systems where response time matters (e.g., fraud detection APIs, chatbots). --- C) Area Under the ROC Curve (AUC) AUC measures the ability of a model to distinguish between classes across thresholds. ❌ N...

Author: Evelyn · Last updated Jul 7, 2026

A music streaming company constantly streams song ratings from an application to an Amazon S3 bucket. The company wants to use the ratings as an input for training and inference of an Amazon SageMaker AI model. The company has an AWS Glue Data Catalog that is configured with the S3 bucket as the source. An ML engineer needs to implement a solution to create a repository for ...

Correct Answer: A) Ingest data into SageMaker Feature Store from the S3 bucket. Apply tags and indexes. --- Why Option A is correct (Key reasoning) The requirement is to: Use streaming song ratings stored in Amazon S3 Support both batch training and real-time inference Ensure data stays synchronized This is exactly what Amazon SageMaker Feature Store is designed for. Key factors: Dual storage (online + offline store) Offline store (S3-backed) → used for training Online store → used for real-time inference Built-in synchronization Automatically keeps feature values consistent across training and inference Low-latency access for inference Online store is optimized for millisecond retrieval Supports streaming ingestion Ideal for continuously arriving ratings data Governance via tagging/indexing Helps manage feature discovery and reuse When to use this option: Use SageMaker Feature Store when: You need a central feature repository You require real-time + batch ML workflows Data must be consistent across training and inference You are building production ML systems --- Why other options are incorrect B) Amazon Athena + CTAS Amazon Athena is a query engine, not a feature repository. CTAS creates derived tables but: No real-time inference support No feature synchronization ...

Author: Noah · Last updated Jul 7, 2026

SNAPSHOT - A company is using Amazon SageMaker to deploy a new version of its ML model. Select the correct SageMaker traffic shifting strategy from the following list for each use case. Each traffic shifting strategy should be selected one ti...

Author: Ravi Patel · Last updated Jul 7, 2026

A company wants to use Amazon SageMaker to host an ML model that runs on CPU for real-time predictions. The model will have intermittent traffic during business hours and will have periods of no traffic after business hours. The company needs a solution that will...

The key requirements are: Real-time predictions (low latency) CPU-based model Intermittent traffic during business hours Zero or near-zero traffic after hours Most cost-effective solution The most important cost driver here is avoiding charges when there is no traffic after business hours, so the solution should ideally scale to zero when idle. --- Option A: SageMaker Real-time endpoint + scheduled auto scaling A SageMaker real-time endpoint is always provisioned and running. Even with auto scaling, there is always at least one running instance You pay for uptime even when there are no requests Auto scaling helps handle bursts but does NOT eliminate idle cost When it is used: Steady or predictable traffic Strict low-latency requirements with continuous availability ❌ Rejected because it cannot scale to zero → not cost-optimal for idle periods. --- Option B: SageMaker Serverless Inference endpoint + provisioned concurrency SageMaker Serverless Inference is designed for exactly this pattern: Automatically scales based on requests Scales down to zero when idle You only pay per request and compute used Ideal for intermittent workloads However: “Provisioned concurrency” is a Lambda concept, not SageMaker Serverless Inference terminology Despite wording inconsistency, the intent clearly maps to serverless behavior When it is used: Sporadic traffic Unpredictable or bursty workloads Cost-sensitive workloads with idl...

Author: ElectricLionX · Last updated Jul 7, 2026

What is the primary purpose of system prompts in generative AI applications?

The correct answer is C) To define the role and behavioral boundaries of AI. Why C is correct A system prompt is a high-priority instruction given to a generative AI model that defines: Role: Who the AI should act as (e.g., tutor, coding assistant, customer support agent). Behavior: Tone, style, and response guidelines. Boundaries and constraints: What the AI should or should not do. Safety and compliance rules: Restrictions to ensure appropriate outputs. Key factors: Controls AI behavior consistently across conversations. Establishes guardrails and operational rules. Helps align responses with business requirements and user expectations. Improves reliability and predictability of outputs. Example scenario: An AWS-powered customer support chatbot receives a system prompt such as: "You are a professional AWS support assistant. Provide accurate cloud guidance, avoid speculation, and maintain a helpful tone." This defines the chatbot's role and behavior. --- Why A is rejected A) To authenticate user credentials to access responses from AI Why incorrect: Authentication is handled by identity and access management systems (e.g., AWS IAM, Cognito, application authentication mechanisms). System prompts do not verify usernames, passwords, tokens, or permissions. They influence AI behavior, not security access control. When this would be used instead: User login systems. API au...

Author: James · Last updated Jul 19, 2026

A company is using a large language model (LLM) to create a generative AI assistant. The company must choose an AI technique to ensure that the AI assistant generates the most factually correct responses. The company selects the Retrie...

The correct answer is A) Hallucinations. Why A is correct Retrieval Augmented Generation (RAG) improves the factual accuracy of an LLM by retrieving relevant, up-to-date information from trusted external data sources before generating a response. Key idea: LLMs generate answers based on patterns learned during training. If the required information was not in the training data or is outdated, the model may confidently generate incorrect information. This behavior is called hallucination. RAG reduces this problem by: Retrieving relevant documents from a knowledge base. Grounding responses in real, verifiable data. Providing current information without retraining the model. Reducing fabricated facts and unsupported claims. Key factors in reasoning: Question emphasizes "most factually correct responses." RAG's primary purpose is to provide external factual context. Hallucinations occur when an LLM invents or misstates facts. Therefore, RAG is specifically designed to mitigate hallucinations. Example scenario: A company chatbot answers questions about internal HR policies. Policies change frequently and may not exist in the model's training data. RAG retrieves the latest policy documents and uses them to generate accurate answers, reducing hallucinations. --- Why B is rejected B) Security Why incorrect: RAG is not primarily a security mechanism. It does not inherently prevent unauthorized access, data breaches, or prompt injection attacks. Security requires controls such as authentication, authorization, encryption, and guardrails. When Security is the main concern: Protecting confidential company data. Controlling user permissions. Preventing data leakage. Implementing AWS IAM permissions and encryption. Key distinction: RAG improves factual grounding; security controls protect systems and data. --- Why C is rejected C) Nondeterminism What nondeterminism means: An LLM may generate different responses to the same prompt because of ...

Author: Olivia · Last updated Jul 19, 2026

A company is building a job recommendation system based on job posting data and job seeker user profiles. The system shows bias in job recommendations based on gender for user profiles that are otherwise equivalent. Which principle...

The correct answer is D) Fairness. Why D is correct The question states: > "The system shows bias in job recommendations based on gender for user profiles that are otherwise equivalent." This is a classic example of an unfair outcome where individuals with similar qualifications receive different recommendations because of a protected attribute (gender). According to AWS Responsible AI principles, Fairness means: AI systems should treat similar individuals similarly. Bias and discrimination should be identified and mitigated. Outcomes should not unfairly favor or disadvantage groups based on characteristics such as gender, race, age, etc. Key factors in reasoning 1. Bias is explicitly mentioned. 2. Gender-based differences are the issue. 3. Equivalent profiles receive different outcomes. 4. The goal is to ensure equitable treatment and recommendations. These are direct indicators of a Fairness problem. Example scenario A job recommendation engine: Recommends software engineering roles to male candidates. Recommends administrative roles to female candidates. Both candidates have identical skills, experience, and qualifications. This is a fairness issue because gender is influencing recommendations. --- Why A is rejected A) Governance (spelled "Governance") What Governance means: Governance focuses on policies, processes, oversight, compliance, risk management, and accountability for AI systems. Examples: Defining AI usage policies. Model approval workflows. Regulatory compliance. Audit requirements. Why it is incorrect here The problem described is not about: Lack of policies, Compliance failures, Missing oversight. The issue is biased recommendations, which is directly addressed by Fairness rather than Governance. When Governance would be the correct answer If the question asked: > "How should the company establish controls to review AI models before deployment?" Then Governance would be appropriate. --- Why B is rejected B) Explainability What Explainability means: The...

Author: Ethan Smith · Last updated Jul 19, 2026

A company wants its AI models to be transparent and explainable. Which combination of Amazon SageMaker AI fea...

The correct answers are A) SageMaker Model Cards and C) SageMaker Clarify (the service is now known as SageMaker Clarify; some exam questions may still refer to it under SageMaker AI). --- How to identify the answer The question emphasizes: > "transparent and explainable" These are key Responsible AI concepts: Transparency → Documenting model details, intended use, limitations, performance, and risk information. Explainability → Understanding why a model made a prediction and identifying feature importance and bias. Therefore, look for AWS services specifically designed for: 1. Model documentation and transparency. 2. Model explainability and bias detection. --- A) SageMaker Model Cards ✅ Why it is correct SageMaker Model Cards provide standardized documentation about ML models. They help organizations record: Model purpose Training datasets Performance metrics Intended use cases Limitations Risk assessments Approval status How it supports transparency Transparency means stakeholders can understand: What the model does How it was built Its limitations Appropriate usage scenarios Model Cards are specifically designed for this purpose. Example scenario A bank deploys a loan approval model. Auditors ask: What data was used? What accuracy was achieved? What are known limitations? A Model Card provides this information in a structured format. AWS Exam Keyword When you see: Documentation Transparency Governance Model reporting Intended use Think SageMaker Model Cards. --- B) SageMaker Pipelines ❌ What it does SageMaker Pipelines automates ML workflows such as: Data preparation Training Testing Deployment Why it is incorrect Pipelines improve: Automation Reproducibility MLOps But they do not directly provide transparency or explainability. When Pipelines is correct Questions about: CI/CD for ML Automated training workflows Reproducible ML processes MLOps automation --- C) SageMaker Clarify ✅ Why it is correct SageMaker Clarify is AWS's primary explainability and bias detection service. It helps: Explain model predictions Identify feature importance Detect bias in datasets Detect bias in model outputs How it supports explaina...

Author: Ahmed · Last updated Jul 19, 2026

A company is developing a product recommendation application by using a generative AI model. The company must minimize the application's environ...

The correct answer is A) Optimize the deployed model architecture to prioritize computational efficiency during model inference. --- Why A is correct The question asks: > "The company must minimize the application's environmental impact." In AWS Responsible AI and sustainability discussions, environmental impact is primarily related to: Energy consumption Compute utilization Carbon footprint Resource efficiency The largest ongoing environmental cost for a deployed generative AI application is often model inference, because recommendations may be generated continuously for many users. By optimizing the model architecture for computational efficiency, the company can: Reduce CPU/GPU utilization Lower power consumption Reduce infrastructure requirements Lower carbon emissions Improve sustainability while maintaining functionality Key factors in reasoning 1. Goal = minimize environmental impact. 2. Environmental impact is strongly tied to compute usage. 3. Efficient inference requires fewer computational resources. 4. Fewer resources → less energy consumption → lower carbon footprint. Example scenarios Techniques include: Model compression Quantization Distillation Smaller optimized models Efficient inference endpoints For example, replacing a 70B-parameter model with a distilled model that delivers similar recommendation quality can significantly reduce energy usage. AWS Exam Keyword When you see: Sustainability Environmental impact Carbon footprint Energy efficiency Green AI Think: Optimize model size and computational efficiency. --- Why B is rejected B) Adopt a distributed inference approach by using multiple smaller models across multiple Availability Zones. Why it sounds tempting People often associate smaller models with efficiency. Why it is incorrect The key issue is: Running models across multiple Availability Zones introduces additional infrastructure. More networking overhead. More synchronization and data transfer. More servers active simultaneously. This architecture is typically chosen for: High availability Fault tolerance Geographic resilience Not specifically for minimizing environmental impact. When B would be appropriate If the requirement were: High availability Disaster recovery Fault tolerance Resilient inference services then distributing across AZs would make sense. Key AWS exam clue "Sust...

Author: StarryEagle42 · Last updated Jul 19, 2026

A company uses Amazon SageMaker AI to generate article summaries in multiple languages. The company needs a metric to evaluate the quality of the summary translations in mu...

The correct answer is B) Bilingual Evaluation Understudy (BLEU). --- Why B is correct The question states: > "generate article summaries in multiple languages" > > "evaluate the quality of the summary translations" The key phrase is "translations in multiple languages." BLEU (Bilingual Evaluation Understudy) is the standard metric used to evaluate machine translation quality by comparing a generated translation against one or more human reference translations. What BLEU measures BLEU evaluates: Word overlap between generated and reference translations. N-gram matching (phrases and word sequences). Translation accuracy relative to a human-translated reference. Key factors in reasoning 1. The task involves multiple languages. 2. The output is a translation. 3. BLEU was specifically designed for machine translation evaluation. 4. AWS exam questions often associate: Translation → BLEU Summarization → ROUGE Since the question focuses on translated summaries, the translation aspect is the deciding factor. Example scenario Reference translation: > "The company announced new products today." Model output: > "The company unveiled new products today." BLEU measures how closely the generated translation matches the reference translation. --- Why A is rejected A) Recall-Oriented Understudy for Gisting Evaluation (ROUGE) What ROUGE is used for ROUGE is primarily used for: Text summarization Comparing generated summaries with reference summaries Measuring content overlap and coverage Why it is incorrect here The question is not asking: > "How good is the summary?" It is asking: > "How good is the summary translation across languages?" Translation quality is better evaluated using BLEU. When ROUGE would be correct Example: A model summarizes a 10-page article into 3 paragraphs in the same language. You want to compare: Generated summary Human-written summary ROUGE is the preferred metric. AWS Exam Shortcut Summarization → ROUGE Translation → BLEU When both appear in a question, identify the primary task being evaluated. Here, it is translation quality. --- Why C is rejected C) Ar...

Author: RadiantPhoenixX · Last updated Jul 19, 2026

A research company is growing microbiological cultures. The company captures images of the cultures without any prior labeled data about growth areas. The company needs to identify the regions of th...

We are given a key constraint: no prior labeled data about growth areas. That immediately rules out supervised learning methods and pushes us toward unsupervised techniques. The goal is also specific: identify regions in images showing culture growth, which is essentially a form of image segmentation or grouping similar pixels/regions based on visual patterns. --- Option analysis A) Logistic Regression — ❌ Not suitable Logistic regression is a supervised classification algorithm. Requires labeled training data (e.g., “growth” vs “no growth” labels for each pixel/region) Outputs class probabilities, not natural region grouping Not designed for raw image region discovery without labels When it would be used: Binary or multi-class classification problems like spam detection, fraud detection, or predicting disease presence from labeled features. --- B) Decision Tree — ❌ Not suitable Decision trees are also supervised learning models. Need labeled target variables Work well on structured/tabular data, not raw image region discovery Do not inherently perform segmentation or clustering of unlabeled image data When it would be used: Credit risk scoring, customer churn prediction, or classification tasks with clear labeled outcomes. --- C) Clustering — ✅ Correct Clustering is an unsupervised learning technique, making it ideal here. ...

Author: John · Last updated Jul 19, 2026

A company that streams media is selecting an Amazon Nova foundation model (FM) to process documents and images. The company is comparing Nova Micro and Nova Lite. The company wants to minimize costs. ...

We need to evaluate Amazon Nova foundation models (FM) — Nova Micro vs Nova Lite with the key requirement: minimize cost while processing documents and images. So the deciding factor is capability vs cost efficiency, especially what modalities each model supports. --- Option analysis A) Nova Micro uses transformer-based architectures. Nova Lite does not use transformer-based architectures. — ❌ Incorrect This is not a meaningful differentiator for model selection in AWS practice. Both Nova Micro and Nova Lite are foundation models built on modern deep learning architectures (transformer-based is standard for FMs) Architecture detail is not used as a selection criterion in AWS service comparisons Does not relate to cost or multimodal capability When such info would matter: Research discussions on model internals Academic comparison of architectures (not AWS solution design) --- B) Nova Micro supports only text data. Nova Lite is optimized for numerical data. — ❌ Incorrect This is misleading and not aligned with AWS Nova model positioning. Nova Lite is not “numerical-only” Foundation models are generally multimodal or text-centric, not restricted to numerical data Media streaming + document/image processing requires multimodal understanding, not numerical optimization When numerical optimization matters: Forecasting models, time-series ML (not FM selection for media/document understanding) --- C) Nova Micro supports only text. Nova Lite supports images, videos, and text. — ❌ Incorrect (key idea reversed for cost reasoning) This option incorrectly frames capabilities. Nova Micro is designed as a lower-cost, text-focused model Nova Lite is typically more capable (including multimodal support) but higher cost than Micro The statement is partially aligned with multimodal idea, but the model roles are inaccurate and therefore not reliable for AWS decision-making When multimodal models are used: Image + text unde...

Author: Kunal · Last updated Jul 19, 2026

A company stores customer data in OpenSearch. The company wants an AI solution to retrieve specific customer information from the stored data. The AI solution must convert queries into data requests and generate CSV files from the results. Then, the AI solution must uploa...

We need the most operationally efficient AWS-native AI approach that can: 1. Take natural language queries 2. Convert them into structured OpenSearch queries 3. Retrieve results 4. Generate CSV files 5. Upload them to Amazon S3 So this is not just “generate text” — it is a multi-step workflow involving tool use, data retrieval, and file operations. --- Option analysis A) Create an AI agent to perform the required steps. — ✅ Correct An AI agent (e.g., using Amazon Bedrock Agents) is specifically designed for: Breaking down a user request into multiple steps Calling external tools/APIs (like OpenSearch queries) Transforming outputs (e.g., JSON → CSV) Interacting with AWS services (like Amazon S3 uploads) Maintaining orchestration without custom glue code Why this fits best: Converts natural language → structured OpenSearch query (via FM reasoning) Executes retrieval from OpenSearch Processes and formats data into CSV Automatically uploads results to S3 Minimizes operational overhead (no custom orchestration pipelines needed) Fully managed and scalable When this is used: Multi-step enterprise workflows Systems requiring tool use + API chaining RAG + data transformation pipelines Automation across AWS services --- B) Use a single foundation model (FM) with few-shot prompting. — ❌ Not sufficient A standalone FM: Can convert queries into OpenSearch DSL (maybe) Can format output as CSV text But it cannot: Execute OpenSearch queries Retrieve live data Upload files to S3 Orchestrate multi-step workflows reliably When it is us...

Author: Matthew · Last updated Jul 19, 2026

Which foundation model (FM) in Amazon Bedrock can be fine-tuned for text, image, and video comprehen...

To answer this, we need to match the requirement carefully: Requirement: A foundation model (FM) in Amazon Bedrock that can be fine-tuned for text, image, and video comprehension. That implies three key capabilities: Multimodal input support (text + image + video) Comprehension / reasoning (not just embeddings) Customization / fine-tuning support in Bedrock --- Option analysis A) Amazon Nova Pro — ✅ Correct choice Designed as a multimodal foundation model in Amazon Bedrock. Supports text, image, and video understanding, enabling cross-modal reasoning (e.g., describing videos, analyzing images with text queries). Suitable for advanced comprehension tasks like: Video scene understanding Image Q&A Multimodal document analysis Supports model customization/fine-tuning (or enterprise adaptation features in Bedrock depending on region/availability). When to use: Multimodal assistants (text + image + video chatbots) Video summarization or scene description Image-based Q&A systems with contextual reasoning --- B) Amazon Titan Multimodal Embeddings G1 — ❌ Incorrect Produces embeddings, not generative outputs or comprehension responses. Used for: Searc...

Author: Daniel · Last updated Jul 19, 2026

A company wants to generate synthetic data responses for multiple prompts from a large volume of data. The company wants to use an API method to generate the responses. The company does not need to generate the respon...

We need to identify the solution that: Generates synthetic data responses for many prompts Works on a large volume of data Uses an API-based approach Does not require immediate response (asynchronous is fine) Requires least development effort --- Key idea This is a classic large-scale, non-real-time generation workload, which strongly favors batch processing in Amazon Bedrock. --- Option analysis A) Real-time inference ❌ Uses synchronous API calls (InvokeModel / InvokeModelWithResponseStream). Designed for interactive, low-latency requests. Not efficient for large-scale prompt processing. Requires more orchestration logic for bulk jobs. Why rejected: Not optimized for bulk synthetic data generation. Higher cost and operational overhead for large datasets. Use case: Chatbots Live user queries Real-time content generation --- B) Amazon Bedrock batch inference ✅ (Correct) Specifically designed for large-scale, asynchronous processing. You submit a dataset of prompts and Bedrock processes them in bulk. Least development effort because: No custom loop logic No infrastructure management No manual API orchestration Outputs are stored and retrieved after processing completes. Why it fits b...

Author: Amelia · Last updated Jul 19, 2026

Which statement accurately describes Retrieval Augmented Generation (RAG)?

We need to identify the statement that correctly defines Retrieval Augmented Generation (RAG) in the context of AWS and LLM architectures. --- Key concept: What RAG actually is RAG (Retrieval Augmented Generation) is an approach where: The model does not rely only on its training data It retrieves relevant information from external knowledge sources (databases, documents, vector stores) That retrieved context is then used to generate more accurate, up-to-date responses Importantly: no retraining of the LLM is required In AWS, this is commonly implemented using: Amazon Bedrock Knowledge Bases for Amazon Bedrock Vector databases like Amazon OpenSearch Service or Amazon Kendra --- Option analysis A) ❌ Incorrect > Uses large amounts of new data to train LLMs This describes model training or fine-tuning, not RAG. RAG does NOT retrain or update model weights. It only retrieves external data at inference time. Use case (actual concept described): Fine-tuning LLMs on domain-specific datasets Custom model training pipelines --- B) ✅ Correct > LLMs reference external authoritative knowledge bases to enhance responses without re-training This is the exact definition of RAG Key characteristics matched: Uses external knowledge sources Improves accuracy and relev...

Author: Lucas Carter · Last updated Jul 19, 2026

A company must comply with regulatory standards to develop and use trustworthy AI management solutions. ...

The requirement is about regulatory compliance and trustworthy AI management, which focuses on responsible AI governance across the entire lifecycle—not performance optimization or technical exclusivity. Correct approach: D) Ensure fairness, transparency, accountability, and security throughout the lifecycle of each AI solution. This aligns with AWS’s responsible AI and governance principles, which emphasize: Fairness: preventing bias in model outcomes Transparency: explainability and traceability of decisions Accountability: clear ownership and auditability Security: protecting data, models, and access controls Lifecycle governance: applying controls from design → training → deployment → monitoring These are essential for meeting regulatory standards (e.g., GDPR-style requirements, model governance policies, auditability expectations). --- Why other options are rejected A) Optimize model inference time by using high-powered GPUs for faster processing Why incorrect: This is a performance optimization goal, not a compliance or governance requirement. When it is used: Suitable for latency-sensitive applications like real-time fraud detection or recommendation systems. Missing: No mention of ethics, ...

Author: Rohan · Last updated Jul 19, 2026

An AI practitioner is developing a prompt for large language models (LLMs) in Amazon Bedrock. The AI practitioner must ensure that the prompt works across all Am...

The key requirement is cross-model prompt portability across Amazon Bedrock foundation models, meaning we must identify what varies between models in a way that affects prompt behavior. --- Correct answer: A) Maximum token count Why A is correct In Amazon Web Services Amazon Bedrock, different foundation models (e.g., Anthropic Claude, Meta Llama, Amazon Titan) have different context window limits, which define: Maximum input tokens (prompt size) Maximum output tokens (response length) This directly affects whether a prompt: Fits entirely in the model context window Requires truncation or chunking Impacts long-document summarization or multi-turn reasoning So even if a prompt is well-designed, it may behave differently or fail entirely depending on token limits across models. --- Why other options are incorrect B) On-demand inference parameter support Why incorrect: In Bedrock, most core inference parameters (like temperature, top-p, max tokens) are generally standardized across models, though some models may expose slight variations. This is not the primary factor affecting cross-model prompt compatibility. When it matters: If using advanced model-s...

Author: David · Last updated Jul 19, 2026

SNAPSHOT - A company wants to use ML to increase customer engagement and sales. The company has collected a large dataset that includes customer demographics, purchase history, browsing patterns, and product ratings. Select ...

Author: Olivia · Last updated Jul 19, 2026

SNAPSHOT - A company wants to build a new ML solution. The company already has data. The company needs to understand the ML lifecycle before building the solution. Select and order the steps fro...

Author: Manish · Last updated Jul 19, 2026

A company is using Amazon Q Business to create an AI assistant. The company needs to restrict user interactions with the AI assistant to company-...

The requirement is to restrict user interactions to company-approved topics in an Amazon Q Business AI assistant. This is fundamentally about controlling what the assistant is allowed to answer, not about indexing scope or data access. --- Correct answer: C) Amazon Q Business application guardrails In Amazon Web Services Amazon Q Business, application guardrails are specifically designed to: Restrict conversations to approved topics Block out-of-scope or irrelevant queries Enforce company policies on AI responses Control what the assistant can and cannot discuss This directly matches the requirement of limiting user interactions to company-approved domains or topics, making guardrails the correct governance mechanism. --- Why other options are incorrect A) Amazon Q Business Enterprise index Why incorrect: The enterprise index defines the data sources and documents the AI can retrieve from, not conversational restrictions. It helps the model answer questions from enterprise knowledge but does not strictly enforce topic-level interaction control...

Author: Deepak · Last updated Jul 19, 2026

A data engineer uses Amazon Kinesis Data Streams to ingest and process records that contain user behavior data from an application every day. The data engineer notices that the data stream is experiencing throttling because hot shards receive much ...

To resolve the throttling issue in Amazon Kinesis Data Streams due to hot shards, we must understand the cause: records are unevenly distributed among shards, meaning some shards are overloaded while others are underutilized. This results in throttling errors like `ProvisionedThroughputExceededException`. --- Option A: Use a random partition key to distribute the ingested records. ✅ This is a strong option. Why it's effective: Kinesis uses the partition key (hashed using MD5) to determine which shard a record goes to. If the same key is reused (e.g., user ID), it can send many records to a single shard, creating a hot shard. Using a random partition key helps evenly distribute records across all shards, reducing throttling. Best used when: The application can tolerate random partitioning and doesn’t need records to be grouped by key for ordering or session state. Limitation: If downstream consumers rely on ordered records by key (e.g., all actions by a user), this could break that logic. --- Option B: Increase the number of shards in the data stream. Distribute the records across the shards. ✅ Also a good option, especially in combination with Option A. Why it can help: Adding more shards increases capacity, allowing more records to be ingested per second. However: Simply increasing shards without ensuring even record distribution (e.g., if partition key logic isn’t fixed) can still result in hot shards. Best used when:...

Author: Liam · Last updated Jul 10, 2026

A company has a data processing pipeline that includes several dozen steps. The data processing pipeline needs to send alerts in real time when a step fails or succeeds. The data processing pipeline uses a combination of Amazon S3 buckets, AWS Lambda functions, and AWS Step Functions state machine...

To determine the best option for real-time alerting in a data processing pipeline using Amazon S3, AWS Lambda, and AWS Step Functions, we must consider real-time responsiveness, integration capabilities, reliability, and appropriate service usage. --- Option Analysis --- A) Configure the Step Functions state machines to store notifications in an Amazon S3 bucket when the state machines finish running. Enable S3 event notifications on the S3 bucket. Why reject it: S3 is not optimized for real-time alerting; it introduces latency. S3 event notifications only trigger on object operations (e.g., PUT), not specifically tied to state machine outcomes. It adds unnecessary complexity and delay to the notification flow. When it could be used: For logging or archiving results, not real-time alerting. --- B) Configure the AWS Lambda functions to store notifications in an Amazon S3 bucket when the state machines finish running. Enable S3 event notifications on the S3 bucket. Why reject it: Similar issues as (A) — delayed notifications via S3. Adds more overhead by involving Lambda to generate and store notifications unnecessarily. Still relies on S3 event triggers, which aren't purpose-built for status change notifications. When it could be used: Suitable if you want to retain a log of...

Author: Rahul · Last updated Jul 10, 2026

A company has an application that uses an Amazon API Gateway REST API and an AWS Lambda function to retrieve data from an Amazon DynamoDB instance. Users recently reported intermittent high latency in the application's response times. A data engineer finds that the Lambda function experiences frequent throttling when the company's other Lambda functions experience increased invocations. The com...

To determine the most cost-effective solution that ensures one Lambda function is not impacted by the load of other Lambda functions, let's analyze each option: --- ❌ A) Increase the number of read capacity units (RCU) in DynamoDB Why it's rejected: Key factor: The issue is Lambda throttling, not DynamoDB read performance. Increasing RCU would help if the bottleneck was DynamoDB read throughput (e.g., `ProvisionedThroughputExceededException`), but that is not the case here. It doesn’t prevent Lambda function throttling due to concurrency limits or resource contention. When to use: Use this option if DynamoDB read requests are throttled (CloudWatch metrics: `ReadThrottleEvents`). --- ✅ B) Configure provisioned concurrency for the Lambda function Why it’s not selected (even though it can help): Provisioned concurrency keeps a set number of Lambda instances initialized and ready to serve requests. It helps reduce cold starts and latency, but does not isolate concurrency from other functions or prevent throttling due to overall account concurrency limits. When to use: Use this when the concern is cold start latency and you want consistent performance. Drawback: More expensive than reserved concurrency ...

Author: Daniel · Last updated Jul 10, 2026

A company has as JSON file that contains personally identifiable information (PII) data and non-PII data. The company needs to make the data available for querying and analysis. The non-PII data must be available to everyone in the company. The PII data must be available on...

To determine the best solution with the least operational overhead, let's analyze each option carefully using the key factors: Key Factors to Consider Operational Overhead: Effort required to implement and maintain the solution. Data Access Control: Fine-grained access control over PII vs non-PII. Scalability: Ability to scale as data or users grow. Automation & Governance: Ability to manage data cataloging, classification, and permissioning centrally. Cost-efficiency: Avoiding unnecessary services or duplication of data. --- Option A: S3 + AWS Glue to split PII and non-PII, separate buckets with IAM access Pros: Clear separation of data (PII vs non-PII). S3 is scalable and cost-effective. Cons: Requires custom ETL logic in Glue to split the data. Maintaining and updating the splitting logic increases operational overhead. Managing separate buckets and ensuring data consistency between them adds complexity. > Use Case Fit: Suitable if manual control of data partitioning is acceptable and custom transformation logic is required. But not ideal for low operational overhead. --- Option B: S3 + Amazon Macie to identify PII and manage access Pros: Macie automatically detects PII data. Cons: Macie is primarily a PII discovery tool, not a data access control or enforcement tool. Does not grant or restrict access directly — you'd still need to set up policies manually. High cost and unnecessary if access control is the primary need. > Use Case Fit: Best fo...

Author: William · Last updated Jul 10, 2026

A company uses AWS Key Management Service (AWS KMS) to encrypt an Amazon Redshift cluster. The company wants to configure a cross-Region snapshot of the Redshift cluster as part of disaster recovery (DR) strategy. A data engineer needs to use the AWS CLI...

To determine the correct combination of steps for setting up cross-Region snapshot copying for an Amazon Redshift cluster encrypted with AWS Key Management Service (KMS), we need to focus on how snapshot encryption and cross-Region snapshot copy work in Redshift. --- Key Concepts to Consider: Redshift automated/manual snapshots can be automatically copied across Regions for DR purposes. If KMS encryption is used, Redshift must be granted permission to use a KMS key in the destination Region for encrypting the copied snapshot. This is done using a snapshot copy grant, which is created in the source Region to refer to the destination Region's KMS key. Multi-AZ is not supported in Redshift; it's only available in services like RDS, so such options are invalid. --- Option Analysis: ✅ A) Create a KMS key and configure a snapshot copy grant in the source AWS Region. ✅ Valid. You must create a snapshot copy grant in the source Region, even though the KMS key it refers to is in the destination Region. This grant allows Redshift to encrypt snapshots during the copy process using the KMS key in the destination Region. Key factor: Snapshot copy grants are created in the source Region. ❌ B) In the source AWS Region, enable snapshot copying. Specify the name of the snapshot copy grant...

Author: Jack · Last updated Jul 10, 2026

A company is using Amazon S3 to build a data lake. The company needs to replicate records from multiple source databases into Apache Parquet format. Most of the source databases are hosted on Amazon RDS. However, one source database is an on-premises Microsoft SQL Server Enterprise instance. The company needs to implement a solution to replicate existi...

To determine the most cost-effective and technically appropriate solution for replicating existing data and future changes from both Amazon RDS and an on-premises Microsoft SQL Server to Amazon S3 in Parquet format, let’s evaluate each option against the key factors: --- ✅ Key Requirements 1. Replicate existing data and future changes 2. Source: Amazon RDS and on-premises SQL Server 3. Target: Amazon S3 in Apache Parquet format 4. Solution must be cost-effective 5. Should work with heterogeneous sources 6. Should support ongoing replication (CDC) --- Option A) Use one AWS Glue job to replicate existing data. Use a second AWS Glue job to replicate future changes. ❌ AWS Glue is great for ETL and transformation, but: It does not natively support CDC (Change Data Capture). Glue jobs are batch-oriented and not designed for near-real-time replication. Connecting to on-premises SQL Server adds network and security complexity. Would require custom polling logic to track changes — increases cost and complexity. ✅ Works for batch ETL of existing data, but not for future changes efficiently. > ✅ Use case: One-time bulk ETL tasks, or periodic data loads where latency is acceptable. --- Option B) Use AWS Database Migration Service (AWS DMS) to replicate existing data. Use AWS Glue jobs to replicate future changes. ❌ AWS DMS is designed for both full load and CDC. Using it only for existing data underutilizes its full capability. ❌ Future changes using AWS Glue (as above) suffer from the same issues — no native CDC, not designed for streaming changes. ❌ Adds operational overhead and cost by splitting logic between two services. > ✅ Use case: Transitional architectures where you phase in CDC later using another tool — but not cost-effective long-term. --- ✅ Option C) Use AWS Database Migration Service (AWS DMS) to replicate existing data and future changes. ✅ DMS is specifically ...

Author: Isabella · Last updated Jul 10, 2026

A data engineer needs to optimize the performance of a data pipeline that handles retail orders. Data about the orders is ingested daily into an Amazon S3 bucket. The data engineer runs queries once each week to extract metrics from the orders data based the order date for multiple date ranges. The data engineer needs an optimization solution tha...

To determine the most cost-effective and scalable solution for optimizing query performance on a growing dataset of retail orders ingested daily into Amazon S3, we need to evaluate each option on several key factors: --- 🔑 Key Factors to Consider: Query pattern: Weekly queries based on order date ranges. Data ingestion location: Amazon S3. Scalability & performance: Must maintain good performance as data volume grows. Cost-effectiveness: Must be optimized for cost. Partitioning strategy: Must match query patterns (i.e., order date). --- ✅ Option A: Partition the data based on order date. Use Amazon Athena to query the data. Why it's suitable: Athena is serverless and cost-effective (pay-per-query). Partitioning by order date aligns directly with query patterns, improving performance and reducing scanned data. Supports direct querying of data in S3 (no data movement). Scales automatically with data volume. Best Use Case: Ad hoc analytics or scheduled queries on large datasets stored in S3, especially when you don’t want to manage infrastructure. ✅ Meets all requirements effectively and cost-efficiently. --- ❌ Option B: Partition the data based on order date. Use Amazon Redshift to query the data. Why it's less suitable: Redshift is a data warehouse and may require ETL to ingest data from S3, adding complexity and cost. Although it performs well...

Author: Ryan · Last updated Jul 10, 2026

A data engineer has two datasets that contain sales information for multiple cities and states. One dataset is named reference, and the other dataset is named primary. The data engineer needs a solution to determine whether a specific set of values in the city and state columns of the primary dataset exactly match the same specific values in the reference datas...

Let's carefully analyze the problem and the options provided: --- Problem Recap: Two datasets: primary and reference. Columns: `city` and `state` in both datasets (reference columns named `ref_city`, `ref_state`). Need to check if a specific set of values in primary’s city and state exactly match the same values in reference. Want to use Data Quality Definition Language (DQDL) rules in an AWS Glue Data Quality job. Check if the city-state pairs in primary are exactly present in reference. --- Key concepts: DatasetMatch rule: Compares entire datasets or subset of columns between two datasets for similarity/match ratio. Useful for row-level or full dataset similarity checks. ReferentialIntegrity rule: Validates that foreign keys in one dataset (primary) exist in the referenced dataset (reference), i.e., for each (city,state) in primary, the same pair exists in reference. This is a typical referential integrity check. The value `1.0` or `100` indicates the threshold for the rule to be considered passing — meaning 100% matches. --- Now, let's analyze each option: --- Option A: `DatasetMatch "reference" "city->ref_city, state->ref_state" = 1.0` Checks if the primary dataset matches the reference dataset in city and state columns with a similarity threshold of 1.0 (100%). Problem: DatasetMatch compares the whole datasets; it’s not designed for referential integrity checks (checking existence of values from primary in reference). The syntax seems off (the value 1.0 vs 100), and it's ambiguous if the percentage is properly interpreted. When to use: Checking if datasets are identical in specific columns. Why rejected: We want to check if city-state pairs from primary exist in reference, not that the entire datasets match exactly. DatasetMatch is not ideal here. --- Option B: `ReferentialIntegrity "city,state" "reference.{ref_city,ref_state}" = 1.0` ReferentialIntegrity rule that checks if every (city,state) pair in primary dataset exists in reference dataset's (ref\_city...

Author: Olivia Johnson · Last updated Jul 10, 2026

A company has an on-premises PostgreSQL database that contains customer data. The company wants to migrate the customer data to an Amazon Redshift data warehouse. The company has established a VPN connection between the on-premises database and AWS. The on-premises database is continuously updated. The co...

To determine the best solution for migrating continuously updated data from an on-premises PostgreSQL database to Amazon Redshift, we must evaluate each option based on data freshness, latency, automation, and ongoing sync support. --- ✅ Option B) Create an AWS Database Migration Service (AWS DMS) full-load task. Set Amazon Redshift as the target. Configure the task to use the change data capture (CDC) feature. Why it's selected: Key Feature: Change Data Capture (CDC): Allows near real-time replication of changes from PostgreSQL to Redshift. Low latency: Changes are replicated continuously after initial full load. Automation: Fully managed and integrated with AWS services. Supports continuous synchronization: Ideal for scenarios where the source database is frequently updated. Recommended use case: For near real-time replication and minimum lag between source and target systems. --- ❌ Option A) Use pg\_dump + AWS SCT + nightly cron job Why it’s rejected: Batch-based, not real-time: `pg_dump` is a snapshot-based tool; it doesn’t support incremental or continuous sync. Latency: Updates happen only nightly—doesn't meet the "as quickly as possible" requirement. Operational complexity: Requires scheduling, scripting, and manual setup. Recommended use case: One-time migrations or nightly reporting with tolerance for stale data. --- ❌ Option C) pg\_dump + S3 + C...

Author: Amira99 · Last updated Jul 10, 2026

A company has several new datasets in CSV and JSON formats. A data engineer needs to make the data available to a team of data analysts who will analyze the data by using SQL queries....

Let's break down each option with key factors: cost-effectiveness, ease of SQL querying, data format support, and analyst accessibility. --- Option A: Create an Amazon RDS MySQL cluster, use AWS Glue to transform and load CSV/JSON into tables, and provide access. Pros: Structured, relational database allows efficient SQL queries. Supports complex joins and indexing for faster queries. Cons: High operational cost: managing RDS instances (compute, storage, backups). Requires ETL process (AWS Glue) to transform and load data into RDS, adding complexity and cost. Less flexible with semi-structured JSON data; often requires flattening or schema design. When to use: When data requires frequent transactional updates or complex relational integrity. When analysts need very low-latency querying on a relational DB. Why rejected: Not the most cost-effective for new datasets stored as CSV/JSON in S3. Setup and maintenance overhead are significant compared to serverless querying options. --- Option B: Create an AWS Glue DataBrew project and share with analysts. Pros: Great for data cleaning, profiling, and transformation without coding. Provides a UI for data wrangling. Cons: Not primarily a SQL query engine — analysts cannot run arbitrary SQL queries easily. Licensing and usage costs can grow with scale. Meant for data preparation, not for direct querying and analysis. When to use: When analysts or data engineers need to clean and prepare data interactively before analysis. Why rejected: Does not fulfill the requirement for analysts to analyze data by using SQL queries. Not cost-effective as a direct query platform. --- Option C: Store data in S3, use AWS Glue crawler to catalog data as tables, and query with Amazon Athena. Pros: Serverless and pay-per-query → very cost-effective. Supports querying CSV and JSON natively via Athena using SQL. Glue crawler automatically discovers schema, so no manual...

Author: Kunal · Last updated Jul 10, 2026

A retail company stores order information in an Amazon Aurora table named Orders. The company needs to create operational reports from the Orders table with minimal latency. The Orders table contains billions of rows, and over 100,000 transactions can occur each second. A marketing team needs to join the Orders data with an Amazon Redshift table named Campaigns in the marke...

Let's analyze the scenario and each option carefully. --- Scenario Summary: Source: Amazon Aurora with billions of rows, high transaction volume (>100,000 TPS). Target: Amazon Redshift (marketing data warehouse). Goal: Create operational reports from the Orders table with minimal latency. Marketing team wants to join Orders data with Campaigns in Redshift. Operational Aurora database must not be affected by reporting workloads. Least operational effort is desired. --- Key Factors for decision: 1. Latency & freshness: Operational reports require near real-time or low latency data access. 2. High transaction throughput: >100,000 transactions per second means heavy load on Aurora. 3. Minimal impact on Aurora: Reporting must not degrade OLTP performance. 4. Ease of setup and maintenance: Minimize operational effort. 5. Data volume: Billions of rows, so solutions need to be scalable. --- Option Analysis: A) Use AWS DMS Serverless to replicate Orders to Redshift + materialized view in Redshift Pros: DMS is designed for continuous replication with minimal lag. Serverless DMS reduces operational overhead (no infrastructure to manage). Data is replicated asynchronously, so Aurora OLTP workload is not impacted. Once data is in Redshift, creating materialized views is straightforward. Cons: Some latency is introduced due to replication (usually seconds or more). DMS is not designed for ultra-low latency real-time reporting but is quite efficient. Use case: Near real-time reporting with minimal Aurora impact and moderate operational effort. B) Use Aurora zero-ETL integration with Redshift + materialized view in Redshift Pros: Zero-ETL is a new Aurora feature to directly replicate changes to Redshift with minimal latency. No manual ETL or replication infrastructure to maintain. Very low latency replication. Minimal operational effort. Cons: Currently, zero-ETL integration may have limitations on scale or availability depending on A...

Author: Ming · Last updated Jul 10, 2026

A data engineer is using an AWS Glue ETL job to remove outdated customer records from a table that contains customer account information. The data engineer is using the following SQL command to remove customers that exist in a table named monthly_accounts_update table from the customer accounts table: MERGE INTO accounts t USING monthly_a...

Let's analyze the SQL command and how it behaves in the context of AWS Glue ETL jobs, particularly focusing on support for the MERGE statement. --- SQL Command (from the question): ```sql MERGE INTO accounts t USING monthly_accounts_update s ON t.customer = s.customer WHEN MATCHED THEN DELETE ``` The logic here is: If a customer exists in both `accounts` and `monthly_accounts_update`, delete it from `accounts`. --- Key Considerations: 1. MERGE statement support in AWS Glue: As of AWS Glue 3.0+ with Apache Spark and Spark SQL, MERGE INTO is supported only on tables stored in Apache Hudi, Iceberg, or Delta Lake format. It is not supported for traditional Glue tables stored in formats like plain Parquet or CSV in S3 (i.e., non-transactional formats). Also, some syntax errors in the question (like `=3D` instead of `=`, and hyphens `-` where line breaks are intended) might break execution unless cleaned up. 2. MERGE INTO ... WHEN MATCHED THEN DELETE is valid SQL (Spark 3+) — if used on supported table formats. But the question does not specify if `accounts` is stored in Hudi/Iceberg/Delta — so we must assume it's a typical Glue table (e.g., backed by S3 in Parquet or similar), which does not support such operations natively. --- Option Analysis: --- A) All customer records that exist in both the customer accounts table and the monthly\_accounts\_update table will be deleted from the accounts table. This would be true if the `MERGE INTO ... DELETE` operation is supported and the data source format allows it. But in standard AWS Glue tables, `MERGE` is not valid unless using supported formats (Hudi, Iceberg, De...

Author: Ming · Last updated Jul 10, 2026

A company builds a new data pipeline to process data for business intelligence reports. Users have noticed that data is missing from the reports. A data engineer needs to add a data quality check for columns that contain null values and for referential integrity at a stage ...

Let's analyze each option carefully considering the requirements: Requirements: Add data quality checks for null values in columns. Add referential integrity checks (i.e., foreign key constraints or matching references). Checks must happen before data is stored. Solution should have least operational overhead (easy to maintain, automated, minimal custom coding). --- Option A) Use Amazon SageMaker Data Wrangler to create a Data Quality and Insights report. Pros: Data Wrangler is good for data exploration, profiling, and creating reports. Cons: It's primarily a manual or semi-automated tool designed for data scientists to profile and transform data interactively. It’s not optimized for automated, continuous validation in a production pipeline before storage. Operational overhead can be high because it may require manual intervention or custom automation workflows. Use case: Best for exploratory data analysis and data profiling in a manual or semi-automated setting, less ideal for automated production quality checks. --- Option B) Use AWS Glue ETL jobs to perform a data quality evaluation transform on the data. Use an IsComplete rule on the requested columns. Use a ReferentialIntegrity rule for each join. Pros: Glue Data Quality feature supports built-in rules like `IsComplete` (for null checks) and `ReferentialIntegrity` (for foreign key validation). These are declarative, easy to implement, and designed specifically for data quality validation in Glue pipelines. This allows automated, repeatable quality checks inside ETL jobs before storing data. Low operational overhead because Glue manages the execution, scaling, and monitoring. Use case: Best for automated production-grade ...

Author: Max · Last updated Jul 10, 2026

A company is setting up a data pipeline in AWS. The pipeline extracts client data from Amazon S3 buckets, performs quality checks, and transforms the data. The pipeline stores the processed data in a relational database. The company will use th...

Let's analyze each option based on the requirements and key factors: Requirements Recap: Extract client data from S3 buckets. Perform quality checks. Transform data. Store processed data in a relational database (for future queries). Be cost-effective. --- Option A: AWS Glue ETL for extraction and transformations. AWS Glue Data Quality for quality checks. Store processed data and quality check results in Amazon RDS for MySQL. Pros: AWS Glue ETL is a fully managed, serverless ETL service that integrates well with S3 and RDS. Glue Data Quality can enforce data quality rules natively in the ETL pipeline. Storing both processed data and quality results in RDS centralizes data for querying. Cons: Glue Data Quality is a newer service; it may incur additional costs. Storing quality check results in RDS might increase database storage and management overhead, possibly increasing costs. Use Case: Good when tight integration between ETL and data quality is needed, and you want all results in a relational DB for easy querying. --- Option B: AWS Glue Studio for extraction. AWS Glue DataBrew for transformations and quality checks. Load processed data into RDS, quality check results into S3. Pros: Glue Studio provides visual ETL development with ease. DataBrew specializes in visual data prep, transformations, and quality checks with low-code/no-code. Separating quality check results into S3 can save RDS storage costs. This can be cost-effective if quality check results are large or less frequently queried. Cons: Using two separate tools (Glue Studio and DataBrew) adds complexity. DataBrew can be more expensive for large-scale processing compared to Glue ETL. Having quality data separate might complicate combined analysis. Use Case: When you want easy, visual data prep with separation of data and quality logs, especially if quality logs are large or accessed separately. --- Option C: AWS Glue ETL for extraction and transformations. AWS Glue DataBrew for quality checks. Load processed data and quality check results into S3. Pros: Fully serverless, managed, and decoupled solution. Cost-effective since S3 storage is cheaper than RDS. Good for large datasets or when relational querying is not immediately needed. Cons: Does not meet the requirement of storing processed data in a relational database for future queries. Querying data in S3 may require Athena or Redshift Spectrum, adding complexity an...

Author: Elizabeth · Last updated Jul 10, 2026

A company uses Amazon Redshift as a data warehouse solution. One of the datasets that the company stores in Amazon Redshift contains data for a vendor. Recently, the vendor asked the company to transfer the vendor's data ...

Let's analyze each option based on the key factors: Requirement Recap: Data is stored in Amazon Redshift. Vendor wants their dataset exported weekly to their Amazon S3 bucket. The solution should transfer data from Redshift to S3 on a schedule. --- Option A: AWS Lambda with Redshift COPY command COPY command in Redshift is used to load data from S3 into Redshift, not to export data out. Lambda can run scheduled tasks, but COPY is the wrong command for export. This option is invalid because COPY is for ingestion into Redshift, not export. Conclusion: Reject because COPY cannot export data to S3. --- Option B: AWS Glue job using Redshift UNLOAD command The UNLOAD command is specifically designed to export data from Redshift to S3. AWS Glue can be scheduled, can connect to Redshift, and run the UNLOAD command on a schedule. This is a fully managed ETL solution that can handle transforming and exporting data. Fits the requirement perfectly: scheduled export from Redshift to vendor's S3 bucket. Conclusion: Suitable and meets the requirement directly. --- Option C: Amazon Redshift Data Sharing with S3 bucke...

Author: Michael · Last updated Jul 10, 2026